WO1993003395A1

WO1993003395A1 - Method for detecting transient signals, particularly in acoustic signals

Info

Publication number: WO1993003395A1
Application number: PCT/FR1992/000760
Authority: WO
Inventors: Alain Lemer; Jean-Marie Nicolas; Jean-Pierre Pignon
Original assignee: Thomson-Csf
Priority date: 1991-08-02
Filing date: 1992-07-31
Publication date: 1993-02-18
Also published as: FR2680006A1; AU657792B2; CA2114632A1; EP0597040A1; FR2680006B1; AU2470892A

Abstract

The invention relates to methods allowing to detect in a continuous acoustic signal impulse acoustic signals. It provides for the determination of parameters (ai) of a filter by apprenticeship (109) by means of modeling by linear self-regressive prediction. Said parameters are then applied to a filter either of the adapted average type (110) or of the self-regressive type (111.1), which receives the reception signals and delivers a signal comprising a low level continuous portion and a high level impulse portion which represents the signals to be detected. Alternatively, a plurality of self-regressive filters (111.1)-(111.n) arranged in series may be used. The invention facilitates the detection of engine breakdowns.

Description

PROCESS ^f ADDE DETECTION SIGNAL TRANSITION, PARTICULARLY IN ACOUSTIC SIGNALS

The present invention relates to methods which make it possible to detect, and possibly identify, the transient acoustic signals characteristic of various phenomena which it is desired to study. It can be applied as well to aerial acoustic signals as to underwater acoustic signals, for applications as diverse as the identification of breakdowns in an automobile and the identification of ships.

Listening to acoustic signals, that is to say essentially noises, is very useful for spotting and identifying all kinds of things. In a trivial way, the driver of an automobile is often warned of an ongoing breakdown by mechanical noises that alert him. He then entrusts his car to the mechanic who is better able to identify with certainty the origin of the fault and to remedy it. Less commonly, sonar observers on board vessels listen to radiated noise in the marine environment and, when they detect something suspicious, entrust the task of final identification to an analyst with considerable experience. In both cases, the suspect noise must be sufficiently distinct, in amplitude and frequency, from the ambient noise to be able to attract the attention of either the motorist or the sonar operator.

As far as motorists are concerned, there are currently no such devices on the market, but as far as sonar operators are concerned, it is known to use sets of filters more or less adapted to the problems posed. These filters, whether low-pass, high-pass or band-pass, are studied to bring out the noises which one can think that they will appear, or possibly to eliminate them. In practice this only works well for noises perrrlanents. Reference has been made to such processing systems, as well as the possibility of repetitively listening to a recording of the signals received, in patent application No. 8715G73 filed on November 13, 1987 by the applicant on an invention relating to essentially a device for locating in direction by stereophonic listening of the signals received.

When the phenomenon is progressive, in particular when it is impulse noise, these filtering systems are very difficult, if not impossible, to implement because the operator cannot adjust quickly enough and sufficiently the ilter cutoff frequencies correctly.

To solve this problem, the invention proposes a method for assisting in the detection of transient signals, in particular in acoustic signals, mainly characterized in that in a first step the parameters of a filter are determined making it possible to modify the received signal to obtain a signal formed by a first substantially continuous and low-level part and a second impulse part having a level significantly higher than this low level and representing the transient signals to be detected, and that in a second step the signal is filtered reception by a filter having the parameters thus determined. Other features and advantages of the invention will appear clearly in the following description given with reference to the appended figure which represents the block diagram of a device making it possible to implement the method according to the invention. In the device represented in the appended figure, the electrical signal representative of the audio signal received, obtained for example from a hydrophone, is applied to an anti-aliasing filter 101 adapted to the maximum frequency received, to the maximum frequency audible by the operator (approximately 20 KHz) and at the sampling frequency of a converter analog-digital (102) located after this filter 101. This analog-digital converter delivers a digital signal Y (T) which is applied to several processing chains whose outputs lead to a switch 103. One of the inputs of this switch is directly connected to converter 102 in order to allow the operator to listen directly to the received signal. The output of the switch 103 is connected to the input of a digital-analog converter 104, followed by a low-pitch filter 105 and an amplifier 106 which supplies headphones 107. When the link between the converters 102 and 104 is direct via the switch 103, the operator provided with the headset 107 hears the same signal as that received by the hydrophone, the passage by the digital channel not causing significant distortion of this signal. To analyze this received signal and bring out anomalies which do not appear to it to direct listening, the operator directs the output signal of the converter 102, using a switch 108, to a learning module 109 This training module makes it possible to model the signal listened to according to a stationary broadband component synthesized by a predictive model, and a prediction residue corresponding to the difference between the signal predicted by the predictive model and the true signal. This modeling will preferably be carried out by the known method of autoregressive modeling of the signal with a P order. The order P of such an autoregressive model corresponds to the P temporal previous samples of the output of the autoregressive filter

y ((k-1) T _e ) _,. . . , y ((kP) T _e )

used to predict the current output y (kT) of the filter. In this preferred embodiment, these P predictor coefficients can be obtained in a known manner by solving the Yule-Walker equations using the Levinson algorithm. This algorithm converges quickly and we can get the P coefficients in the space of a few seconds during which the operator closes the switch 108. Optionally, this selection of the learning sequences can be carried out automatically, either repeatedly (for example 1 second every 10 seconds), or by detection of an abnormal phenomenon identified in any way. This gives a vector formed of P coefficients a., Which can be used in different ways, for example by reverse filtering or by so-called "caricatural" filtering. In the case of the use of reverse filtering, the coefficients a. are applied to a member 110, called the adapted average filter (MA) and which performs on the signal y (kT) delivered by the converter 102 the operation defined by the formula: P e (kT _e ) = y (KT _e ) - ∑ a. * y ((ki) T _e ) i = l The signal e (kT) obtained at the output of this adapted average filter MA is applied via the switch 103 to the converter 104, and the processed signal finally arrives on the headphones 107.

In the case of so-called caricatural filtering, recursive filters of the self-regressive type called AR are used which receive the coefficients a. and apply a processing defined by the formula to the input signal:

c _χ (kT _e ) = ∑ _i κ o ₁ ((ki) T _e ) + y (kT _e ) 1 = 1

According to the invention, in a first channel the signal at the output of converter 102 is applied to a first AR filter

111.1, the output of which is applied to the converter 104 via the switch 103. The resulting signal then arrives on the headphones 107. The inventors have proposed to call signal c. (kT) thus obtained, order 1 caricature of the signal s (t) because it has a power spectral density substantially equal to the square of the power spectral density of this signal s (t), which amounts to saying that the permanent noise contained in the signal s (t) has been "flattened" by bringing out the impulse noise contained in this same signal. This impulse noise is thus exaggerated by reducing the permanent noise, which corresponds pictorially to a caricature of the signal s (t). In this way the operator who listens to the signal resulting from this processing in the helmet 107 more easily perceives this impulse signal, which emerges better from the permanent signal.

The invention further proposes to further increase the influence of this processing by applying the signal at the output of the first AR filter in a second AR filter 111.2 then possibly in a succession of AR filters put in series up to an n th AR filter 111. n. The signal cj thus obtained at the output of the AR filter 111. j will quite naturally be called caricature of order j of the input signal, and its power spectral density is substantially equal to the power n of the power spectral density of the signal s . The treatment thus carried out is defined by the formula:

P c. (kT _e ) = ∑ a. xc ^ Uk-i) ^) + c ^ fl T ^ i = l

The outputs of the AR filters are respectively connected to inputs of the switch 103 and the operator can thus select, in addition to the direct signal and the signal filtered by reverse filtering using the MA filter, a caricature of order between 1 and n of the input signal. In practice, the operator after having learned will successively select the different outputs and then come back to the one which seems to him to have particularly interesting characteristics. Indeed the passage in such a system, whether by reverse filtering or by j-order caricature, deeply distorts the signal, since in the best of cases there is a continuous signal of very low level barely audible from which come out more or less regular pulses which are clearly perceptible to the operator but which are distorted with respect to the initial signal received. In itself this signal is very difficult to analyze, even by a trained operator, to obtain precise information on the source of the disturbing signal, because the deformations which are made to this disturbing signal, if they make it more detectable, distort it too much compared to the original signal to allow effective identification. Indeed the few particularly trained operators who can easily identify an undistorted signal, even if it is very weak, have such training on this natural signal that it is very difficult, if not impossible, for them to recognize a modified signal of this way that changes all their habits. In addition, these modifications are essentially variable according to the modalities brought, within the framework of the invention, to the processing chains, for example the order P of the auto-regression or the order j of the caricature.

The method according to the invention therefore essentially serves as a detection aid for an operator, who, after having identified the presence of a disturbing signal, may return for example to direct listening without filtering to concentrate his attention on the identification of the source of this disturbing signal. Possibly he can pass the baton to a more trained operator who will more easily carry out this identification.

Different variants of the process according to the invention can be used:

Thus with regard to the order of the model, it has been determined experimentally that an order 30, corresponding to samples of a few milliseconds, gives good results. To go further we can select this order using known criteria such as AKAIKE, RISSANEN or HANNAN.

As mentioned above, the invention has been described in the case where the selection of the learning range is made by the operator, and the possibility of using automatic criteria based on scales has been mentioned. of time. One could also use an automatic detector operating on the basis of other criteria, or other signals, or else continuous learning using, for example, self-recursive modeling algorithms in time such as that of MORF.

Similarly, we can use a different learning algorithm than that of LEVINSON, for example those of BURG, ITAKURA, GUEGUEN-LEROUX, MORF and FALCONER, or MORF and LEE. . .

Finally, even more generally, the invention extends beyond self-regressive analysis to any technique for modeling the initial signal which allows a decomposition into a set of parameters characteristic of the chosen model and an "innovation" ( difference between the prediction provided by the model and the true signal). We will cite for example the ARMA model, the KALMAN filter, neural networks. . .

Finally, in addition to assistance to a human operator, it is possible to integrate a device applying the method according to the invention, in order to obtain an automatic function making it possible to improve the performance of a device for detecting and automatically classifying acoustic signals. For example, it is possible to use the adapted medium filter to produce a pre-processing module making it possible to whiten and minimize the energy of a stationary signal, in order to favor the automatic detection of transient signals contained in an acoustic signal, by controlling for example the learning phases by a clock validated by a transient non-detection signal.

Claims

R E V E N D I C A T I O N S

1. Method for assisting in the detection of transient signals, in particular in acoustic signals, characterized in that in a first step (109), the parameters (a.) Of a filter (110) are used to modify the signal received to obtain a signal formed by a first substantially continuous and low-level part and a second impulse part having a level significantly higher than this low level and representing the transient signals to be detected, and that in a second step the reception signal by a filter (110) having the parameters thus determined.

2. Method according to claim 1, characterized in that to determine (109) the parameters of the filter, a predictive model with autoregressive analysis of order P is used. 3. Method according to claim 2, characterized in that the the P producing coefficients of the regressive model are determined by the Yule-alker equations with the Levinson algorithm.

4. Method according to any one of claims 2 and 3, characterized in that one uses a suitable average filter (110).

5. Method according to any one of claims 2 and 3, characterized in that one uses at least one auto-regressive filter (111.1). β. Method according to Claim 5, characterized in that several self-regressive filters (111.1-111. N) are used in series and having the same filtering coefficients.

7. Method according to any one of claims 1 to 6, characterized in that the duration during which the filter parameters are determined (109) is fixed manually (108). 8. Method according to any one of claims 1 to 6, characterized in that the duration during which the filter parameters are determined (109) is fixed automatically.

9. Method according to any one of claims 1 to 8, characterized in that an audio listening (107) of the filtered signal is used.