Title of Invention	"AN APPARATUS FOR IDENTIFICATION AND SEPARATION OF COMPLEX COMPOSITE SIGNALS INTO ITS DETERMINISTIC AND NOISY COMPONENTS"
Abstract	The present invention provides an apparatus for identification and /or separation of complex composite signals into its deterministic and noisy components which comprises source means for obtaining a signal to be identified and/or separated from suitable sensors appropriately located in said apparatus/equipment;means for digitizing the said composite signal, obtained as sequential or time-series data relating to a variation in a system property and various computing means . The apparatus has application in a variety of fields where digitalized data is made available via an apparatus which converts the monitored signals

Title of Invention

"AN APPARATUS FOR IDENTIFICATION AND SEPARATION OF COMPLEX COMPOSITE SIGNALS INTO ITS DETERMINISTIC AND NOISY COMPONENTS"

Abstract

The present invention provides an apparatus for identification and /or separation of complex composite signals into its deterministic and noisy components which comprises source means for obtaining a signal to be identified and/or separated from suitable sensors appropriately located in said apparatus/equipment;means for digitizing the said composite signal, obtained as sequential or time-series data relating to a variation in a system property and various computing means . The apparatus has application in a variety of fields where digitalized data is made available via an apparatus which converts the monitored signals

Full Text	The invention relates to an apparatus for the identification and/or separation of complex composite signals into its deterministic and noisy components. More particularly this invention relates to a method for separating complex composite signals made available in the form of digitized data by an apparatus which appropriately converts the monitored signals obtained from medical diagnostic equipment, molecular spectroscopy, seismographic analysis, tomography, image cleaning, chemical reactors/reaction data etc. into its deterministic (true signal) and noisy components. Detecting and/or separation of the underlying deterministic dynamics from noisy data obtained from complex systems is of considerable importance. Estimating the deterministic content of a monitored digital sequence (usually sequential or time-series) continues to evade satisfactory solution, despite considerable efforts over the years. This problem is central to most measured data and is especially important in applications when data is collected from medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors/reactions etc. The central problem is therefore to identify and/or separate from the composite signal, its deterministic and noise components. One of the methods for the detection and/or separation of the deterministic components of the signal is by band-pass filtering using Fast Fourier Transform (FFT) (Cohen, L., Time-frequency analysis, Prentice Hall, Englewood Cliffs, 1995) where assumptions about the frequency distribution of the noise can be made and thereby suppress those frequency components before inversion to the time domain. Thus for example, low pass filters cut off the high frequency components in the signal. This is disadvantageous when all-frequency noise is present in the signal or when the signal component itself is broadband. Fourier filtering is especially unsatisfactory when the signal dynamics originate from systems exhibiting deterministic chaos. Another method uses kernel estimators or spline estimators but a major disadvantage of these standard smoothing techniques is the fact that they do not resolve local structures well enough (Hardle, W., Applied nonparametric regression, Econometric Society Monographs, Cambridge University Press, 1990). Another method used for the detection and separation of the composite signals is singular value analysis to reconstruct the dynamics (Broomhead, D. and King, G.P., Physica D, 20 217, 1986, Cawley, R and Hsu,G.-H, Phys. Rev. A. 46 3057, 1992). Here the original time-series is projected onto a subspace spanned by the corresponding singular vectors, namely those spanning the largest fraction of the total variance of the data. The remaining orthogonal directions then contain most of the noise. However, these methods are nontrivial and difficult to implement because of the abstract notions involved. Yet another method which is very commonly employed in the recent times for the detection and separation of the deterministic components of signal is based on the theory of wavelet transforms (WT) (Vettererli, M. & Kovacevic, Wavelets and subband coding, Prentice Hall, PTR, Englewood Cliffs, 1995). Wavelet transform methods are increasingly used as tools for studying multiscale, nonstationary processes in various fields. General methodologies for multiresolution signal processing, sub-band coding and also mutigrid techniques have led to applications in image and speech processing, data compression, communication, quantum mechanics, medicine, spectroscopy, etc. Wavelet transforms often complement Fourier transforms (FT) techniques because the spectrum of frequencies present in the signal may be analyzed at various scales with the added advantage of time localization. Methods based on the WT have proved to be efficient and easy to implement. Wavelets are derived from rapidly oscillating functions with mean zero and obtained by suitable scaling of analyzing function to match the desired frequencies with simultaneous translations in time. A wide variety of analyzing functions amenable for discrete or continuous time analysis are known. (Strang, G and Nguyen, T Wavelets and Filter Banks, Wellesley-Cambridge Press, Wellesley, MA, 1996). Typical wavelet examples are: the discrete Haar (following a box function), the continuous Morlet (a sine function modulated by a Gaussian envelope), Mexican hat (second derivative of a Gaussian), Daubechies spanning a wide range of discrete and continuous properties (including compact support) depending on the number of wavelet filter coefficients, Lemarie, biorthogonal spline, Malvar, Lemarie, Coiflet, Meyer, Symlet, etc. Concisely stated, WT is a generalization of the FT and is obtained by taking the inner product of a set of basis wavelets (dependent on the resolution, translation and choice of the wavelet) with the available data X. Systematic applications for varying oscillating scales a of the wavelet and also their localization b yields scale-frequency information in a scalogram of wavelet coefficients W, rather than the frequency periodogram in a FT analysis. The hitherto used wavelet methods for the detection and separation of the composite signals are mainly based on the concept of thresholding the wavelet coefficients obtained from a single transformation. Hard thresholding cuts off coefficients below a certain threshold A while soft-thresholding reduces all coefficients at various scales by the threshold value (Donoho, D.L., Johnstone, I.M., Kerkyacharian, Picard, D., J. Royal Stat. Soc., Series B SI, 301). The threshold value is determined by statistical calculations and is dependent on the standard deviation of the noise (Nason, G.P., Wavelet regression by cross-validation, Dept. of Mathematics, Univ. of Bristol, 1994) and the length of the signal. The specified threshold value may also be used to evaluate special cost functions or entropy functional for generating appropriate basis. However, considerable improvements are still needed for applications in precision applications like medical diagnostics, seismographic analysis, image analysis etc., thereby resulting in incorrect inferences. This is because it is known that noise can be present at all frequencies and eliminating components based on statistical thresholding may not be adequate enough. It will be advantageous and beneficial to have a process by means of which noise at all scales will be identified in a systematic and rational manner without thresholding based on properties of the signal. The present invention discloses an improved and systematic method for the identification and/or separation of composite signals into its deterministic and noisy components without the need for threshold values to be calculated and based on scientific rationale. The results of analysis also show precision and accuracy and lead to correct inferences. The application of WT to reducing ultrasound speckle noise has been described in US patent no. 5497777 (General Electric Company, 1996). In another continuing with the same topic (US patent No. 5619998, General Electric Company, 1997), a procedure whereby a coherent imaging system signal is reduced of speckle noise by nonlinear adaptive thresholding of wavelet coefficients by WT has been described. The resulting image has an improved signal-to-noise ratio. The method followed thresholds the wavelet coefficients from the finest scale to the coarsest scale after dividing the imaging system signal into a number of subinterval signals of equal length. Coefficients in the various scales of the wavelet transform related to noise are thereby identified in each subinterval and these are eliminated before an inverse discrete wavelet transformation. It may be noted that the procedure followed is based on analyzing the scalogram of wavelet coefficients obtained by a single WT of data contained in a subinterval. The present invention is based on the principle of recursive wavelet transformations, wherein the components are identified and separated based on their dimensionality and extent of correlatedness. It is generally known that noise is high dimensional and uncorrelated whereas the deterministic signal is usually of lower dimensional and correlated. The difference with respect to correlation in respect of noise and the deterministic signals comes out distinctly on taking recursive wavelet transformations of the wavelet coefficients at each scale. The recursive wavelet transforms result in separation of the signals into its noisy and deterministic parts. The process of separation occurs because the relative power distribution (RPD) remains constant for the noise while for signals with even short range correlations (e.g., deterministic chaotic signals) show considerable variations. For a composite signal containing both noise and deterministic components, eliminating those wavelet coefficients at each scale (identified by recursive transformations) which correspond to the noisy part yields the pure deterministic signal on recursively inverting using discrete wavelet transform. The proposed method, therefore, involves taking wavelet transform of the data at different scales and this is resolved further using recursive wavelet transformation into a plurality of scales with time localization. The differences in the correlatedness and dimensionality of the two parts of the composite signal allows easy identification of those coefficients which belong to noisy part and those that belong to the true signal when subjected to recursive transformations. The identified wavelet coefficients that belong to noise are then eliminated and the inverse operation carried out to recover the true signal. The procedure guarantees identification of noisy components present at all scales and thus is devoid of any local/global approximations and thresholding. Removal of noise is thus possible. It is therefore the main object of the present invention to provide an apparatus for the identification and/or separation of composite signals into its deterministic and noisy parts. Another object of the present invention is to provide an apparatus . for the identification and/or separation of composite signal into its deterministic and noisy parts utilizing the concept of recursive wavelet transformation. Still another object of the present invention is to provide an apparatus for the identification and/or separation of composite signals that can lead to better and more accurate inferences. DESCRIPTION OF DRAWINGS FIG. 1. Block diagram showing the apparatus and methodology. FIG 2. Flow diagram showing the processor for the identification and/or separation of composite signals into its deterministic and noisy components. FIG. 3. A schematic of the behavior for noise (thin line bars) and a chaotic signal (thick line bars) in Relative Power Distribution (RPD) and Fractional Power Distribution (FPD) p(k). i p(')wr ^en the wavelet coefficients of a dyadic scale j is recursively transformed k times, (a) The magnitude of RPD (thin line bars) for noise remains level due to its uncorrelated structure while for the deterministic signal it varies with k even for the short term correlation that is characteristic of chaotic dynamics, (b) The FPD shows a different decline for white noise than for the signal suggesting that this sensitivity can form a basis for their separation. FIG. 4. Distinguishing noise and a deterministic chaotic signal by the outlined method, (a) The bottom panel shows the Gaussian white noise [1024 data-points (bottom panel) with mean zero and variance unity] analyzed. The top panel shows the WT resolution at each dyadic scale j and obtained from the wavelet coefficient matrix W(l) (by inverting the wavelet coefficients for each scale W^> back to the time domain.). The decomposition by WT resolves the noise at different frequency components at all time /. (b) For each W^y (now treated as a data) the next generation of wavelet coefficients W2) was obtained. The coefficients of those scales in W transforms which results in \V(/(I) = 0. (c) WT of a short chaotic time-series data from a CSTR and analyzed similar to panel a), d) The signal in c) is detected as a signal arising from a deterministic process because the resulting W,,(I)=W(1). FIG.5. Signal detection with noise reduction for two representative signals, (a) Filtering short noisy data (1024 points) X (black) from the nonlinear logistic map f-Vj,, =/ljr('(l-jr.) + jrn;A =3.2\| eliminates the scatter completely to yield the filtered data X(, (light gray) with the \tl completely synchronizing with the underlying 2-period dynamics (gray) \|.v.(l =A.Y (l-x,);! = 3.21. (h) Noise reduction capabiliity for short and noisy time-series data (1024 points) obtained from a CSTR exemplifying chaotic dynamics. Noise significantly distorts the signal as seen in the phase-plane plot of concentration CA(/) and temperature T(t) (black). The filtered signal (light gray) obtained after limiting the procedure to two generations shows remarkable recovery (light gray) of the sensitive deterministic dynamics (gray). Accordingly the present invention provides for the identification and/or separation of composite signal into its deterministic and noisy parts which comprises: An apparatus for identification and/or separation of composite signals obtained from an instrument/equipment recording the variations in a system property as sequential or time-series data from the said instruments/equipments selected from medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors/reactions, into its deterministic (true signal) and noisy parts which comprises, (a) source means for obtaining the signal to be identified and/or separated from suitable sensors appropriately located in an apparatus/equipment; (b) means for digitizing the said composite signal, obtained as sequential or time-series data relating to a variation in a system property, (c) computing means for subjecting the said digitized data obtained from step (b) above to wavelet transform to obtain a scalogram in terms of wavelet coefficients; (d) computing means for organizing the said resulting wavelet coefficients in each of the scales to form new sets of data; (e) computing means for taking each of the above said data sets and carry out their wavelet transform to obtain another scalogram in terms of wavelet coefficients; (f) computing means for implementing steps $}€ recursively with testing for the constancy in the power distribution, the said power distribution being the ratio of the power in a particular scale with respect to the total power in all the scales of that scalogram for two consecutive recursive scalograms and thereby identifying the recursive wavelet scales contributing to noise in the signal; (g) computing means for eliminating the above said wavelet coefficients in the recursive wavelet scales contributing to noise by setting them to zero; (h) computing means for inverting above said the wavelet coefficients by inverse recursive wavelet transformation and thereby determining the deterministic signal component, the said signal component being the true signal is digitized form separated from the noise component. (i) means for converting the above said digitized deterministic signal component to an analog signal using D to A converter The signal to be identified and or separated according to the process of the present invention should be obtained by the use of suitable sensors appropriately located in the apparatus equipment. In an embodiment of the present invention, we include conversion by analog-to-digital (A/D) apparatus of signals to yield rescaled digitized numbers that can be fed to a programmed device or a programmed general purpose digital computer, with a storage medium, capable of executing the steps of the said method for the purposes as claimed in claim 1 and 2. In an embodiment of the present invention an improved method is claimed for data obtained even from deterministic systems exhibiting chaotic dynamics. It is also understood that the process of identification and/or separation of composite signals may be used for sequential data or time-series data obtained as variations in a system property in space or time; dividing the signal into a number of subinterval signals and applying recursive wavelet transformation to each subinterval signal. In an embodiment of the present invention it is also understood that the process of identification and/or separation of composite signals may be used for discrete, biorthogonal, continuous wavelet (Haar, Morlet, Mexican hat, biorthogonal spline, Daubechies, Malvar, Lemarie, Coiflet, Meyer, Symlet, etc.) more preferably by Haar, and also the alternate methodologies for obtaining the wavelet transform and its inverse (fast wavelet transform algorithm, convolution, etc.). The Haar wavelet is simple to understand and possesses features like signal reconstruction capability. Although other waveforms are known we shall use the Haar wavelet as a reference for its ease in discussion. The Haar wavelets are scaled forms of a square-wave function [#(/)=+! for 0 and#(/)=0 otherwise] and is given by 2j/2H(2jt-i) where 7'= 0,1 p-\ (for a data set with N=2 p points). It is convenient to choose the wavelets for obtaining the WT to follow dyadic scales j, i.e., a=2~J with translations b=2'Ji (/ = 0,1,...,2;-1). The Haar matrix HN is (N x N) dimensional and forms an orthonormal basis with each row corresponding to a dyadic scale j and shift / in discrete time. The Haar WT may then be carried out as to obtain the N wavelet coefficients W (via,. W = 1 / ^/~N(HN • X) ; where X is the column vector of data and 1 / JN the normalizing factor) at (p - 1) dyadic scales with the last N12 coefficients for the smallest scale, the preceding N14 coefficients for the next scale etc. Note that the first coefficient relates to the average of the signal. Obtaining a reconstruction of the signal from W is simple inversion (i.e., Xr = reconstruction X = Xr. For a noisy signal X=Xd+Xn where Xd is the deterministic -v signal and Xn the noise (say, gaussian white noise). The aim in noise reduction would be to modulate the coefficients in W to Wd so that on reconstruction with Wd by inverse WT the deterministic signal Xd is recovered. The residual component corresponds to noise and may be obtained as Xn = X - Xd to obtain the noise content Xn. Viewed in this fashion, wavelets facilitate the separation of noise and signal, provided a convenient and intrinsic property differentiating the deterministic and noisy components is identified. In analogy, we may cite the example of separation by distillation of a mixture of two components A and B. In this process the intrinsic property utilized is the difference in volatility of the individual components. In the case of noise and the signal we suggest that the behavioral patterns of the power distribution at each scale show distinct trends on taking the WT of the wavelet coefficients, W, at each scale j. This feature is discussed more specifically next. From a parent WT with coefficients Wll), it is possible to obtain a second generation recursive transform of its 7 scale coefficients, say, W(2-;). Note that in principle later generation matrices of wavelet coefficients, W(M, k = 2,3,.... can be obtained and structured so as to obtain a tree of scalograms. The number of generations allowed depends on the number of initial data points N and the chosen starting j* scale from the parent W(l). It is interesting to note that for an orthonormal wavelet family, such as the Haar, the signal energy is similar to expressions in fourier theory with the power in each scale j given by P, =^ w>(0\|2 with w>(0. they scale elements of W(I). Similarly, the total power in the signal is therefore Ptot =^\w(i)\2;i = l,2,...,N . In general, the power P}m for a £* generation wavelet transform in the y'"1 scale is conserved in the total power of generation, i.e., P^'^ = /j(i+1); = P(t);-. Now, calculations were carried out for successive generations k and different scales j for the Relative Power Distribution (RPD), viz., P(k]j I P(k),ot and also Fractional Power Distribution (FPD) at a particular scale and recursive generation number k, viz., P^j I P^M for both gaussian white noise and chaotic signals. The results obtained, qualitatively sketched in Fig. 3, bring out two interesting observations. Firstly, it was observed that the RPD remained level for noise while those for the chaotic signal showed considerable variations (Fig. 3). Secondly, it was seen that the FPD at particular scales is different for noise when compared to the deterministic signal (Fig. 3) forming a basis for separation. This constancy in the RPD among scales along with a different decline in the FPD for uncorrelated noise for succeeding generations of the WT may be exploited profitably to reduce the effects of noise. The aim here would be to obtain estimates of the wavelet coefficients, W(';), in later generations obtained by recursion and scales (i.e., starting from the noise affected coefficients in the parent W(1)) such that recursive inversion results in W(1)d as the set of filtered coefficients. The method to obtain W(1) needs to ensure that constancy in RPD is seen at every scale for the tree of scalograms. At each stage when constancy is seen, the residual wavelet coefficients contribute to noise and these are eliminated from the immediate prior generation before inversion. Strictly followed noise may then be completely eliminated, but this would require a data set of sufficient length to be analyzed. However, in practice, when the data set is short, a practical approach would be to limit the analysis to a specified number of generations and filter out those scales which do not cause abrupt changes in FPD on taking successive transforms. The procedure can be carried out for every frequency scale of W^1' in a systematic manner and is easily coded for automation. The method is novel because of recursive wavelet transformation used in the said method. Note that the method of identification and/or separation of composite signal into its deterministic component and noise component is devoid of hard and soft thresholding of the wavelet coefficients at various scales based on statistical analysis of the wavelet coefficients. The results obtained using the above formalism for short time-series data of 1024 data points for noisy, chaotic and composite signals are shown in Figs. 4,5. The invention is described in details with reference to the examples given below which are provided to illustrate the invention and therefore should not be construed to limit the scope of the invention. Example 1: This example is provided to detect a signal solely consisting of noise Using a random number generator digital data in the form of white noise with Gaussian Distribution (Press, W.H., Flannery, B.P., Teukolsky, S.A., Vetterling, W.T., Numerical Recipes, Cambridge University Press, 1986, p 191-203) was generated. This sequential data was subjected to recursive Wavelet Transformation using the Haar wavelet basis function. A near constant Relative Power Distribution was immediately reached, thereby indicating the signal employed consists only of noise. A schematic of the behavior for noise (thin line box) in Relative Power Distribution (RPD) P(k]i I P(k]tot and Fractional Power Distribution (FPD) P(k)/ I />(I)to, when the wavelet coefficients of a dyadic scale j is repeatedly transformed k times is shown in Fig. 3. Note that the magnitude of RPD (thin line box) for noise remains level due to * its uncorrelated structure The FPD for (thin line box) also shows a decline for the white noise signal for recursive transformations. The results of experiments using the invention with signal having only noise component is shown in Fig. 4(a) and Fig. 4(b) of the drawing accompanying this specification. In Fig. 4(a) the signal is shown in the bottom panel while the wavelet transformation showing decomposition of the signal at dyadic scales and localized in time are displayed as a scalogram in the upper panel. Fig. 4(b) shows the resultant scalogram upon recursive wavelet transformation (up to k = 2 when near constancy in RPD was obtained) and then followed by recursive inverse wavelet transformation. The results clearly reveal that the method of the invention has precisely identified the entire signal employed as consisting only of noise and does not possess any deterministic component. Figure 4b shows the cleaned signal with the noise removed for transforms up to the second generation (i.e., k = 2 ) considered, i.e., Example 2: This example is provided to detect the signal solely consisting of true and deterministic signals. A signal was generated corresponding to a first-order nonisothermal exothermic endothermic reactions taking place in a continuously stirred tank reactor (CSTR) (Kahlert, C; Rossler, O.E.; Varma, A; Springer Ser. in Chem. Phys., 18, 355 (1981). The system parameters and their values used for obtaining the time-series data are Z)a=0.26;£=0.0;S=0.5;/c = 1.0;5=57.77;a = 0.42;x3c = 0.0;/?=8.0, where, Da= Damkohler number ; e = activation energy grouping ; K = ratio of activation energies for the exothermic and endothermic reaction steps ; B = adiabatic temperature rise grouping; a = ratio of heat effects; x3c = dimensionless coolant temperature; /? = heat transfer grouping. The system variables are the dimensionless temperature T\ and the concentrations C'A,C'B of the reactant A and intermediate product B, respectively. For these parameter values the dynamics of the CSTR exhibits low dimensional deterministic chaos. The time-series data in temperature T is a dynamically fluctuating signal in time. This time-series signal was subjected to recursive wavelet transformation using the Haar wavelet basis function. The results obtained are given below. Figure 3 schematically shows the Relative Power Distribution (RPD) P(k]j I P(k\ot and Fractional Power Distribution (FPD) P(k]j I P(l\ot when the wavelet coefficients of a dyadic scale j is recursively transformed k times. From Fig. 3, the following observations may be made. Firstly, it is seen that the magnitude of RPD for the deterministic signal (thick line bars, Fig. 3 a) varies with k because of the short term correlations that is characteristic of chaotic dynamics and may be compared with that for the noise signal (thin line bars, Fig. 3 a) which remains level due to its uncorrelated structure. Secondly, the FPD for the chaotic deterministic signal (thick line bars, Fig. 3b) shows a different rate of decline from that for white noise (thin line bars, Fig. 3b) than for the signal, These features indicate the system sensitivity to these properties and show that they may be profitably exploited for identification/separation of deterministic and noisy components of a signal. The results of experiments using the invention with the time-series signal having only the deterministic chaos component is shown in Fig. 4(c) and Fig. 4(d) of the drawing accompanying this specification. In Fig. 4(c) the time-series signal is shown in the bottom panel while the wavelet transformation showing decomposition of the signal at dyadic scales and localized in time are displayed as a scalogram in the upper panel. Fig. 4(d) shows the resultant scalogram upon recursive wavelet transformation and then followed by recursive inverse wavelet transformation. The results correctly reveal that the method of the invention has precisely identified the entire signal employed as consisting only of deterministic signal and does not possess any noise component. The results for the analysis of short time-series data from a system exhibiting deterministic chaos [viz., a continuously fed stirred tank reactor (CSTR) with an exothermic reaction show W(1)«/=W(I). In other words, the constituent components at all scales j are not modulated whereby the processed time-series retains its original behavior as desired. This result may be viewed with that obtained in example 1 where W^1' * W/1' =0. where the constituent components at all scales j are appropriately modulated whereby the processed time-series is shown to be noisy as desired. Example 3: This example illustrated the identification and separation of combination of deterministic and noisy signals. Composite signals with both deterministic and noise components were generated for two representative cases, viz.: a) biperiodic deterministic signal of a process following the logistic map (yielding sequential data) (Rasband, S.N., Chaotic dynamics of nonlinear systems, John Wiley & Sons, 1990) as its deterministic component with gaussian white noise (mean = 0; variance = 0.01) component, and b) deterministic chaotic dynamics from a process wherein a first order exo-endothermic reaction in a continuously stirred tank reactor (CSTR) as specified in Example 2 with gaussian white noise component (mean = 0; variance = 0.04). These data sets were subjected to recursive wavelet transformation using the Haar wavelet basis function. The denoising capability by the process of identification and/or separation of the deterministic and noise components by recursive wavelet transformation is shown below. The resultant signals were subjected to recursive wavelet transformation till k =2, i.e., using two recursive transformations. The wavelet coefficients were then inverted using the inverse recursive wavelet transformation. Both the examples show the remarkable robustness and reliability for noise removal from even short signals as given below and in Fig. 5. Figure 5 shows the capability of the approach to simultaneously detect and reduce noise in the two specified and noisy signals of this example. Note that the signals are of varying complexity. The first signal typifies periodic data from the logistic map corrupted with noise while the other represents highly sensitive chaotic dynamics from the CSTR. The accurate recovery of the deterministic signal for the periodic signal is seen in Fig. 5a. Note that this approach is devoid of statistical inferences and local/global fitting procedures. Even with the extreme case of sensitive chaotic dynamics (where small errors can grow unbounded), Fig. 5b shows efficiency in signal recovery and noise reduction. In either case with just two recursive transformations the process of separation of the deterministic and noisy components could be effected to over 90% (Calculated via, Extent of separation = (P} - P2 jl P\ ; where P, is the power of the noise component, and P2 is the power of the noise left in the signal after the process of separation by recursive wavelet transformation). Advantages of the invention: 1. The method of the invention is simple and accurate because of which precise inferences can be arrived at; 2. The identification and/or separation of the signals can be achieved efficiently by limited number of recursive wavelet transformations for signals of short data length; 3. The method is not based on thresholding of wavelet coefficients based on statistical inferences; 4. The method of the present invention is based on a procedure having a rational scientific basis. While only certain important features of the invention have been illustrated and described, many modifications and changes can be implemented by those skilled in the art. It is, therefore, to be understood that the appended claims are intended to cover all such modifications and changes as fall within the spirit of the invention. Thus, it is understood that the process of identification and/or separation of composite signals described may be used for discrete, biorthogonal, continuous wavelet types (Haar, Morlet, Mexican hat, biorthogonal spline, Daubechies, Malvar, Lemarie, Coiflet, Meyer, Symlet, etc.); alternate methodologies for obtaining the WT (Mallat fast wavelet transform, convolution etc.); data obtained from various sources (medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors/reactions, etc) or for variation in a system property in space; dividing the signal into a number of subinterval signals and applying recursive wavelet transformation to each subinterval signal, etc. We Claim: 1. An apparatus for identification and/or separation of composite signals obtained from an instrument/equipment recording the variations in a system property as sequential or time-series data from the said instruments/equipments selected from medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors/reactions, into its deterministic (true signal) and noisy parts which comprises, (a) source means for obtaining the signal to be identified and/or separated from suitable sensors appropriately located in an apparatus/equipment; (b) means for digitizing the said composite signal, obtained as sequential or time-series data relating to a variation in a system property, (c) computing means for subjecting the said digitized data obtained from step (b) above to wavelet transform to obtain a scalogram in terms of wavelet coefficients; (d) computing means for organizing the said resulting wavelet coefficients in each of the scales to form new sets of data; (e) computing means for taking each of the above said data sets and carry out their wavelet transform to obtain another scalogram in terms of wavelet coefficients; (f) computing means for implementing steps ctf e. recursively with testing for the constancy in the power distribution, the said power distribution being the ratio of the power in a particular scale with respect to the total power in all the scales of that scalogram for two consecutive recursive scalograms and thereby identifying the recursive wavelet scales contributing to noise in the signal; (g) computing means for eliminating the above said wavelet coefficients in the recursive wavelet scales contributing to noise by setting them to zero; (h) computing means for inverting above said the wavelet coefficients by inverse recursive wavelet transformation and thereby determining the deterministic signal component, the said signal component being the true signal in digitized form separated from the noise component. (i) means for converting the above said digitized deterministic signal component to an analog signal using D to A converter 2. An apparatus as claimed in claim 1 wherein the means for wavelet transformation is employed may be selected for discrete, biorthogonal, or continuous wavelet types. 3. An apparatus as claimed in claim 1 wherein the means for wavelet types employed may be selected from Haar, Mlrlet, Mexican hat, biorthogonal spline, Daubechies, Malvar, Lemarie, Coiflet, Meyer or Symlet. 4. An apparatus as claimed in claim 1 wherein source for obtaining sequential or time- series data is selected from medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors or reactions. 5. An apparatus as claimed in claim 1 wherein the source for obtaining sequential or time-series data may also be selected from deterministic systems exhibiting chaotic dynamics. 6. An apparatus as claimed in claim 1 wherein means for hard and soft thresholding of the wavelet coefficients at various scales is not required because of recursive wavelet transformations used. 7. An apparatus as claimed in claim 1 wherein the digitizing means employed is an analogue to digital converter. 8. An apparatus as claimed in claim 1 wherein converting means employed is a digital to analogue converter. 9. An apparatus as claimed in claim wherein the computing means employed is a device or a programmed general purpose digital computer, with a storate medium, capable of executing the steps of the said method. 10. An apparatus for identification and/or separation of composite signals obtained from an instrument/equipment recording the variations in a system property as sequential or time-series data from medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors/reactions, into its deterministic (true signal) and noisy parts.

Full Text

The invention relates to an apparatus for the identification and/or separation of complex composite signals into its deterministic and noisy components. More particularly this invention relates to a method for separating complex composite signals made available in the form of digitized data by an apparatus which appropriately converts the monitored signals obtained from medical diagnostic equipment, molecular spectroscopy, seismographic analysis, tomography, image cleaning, chemical reactors/reaction data etc. into its deterministic (true signal) and noisy components. Detecting and/or separation of the underlying deterministic dynamics from noisy data obtained from complex systems is of considerable importance. Estimating the deterministic content of a monitored digital sequence (usually sequential or time-series) continues to evade satisfactory solution, despite considerable efforts over the years. This problem is central to most measured data and is especially important in applications when data is collected from medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors/reactions etc. The central problem is therefore to identify and/or separate from the composite signal, its deterministic and noise components.
One of the methods for the detection and/or separation of the deterministic components of the signal is by band-pass filtering using Fast Fourier Transform (FFT) (Cohen, L., Time-frequency analysis, Prentice Hall, Englewood Cliffs, 1995) where assumptions about the frequency distribution of the noise can be made and thereby suppress those frequency components before inversion to the time domain. Thus for example, low pass filters cut off the high frequency components in the signal. This is disadvantageous when all-frequency noise is present in the signal or when the signal component itself is broadband. Fourier filtering is especially unsatisfactory when the signal dynamics originate from systems exhibiting deterministic chaos.
Another method uses kernel estimators or spline estimators but a major disadvantage of these standard smoothing techniques is the fact that they do not resolve local structures well enough (Hardle, W., Applied nonparametric regression, Econometric Society Monographs, Cambridge University Press, 1990).
Another method used for the detection and separation of the composite signals is singular value analysis to reconstruct the dynamics (Broomhead, D. and King, G.P., Physica D, 20 217, 1986, Cawley, R and Hsu,G.-H, Phys. Rev. A. 46 3057, 1992). Here the original time-series is projected onto a subspace spanned by the corresponding singular vectors, namely those spanning the largest fraction of the total variance of the data. The remaining orthogonal directions then contain most of the noise. However, these methods are nontrivial and difficult to implement because of the abstract notions involved.
Yet another method which is very commonly employed in the recent times for the detection and separation of the deterministic components of signal is based on the theory of wavelet transforms (WT) (Vettererli, M. & Kovacevic, Wavelets and subband coding, Prentice Hall, PTR, Englewood Cliffs, 1995).
Wavelet transform methods are increasingly used as tools for studying multiscale, nonstationary processes in various fields. General methodologies for multiresolution signal processing, sub-band coding and also mutigrid techniques have led to applications in image and speech processing, data compression, communication, quantum mechanics, medicine, spectroscopy, etc. Wavelet transforms often complement Fourier transforms (FT) techniques because the spectrum of frequencies present in the signal may be analyzed at various scales with the added advantage of time localization. Methods based on the WT have proved to be efficient and easy to implement.
Wavelets are derived from rapidly oscillating functions with mean zero and obtained by suitable scaling of analyzing function to match the desired frequencies with simultaneous translations in time. A wide variety of analyzing functions amenable for discrete or continuous time analysis are known. (Strang, G and Nguyen, T Wavelets and Filter Banks, Wellesley-Cambridge Press, Wellesley, MA, 1996). Typical wavelet examples are: the discrete Haar (following a box function), the continuous Morlet (a sine function modulated by a Gaussian envelope), Mexican hat (second derivative of a Gaussian), Daubechies spanning a wide range of discrete and continuous properties (including compact support) depending on the number of wavelet filter coefficients,
Lemarie, biorthogonal spline, Malvar, Lemarie, Coiflet, Meyer, Symlet, etc. Concisely stated, WT is a generalization of the FT and is obtained by taking the inner product of a set of basis wavelets (dependent on the resolution, translation and choice of the wavelet) with the available data X. Systematic applications for varying oscillating scales a of the wavelet and also their localization b yields scale-frequency information in a scalogram of wavelet coefficients W, rather than the frequency periodogram in a FT analysis.
The hitherto used wavelet methods for the detection and separation of the composite signals are mainly based on the concept of thresholding the wavelet coefficients obtained from a single transformation. Hard thresholding cuts off coefficients below a certain threshold A while soft-thresholding reduces all coefficients at various scales by the threshold value (Donoho, D.L., Johnstone, I.M., Kerkyacharian, Picard, D., J. Royal Stat. Soc., Series B SI, 301). The threshold value is determined by statistical calculations and is dependent on the standard deviation of the noise (Nason, G.P., Wavelet regression by cross-validation, Dept. of Mathematics, Univ. of Bristol, 1994) and the length of the signal. The specified threshold value may also be used to evaluate special cost functions or entropy functional for generating appropriate basis. However, considerable improvements are still needed for applications in precision applications like medical diagnostics, seismographic analysis, image analysis etc., thereby resulting in incorrect inferences. This is because it is known that noise can be present at all frequencies and eliminating components based on statistical thresholding may not be adequate enough.
It will be advantageous and beneficial to have a process by means of which noise at all scales will be identified in a systematic and rational manner without thresholding based on properties of the signal. The present invention discloses an improved and systematic method for the identification and/or separation of composite signals into its deterministic and noisy components without the need for threshold values to be calculated and based on scientific rationale. The results of analysis also show precision and accuracy and lead to correct inferences.
The application of WT to reducing ultrasound speckle noise has been described in US patent no. 5497777 (General Electric Company, 1996). In another continuing with the same topic (US patent No. 5619998, General Electric Company, 1997), a procedure whereby a coherent imaging system signal is reduced of speckle noise by nonlinear adaptive thresholding of wavelet coefficients by WT has been described. The resulting image has an improved signal-to-noise ratio. The method followed thresholds the wavelet coefficients from the finest scale to the coarsest scale after dividing the imaging system signal into a number of subinterval signals of equal length. Coefficients in the various scales of the wavelet transform related to noise are thereby identified in each subinterval and these are eliminated before an inverse discrete wavelet transformation. It may be noted that the procedure followed is based on analyzing the scalogram of wavelet coefficients obtained by a single WT of data contained in a subinterval. The present invention is based on the principle of recursive wavelet
transformations, wherein the components are identified and separated based on their dimensionality and extent of correlatedness. It is generally known that noise is high dimensional and uncorrelated whereas the deterministic signal is usually of lower dimensional and correlated. The difference with respect to correlation in respect of noise and the deterministic signals comes out distinctly on taking recursive wavelet transformations of the wavelet coefficients at each scale. The recursive wavelet transforms result in separation of the signals into its noisy and deterministic parts. The process of separation occurs because the relative power distribution (RPD) remains constant for the noise while for signals with even short range correlations (e.g., deterministic chaotic signals) show considerable variations. For a composite signal containing both noise and deterministic components, eliminating those wavelet coefficients at each scale (identified by recursive transformations) which correspond to the noisy part yields the pure deterministic signal on recursively inverting using discrete wavelet transform.
The proposed method, therefore, involves taking wavelet transform of the data at different scales and this is resolved further using recursive wavelet transformation into a plurality of scales with time localization. The differences in the correlatedness and dimensionality of the two parts of the composite signal allows easy identification of those coefficients which belong to noisy part and those that belong to the true signal when subjected to recursive transformations. The identified wavelet coefficients that belong to noise are then eliminated and the inverse operation carried out to recover the true signal. The procedure guarantees identification of noisy components present at all scales and thus is devoid of any local/global approximations and thresholding. Removal
of noise is thus possible.
It is therefore the main object of the present invention to provide an apparatus
for the identification and/or separation of composite signals into its deterministic and noisy parts.
Another object of the present invention is to provide an apparatus . for the identification and/or separation of composite signal into its deterministic and noisy parts utilizing the concept of recursive wavelet transformation.
Still another object of the present invention is to provide an apparatus for the identification and/or separation of composite signals that can lead to better and more accurate inferences.
DESCRIPTION OF DRAWINGS
FIG. 1. Block diagram showing the apparatus and methodology.
FIG 2. Flow diagram showing the processor for the identification and/or
separation of composite signals into its deterministic and noisy components.
FIG. 3. A schematic of the behavior for noise (thin line bars) and a chaotic
signal (thick line bars) in Relative Power Distribution (RPD)
and Fractional Power Distribution (FPD)

p(k). i p(')wr ^en the wavelet coefficients of a dyadic scale j is recursively transformed k times, (a) The magnitude of RPD (thin line bars) for noise remains level due to its uncorrelated structure while for the deterministic signal it varies with k even for the short term correlation that is characteristic of chaotic dynamics, (b) The FPD shows a different decline for white noise than for the signal suggesting that this sensitivity can form a basis for their separation.
FIG. 4. Distinguishing noise and a deterministic chaotic signal by the outlined
method, (a) The bottom panel shows the Gaussian white noise [1024 data-points (bottom panel) with mean zero and variance unity] analyzed. The top panel shows the WT resolution at each dyadic scale
j and obtained from the wavelet coefficient matrix W(l) (by inverting
the wavelet coefficients for each scale W^> back to the time domain.). The decomposition by WT resolves the noise at different frequency components at all time /. (b) For each W^y (now treated as a data) the next generation of wavelet coefficients W*2) was obtained. The coefficients of those scales in W
transforms which results in \V(/(I) = 0. (c) WT of a short chaotic time-series data from a CSTR and analyzed similar to panel a), d) The signal in c) is detected as a signal arising from a deterministic process because
the resulting W,,(I)=W(1).
FIG.5. Signal detection with noise reduction for two representative signals, (a)
Filtering short noisy data (1024 points) X (black) from the nonlinear logistic map f-Vj,, =/ljr('(l-jr.) + jrn;A =3.2| eliminates the scatter
completely to yield the filtered data X(, (light gray) with the \tl
completely synchronizing with the underlying 2-period dynamics (gray)
|.v.(l =A.Y (l-x,);! = 3.21. (h) Noise reduction capabiliity for short
and noisy time-series data (1024 points) obtained from a CSTR exemplifying chaotic dynamics. Noise significantly distorts the signal as seen in the phase-plane plot of concentration CA(/) and temperature T(t) (black). The filtered signal (light gray) obtained after limiting the procedure to two generations shows remarkable recovery (light gray) of the sensitive deterministic dynamics (gray).
Accordingly the present invention provides for the identification and/or separation of composite signal into its deterministic and noisy parts which comprises: An apparatus for identification and/or separation of composite signals obtained from an instrument/equipment recording the variations in a system property as sequential or time-series data from the said instruments/equipments selected from medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors/reactions, into its deterministic (true signal) and noisy parts which comprises,
(a) source means for obtaining the signal to be identified and/or separated from suitable
sensors appropriately located in an apparatus/equipment;
(b) means for digitizing the said composite signal, obtained as sequential or time-series
data relating to a variation in a system property,
(c) computing means for subjecting the said digitized data obtained from step (b) above
to wavelet transform to obtain a scalogram in terms of wavelet coefficients;
(d) computing means for organizing the said resulting wavelet coefficients in each of
the scales to form new sets of data;
(e) computing means for taking each of the above said data sets and carry out their
wavelet transform to obtain another scalogram in terms of wavelet coefficients;

(f) computing means for implementing steps $}€ recursively with testing for the
constancy in the power distribution, the said power distribution being the ratio of the
power in a particular scale with respect to the total power in all the scales of that
scalogram for two consecutive recursive scalograms and thereby identifying the
recursive wavelet scales contributing to noise in the signal;
(g) computing means for eliminating the above said wavelet coefficients in the
recursive wavelet scales contributing to noise by setting them to zero;
(h) computing means for inverting above said the wavelet coefficients by inverse recursive wavelet transformation and thereby determining the deterministic signal component, the said signal component being the true signal is digitized form separated from the noise component.
(i) means for converting the above said digitized deterministic signal component to an analog signal using D to A converter
The signal to be identified and or separated according to the process of the present invention should be obtained by the use of suitable sensors appropriately located in the apparatus equipment.
In an embodiment of the present invention, we include conversion by analog-to-digital (A/D) apparatus of signals to yield rescaled digitized numbers that can be fed to a programmed device or a programmed general purpose digital computer, with a storage medium, capable of executing the steps of the said method for the purposes as claimed in claim 1 and 2.
In an embodiment of the present invention an improved method is claimed for data obtained even from deterministic systems exhibiting chaotic dynamics. It is also understood that the process of identification and/or separation of composite signals may be used for sequential data or time-series data obtained as variations in a system property in space or time; dividing the signal into a number of subinterval signals and applying recursive wavelet transformation to each subinterval signal.
In an embodiment of the present invention it is also understood that the process of identification and/or separation of composite signals may be used for discrete, biorthogonal, continuous wavelet (Haar, Morlet, Mexican hat, biorthogonal spline, Daubechies, Malvar, Lemarie, Coiflet, Meyer, Symlet, etc.) more preferably by Haar, and also the alternate methodologies for obtaining the wavelet transform and its inverse (fast wavelet transform algorithm, convolution, etc.).
The Haar wavelet is simple to understand and possesses features like signal reconstruction capability. Although other waveforms are known we shall use the Haar wavelet as a reference for its ease in discussion. The Haar wavelets are scaled forms of a square-wave function [#(/)=+! for 0 and#(/)=0 otherwise] and is given by 2j/2H(2jt-i) where 7'= 0,1 p-\ (for a
data set with N=2 p points). It is convenient to choose the wavelets for obtaining the WT to follow dyadic scales j, i.e., a=2~J with translations b=2'Ji (/ = 0,1,...,2;-1). The Haar matrix HN is (N x N) dimensional and forms an orthonormal basis with each row corresponding to a dyadic scale j and shift / in discrete time. The Haar WT may then be carried out as to obtain the N wavelet coefficients W (via,. W = 1 / ^/~N(HN • X) ; where X is the column vector of data and
1 / JN the normalizing factor) at (p - 1) dyadic scales with the last N12 coefficients for the smallest scale, the preceding N14 coefficients for the next scale etc. Note that the first coefficient relates to the average of the signal. Obtaining a reconstruction of the signal from W is simple inversion (i.e., Xr = reconstruction X = Xr. For a noisy signal X=Xd+Xn where Xd is the deterministic
-v
signal and Xn the noise (say, gaussian white noise). The aim in noise reduction would be to modulate the coefficients in W to Wd so that on reconstruction with Wd by inverse WT the deterministic signal Xd is recovered. The residual component corresponds to noise and may be obtained as Xn = X - Xd to obtain the noise content Xn. Viewed in this fashion, wavelets facilitate the separation of noise and signal, provided a convenient and intrinsic property differentiating the deterministic and noisy components is identified. In analogy, we may cite the example of separation by distillation of a mixture of two components A and B. In this process the intrinsic property utilized is the difference in volatility of the individual components. In the case of noise and the signal we suggest that the behavioral patterns of the power distribution at each scale show distinct trends on taking the WT of the wavelet coefficients, W, at each scale j. This feature is discussed more specifically next.
From a parent WT with coefficients Wll), it is possible to obtain a second generation recursive transform of its 7* scale coefficients, say, W(2-;). Note that in
principle later generation matrices of wavelet coefficients, W(M, k = 2,3,.... can be obtained and structured so as to obtain a tree of scalograms. The number of generations allowed depends on the number of initial data points N and the chosen starting j* scale from the parent W(l). It is interesting to note that for an orthonormal wavelet family, such as the Haar, the signal energy is similar to expressions in fourier theory with the power in each scale j given by P, =^ w>(0|2 with w>(0. they
scale elements of W(I). Similarly, the total power in the signal is therefore Ptot =^\w(i)\2;i = l,2,...,N . In general, the power P}m for a £* generation
wavelet transform in the y'"1 scale is conserved in the total power of generation, i.e., P^'^ = /j(i+1); = P(t);-. Now, calculations were carried out for
successive generations k and different scales j for the Relative Power Distribution (RPD), viz., P(k]j I P(k),ot and also Fractional Power Distribution (FPD) at a particular scale and recursive generation number k, viz., P^j I P^M for both gaussian white noise and chaotic signals. The results obtained, qualitatively sketched in Fig. 3, bring out two interesting observations. Firstly, it was observed that the RPD remained level for noise while those for the chaotic signal showed considerable variations (Fig. 3). Secondly, it was seen that the FPD at particular scales is different for noise when compared to the deterministic signal (Fig. 3) forming a basis for separation. This constancy in the RPD among scales along with a different decline in the FPD for uncorrelated noise for succeeding generations of the WT may be exploited profitably to reduce the effects of noise.
The aim here would be to obtain estimates of the wavelet coefficients, W(*';), in later generations obtained by recursion and scales (i.e., starting from the noise affected coefficients in the parent W(1)) such that recursive inversion results in W(1)d as the set of filtered coefficients. The method to obtain W(1) needs to ensure that constancy in RPD is seen at every scale for the tree of scalograms. At each stage when constancy is seen, the residual wavelet coefficients contribute to noise and these
are eliminated from the immediate prior generation before inversion. Strictly followed noise may then be completely eliminated, but this would require a data set of sufficient length to be analyzed. However, in practice, when the data set is short, a practical approach would be to limit the analysis to a specified number of generations and filter out those scales which do not cause abrupt changes in FPD on taking successive
transforms. The procedure can be carried out for every frequency scale of W^1' in a systematic manner and is easily coded for automation. The method is novel because of recursive wavelet transformation used in the said method. Note that the method of identification and/or separation of composite signal into its deterministic component and noise component is devoid of hard and soft thresholding of the wavelet coefficients at various scales based on statistical analysis of the wavelet coefficients. The results obtained using the above formalism for short time-series data of 1024 data points for noisy, chaotic and composite signals are shown in Figs. 4,5.
The invention is described in details with reference to the examples given below which are provided to illustrate the invention and therefore should not be construed to limit the scope of the invention.
Example 1: This example is provided to detect a signal solely consisting of noise
Using a random number generator digital data in the form of white noise with Gaussian Distribution (Press, W.H., Flannery, B.P., Teukolsky, S.A., Vetterling, W.T., Numerical Recipes, Cambridge University Press, 1986, p 191-203) was generated. This sequential data was subjected to recursive Wavelet Transformation using the Haar
wavelet basis function. A near constant Relative Power Distribution was immediately reached, thereby indicating the signal employed consists only of noise. A schematic of the behavior for noise (thin line box) in Relative Power Distribution (RPD)
P(k]i I P(k]tot and Fractional Power Distribution (FPD) P(k)/ I />(I)to, when the wavelet coefficients of a dyadic scale j is repeatedly transformed k times is shown in Fig. 3. Note that the magnitude of RPD (thin line box) for noise remains level due to
**
its uncorrelated structure The FPD for (thin line box) also shows a decline for the white noise signal for recursive transformations.
The results of experiments using the invention with signal having only noise component is shown in Fig. 4(a) and Fig. 4(b) of the drawing accompanying this specification. In Fig. 4(a) the signal is shown in the bottom panel while the wavelet transformation showing decomposition of the signal at dyadic scales and localized in time are displayed as a scalogram in the upper panel. Fig. 4(b) shows the resultant scalogram upon recursive wavelet transformation (up to k = 2 when near constancy in RPD was obtained) and then followed by recursive inverse wavelet transformation. The results clearly reveal that the method of the invention has precisely identified the entire signal employed as consisting only of noise and does not possess any deterministic component. Figure 4b shows the cleaned signal with the noise removed for transforms up to the second generation (i.e., k = 2 ) considered, i.e.,
Example 2: This example is provided to detect the signal solely consisting of true and deterministic signals.
A signal was generated corresponding to a first-order nonisothermal exothermic endothermic reactions taking place in a continuously stirred tank reactor (CSTR) (Kahlert, C; Rossler, O.E.; Varma, A; Springer Ser. in Chem. Phys., 18, 355 (1981).
The system parameters and their values used for obtaining the time-series data are Z)a=0.26;£=0.0;S=0.5;/c = 1.0;5=57.77;a = 0.42;x3c = 0.0;/?=8.0, where, Da= Damkohler number ; e = activation energy grouping ; K = ratio of activation energies for the exothermic and endothermic reaction steps ; B = adiabatic temperature rise grouping; a = ratio of heat effects; x3c = dimensionless coolant temperature; /? = heat transfer grouping. The system variables are the dimensionless temperature T\ and the concentrations C'A,C'B of the reactant A and intermediate product B, respectively.
For these parameter values the dynamics of the CSTR exhibits low dimensional deterministic chaos. The time-series data in temperature T is a dynamically fluctuating signal in time. This time-series signal was subjected to recursive wavelet transformation using the Haar wavelet basis function. The results obtained are given below.
Figure 3 schematically shows the Relative Power Distribution
(RPD) P(k]j I P(k\ot and Fractional Power Distribution (FPD) P(k]j I P(l\ot when the wavelet coefficients of a dyadic scale j is recursively transformed k times. From Fig. 3, the following observations may be made. Firstly, it is seen that the magnitude of RPD for the deterministic signal (thick line bars, Fig. 3 a) varies with k because of the short term correlations that is characteristic of chaotic dynamics and may be compared with that for the noise signal (thin line bars, Fig. 3 a) which remains level due to its uncorrelated structure. Secondly, the FPD for the chaotic deterministic signal (thick line bars, Fig. 3b) shows a different rate of decline from that for white noise (thin line bars, Fig. 3b) than for the signal, These features indicate the system sensitivity to these properties and show that they may be profitably exploited for identification/separation of deterministic and noisy components of a signal.
The results of experiments using the invention with the time-series signal having only the deterministic chaos component is shown in Fig. 4(c) and Fig. 4(d) of the drawing accompanying this specification. In Fig. 4(c) the time-series signal is shown in the bottom panel while the wavelet transformation showing decomposition of the signal at dyadic scales and localized in time are displayed as a scalogram in the upper panel. Fig. 4(d) shows the resultant scalogram upon recursive wavelet transformation and then followed by recursive inverse wavelet transformation. The results correctly reveal that the method of the invention has precisely identified the entire signal employed as consisting only of deterministic signal and does not possess any noise component. The results for the analysis of short time-series data from a system exhibiting deterministic chaos [viz., a continuously fed stirred tank reactor (CSTR) with an exothermic reaction show W(1)«/=W(I). In other words, the
constituent components at all scales j are not modulated whereby the processed time-series retains its original behavior as desired.
This result may be viewed with that obtained in example 1 where W^1' * W/1' =0. where the constituent components at all scales j are appropriately modulated whereby the processed time-series is shown to be noisy as desired.
Example 3: This example illustrated the identification and separation of combination of deterministic and noisy signals.
Composite signals with both deterministic and noise components were generated for two representative cases, viz.: a) biperiodic deterministic signal of a process following the logistic map (yielding sequential data) (Rasband, S.N., Chaotic dynamics of nonlinear systems, John Wiley & Sons, 1990) as its deterministic component with gaussian white noise (mean = 0; variance = 0.01) component, and b) deterministic chaotic dynamics from a process wherein a first order exo-endothermic reaction in a continuously stirred tank reactor (CSTR) as specified in Example 2 with gaussian white noise component (mean = 0; variance = 0.04). These data sets were subjected to recursive wavelet transformation using the Haar wavelet basis function. The denoising capability by the process of identification and/or separation of the deterministic and noise components by recursive wavelet transformation is shown below.
The resultant signals were subjected to recursive wavelet transformation till
k =2, i.e., using two recursive transformations. The wavelet coefficients were then inverted using the inverse recursive wavelet transformation. Both the examples show the remarkable robustness and reliability for noise removal from even short signals as given below and in Fig. 5.
Figure 5 shows the capability of the approach to simultaneously detect and reduce noise in the two specified and noisy signals of this example. Note that the signals are of varying complexity. The first signal typifies periodic data from the logistic map corrupted with noise while the other represents highly sensitive chaotic dynamics from the CSTR. The accurate recovery of the deterministic signal for the periodic signal is seen in Fig. 5a. Note that this approach is devoid of statistical inferences and local/global fitting procedures. Even with the extreme case of sensitive chaotic dynamics (where small errors can grow unbounded), Fig. 5b shows efficiency in signal recovery and noise reduction. In either case with just two recursive transformations the process of separation of the deterministic and noisy components
could be effected to over 90% (Calculated via, Extent of separation = (P} - P2 jl P\ ; where P, is the power of the noise component, and P2 is the power of the noise left in the signal after the process of separation by recursive wavelet transformation).
Advantages of the invention:
1. The method of the invention is simple and accurate because of which precise
inferences can be arrived at;
2. The identification and/or separation of the signals can be achieved efficiently by

limited number of recursive wavelet transformations for signals of short data length;
3. The method is not based on thresholding of wavelet coefficients based on
statistical inferences;
4. The method of the present invention is based on a procedure having a rational
scientific basis.
While only certain important features of the invention have been illustrated and described, many modifications and changes can be implemented by those skilled in the art. It is, therefore, to be understood that the appended claims are intended to cover all such modifications and changes as fall within the spirit of the invention. Thus, it is understood that the process of identification and/or separation of composite signals described may be used for discrete, biorthogonal, continuous wavelet types (Haar, Morlet, Mexican hat, biorthogonal spline, Daubechies, Malvar, Lemarie, Coiflet, Meyer, Symlet, etc.); alternate methodologies for obtaining the WT (Mallat fast wavelet transform, convolution etc.); data obtained from various sources (medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors/reactions, etc) or for variation in a system property in space; dividing the signal into a number of subinterval signals and applying recursive wavelet transformation to each subinterval signal, etc.

We Claim:
1. An apparatus for identification and/or separation of composite signals obtained from an instrument/equipment recording the variations in a system property as sequential or time-series data from the said instruments/equipments selected from medical diagnostic and scanning equipment, seismographic instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors/reactions, into its deterministic (true signal) and noisy parts which comprises,
(a) source means for obtaining the signal to be identified and/or separated from suitable
sensors appropriately located in an apparatus/equipment;
(b) means for digitizing the said composite signal, obtained as sequential or time-series
data relating to a variation in a system property,
(c) computing means for subjecting the said digitized data obtained from step (b) above
to wavelet transform to obtain a scalogram in terms of wavelet coefficients;
(d) computing means for organizing the said resulting wavelet coefficients in each of the
scales to form new sets of data;
(e) computing means for taking each of the above said data sets and carry out their
wavelet transform to obtain another scalogram in terms of wavelet coefficients;
(f) computing means for implementing steps ctf e. recursively with testing for the
constancy in the power distribution, the said power distribution being the ratio of the
power in a particular scale with respect to the total power in all the scales of that
scalogram for two consecutive recursive scalograms and thereby identifying the recursive
wavelet scales contributing to noise in the signal;
(g) computing means for eliminating the above said wavelet coefficients in the recursive
wavelet scales contributing to noise by setting them to zero;
(h) computing means for inverting above said the wavelet coefficients by inverse
recursive wavelet transformation and thereby determining the deterministic signal
component, the said signal component being the true signal in digitized form separated
from the noise component.
(i) means for converting the above said digitized deterministic signal component to an
analog signal using D to A converter
2. An apparatus as claimed in claim 1 wherein the means for wavelet transformation is
employed may be selected for discrete, biorthogonal, or continuous wavelet types.
3. An apparatus as claimed in claim 1 wherein the means for wavelet types employed
may be selected from Haar, Mlrlet, Mexican hat, biorthogonal spline, Daubechies,
Malvar, Lemarie, Coiflet, Meyer or Symlet.
4. An apparatus as claimed in claim 1 wherein source for obtaining sequential or time-
series data is selected from medical diagnostic and scanning equipment, seismographic
instruments, tomography, image analyzers, molecular spectroscopy, chemical reactors or
reactions.

5. An apparatus as claimed in claim 1 wherein the source for obtaining sequential or
time-series data may also be selected from deterministic systems exhibiting chaotic
dynamics.
6. An apparatus as claimed in claim 1 wherein means for hard and soft thresholding of
the wavelet coefficients at various scales is not required because of recursive wavelet
transformations used.
7. An apparatus as claimed in claim 1 wherein the digitizing means employed is an
analogue to digital converter.
8. An apparatus as claimed in claim 1 wherein converting means employed is a digital to
analogue converter.
9. An apparatus as claimed in claim wherein the computing means employed is a device
or a programmed general purpose digital computer, with a storate medium, capable of
executing the steps of the said method.
10. An apparatus for identification and/or separation of composite signals obtained from
an instrument/equipment recording the variations in a system property as sequential or
time-series data from medical diagnostic and scanning equipment, seismographic
instruments, tomography, image analyzers, molecular spectroscopy, chemical
reactors/reactions, into its deterministic (true signal) and noisy parts.

Documents:

1299-del-1998-abstract.pdf

1299-del-1998-claims.pdf

1299-del-1998-correspondence-others.pdf

1299-del-1998-correspondence-po.pdf

1299-del-1998-description (complete).pdf

1299-del-1998-drawings.pdf

1299-del-1998-form-1.pdf

1299-del-1998-form-19.pdf

1299-del-1998-form-3.pdf

1299-del-1998-petition-138.pdf

« Previous Patent

Next Patent »

Patent Number

214400

Indian Patent Application Number

1299/DEL/1998

PG Journal Number

08/2008

Publication Date

22-Feb-2008

Grant Date

11-Feb-2008

Date of Filing

15-May-1998

Name of Patentee

NARENDRA MADHUKAR DIXIT

Applicant Address

NATIONAL CHEMICAL LABORATORY, PUNE, MAHARASHTRA, INDIA.

Inventors:

#	Inventor's Name	Inventor's Address
1	VIRUTHIAMPARAMBATH RAVI KUMAR	NATIONAL CHEMICAL LABORATORY, PUNE, MAHARASHTRA, INDIA.
2	BHASKAR DATTATRAYA KULKARNI	NATIONAL CHEMICAL LABORATORY PUNE, MAHARASHTRA,
3	NARENDRA MADHUKAR DIXIT	NATIONAL CHEMICAL LABORATORY PUNE, MAHARASHTRA,
4	NITIN VAISH	NATIONAL CHEMICAL LABORATORY PUNE, MAHARASHTRA,

PCT International Classification Number

A61B 8/00

PCT International Application Number

N/A

PCT International Filing date

PCT Conventions:

#	PCT Application Number	Date of Convention	Priority Country
1			NA