Method for characterizing defects in a cable using machine learning

Abstract

A computer-implemented method for training an artificial intelligence model configured to characterize a fault on a transmission line, the method comprising the steps of: Receiving (501) a set of signals constituting a training signal base of the model, each signal corresponding to a simulation or measurement of a compressed reflectometry signal for a transmission line having a fault characterized by an impedance and / or a position, A compressed reflectometry signal resulting from the execution of the steps of: Injecting a predefined reflectometry signal into the transmission line, Carrying out a compressed acquisition of said signal comprising at least the multiplication of said signal taken from the line by a predefined pseudo-random sequence,Train (502) the artificial intelligence model from the training signal base to perform at least one task among: a task of classifying a type of fault present on the line, a task of predicting the position and / or the impedance of a fault present on the line. Figure 5,

Description

Title of the invention: Method for characterizing defects in a cable by machine learning

[0001] The invention relates to the field of wired diagnostic systems based on the principle of reflectometry for identifying and characterizing electrical faults on cables or more generally transmission lines.

[0002] The invention relates more specifically to a method combining acquisition by compressed sampling of a reflectometry signal taken from a cable with an artificial intelligence model trained to classify different types of faults and / or to predict a physical value characterizing a fault such as its position, its characteristic impedance or its reflection coefficient.

[0003] Cables are omnipresent in all electrical systems, for power supply or information transmission. These cables are subject to constraints and can be prone to failures. It is therefore necessary to be able to analyze their condition and provide information on the detection of faults that impact these cables, this information including the existence of faults but also their location and type. Fault analysis helps with cable maintenance. Standard reflectometry methods allow this type of analysis.

[0004] Reflectometry methods use a principle close to that of radar: an electrical signal, the probe signal or reference signal, is injected into one or more locations of the cable to be tested. The signal propagates in the cable or cable network and returns part of its energy when it encounters an impedance discontinuity. An impedance discontinuity can result, for example, from a connection, the end of the cable or a fault or more generally from a break in the signal propagation conditions in the cable. It results from a fault which locally modifies the characteristic impedance of the cable by causing a discontinuity in its linear parameters.

[0005] The analysis of the signals returned to the injection point makes it possible to deduce information on the presence and location of these discontinuities, and therefore any possible defects. An analysis in the time or frequency domain is usually carried out. These methods are designated by the acronyms TDR coming from the English expression "Time Domain Reflectometry" and FDR coming from the English expression "Frequency Domain Reflectometry".

[0006] Reflectometry diagnostic systems encounter a notable limitation linked to the sampling frequency of the analog-digital converter, which is essential to comply with the Nyquist-Shannon theorem. This constraint impacts directly the spatial resolution, more precisely the fineness of the defect detection. In order to improve this resolution in the detection system, the use of a high-frequency analog-to-digital converter is necessary. However, this approach has disadvantages such as high implementation cost, high power consumption, considerable storage memory requirement, and high noise level.

[0007] The use of compressed sampling proves to be an innovative solution to address these issues, by eliminating the need for a high-sample-rate analog-to-digital converter (ADC). This approach significantly simplifies implementation while encouraging the use of a low-sample-rate converter. Thus, it opens the way to the design of systems using readily available, robust components characterized by low power consumption. However, the use of compressed sampling adds an additional step in the process of fault detection with reflectometry, which is the step of compressed signal reconstruction. This step requires the use of reconstruction algorithms that are very expensive in terms of computation time and energy consumption and give degraded results for high compression rates.

[0008] There is therefore a general need to further improve reflectometry diagnostic systems based on the implementation of a compressed sampling technique in order to provide a global solution with low implementation cost which makes it possible to characterize a fault impacting a cable by its nature (short circuit, open circuit, non-clear fault) but also by its position on the cable and its physical properties such as its characteristic impedance or its reflection coefficient.

[0009] The Applicant's patent application FR3127581 describes a method for designing a multicarrier reflectometry signal of the OMTDR type having the form of a "chirp". It also describes the principle of compressed acquisition of such a signal.

[0010] The principle of compressed acquisition is also described in documents [1] and [2],

[0011] The method consists of multiplying the sampled signal, in the analog domain, by a pseudo-random sequence, applying a low-pass filter or an integrator to the result and then digitizing the signal at a sampling frequency lower than the Nyquist frequency. The method is advantageously applied to sparse reflectometry signals such as OMTDR signals having a "chirp" shape. Thus, this method makes it possible to digitize the signal at a lower frequency, thus limiting the complexity of the analog-to-digital converter while allowing high fault detection resolution.

[0012] However, a disadvantage of this solution is that it requires a step of reconstructing the reflectometry signal which has been modified in particular via multiplication with a pseudo-random sequence. This reconstruction step is generally carried out by means of iterative algorithms which consume computing time and memory resources.

[0013] There is therefore a need for an alternative solution making it possible to reduce the complexity inherent in the reconstruction algorithm when the signal has been obtained via compressed acquisition.

[0014] The invention proposes a new method for characterizing a defect from a signal measured via a compressed acquisition technique involving the multiplication of the signal with a pseudo-random sequence.

[0015] The invention makes it possible to avoid the step of reconstructing the digitized signal by using an artificial intelligence model trained on signals corresponding to an identical processing chain but to different types of faults.

[0016] In particular, the invention proposes a method for automatic classification of the type of fault as well as a method for automatic prediction of the position of the fault and the characteristic impedance of the fault (or its reflection coefficient).

[0017] The invention applies to any type of sparse reflectometry signals, in particular to CHIRP-OMTDR signals as proposed in the Applicant's patent application FR3127581 but also to other multi-carrier signals such as MCTDR (Multi Carrier Time Domain Reflectometry) signals.

[0018] The subject of the invention is a method, implemented by computer, for training an artificial intelligence model configured to characterize a fault on a transmission line, the method comprising the steps of: - Receive a set of signals constituting a base of training signals for the model, each signal corresponding to a simulation or a measurement of a compressed reflectometry signal for a transmission line having a fault characterized by an impedance and / or a position, - A compressed reflectometry signal resulting from the execution of the steps of: i. Inject a predefined reflectometry signal into the transmission line, ii. Carry out a compressed acquisition of said signal comprising at least the multiplication of said signal taken from the line by a predefined pseudo-random sequence, - Train the artificial intelligence model from the training signal base to perform at least one of the following tasks: a task of classification of a type of fault present on the line, a task of predicting the position and / or impedance of a fault present on the line.

[0019] According to a particular aspect of the invention, the training of the artificial intelligence model is supervised, each training signal being annotated by a fault type and / or a fault position value and / or a fault impedance value.

[0020] The invention also relates to a method for characterizing faults in a transmission line comprising the steps of: - Measure a signal, at a point on the line, characteristic of the reflection of a predefined reflectometry signal, previously injected into the line, - Carry out a compressed acquisition of the measured signal including at least the multiplication of the measured signal by a predefined pseudo-random sequence to obtain a digitized signal, - Executing an artificial intelligence model trained using the training method according to any one of the preceding claims on the digitized signal to characterize a fault on the transmission line by classifying the type of fault present on the line and / or by predicting the position and / or impedance of said fault present on the line.

[0021] According to a particular aspect of the invention, the artificial intelligence model is trained to classify a type of fault among: a short circuit, an open circuit, a non-clear fault, the absence of fault.

[0022] According to a particular aspect of the invention, the artificial intelligence model is trained to predict the value of a reflection coefficient of the signal on a fault and then deduce the value of the impedance of the fault from the value of the reflection coefficient and the characteristic impedance of the line.

[0023] According to a particular aspect of the invention, the artificial intelligence model is single-task and taken from: an automatic classification model or an automatic prediction model.

[0024] According to a particular aspect of the invention, the artificial intelligence model is multi-task and is configured to simultaneously perform an automatic classification task and an automatic prediction task.

[0025] According to a particular aspect of the invention, the predefined reflectometry signal is a multi-carrier signal having a “chirp” signal shape.

[0026] According to a particular aspect of the invention, the compressed acquisition of the measured signal comprises at least the steps of: - Take the signal at a point on the transmission line, - Multiply the signal with the predefined pseudo-random sequence, - Apply a low-pass filter or an integrator to the signal obtained, - Digitize the filtered signal at a frequency lower than the Nyquist frequency of the signal

[0027] According to a particular aspect of the invention, the artificial intelligence model is taken from: a k nearest neighbors algorithm or a convolutional neural network or a support vector machine algorithm.

[0028] The subject of the invention is a device for detecting non-clear faults in a cable comprising an apparatus for measuring, at a point on the cable, a signal back-propagated in the cable and a processing unit configured to execute the fault characterization method according to the invention.

[0029] The subject of the invention is a reflectometry system comprising a device for injecting, at a point on the cable, a predefined reflectometry signal and a device for detecting non-clear defects in a cable according to the invention.

[0030] The subject of the invention is a computer program comprising instructions for executing the method for training an artificial intelligence model according to the invention, when the program is executed by a processor and a computer-readable recording medium on which the computer program according to the invention is recorded.

[0031] Other characteristics and advantages of the present invention will appear better on reading the description which follows in relation to the following appended drawings.

[0032] [Fig-1] represents a diagram of a fault diagnosis system by reflectometry according to a first embodiment of the prior art,

[0033] [Fig.2] represents a diagram of a fault diagnosis system by reflectometry according to a second embodiment of the prior art,

[0034] [Fig.3] represents a diagram of a fault diagnosis system by reflectometry according to one embodiment of the invention,

[0035] [Fig.4] represents a diagram illustrating the appearance of several signals obtained via a compressed acquisition for four different types of defects,

[0036] [Fig.5] represents a flowchart detailing the steps of implementing a method of training an artificial intelligence model to characterize defects using the system of [Fig.3], according to one embodiment of the invention,

[0037] [Fig.6] represents a flowchart detailing the steps of implementing a method for characterizing defects using the system of [Fig. 3], according to one embodiment of the invention,

[0038] [Fig.7] represents an example of a convolutional neural network model capable of be used by the invention to carry out a task of classifying types of defects,

[0039] [Fig.8] represents an example of a convolutional neural network model capable of be used by the invention to perform a task of predicting the position and / or characteristic impedance of a defect.

[0040] [Fig.l] represents a diagram of a fault diagnosis system on a cable according to a reflectometry technique known from the prior art.

[0041] The system comprises a digital reflectometry signal generator 101. The generated signal can be a time pulse or a multi-carrier signal or any suitable signal. The generated signal is converted into an analog signal via a digital-to-analog converter DAC and then it is injected into a cable or more generally a transmission line L whose state is to be characterized. The signal injected into the line (for example by means of a coupler not shown) propagates in the line and encounters a fault DF characterized by an impedance discontinuity. Part of the signal is reflected on this discontinuity to be backpropagated to a measurement point. The reflected signal is measured (via a coupler not shown) and then it is digitized via an analog-to-digital converter CAN. A correlator COR performs the cross-correlation between the generated signal and the received signal in order to produce a time reflectogram R(t).If the reflectometry signal used is a simple time pulse, the correlator COR can be made optional.

[0042] The time domain reflectogram R(t) includes an amplitude peak characteristic of the DF fault at a time abscissa which is linked to the distance between the signal measurement point and the DF fault and to the speed of propagation of the signal in the cable.

[0043] As is known in the field of reflectometry diagnostic methods, the position dDNF of the non-clear fault on the cable, in other words its distance from the signal injection point, can be directly obtained from the measurement, on the time domain reflectogram, of the duration tDNF between the first amplitude peak recorded on the reflectogram and the amplitude peak corresponding to the signature of the fault.

[0044] Various known methods can be envisaged for determining the position dDNF. A first method consists of applying the relationship linking distance and time: dDNF = V.tDNF / 2 where V is the speed of propagation of the signal in the cable. Another possible method consists of applying a proportionality relationship of the type dDNF / tDNF = L / t0 where L is the length of the cable and t0 is the duration, measured on the reflectogram, between the amplitude peak corresponding to the impedance discontinuity at the injection point and the amplitude peak corresponding to the reflection of the signal on the end of the cable.

[0045] A disadvantage of this type of system is that the spatial resolution with which a defect can be detected (the detection accuracy) depends on the spectral band of the signal. The wider the signal bandwidth, the better the detection resolution. However, to digitize a wideband signal, this involves using an analog-to-digital converter that is expensive to implement, consumes a lot of energy and has a high noise level.

[0046] A second solution known from the state of the art and described in [Fig.2] consists of replacing such an analog-digital converter with a compressed sampling acquisition architecture as described in references [2] and [3].

[0047] A comparative diagram of the two prior art reflectometry systems is presented in [Fig. 2]. The architecture implementing compressed sampling acquisition is shown in the lower part 200 of the figure.

[0048] In this architecture, the analog-to-digital converter of the first system 100 shown in [Fig.l] is replaced by a demodulator DEM which comprises an analog multiplier MUL with a pseudo-random sequence pc(t) with values ​​1 and -1. The modulated signal at the output of the multiplier MUL is then filtered via a low-pass filter FPB and is then digitized using an analog-to-digital converter CAN which can operate at a sampling frequency lower than that which would be used by the system of [Fig.l]. Alternatively, the low-pass filter FPB can be replaced by an integrator.

[0049] The digitized signal YCs is then supplied to a reconstruction module REC which implements an iterative signal reconstruction algorithm. An example of possible implementation consists of applying a greedy algorithm making it possible to iteratively construct a sparse approximation of the signal.

[0050] The reconstructed signal is then supplied to a correlator COR which performs an intercorrelation calculation with the signal injected into the cable to generate a reflectogram R(t).

[0051] A disadvantage of this second solution is the need to reconstruct the signal via an iterative algorithm which adds additional implementation complexity.

[0052] [Fig. 3] represents a diagram of a reflectometry system according to an embodiment of the invention. The system of [Fig. 3] comprises the elements common to the system of [Fig. 2], these elements having the same references. More precisely, the processing chain is identical except that it does not comprise a signal reconstruction module REC or a correlator COR. These two elements are replaced by an artificial intelligence module IA which implements a machine learning model trained to characterize a DF defect present on the line L from the YCs signal at the output of the demodulator DEM.

[0053] [Fig.4] illustrates, for an example, the shape of the YCs signal obtained at the output of the DEM demodulator. On the diagram of [Fig.4], four signals 401, 402, 403, 404 are represented, respectively obtained for 4 different types of faults. In all cases, the fault is positioned at a distance of 1 m from the signal injection point. The injected signal is a CHIRP-OMTDR type signal. Signal 401 corresponds to a short-circuit type fault. Signal 402 corresponds to a circuit type fault. open. Signal 403 corresponds to a resistive fault type fault. Signal 404 corresponds to a cable with no faults.

[0054] [Fig.4] illustrates the fact that the YCs signals obtained at the output of the compressed acquisition demodulator have a random appearance which cannot be directly used to identify a fault, unlike a conventional time domain reflectogram for which faults are identified by the presence of an autocorrelation peak. This is why a signal reconstruction algorithm is necessary.

[0055] The proposed invention consists of implementing an artificial intelligence model trained on signals of the type described in [Fig.4] to carry out an automatic characterization of a defect.

[0056] [Fig.5] represents a flowchart detailing the steps of implementing a method for training a model for automatic characterization of a defect according to an embodiment of the invention.

[0057] The method comprises a first step 501 of designing a training signal base for training an artificial intelligence model. According to a first embodiment variant, the training signals are obtained via different measurements of YCs signals at the output of a compressed acquisition demodulator for the same reflectometry signal injected into a cable, for the same cable but for different types of faults positioned at different locations on the cable.

[0058] This first variant requires carrying out numerous measurements and artificially generating faults on the same cable, which has the disadvantage of being complex and tedious to implement.

[0059] According to a second embodiment variant, these training signals are obtained by means of a software simulator configured to simulate a reflectometry processing chain as described in [Fig. 3] from end to end. This second variant has the advantage of being able to more easily simulate a large number of faults on the same cable.

[0060] The training signal database generated in step 501 is for example divided into three sets: a set used for learning comprising approximately 80% of the signals, a set used for validation comprising approximately 10% of the randomly drawn signals and a set used for testing comprising approximately 10% of the randomly drawn signals.

[0061] Each drive signal is annotated with information characterizing the fault. For example, this information includes the nature of the fault (short circuit, open circuit, non-clear fault or no fault), the position of the fault on the cable, the characteristic impedance of the fault or the reflection coefficient of the signal on the fault.

[0062] In step 502, the artificial intelligence model is then trained in a supervised manner from the training signals to characterize the defects associated with each signal.

[0063] According to a first embodiment variant, the model is trained to perform a task of classifying signals according to the nature of the fault. For example, four classes of faults are envisaged: short circuit, open circuit, non-clear fault or absence of fault.

[0064] According to a second embodiment variant, the model is trained to perform a task of predicting a position value of the fault on the line or a value of the propagation time of the signal between the injection point in the cable and its reflection on the fault or a characteristic impedance value of the fault or a value of the reflection coefficient of the signal on the fault.

[0065] Advantageously, the classification and prediction tasks are performed sequentially. In other words, the prediction task is only activated if the classification task has previously made it possible to detect a fault. This means that the classification model is trained on all signals including those corresponding to a cable without a fault. The prediction model is trained only on signals corresponding to a fault present on the cable.

[0066] According to a third embodiment variant, the model is jointly trained to perform both the classification task and the prediction task.

[0067] Different artificial intelligence models can be considered to implement the invention.

[0068] In a first embodiment, the model used is based on a K nearest neighbors algorithm. In this case, the validation set is used to determine the parameter K of the model which corresponds to the number of neighbors.

[0069] In another embodiment, the model used is based on a support vector machine (SVM) algorithm.

[0070] In another embodiment, the model used is based on one or more convolutional neural networks.

[0071] [Fig.7] illustrates a diagram of a convolutional neural network configured to perform a task of classifying signals according to the nature of the defect according to an embodiment of the invention. The example of [Fig.7] is given for illustrative and non-limiting purposes. It relates to a neural network comprising four convolution layers 101, 102, 103, 104, two fully connected layers 105, 106 and an activation layer 107 implementing the SoftMax activation function.

[0072] Each convolution layer 101,102,103,104 comprises different filters of different dimensions. The first fully connected layer 105 comprises a number of neurons equal to the number of samples of the signal provided as input (equal to 512 on the non-limiting example of [Fig.7]). The last fully connected layer 106 comprises a number of neurons equal to the number of classes of defect types (4 classes in the non-limiting example of [Fig.7]).

[0073] The loss function used for training the network is for example the cross entropy function or any other cost function suitable for optimizing a classification network.

[0074] [Fig.8] illustrates a diagram of a convolutional neural network configured to perform a task of predicting the position of a fault and / or the value of the characteristic impedance of the fault according to a non-limiting exemplary embodiment.

[0075] The neural network shown comprises seven convolution layers 101-107 and two fully connected layers 108, 109. The first fully connected layer 108 comprises a number of neurons equal to the number of samples of the signal provided as input (equal to 512 in the non-limiting example of [Fig.8]). The last fully connected layer 109 comprises a neuron whose output corresponds to the desired prediction.

[0076] The loss function used for training the prediction network is for example a mean square error or mean absolute error function or any other function based on an error calculation between the annotated value of the prediction (position or impedance) and the predicted value.

[0077] According to another exemplary embodiment, a multi-task convolutional neural network is used to simultaneously perform the classification task and the prediction task.

[0078] In a particular embodiment, the neural network used for prediction (or the prediction branch of the multi-task network) is trained to predict the value of the reflection coefficient of the signal on the impedance discontinuity caused by the fault and then the characteristic impedance of the fault is determined using the following relationship:

[0079] r ' d~z^zc

[0080] With Zd and Zc which correspond respectively to the fault impedance and the characteristic impedance of the cable considered and Ij is the reflection coefficient predicted by the neural network.

[0081] The nature of the fault in the cable is determined according to the value of the reflection coefficient which depends directly on Zd because Zc is constant.

[0082] If rd = -1, then the fault type corresponds to a short circuit, Zd =0 Q

[0083] If rd = 1, then the fault type corresponds to an open circuit, Zd tends towards infinity.

[0084] If rd = 0, in this case there is no fault, Zd = Zc

[0085] If < 1 then the defect type corresponds to a non-clear defect, 2d = Z, x

[0086] An advantage of this embodiment is that the characteristic impedance potentially varies over a very high range of values ​​(up to infinity for an open circuit) while the reflection coefficient varies over the interval [-1; 1].

[0087] Advantageously, the architecture of the predictive neural network of [Fig.8] is completed in this case by adding a hyperbolic tangent activation layer in order to obtain an output bounded between -1 and 1.

[0088] According to one embodiment, the two values ​​(position and characteristic impedance or reflection coefficient) can be predicted simultaneously via a two-component vector.

[0089] An advantage of using convolutional neural networks (CNNs) is that they are easier to integrate into embedded systems than a k-nearest neighbor (KNN) algorithm. Indeed, after training a convolutional neural network, it is sufficient to run the model in inference. Conversely, the KNN algorithm assumes the complete storage of the training data set. Thus, convolutional neural networks offer a more practical and efficient solution for embedded systems, meeting memory and power constraints while maintaining high performance in processing complex data.

[0090] In addition, convolutional neural networks (CNNs) allow multi-task learning to simultaneously learn to solve multiple tasks and share relevant common information between these tasks. This allows for better generalization and improvement of the overall performance of the model by exploiting the relationships between tasks. This property is also convenient for integration into an embedded architecture for fault detection and localization, because multi-task learning allows for a single model to be embedded performing multiple tasks, which is more advantageous than embedding two models separately.

[0091] Finally, neural networks allow for transfer learning. This learning technique allows a neural network previously trained on a specific task to benefit from its knowledge for carrying out a similar task. Transfer learning is particularly practical in cases where there is not much training data available. By reusing pre-trained models as a starting point, the learning process on new tasks is accelerated in terms of time and data required for learning.

[0092] Conversely, an advantage of using the KNN algorithm lies in its better classification and prediction performance.

[0093] The use of this algorithm is efficient because it is based on the calculation of distances for the choice of neighbors and thanks to the restricted isometry property of Compressed sampling: The distances between the data before compression and after compression are preserved, which means that samples of the same class will necessarily have neighbors of the same class, which makes it easier to find neighbors for different samples. This is also present in the case of position prediction. When the model is trained, it can be used in inference to perform fault detection from a new compressed signal measurement made on a cable with similar properties and for a reflectometry signal identical to those used for training.

[0094] This process is shown diagrammatically in [Fig.6]. In step 601, a new measurement of a compressed acquisition signal is carried out. In step 602, the model trained on this measurement is executed in order to detect the presence of a defect and to determine its position and its characteristic impedance. References

[0095] [1] Tzila AJAMIAN. “Exploration of Compressed Acquisition applied to the Reflectometry”. Doctoral thesis. CENTRAL SCHOOL OF NANTES, 2019.

[0096] [2] Gargouri, Yosra, et al. “Enhancing Reflectometry Systems with CHIRP-OMTDR and Compressed Sensing: A Study on Signal Recovery Quality." 2023 30th IEEE International Conference on Electronics, Circuits and Systems (ICECS). IEEE, 2023.

Claims

Claims

1. A computer-implemented method of training an artificial intelligence model configured to characterize a fault on a transmission line, the method comprising the steps of: - Receiving (501) a set of signals constituting a training signal base of the model, each signal corresponding to a simulation or measurement of a compressed reflectometry signal for a transmission line having a fault characterized by an impedance and / or a position, - A compressed reflectometry signal resulting from the execution of the steps of: i. Injecting into the transmission line a predefined reflectometry signal, ii.Carry out a compressed acquisition of said signal comprising at least the multiplication of said signal taken from the line by a predefined pseudo-random sequence, - Train (502) the artificial intelligence model from the training signal base to carry out at least one task among: a task of classifying a type of fault present on the line, a task of predicting the position and / or the impedance of a fault present on the line.

2. A method according to claim 1 wherein the training of the artificial intelligence model is supervised, each training signal being annotated by a fault type and / or a fault position value and / or a fault impedance value.

3. Method for characterizing defects in a transmission line comprising the steps of: - Measuring (601) a signal, at a point on the line, characteristic of the reflection of a predefined reflectometry signal, previously injected into the line, - Carrying out (601) a compressed acquisition of the measured signal comprising at least the multiplication of the measured signal by a predefined pseudo-random sequence to obtain a digitized signal, - Execute (602) an artificial intelligence model trained using the training method according to any one of the preceding claims on the digitized signal to characterize a fault on the transmission line by classifying the type of fault present on the line and / or by predicting the position and / or the impedance of said fault present on the line.

4. Method according to any one of the preceding claims in which the artificial intelligence model is trained to classify a type of fault among: a short circuit, an open circuit, a non-clear fault, the absence of a fault.

5. Method according to any one of the preceding claims in which the artificial intelligence model is trained to predict the value of a reflection coefficient of the signal on a fault and then deduce the value of the impedance of the fault from the value of the reflection coefficient and the characteristic impedance of the line.

6. Method according to any one of the preceding claims in which the artificial intelligence model is single-task and taken from: an automatic classification model or an automatic prediction model.

7. A method according to any one of claims 1 to 5 wherein the artificial intelligence model is multi-task and is configured to simultaneously perform an automatic classification task and an automatic prediction task.

8. A method according to any preceding claim wherein the predefined reflectometry signal is a multi-carrier signal having a chirp signal shape.

9. Method according to any one of the preceding claims in which the compressed acquisition of the measured signal comprises at least the steps of: - Taking the signal at a point on the transmission line, - Multiplying the signal with the predefined pseudo-random sequence, - Applying a low-pass filter or an integrator to the signal obtained, - Digitize the filtered signal at a frequency lower than the Nyquist frequency of the signal.

10. A method according to any preceding claim wherein the artificial intelligence model is taken from: a k-nearest neighbor algorithm or a convolutional neural network or a support vector machine algorithm.

11. Device for detecting non-clear faults in a cable comprising a measuring device (DEM), at a point on the cable, of a signal back-propagated in the cable and a processing unit (IA) configured to execute the fault characterization method according to any one of claims 3 to 10.

12. Reflectometry system comprising a device for injecting, at a point on the cable, a predefined reflectometry signal and a device for detecting non-clear defects in a cable according to claim 11.

13. A computer program comprising instructions for executing the method of training an artificial intelligence model according to any one of claims 1 or 2, when the program is executed by a processor.

14. A computer-readable recording medium on which the computer program according to claim 13 is recorded.

Images