SYSTEM AND METHOD FOR STATISTICALLY SEPARATING AND CHARACTERIZING NOISE WHICH IS ADDED TO A SIGNAL OF A MACHINE OR A SYSTEM
Methods for determining variance properties of a noise component of a raw signal of a machine or a system. An example method includes recording a signal using a noise estimation unit, numerically differentiating the signal using a first module of the noise estimation unit to obtain a differentiated signal, identifying, using a second module of the noise estimation unit, a histogram which corresponds to the differentiated signal, and determining using the histogram, a variance property of the noise component of the signal.
Latest Nemor Properties LLC Patents:
This application is a continuation of pending U.S. patent application Ser. No. 12/471,196, filed May 22, 2009, which application is a continuation of U.S. patent application Ser. No. 11/109,061, filed Apr. 18, 2005, U.S. Pat. No. 7,552,154, which claims foreign priority benefits from Israeli Patent Application No. 166837, filed Feb. 10, 2005. The entire content of these applications and patent are incorporated herein by reference, in their entirety, for any purpose.
FIELD OF THE INVENTIONThe present invention relates to the field of estimating and characterizing noise which is added to a signal of a machine or a system. More particularly, the invention relates to a method and system for estimating and characterizing the component of the added noise of a signal. The method of the invention enables the finding of both the statistical nature and the type of the probability distribution or density functions of the noise component as well as its variance.
BACKGROUND OF THE INVENTIONThe importance of the knowledge of the fundamental properties of stochastic systems and processes has been recently acknowledged by a growing portion of the scientific and engineering community. Among other properties of stochastic processes, the nature of the noise component which contaminates the pure signal of the system is of major importance. The term “noisy signal” or “raw signal” whenever referred to in this application, refers to a signal which comprises a noise component and a pure signal which are inseparable. Throughout this application, the term “noise” refers to any random or unknown component whose exact behavior cannot be exactly predicted, but knowing its probability density function is highly valuable. Also, the term “variance” relates to the second moment of the probability density function and is used as is common in the art of Statistics and Probability theories. Moreover, throughout this application the terms “machine”, “system” and “process” are used interchangeably with respect to the method of the invention. An accurate estimation of the noise properties can provide to the system designer very important tools for improving the system behavior. An accurate determination of the noise properties is particularly important for dynamical systems where non-linear behavior is expected and in which the noise may seriously alter any estimation of the states of the system, if not to cause a total divergence of the parameters of the system model. Such conditions are particularly common in non-linear systems when modeled by recursive or adaptive methods such as Weiner or Kalman filtering. The principles and theory of Kalman and Weiner filtering are described, for example, in Gelb, A., “Applied Optimal Estimation”, Chapter 1, pp. 1-7, The MIT Press, Cambridge, Mass., 1974.
The following United States patents are believed to represent the state of the art for Signal estimation, noise characteristics, and Kalman and adaptive filtering in applicable systems: U.S. Pat. Nos. 6,829,534; 6,740,518; 6,718,259; 6,658,261; 6,836,679; 6,754,293; and 6,697,492.
The theory of non-linear filtering and its applications are discussed in:
- (a) Grewal, M. S. et al., Kalman Filtering, Prentice-Hall, 1993;
- (b) Jazwinski, A. H., Stochastic Processes and Filtering Theory, Academic Press, New York, 1970, chapters 1 and 2, pp. 1-13;
- (c) Gelb, A., Applied Optimal Estimation, The MIT Press, Cambridge, Mass., 1974 Chapter 1, pp. 1-7; and
- (d) Wiener, N., Journal of Mathematical and Physical Sciences 2, 132 (1923).
The art of signal processing, probability and stochastic processes and noise characteristics are also discussed in:
- (a) Bruno Aiazzi et al., IEEE Signal Processing Lett. 6 138 (1999);
- (b) R. Chandramouli et al., “Probability, Random Variables and Stochastic Processes”, A. Papoulis, McGraw-Hill USA, (1965);
- (c) IEEE Signal Processing Lett. 6 129;
- (d) Zbyszek P. Karkuszewski, Christopher Jarzynski, and Wojciech H. Zurek, Phys Rev. Lett. 89, 170405 (2002);
- (e) A. F. Faruqi and K. J. Turner Applied Mathematics and Computation, 115, 213 (2000);
- (f) J. P. M. Heald and J. Stark, Phys. Rev. Lett. 84, 2366 (2000);
- (g) A. A. Dorogovtsev, Stochastic Analysis and Random Maps in Hilbert Space, VSP Publishing, The Netherlands, (1994) (in particular see the consideration for high-order stochastic derivative in chap. 1);
- (h) H. Kleinert and S. V. Shabanov, Phys. Lett. A, 235, 105, (1997);
- (i) Elachi, C., Science, 209, 1073-1082, (1980);
- (j) Valeri Kontorovich et al., IEEE Signal Processing Lett. 3, 19 (1996);
- (k) Steve Kay., IEEE Signal Processing Lett. 5, 318 (1998);
- (l) Michael I. Tribelsky, Phys. Rev. Lett. 89, 070201 (2002).
The theory of curve fitting, differentiation and high order derivatives is discussed in:
- (a) G. Di Nunno, Pure Mathematics 12, 1, (2001); and
- (b) K. Weierstrass, Mathematische Werke, Bd. III, Berlin 1903, pp. 1-17.
It is an object of the present invention to provide a method for the statistical separation and determination of the noise properties from the noisy signal.
It is another object of the invention to provide such a method that can be performed in real-time.
It is still another object of the present invention to provide such a method for characterizing the noise which is adaptive.
It is still another object of the present invention to provide said method for characterizing the noise that can determine not only the variance of the noisy signal, but also the type of the probability density function (pdf) of the noise component.
It is still another object of the present invention to provide said method for characterizing the noise that does not depend on a priori knowledge of the structure of the pure signal.
It is still another object of the present invention to provide said method for characterizing the noise that does not depend on the structure of the pure signal.
It is still another object of the present invention to provide said method for characterizing the noise that involves defining a window of the analyzed signal, and given said window, the method does not depend on any accumulative information outside said window boundaries.
Other objects and advantages of the present invention will become apparent as the description proceeds.
SUMMARY OF THE INVENTIONThe present invention refers to a method for finding the probability density function and the variance properties of the noise component N of a raw signal S of a machine or a system, said raw signal S being combined of a pure signal component P and said noise component N, the method comprising the steps of: (a) defining a window within said raw signal; (b) recording the raw signal S; (c) numerically differentiating the raw signal S within the range of said window at least a number of times m to obtain an m order differentiated signal; (d) finding a histogram that best fits the m order differentiated signal; (e) finding a probability density function type that fits the distribution of the histogram; (f) determining the variance (or any equivalent parameter, depending on the specific said pdf type) of the histogram, said histogram variance being essentially the m order variance σ2(m) of the noise component N; and (g) knowing the histogram distribution type, and the m order variance or σ2(m) of the histogram, transforming the m order variance σ2(m) to the zero order variance σ2(0), said σ2(0) being the variance of the pdf of the noise component N, and wherein the histogram type as found in step (e) being the probability density function type of the noise component N.
Preferably, the method is repeatedly performed as the raw signal S progresses.
Preferably, the method is performed in real-time.
Preferably, the probability density function type that fits the distribution of the histogram is the one from a list that found to be best fitting the distribution of the histogram.
Preferably, the list comprises only one probability density function type.
Preferably, the one probability density function type is the Gaussian type.
Preferably, the transformation is performed by means of a specific expression suitable for the said fitted probability density function, wherein said specific expression is derived from the following general expression
Preferably, when the fitted probability distribution function is Gaussian, the transform is performed by means of the following specific expression
The present invention also relates to an apparatus for determining the probability density function type and the variance properties of the noise component N of a raw signal S of a machine or a system, said raw signal S being combined of a pure signal component P and said noise component N, the system comprises: (a) differentiating module, for receiving and numerically differentiating the raw signal S within the range of a predefined window at least a number of times m to obtain an m order differentiated signal; (b) a module for finding a histogram that best fits the m order differentiated signal; (c) a list containing at least one type of predefined probability density function; (d) a module for finding one probability density function type from said list that best fits the distribution of the histogram; (e) a module for determining the variance of the histogram, said histogram variance being essentially the m order variance σ2(m) of the noise component N; and (f) a module for, given the histogram distribution type and the m order variance σ(m) of the histogram, transforming the m order variance σ2(m) to the zero order variance σ2(0), said σ2(0) being the variance of the pdf of the noise component N, wherein the histogram type as found in step (d) being the probability density function type of the noise component N.
Preferably, the apparatus components operate repeatedly to find the updated probability density function type and the variance properties of the noise component as the signal S progresses.
The present invention also relates to a system for receiving a raw signal S which is combined from a pure signal P and a noise component N, and for outputting a signal which is essentially said pure signal, wherein the system comprises: (a) apparatus as described above for receiving said raw signal and outputting the probability density function and distribution type of said noise component into a filter; and (b) a filter receiving said raw signal and also receiving said probability density function and distribution type of said noise component from said apparatus, and given said received data, processing and outputting a signal which is essentially said pure signal.
Preferably, the filter is an adaptive filter.
Preferably, the filter is a Kalman Filter.
Preferably, the system of the invention as described above operates continuously in real time.
As said, the knowledge of the properties of the expected noise component provides to the system designer a very significant tool for improving the system structure and behavior. The essence of this invention is to perform relatively simple numerical calculations on the noisy signal in order to derive both the type of the probability density function of the noise and the properties of said of the probability density function and in particular the variance of said function.
The method of the present invention comprises of the following steps:
- 1. Defining a portion of the raw signal that may dynamically progress according to the development of the signal, hereinafter defined as the “window”, or the “analysis window”, provided that the number of elements in said window is statistically sufficient;
- 2. Recording the noisy signal S which is combined of a pure signal portion P and of a noise component N;
- 3. Numerically differentiating the raw signal at least a number of times m to obtain an m times differentiated signal;
- 4. Finding the histogram of the differentiated signal;
- 5. Providing a list of optional probability density functions, and from said list finding the one probability density function that is best fitted to the histogram distribution;
- 6. Determining from the best fitted probability density function the parameters that characterize that function, said function being the probability density function of the m order differentiated raw signal S, but essentially being a close approximation to the in order differentiated probability density function of the noise component N. The arguments for supporting this assumption are given hereinafter;
- 7. Transforming the parameters of the fitted, in order differentiated probability density function to extract the parameters of the zero order probability density function of the noise component N of signal S. The transformation is performed using an expression which is suitable for the fitted probability density function type (as will be elaborated later, expressions (3) and (4) which are given below are general expressions that are suitable for any type of probability density function, while the simplified expression (5) is suitable for Gaussian probability density function);
Now, the present invention will be described by means of an example.
One of the advantages of the invention as described is the fact that the method can be relatively easily performed in real-time, as the amount of data that is necessary for performing the analysis is relatively small, i.e., only to the extent of statistical validity. Moreover, the method requires the use of very limited amount of memory resources, as no historical data of the signal is advantageous. The only information necessary is that contained in the selected window, and the window in most cases can be narrow.
The present invention is applicable to most types of probability density functions. For each type of pdf one can easily derive the suitable expression as is necessary in step 7 above. Therefore, it is preferably recommended to keep in the list of step 5 above at least one type of probability density function, or preferably more, to keep those functions that are most expected for noise probability density functions.
Theoretical ConsiderationsConsidering a stochastic process ξ(ni), with ni the collection of stochastic events, in a measurable space (state-space) so that variance values of the stochastic variables considered here are finite, a differentiating operator, operating on a signal vector, may be defined with respect to the index of the signal data points in their sequenced order (or equivalently, treating the signal as a time series vector with a unit time step). By doing this, one may realize that a differentiation procedure, of the first order, is equivalent to numerical subtracting the element ni from the element ni+1, in the stochastic signal. Since in such random set of points each point is totally independent of all other points and correlated to any other data point within the set only by the mutual statistics of the sample space, denoted by Ω (i.e. all points (i, j) are uncorrelated where i≠′j), the equivalence to subtracting the element ni from the element ni+1 in the noise signal would be the equivalent of the subtraction of two independent Random Variables with identical statistical distribution (IID).
In contrast with the case of the first derivative, where one could assume that all individual data points were uncorrelated, higher order derivatives involve correlated expressions that lead, in the general case, to non-trivial expressions for the resultant probability functions.
Considering the above definitions and referring to some arbitrary random variable function V(ni, ξ), referred here as the original data signal with ξ as the stochastic random variable, one can now derive the second order derivative index series, V(2)(ni(2),ξ(2)), with ξ(2) refers to the (yet) unknown stochastic random variable corresponding to the second-order derivative vector by realizing that ni(1)=n1−ni−1 and ni+1(1)=ni+1−n1 so that ni(2)=ni+1(1)−ni1=ni+1−2ni+ni−1. These expressions imply that the probability density function of the second order derivative is the equivalent pdf of the sum of three independent, however non-identical, random variables (InID), all with similar, however not identical, probability density functions. Referring now to the general result that given two independent random variables ξ1 and ξ2 on the space Rk with μ and ν their respective distribution functions and f and g denote their respective density functions, than the distribution of the sum ξ1+ξ2 is the convolution μ*ν and the analogue density function of the sum equals the convolution integral denoted by f*g.
Using the notation fξ
fn
Following the above arguments, for higher derivatives, it can now easily be deduced that the m'th derivative of a random variable derived from an arbitrary statistically defined variable can be obtained by noting that the correlation elements that dictate the derivative expressions are given by the matrix (hereinafter: “the Stochastic-Derivative Matrix”):
and are governed by the following expression, denoted here as the Stochastic-Derivative matrix Skm, and given by
where
denote the elements of the binomial coefficients, and the Stochastic-Derivative matrix Skm, as defined above is in fact a variant of Pascal Triangle.
In terms of a summation of the individual elements needed to account for the probability density function of the m'th order numerical derivative, the summation may be written as:
wherein f(z) represents the probability density function of the original random variable. For instance, for the second derivative this is equivalent to fn
Fz(z)=P(Z≦z)=P((ξ1, ξ2, . . . ξm)εDz)
which gives:
Thus, in order to find the distribution probability function of the new random variable ξz, given the distribution functions of the random variables ξj's, one needs to define the range of the validity of the new variable z and to evaluate the integral using the mutual density function.
For the case of independent random variables, the above expression simplifies with the integrand replaced by
Since the density function is the same for all individual elements of the multiplication term under the integral, expression (3) can symbolically be written as:
wherein Fn
that can be easily evaluated to derive the respective density function, recalling that the term
really represents a convolution of the original probability function weighted accordingly.
The following discussion is focused on the case where the probability density function of the noise statistics is Gaussian. For the Gaussian case, the analysis yields a relatively straightforward expression as the Gaussian pdf belongs to the few probability functions that convolve into similar functions. A Gaussian distribution is therefore considered, where ξ is referred to as the random variable, N(0,σ02), i.e. a Gaussian distribution with the first moment equals zero, and the variance is given by σ02 as an illustrative probability (the derivation of the following with mean values other than zero is straightforward).
For the above, it can be found that the following expression (5) explicitly describes the resultant statistics, wherein β(m) is the sum of the squares of the elements of the m+1's row in the Stochastic-Derivative matrix given above, and α(m) is the inverse of the square-root of the sum of the squares of the elements of the m+1's row of the Stochastic-Derivative matrix given above.
Note that for a normal distribution function, as used above, the condition α∝1/√{square root over (β)} is required by the normalization condition.
Using equation (5) and the arguments above, the probability density function of a zero mean normal distribution for the exemplary cases of the first (equation 6), second (equation 7), and fifth (equation 8) derivatives respectively can be derived to be as follows:
This was indeed verified by numerical simulations where a normal distributed random set of 200K elements was generated (
In relation to the above, it should be clear that the histograms of the resultant vectors were then taken and are shown to have Gaussian shapes with variance values compatible with the above results.
Following the above theoretical considerations, it can obviously be concluded that expressions (3) and (4) can be used in the transformation step 7 above, while the simplified expression (5) can be used when the distribution is Gaussian.
To demonstrate one of the proposed motivations for the use of a high-order numerical derivative of a stochastic signal, we now refer to the derivation of the noise-level of an experimental output, where noise, either due to experimental set-up or due to the process itself (or due to both), is added to the signal. It is the aim of the following to demonstrate how to extract a simulated noise component such that the simulated noise is statistically identical to the noise part in the original experimental signal.
For simplicity we assume that the arbitrary noisy, raw signal can be represented by an arbitrary smooth and continuous signal contaminated by noise wherein S=P+N, N being the noise that is added to the pure signal P. Let us further assume that within the interval of validity of P, one can approximate P (for instance, in the Least Mean Square sense) by an m-degree polynomial function that may belong to a complete orthogonal polynomial basis. This can be proven to be possible for any bounded, smoothed and continuous function P (see for example the classical proof by K. Weierstrass, Mathematische Werke, Bd. III, Berlin 1903, pp. 1-17, and can also be found in most textbooks on Functional Analysis), but may be of practical use only when the interval is not too long, as compared to the structure of the signal, and for a relatively low polynomial degree.
Assuming the above, it turns out that
as the m'th derivative of P, under the above assumptions, is constant and thus vanishes for higher orders. For most experimental data, m would not exceed 5 (see the above example). However the present approach holds for any arbitrarily higher order.
Now, if the characteristics of the statistical properties of the high-order derivative of the original noise
is known, i.e. the probability density function that statistically describes the initial noise subject to high-order numerical derivative, in terms of the parameters (assumed to be unknown) of the statistical nature of the noise (assumed to be known), one can obtain the specific parameters of the original noise and thus deduce the noise-level in the original signal S.
It should be noted that the exemplary system of
While some embodiments of the invention have been described by way of illustration, it will be apparent that the invention can be put into practice with many modifications, variations and adaptations, and with the use of numerous equivalents or alternative solutions that are within the scope of persons skilled in the art, without departing from the spirit of the invention or exceeding the scope of the claims.
Claims
1. A method of determining a variance property of a noise component in a signal, the method comprising:
- recording the signal using a noise estimation unit;
- numerically differentiating the signal using a first module of the noise estimation unit to obtain a differentiated signal;
- using a second module of the noise estimation unit to identify a histogram which corresponds to the differentiated signal; and
- determining, using the histogram, a variance property of the noise component of the signal.
2. The method of claim 1, wherein a distribution of the histogram best fits the differentiated signal.
3. The method of claim 2, further comprising determining a probability density function type which best fits the distribution of the histogram.
4. The method of claim 3, wherein said determining a probability density function type comprises selecting the probability density function type from a library of probability density function types accessible by the noise estimation unit.
5. The method of claim 1, further comprising outputting the signal and the variance property to a filter.
6. The method of claim 1, further comprising processing, with a filter, the recorded signal and an output of the noise estimation unit.
7. The method of claim 6, wherein the filter is a Kalman filter or an adaptive filter.
8. The method of claim 1, wherein the signal corresponds to a raw signal comprising a noise component and a pure signal component, and wherein the method further comprises outputting an output signal corresponding to the pure signal component.
9. The method of claim 1, wherein the steps of the method are performed in real time.
10. The method of claim 1, further comprising transforming both the variance property and a probability density function type which best fits the histogram to obtain a zero-order variance.
11. The method of claim 1, wherein the duration of the signal is configurable in length.
12. An apparatus for determining a variance property of a noise component in a signal, the apparatus comprising:
- a noise estimation unit including: a differentiation module programmed to numerically differentiate a signal received by the noise estimation unit; a first module programmed to identify a histogram corresponding to the differentiated signal; and a second module programmed to determine, using the histogram, the variance property of the noise component of the signal.
13. The apparatus of claim 12, further comprising:
- a recording module configured to record the signal; and
- an output module configured to output one or more of the signal and the variance property.
14. The apparatus of claim 12, wherein the signal corresponds to a raw signal comprising a noise component and a pure signal component, and wherein the noise estimation unit is configured to determine a zero-order variance on the raw signal in real time.
15. The apparatus of claim 12, further comprising a transformation module programmed to transform both the variance property and a probability density function type which best fits the histogram to obtain a zero-order variance.
16. The apparatus of claim 12, wherein the first and second modules are the same module.
17. The apparatus of claim 12, wherein the noise estimation unit is configured to be in communication with a library comprising a plurality of probability density functions types, and wherein the noise estimation unit is further configured to select, from the plurality of probability density function types, a probability density function type which best fits the histogram.
18. The apparatus of claim 12, wherein the apparatus is configured to output the signal and the variance property to a filter.
19. The apparatus of claim 18, wherein the apparatus is further configured to output both the signal and the variance property to the filter in real time.
20. The apparatus of claim 12, wherein the duration of the signal is configurable in length.
Type: Application
Filed: Oct 8, 2012
Publication Date: Feb 7, 2013
Applicant: Nemor Properties LLC (Wilmington, DE)
Inventor: Nemor Properties LLC (Wilmington, DE)
Application Number: 13/647,223
International Classification: G06F 17/10 (20060101); G06F 17/14 (20060101); G06F 15/00 (20060101);