Process and apparatus for extracting a useful signal having a finite spatial extension at all times and which is variable with time

Description Submit all comments and votes

TECHNICAL FIELD

The present invention relates to a process and to a device for the extraction of a useful signal having a finite spatial extension at all times and which varies in time.

STATE OF THE ART

The prior art processes making it possible to extract a useful signal from a signal received by one or more sensors or transducers can be differentiated as a function of the number of sensors or transducers used. Thus, a distinction can be made between monodimensional processes and multidimensional processes with or without noise reference.

In monodimensional processes the information from a single sensor only makes it possible to use conventional filtering methods (time or frequency based). The coverage of a monitoring zone can only be obtained by moving the sensor, which can lead to difficultly solvable practical problems.

In particular, the environment close to the sensor can have an interfering effect, because it may be integrated into a transportation system

The extraction of a useful signal is of an optimum nature (with a single signal) using the matched filtering method when the additive noise is white. This rarely fulfilled condition can be approached following a preliminary prewhitening operation. However, the filtering does not use the space coherence property of the noises and remains of a suboptimum nature compared with all multisensor methods which can make use of the spatial predictability properties of the noise.

Monodimensional processes can be illustrated by an application in the magnetic field. The methods and applications described in the Institut National Polytechnique de Grenoble Thesis of 1979 by R. Blanpain entitled "Real time processing of the signal from a magnetometer probe for the detection of magnetic anomalies" carry out a space--time recording of the magnetic field. A single sensor is moved over the area to be monitored. It records the geomagnetic noise, the geological noise (because the probe is moving rapidly here) and a possible useful signal. The well known method of matched filtering is then put into effect in order to eliminate in the best possible way the noises deteriorating the useful signal and perform a detection. This method is of an optimum nature in the case where the noise accompanying the signal is white, which is not the case here. Therefore, prior to any filtering, a prewhitening is necessary. However, prewhitening is difficult for reasons of the non-stationary nature of the noises. Thus, in practice the mean or auto-matching whitening filter only performs a suboptimum operation. Non-white geomagnetic noise residues remain and disturb the matched filtering operations.

In such monodimensional processes, the separation performed is consequently limited, because it cannot perform the spatial filtering in view of the fact that there is only one measuring sensor. The space coherence properties of the geomagnetic fluctuations are not used. This process can be effectively completed by the system proposed in the invention, which makes it possible to perform an effective filtering of the input unprocessed signals using their spatial properties.

There are numerous procedures in connection with multidimensional processes. They are combined within the general theory of processing multidimensional signals, e.g. in the article entitled "Models and processing of multidimensional signals" by J. L. Lacoume ("Traitement du Signal", vol. 5, no. 2, 1988). Consideration will be given to those which would appear to be representative and have an application in the solving of the problem defined hereinbefore. The case of magnetic detection illustrates the applications. The processes are classified in accordance with the presence of a "noise reference". Thus, the fact of knowing one or more sensors only recording noise is an advantage and use is made of this by noise subtraction methods. It is pointed out that up to now magnetic arrays or networks have not often been used and the following processing operations have been really employed on magnetic signals. Other fields such as sound detection/locating use them to a significant extent.

With noise references, said processes use an array of sensors called "noise reference", which only record noise, e.g. in the vicinity of the area to be monitored. It is necessary to have at least the same number of noise references as there are independent noises. The reference sensors are able to measure a physical phenomenon of a different nature to that of the useful signal (it is possible to filter a magnetic signal with the aid of a signal from e.g. a pressure sensor or transducer, if these two signals have a correlation). The transfer functions from the "noise reference" sensors to the useful signal sensors are identified. Therefore the noise is predicted and subtracted from the total signal. These noise subtraction processes are completely described in the article by D. Baudois, C. Serviere and A. Silvent entitled "Noise subtraction--bibliographic analysis and synthesis" ("Traitement du Signal", vol. 6, no. 5, 1989).

They can only rarely be applied in the operational context for array detection, because they assume the knowledge of all the noise only sensors. This hypothesis is not made in the process according to the invention. The path or trajectory of the magnetic target is not known beforehand and the partitioning information of the sensors E is not available. It is also shown that noise subtraction does not withstand errors made in partitioning or subdividing the sensors into noise only sensors and useful sensors. Therefore such a process is not suitable for the set problem, but still remains of an optimum nature when the group of useful sensors E.sub.su and the group of noise sensors E.sub.ref are fixed and known a priori.

Without a noise reference, the second class of signal separation systems is still based on source independence and space coherence properties. When all the sensors receive the useful signal or all the noise only sensors are not known, a priori all the sensors have the same function. Conventional antenna processing processes make it possible to carry out a filtering of the sum of the spatially coherent signals in such a way as to attenuate spatially white noises (i.e. totally incoherent in space). By hypothesis, the signals must be stationary or slowly evolutive. The larger the number of sensors the better the separation obtained. These processes are not applicable in magnetic detection, because the signal to noise ratio gain is too low to be satisfactory, the magnetic networks having few sensors and the useful signal is neither spatially white, nor spatially coherent.

Finally, the processing processes using statistics with orders equal to or higher than two make it possible to carry out a blind separation of a linear combination of filters of q source signals reaching N sensors based solely on the independence property of the sources. They constitute an extension of antenna processing processes to statistical orders higher than two and are still based on the signal stationarity hypothesis. Moreover, the processing of broad band signals requires significant theoretical developments. Consideration is given here to a pulse-type useful signal having a limited time extension, having non-stationary properties and of a broad band nature, which is not appropriate for these methods.

Thus, the processing processes by using statistics of orders equal to or higher than two are not suitable for the processing of magnetic networks because of the small number of sensors used and the extreme non-stationarity of the useful signal.

DESCRIPTION OF THE INVENTION

The object of the invention is to bring about the detection of a time-variable useful signal having a finite spatial extension and the separation of spatially coherent additive noises having a considerable extension compared with that of the useful signal.

To this end it proposes a process for the extraction of a time-variable useful signal of finite spatial extension by an array of N sensors, N being equal to or greater than 3, receiving said useful signal to which have been added q spatially coherent additive noises, q being less than N, said process comprising the following stages:

a stage of acquiring unprocessed signals on the output of each sensor,

a stage of band-pass filtering said signals in order to restrict to the frequency band of the useful signals,

a stage of digitizing said filtered signals, characterized in that it then comprises:

a stage of calculating space prediction error signals of the noise during which:

a) a particular sensor from the array of N sensors is chosen,

b) the remaining N-1 sensors are distributed into groups of the same size having q sensors, whereby the same sensor can belong to more than group, and one group is used for constructing a prediction error signal if the q signals of the group are independent,

c) for each admissible group of q sensors is constructed a spatial prediction of the signal of the sensor chosen in stage a) in the following way:

q transfer functions inherent in the chosen admissible group of q sensors and the sensor chosen in stage a) are constructed with the aid of elements of intersensor transfer functions characteristic of the distribution of the noises at all times and applied respectively to the signals of the sensors of the admissible group of q sensors considered,

the q thus constructed signals are combined for each group in order to construct the prediction signal of the sensor chosen in stage a),

d) the prediction signal of the sensor chosen in stage a) is compared by a comparison operator with the signal on the sensor chosen in stage a) in order to construct a prediction error signal on the sensor chosen in stage a),

a stage of analyzing prediction error signals so as to carry out the detection of the useful signal and its separation from the q additive noises.

Advantageously, the analysis stage comprises:

a stage of calculating detection indexes,

a stage of generating, at all times a subdivision of the array of sensors into an array of sensors receiving the useful signal and noise and an array of sensors only receiving the noise, and weightings corresponding to said subdivision,

a weighted projection stage constituted by two substages:

a first substage of associating the thus calculated weighting with the signal of each corresponding sensor for generating N weighted signals,

a second substage of applying an antenna processing method to the N weighted signals in order to carry out a source space/noise space separation, knowing transfer functions of the noises, the N signals of the noise space being estimates of the useful signal present in each channel of the initial signal.

Advantageously, stages a), b) c) and d) are performed simultaneously for the N sensors and the admissible groups of q sensors of the array and this takes place at all times.

Advantageously, prior to the stage of calculating the space prediction error signals of the noise, there is an estimate with respect to the characteristic transfer functions of the propagation of the noises at all times with the aid of a recording extract of the signal of the array during which each useful signal is present.

In different special embodiments one or more of the following characteristics are encountered:

during the stage of calculating the space prediction error signals of the noise the combination of q signals is an addition,

during the stage of calculating the space prediction error signals of the noise, the comparison of the prediction signal of the chosen sensor with the signal of said chosen sensor is a subtraction,

during the stage of associating the weighting with the signal of the corresponding sensor, use is made of a multiplication.

The process according to the invention makes it possible to use a relevant supplementary information relative to the partitioning or subdivision of the array of sensors into two subarrays, the subarray of the noise sensors only and the subarray of the useful sensors.

The interest of this process is that it is resistant to possible subdivision errors, because it produces a mean value (property of antenna processing), unlike in the case of noise subtraction processes, which require a certain subdivision of the sensors (which is generally unavailable e.g. for magnetic networks) and which have a poor resistance to classification errors.

Compared with conventional antenna processing processes, subdivision or partition makes it possible to avoid rough errors by automatically rejecting sensors which may have received the useful signal. There is a considerable reduction (in proportions dependent on the quality of the estimator of the space prediction functions and the construction of the detection indices) of the defects due to the projection of part of the useful signal into e.g. the geomagnetic noise space. The quality of possible subsequent processings of the useful signal, such as e.g. the application of methods for locating the source of the useful signal, is therefore significantly improved.

Thus, the weighted projection controlled by the expert system is located between the noise subtraction processes, which do not have a good resistance to partitioning errors, but do have a good signal to noise ratio, and conventional antenna processing processes, which have a good resistance, but generate significant defects at the output. The process according to the invention carries out a filtering incorporating at all times the sensors which only receive noise, combining the advantage of a good signal to noise ratio at the output and a good resistance.

The invention also relates to a device for the extraction of a time-variable useful signal with a finite spatial extension from a signal incorporating said useful signal to which are added q spatially coherent additive noises and having a large extension compared with that of the useful signal, characterized in that it comprises an array of N sensors, N being equal to or greater than 3, N strictly exceeding q, said N sensors being adequately spaced so that the useful signal is unable to touch all the sensors at once, the area monitored by each sensor of the array not containing all the other sensors, said N sensors being followed by N filtering modules, N digitizing modules, a module for calculating the space prediction error signals of the noise, a module for calculating the detection indexes, a real time expert system module and a weighted projection module.

Advantageously, the array of sensors can be constituted by sensors of different types. The complementary nature of the measurements thus makes it possible to obtain a better detection of a physical phenomenon by its different effects (pressure, electromagnetic, acoustic, etc.).

Each sensor can be a "gradientmeter", i.e. can be constituted by several slightly spaced sensors of the same type and between which a difference is formed.

Advantageously, the N sensors are adequately spaced to enable the detection perimeter of each sensor of the array not to contain all the sensors at once. The maximum intersensor space can be approximately twice the range of a sensor.

The process and device according to the invention can be performed on all network signal types, provided that they can be modelled in accordance with the above-defined hypotheses. It is then possible to separate the two sources: coherent signal and incoherent signal and limited spatial extension. The expert system can be adapted to all signal types and receive a prior supplementary informations in order to supply a subdivision or partition which is as close as possible to reality. The monosensor networks of all types of sensors (acoustic, seismic, electric, pressure) can be processed. The multisensor networks (several sensor types at the same time) can be processed in accordance with a similar process.

Several useful signals of independent origins can be present on the network at the same time. The reconstructed useful signal is then the sum of the useful signals of the different sources.

It is pointed out that the notion of space coherence can exist for signals of different types. For example, a pressure signal can be linked with a magnetic signal by a linear transfer. When relations between signals do not mathematically exist or are difficult to calculate, they cannot be integrated in the matrix of transfer functions, but they can be incorporated into the expert system.

The process and device according to the invention have numerous industrial applications and in particular:

the detection of magnetic devices moving over a given area,

the detection of the unsatisfactory operation of sensors in a system or array,

the monitoring of industrial sites, airports and places of passage,

the monitoring of volcanic activity,

checking the migration of fluids in geological structures.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram which illustrates a device for extracting the signal according to the invention;

FIG. 2 illustrates an expert system module of the device according to the invention in a preferred embodiment, in the case where the transfer functions are linear and for a number of independent noises q equal to 1;

FIG. 3 is a flow chart illustrating the process of the invention;

FIG. 4 is a graphical diagram showing the output signals produced by an array of five magnetic probes, as a function of the sample number n;

FIG. 5 is a diagram showing the geometric placement of the five probes in FIG. 4 and an exemplary dipole passage trajectory;

FIG. 6 graphs the upper and lower probability pairs for the signal P (channel 16 in FIG. 2);

FIGS. 7A and 7B are graphs of the partition decisions for AN (p) and AN topo(P), respectively,

FIG. 8 graphs the final petition decisions for the index, target (p) for p varying from 1 to N, and

FIG. 9 is a graphical diagram comparing A (nT) as a function of the sample member n for the invention method as compared with conventional antenna processing methods.

DETAILED DESCRIPTION OF EMBODIMENTS

The object of the process according to the invention is to make it possible to process a network signal and separate one or more measurable physical phenomena in the form of a useful signal, which are optionally added to the noises of the network.

The physical phenomena involved can be of a random nature if the following hypotheses with regards to the useful signal and the noises accompanying its measurement are respected:

the useful signal has a finite spatial extension,

the q additive noises are spatially coherent and have a large extension compared with that of the useful signal.

The useful signal must touch a subarray E.sub.su (t) (the subarray being called cardinal "useful signal" v, t designating the time) of the array E of N sensors, E.sub.su (t) being unknown and can vary in time. The additive noises must be spatially coherent, i.e. there are combinations of signals from q sensors calculating the noise reaching the other sensors. In a particular embodiment where the transfer functions are linear filters, these transmission modes are stored in a transfer function matrix. The coherence properties can be stationary and therefore the transfer matrix is constant in time, or may be non-stationary. In the latter case, the device according to the invention must be able to access the transfer matrix, whose evolution must be calculable at all times.

The sensors can be of different types (magnetic, pressure, temperature, etc.) for the same network. A multisensor network or array makes it possible to detect a physical phenomenon by its different effects.

In the remainder of the description, for illustrating the operation of the process of the invention, consideration is given in exemplified manner to the field of the magnetic detection of mobile sources, the considered device being constituted by N magnetic sensors, which are either fixed or moving slowly and located on the area to be monitored, as well as a processing circuit. A multisensor network could complete the magnetic measurement by pressure measurements permitting a more reliable detection.

The process according to the invention makes it possible to separate the signals received on the antenna and designate the sensors which receive the useful signal. In this case the spatial coherence of the noise signals is ensured by the existence of linear prediction filters of the noise between the sensors.

A magnetic object moving in the vicinity of an array of magnetic sensors generates a useful signal, which is added to the natural magnetic signals. Close to the surface of the earth, the measured magnetic field is formed from the superimposing of vector signals generated by three sources:

a space and time-fixed signal concerning the dimensions of the considered areas. It is roughly modelled by the field of a dipole located along an axis having a direction slightly different from that of the earth. However, finer or more accurate models exist. As a function of the required precision, the chosen model is more or less accurate. For the sizes of areas envisaged here, it is considered that said signal is constant in time and space. Its module is very large compared with the other recorded signals;

a space only-variable signal (for the considered time intervals) generated by the local geology (local being the geology--sensor distance of the same order of magnitude as the area to be monitored) and called geological field. The movement of a sensor in this space-variable field generates a time signal called geological noise;

a time and space-variable signal generated by ionospherical currents and called geomagnetic noise or geomagnetic fluctuation.

It is necessary to add to this list the useful signal, which is variable in time and space and whose spatial extension is smaller than the monitored area. The properties of the three sources are given in order to show that there is a network of two signals in accordance with the hypotheses made hereinbefore.

The static earth Gauss field is filtered by a high-pass filter.

It is considered here that the sensor displacement speed is sufficiently close to 0 for the largest possible frequency of the geological noise to be outside the band liable to receive the useful signal. Thus, it can be eliminated by a band-pass frequency filtering.

The third signal, i.e. the geomagnetic noise has the property of being coherent in space. Berdichevski and Zdhanov in an article entitled "Advanced Theory of Deep Magnetic Sounding" (Elsevier, 1984) e.g. demonstrate that there are intersensor transfer functions making it possible to make a space prediction on the geomagnetic fluctuations between individual locations in the area. In favourable locations, these transfer functions or linear filters are identity filters.

For the envisaged dimensions of the networks and the study frequency band, the geomagnetic noise measured at a point r is equal to the sum of the filters of two independent components of this field measured at a point r'.

Geomagnetic fluctuations are similar to the effect produced by a primary plane wave which excites a conductive medium. In the general case, q is the number of degrees of freedom of the primary wave. For example, Egbert, in a thesis entitled "A Multivariate Approach to the Analysis of Geomagnetic Array Dam" (Washington University, 1987) demonstrates that q=2 for geomagnetic fluctuations. The primary wave has two degrees of freedom in the case of plane waves. In certain practical cases and for small distances, it is standard practice to admit that the plane wave is only slightly deformed.

The geomagnetic noise is then considered as identical throughout the study space for the same time. This is a special important case of the preceding model for which the transfer functions are scalar (and not bidimensional) and unitary.

Therefore the sensors measure geomagnetic fluctuations, which have space coherence properties, as well as a possible useful signal, which can only apply to a small number of signals of the N sensors at once. The magnetic detection of mobile sources is therefore a problem in accordance with that which the invention aims to solve with q=2. The transfer function matrix can here be identified by estimating the interspectral matrix of the network, in the absence of a useful signal.

The device and process according to the invention make use of the space coherence properties of geomagnetic fluctuations, as well as the limited spatial extension property of the useful signal. Although the invention is described relative to the example of magnetic detection, it remains a general solution for problems of other natures for as long as the above-defined hypotheses concerning the signal and noises remain valid.

As shown in FIG. 1, the device according to the invention for extracting a time-variable useful signal of finite spatial extension from a signal incorporating said useful signal and to which have been added q spatially coherent additive noises, comprises an array of N sensors, N being larger or equal to 3 and strictly larger than q, followed by N filtering modules 11 and N digitizing modules 12.

A module 13 for calculating space prediction error signals of the noise receives signals from these digitizing modules 12, as well as a module 14 for storing the transfer functions. It is connected to a useful signal detection module 21, which can contain a module for calculating detection indexes, whose shape recognition, derived, integral, proportional output channels for each error signal are inputted into a real time expert system module. This module 21 also receives informations from a module 22 in which is stored the position of the sensors, which varies in time.

A weighted projection module 23 receives informations from the digitizing modules 12, the transfer function storage function 14 and the module 21 for supplying a target signal SC and a geomagnetic noise signal BG. The elementary module of an expert system for detector module 21 is more precisely shown in FIG. 2.

In order to know whether the proportional output P (channel 16) and/or derived D (channel 17) and/or integral I (channel 18) detection indexes perform a detection of a non-zero error signal for a given pair of sensors (p, i), an OR 31 receives the channels P, I, D from the module 15 corresponding to this pair. The same procedure is used for the proportional output (channel 16') and/or derived (channel 17') and/or integral (channel 18') detection indices of the pair (p,i+1). These OR gates 31 supply detection indexes A1(p,i) and A1(p,i+1) and all the pairs A1 which can be formed with the sensor of rank p, each of the indexes A1 making it possible to answer the question: is something happening on the considered pair? These OR gates 31 are connected to a "VOTE" module 32, which forms a geometrical mean supplying a detection index AN(p). The mixed lines 15 illustrate the generalization to other pairs of sensors.

N "INFER" modules 33 receive indexes AN(1) . . . AN(p) . . . AN(k) . . . AN(N) in order to carry out an operation of type lie.sub.-- court (d)=.alpha..sup.3 /.alpha..sup.3 +d.sup.3, .alpha. being dependent on the range of the sensor and the distance d separating the sensor p from the sensor corresponding to the detection index received by the INFER module and are connected to a "JOIN" module 34 for carrying out an aggregation operation bri