WO2012162825A1

WO2012162825A1 - Dynamic clustering of transient signals

Info

Publication number: WO2012162825A1
Application number: PCT/CA2012/050343
Authority: WO
Inventors: François LÉONARD
Original assignee: HYDRO-QUéBEC
Priority date: 2011-05-27
Filing date: 2012-05-25
Publication date: 2012-12-06
Also published as: JP6148229B2; CA2741202A1; JP2014517398A; CA2836421C; US20140100821A1; EP2715289B1; US10378937B2; EP2715289A1; EP2715289A4; CA2836421A1

Abstract

A method for clustering of transient signals is provided. The method comprises the steps of acquiring the transient signals as they come, dynamically building up clusters of similar transient signals in a hyperspace based on comparison and clustering rules so that each new one of the transient signal acquired ends up in a cluster with similar transient signals formerly acquired, analyzing the clusters to determine respective signatures defined by the transient signals gathered in the clusters, and processing the signatures to detect a phenomenon connectable to an intrinsic attribute of the transient signals.

Description

DYNAMIC CLUSTERING OF TRANSIENT SIGNALS FIELD OF THE INVENTION The present invention relates to signal analysis methods, and more particularly to a system and a method for dynamic clustering of transient signals.

BACKGROUND Transient signals are found in various areas such as radars, partial discharges, arcing noises (unsteady discharge in a plasma), stock price fluctuations, fluid cavitation, acoustic emission, telluric waves and imagery.

A recurring problem in many of these areas is that the distinct processing of each transient requires an exaggerated computation effort and targets a noisy signal.

US 6,088,658 (Yazici et al.), US 6,868,365 (Balan et al.), US 7,579,843 (Younsi et al.) and US 2008/0088314 (Younsi et al.) provide examples of prior art systems and methods for analyzing signals, involving time consuming and resource intensive computation and computer-related tasks.

SUMMARY An object of the invention is to provide a system and method of clustering transient signals, which regroup similar transients in a characteristic signature in order to process a reduced number of signatures for the groups, e.g. one signature per group. As there are less signatures, computation time is reduced. A signature being less noisy than its separated components, the processing result is all the more precise and the result is already classified.

According to one aspect of the present invention, there is provided a method for clustering of transient signals, comprising the steps of: acquiring the transient signals as they come;

dynamically building up clusters of similar transient signals in a hyperspace based on comparison and clustering rules so that each new one of the transient signal acquired ends up in a cluster with similar transient signals formerly acquired;

analysing the clusters to determine respective signatures defined by the transient signals gathered in the clusters; and

processing the signatures to detect a phenomenon connectable to an intrinsic attribute of the transient signals.

According to another aspect of the present invention, there is also provided a computer system having a processor and a memory storing executable instructions to be executed by the processor to perform the steps of the method. According to another aspect of the invention, there is provided a non-transitory tangible computer readable storage medium storing executable instructions to be executed by a computer system for performing the steps of the method.

BRIEF DESCRI PTION OF THE DRAWINGS

A detailed description of preferred embodiments will be given herein below with reference to the following drawings:

Figure 1 is a flow chart illustrating the method according to the invention.

Figure 2 is a schematic diagram illustrating two clusters.

Figure 3 is a graph illustrating cluster partial discharge signature of measurements (gray line) and one measurement snapshot (thin black line).

Figure 4 is a histogram illustrating a resolving power. Figure 5 is a histogram illustrating cluster coherency. Figure 6 is a histogram illustrating a number of clusters per sequence.

Figure 7 is a graph illustrating a decrease of the processing time with a c_dj_St parameter.

Figure 8 is a phase-resolved partial discharge diagram of superimposed clusters. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

In the present disclosure, the time or the space can be replaced by any other dimension of any other nature. The present invention is directed to transient signals whose characteristics are to be repetitive for a portion of their population. By repetitive, it is to be understood that a same transient may be observed more than one time in the time or the space, with an amplitude that may vary and also with a low dissimilarity that may be explained by the noise, a measurement error, a time or spatial distortion of the transient carrier or any other modelizable phenomenon (digitally, analytically or statistically).

The present invention proposes to dynamically regroup the transients, i.e. as they are captured by the measurement system or acquired by an appropriate device. By regrouping, it is to be understood that similar transients are assembled into a same group, a given group thus containing at least one transient, and the result yielding at least one group having more than one transient associated therewith. A grouping that is non-dynamic means that a comparison is made with all the transients on hand; although closer to optimality, the computation time of this approach is exhaustive.

Presented in a space ¾^N, where N is a number of temporal or spatial dots characterizing a transient, a grouping appears such as a cloud of dots in this hyperspace. The center of mass of the grouping that corresponds to the means of the transients of the grouping will be called signature.

The transient to transient, transient to signature or signature to signature comparison requires a time or space shift in order to maximize the correlation or minimize the distance between both elements of the comparison. In the case of a distance based comparison criteria, at zero order, the shift is achieved in one block such that

depicts the distance between the transient X and the signature S. It is possible to achieve a first order shift by interpolating the transient (or the signature) so as to stretch or compress it. It is the same for the second order. Furthermore, a dynamic time warping type of method may be considered as a shifting means for the distance computation. Basically, an appropriate shift between both elements of the comparison is to occur in the comparison.

Referring to Figure 1 , the method according to the invention proceeds as follows. Transients are captured as depicted by block 2 and are subjected to a clustering as depicted by block 4 for dynamic cluster construction with cluster signatures. The signatures are analyzed as depicted by block 6 so as to determine signature features usable for further processing related to the nature of the transients under scrutiny as depicted by block 8. The method may be implemented in a computer system having a processor and a memory storing executable instructions to be executed by the processor to perform the above recited steps. The method may also take the form of a non-transitory tangible computer readable storage medium storing executable instructions to be executed by a computer system for performing the steps of the method.

The following provides an example of an embodiment of the invention in the context of partial discharge detection, location and analysis. It should be understood that the invention is not limited to such embodiment and application, and that changes and modifications can be made without departing from the invention.

The method according to the invention may be used for voltage transient classification by a dynamic time clustering. Inside an underground vault of an electrical distribution network, many hundred signal transients may be captured in a few seconds, many of which are partial discharges (PD). Signal processing and pattern recognition for each transient are time consuming. Grouping M transients into / clusters dramatically reduces the processing time and significantly increases the signal-to-noise ratio of the corresponding / signatures. The clustering may be done over many hundred dimensions Λ/, with each dimension corresponding to a signal time sample. Since the time position of a transient is corrupted by a time jitter, the distance function is calculated for T different time alignments. A heuristic similar to the k-means algorithm is explained based on the "sphere hardening" phenomenon and has a 0(7 x N x M x l) complexity for / clusters. Different tools are proposed for assessing the accuracy of the clustering process and optimizing some parameters of the method. The signals to be processed according to the invention may be sampled e.g. at 1 Gs/s, filtered, interpolated and truncated. A few hundred time samples N describe the transient pattern. For a PD emission location "/^'", assuming a constant normalized time signature s, (f) over different amplitudes, there is obtained

x_m (t) = a_m - Si (t - t_m) + n_m(t) (1 ) the realization of the measurement "m" of a transient signature, where a_m is the realization amplitude, t_m the realization delay and n_m the additive noise. The corresponding modeling is

Xmn = ^am ^' Sn-d_m + ⁿmn (2) with discretization. The successive measurements taken over one analog input scale are called a sequence. The a_m dynamic range may be less than 10 dB for a sequence, i.e. the ratio of the clipping level on the trigger setting level. The first clustering step may be performed for a fixed scale. The full dynamic range can then be obtained in a second step by merging the clusters obtained from different A/D scales. In some cases, the a_m dynamic range may exceed 30 dB. Assuming a Gaussian noise and disregarding the a_m dynamic, the ¾^w projection shows a hypersphere centered on the 'ϊ' signature

&i,1 > Si,2 > - - -> Si,N (3) where the measurements

are close to the hypersphere boundary. The boundary thickness is a function of the metric, the measurement signal-to-noise ratio (SNR) and the number of time samples N. For an Euclidian metric, the distance

has the expected mean

r_i = E(D_m, ) = ^NE[n_m ² _n ) (6) for x_m e cluster "/^'" and the standard deviation σ, =V¾). (7)

Referring to Figure 2, the radius and hypersphere boundary thicknesses are respectively illustrated in ¾^w, r, and 2 σ,. The ratio of the boundary thickness on the hypersphere radius tends to 0 when N→∞. This phenomenon is called sphere hardening. Calculated using numerous noise samples, the distance X_m- Si is barely constant. There are no measurements in the hypersphere except near its boundary.

In ¾^w, cluster probability densities appear like distributed shells with similar radius and thickness. With the presence of a significant a_m dynamic, the single point signature is replaced by a rod pointing to the axis origin. The corresponding shell is dilated along the axis of the rod. The shell thickness is increased in the rod's direction. Signatures appear like distributed shells with a dissimilar elongation function of the amplitude dynamic. The use of an appropriate metric can partially overcome this shell distortion. In dynamic clustering, the number of clusters and the cluster centroid location may be adjusted dynamically. The working dimension may be limited by l_max, the maximum allowable signatures and l_min, the minimum allowable signatures. A running estimation of the average distance measurement-to-signature may be done during the process of a sequence. The average distance includes the contribution of the noise and the amplitude dynamic. The maximum allowed distance is defined as the average distance r multiplied by a distance coefficient Cdist. This coefficient may be set at about 1 .5. The process may proceed on the basis of comparison and clustering rules as follows:

- storing the first l_min measurements as new signatures;

- for the next measurements, finding the nearest distance measurement-to-the signatures;

- if the nearest distance is less than the allowed maximum distance,

- then merging the measurement to the nearest signature and re-estimating the average distance r,

- else calculating the signature-to-signature distances and finding the nearest signature-to-signature distance;

- if the distance is less than the allowed maximum distance,

- then merging the two nearest signatures and re-estimating the average distance r,

- else putting the measurement as a new signature or, if l_max is exceeded, forcing the measurement to merge with the nearest signature;

- after the last measurement, the span of time alignment is doubled, for 27^~- 7^~:

1 . calculating the signature-to-signature distances;

2. finding the nearest signature-to-signature distance;

3. if the nearest distance is less than the allowed maximum distance:

4. then, merging the two nearest signatures, calculating the signatures to the new signature distance, inserting the result in the signature-to- signature distance triangle matrix, and going back to step 2; 5. else, stopping.

Sphere-hardening may be used to refine results: measurements distant from the sphere boundary are reprocessed and may be reallocated to another

cluster. The threshold may be fixed proportionally to where var(Oj) is the variance uncertainty on σ, estimation.

Assuming /",^■»/}^■, the average distance r is recursively estimated from the minimum distance D_m calculated during the cluster building. This average distan

includes the noise of the measurement and the signature position error. The right factor, a function of the cluster population P,, takes into account the signature position variance. The b coefficient is a weighting factor > 1 .

The merge process S, u S , → S, ' is

S - - (PiSi + PjSj ) (9)

P, + Pj

for the new signature calculation where P, and P, are the cluster populations. Before the merge, the signatures are time-aligned with respect to the minimum distance. The signature time alignment is also weighted by the cluster population

where d is the distance (Eq. 5) between the signatures, expressed in number of samples. The order of arrival of the measurements slightly affects the final result, but at the end k,n -dj< (1 1 )

Among various metrics, the minimum of the square distance

calculated over T tested time alignments may yield the best results. Note that minimizing the right term is like maximizing correlation. Maximizing only the correlation may be ineffectual since many small noise patterns will correlate with some PD signatures. With this metric, the contribution of the PD amplitude variation is reduced by the second term.

Correlation (i.e. Bravais-Pearson coefficient) can be used in the second clustering performed to merge the clusters generated by the different measuring scales. In this second step, the cluster signatures SNR is high and no mistake can occur between a PD cluster signature and a noise signature.

The calculation complexity is 0(N x T x M x l) for the distance measurement to signatures and 0(N ^χ Τ ^χ M ^χ /(/ - l)/2) for the signature to signature distance triangle matrix. Coefficients c_di_St, l_max and l_min set a compromise between the calculation time and the wrong merge probability. For numerous measurements, when the number of clusters is stabilized, the calculation of the signature-to- signature distances is no longer needed: the overall calculation complexity tends to 0(/V ^χ Τ ^χ Μ ^χ ΐ) .

Accuracy measurements may include some information on cluster dispersion and cluster superimposition. Since the latter information is constant over a data set, the idea is to optimize the process using one or more accuracy estimates that are representative of clustering errors. Proposed estimates are based on two opposite directions. On the one hand, the resolving power, defined as the ratio of the inter-cluster distance over the cluster radius, pertains to the inter- cluster overlapping. On the other hand, the coherency, the ratio of the coherent energy over the total cluster energy, pertains to the cluster itself.

The resolving power ¾ is a measure of the cluster-resolving ability. The resolvin power

corresponds to a signal-to-noise ratio where the RMS cluster inter-distance is the signal and RMS cluster radius is the noise. The equality

0.5 - l(l -Ό

explains the denominator of the first term and ∑P_j = M .

The coherency

of the cluster 'ϊ' is calculated from the P, members of this cluster. The mean coherency

is defined for a sequence considering all the contributing clusters. The following hypothesis may be used: the resolving power and the mean coherency are at their maximum levels for the best solution and decrease with the accumulating errors in the clustering process.

Figure 3 shows a typical PD signature. The results have been obtained for 1 730 sequences (1 18 932 measurements) taken in an underground distribution network. The first lobe is the sharpest and sets the PD polarity to a negative value here. The range of the resolving power values, 1 to 10 000, calls for logarithm (dB) scaling. The distance coefficient c_di_St must be set less than the resolving power. To the left of Figure 4, the 20 counts at 0 dB for c_d;_sf=10 are explained as clustering errors since c_di_St> resolving power.

The cluster coherency is altered mainly for sequences with small resolving power. On Figure 5, most of the sequences have a large resolving power: shifting the value of c_dist to 10 has little effect on coherency. However, the same shift significantly decreases the cluster population of some of the sequences (Figure 6). Figure 6 illustrates that the maximum allowable signatures l_max can be set at 30. Figure 7 shows the processing time decreasing with c_di_St- The compromise between the processing time and clustering error probability is set by the data itself and the desired confidence for diagnostic purposes. Based on these latter figures and other results, set Q;_sf=1 .5 targets more than 99% of the data population and yields a processing time close to the measuring time (« 8 s).

The time domain transient clustering contribution is recognized in the PRPD diagrams with a cluster discrimination. Figure 8 highlights a cluster of 65 measurements among nine other clusters for a total of 1 182 measurements. The clustering of transient signals in a time domain as the first step analysis, with further signal processing applied on the cluster signatures, is advantageous at least in those ways: (1 ) the information is reduced to some signatures instead of numerous measurements; (2) the SNR signature increases with the cluster population; and (3) the post-processing time is reduced. Moreover, the superimposed clusters are discriminated in the PRPD diagram. The explained suboptimal heuristic appears as fast and accurate. Testing using numerous field data may be achieved to tune the parameters and set the choice of metrics in the method according to the invention.

Claims

CLAIMS:

1 . A method for clustering of transient signals, comprising the steps of: acquiring the transient signals;

analyzing the clusters to determine respective signatures defined by the transient signals gathered in the clusters; and

2. The method according to claim 1 , wherein the building up of the clusters comprises realigning the acquired transient signals and carrying out at least one of a transient-to-transient, transient-to-signature, and signature-to-signature distance comparison.

3. The method according to claim 1 , wherein the building up of the clusters begins with clusters formed of a number of initial transient signals acquired.

4. The method according to claim 1 , wherein the building up of the clusters comprises merging similar transient signals and similar clusters and computing new parameters of the comparison and clustering rules applicable to the clustering of new transient signals acquired.

5. The method according to claim 4, wherein the parameters comprise distance calculation metrics.

6. A computer system having a processor and a memory storing executable instructions to be executed by the processor to perform the steps of the method according to any one of claims 1 to 5.

7. A non-transitory tangible computer readable storage medium storing executable instructions to be executed by a computer system for performing the steps of the method according to any one of claims 1 to 5.