Method for detecting the presence of an electric arc, and associated device

Abstract

The invention relates to a computer-implemented method for detecting the presence of an electric arc on a power line (Ll) using a machine learning algorithm, carried out using a time-correlated measurement of the arc voltage Varc, of the current I and of the source voltage Vsource and also a knowledge model for linking the arc voltage Varc and the current I to constants depending on the knowledge model, implemented after a learning phase that involves adjusting a learning model that takes the current I and the source voltage Vsource at input and provides the voltage Varc and the constants of the knowledge model at output; the method comprising continuously detecting the presence of an arc using a decision function depending on a cost function used in the learning phase.

Description

DESCRIPTION TITLE: Method for detecting the presence of an electric arc and associated device TECHNICAL FIELD OF THE INVENTION

[0001] The technical field of the invention is that of detecting the presence of an electric arc using an artificial learning method.

[0002] The present invention relates to a method and a device for detecting the presence of an electric arc using a learning method and in particular a method in which a learning model and a knowledge model are used jointly. TECHNOLOGICAL BACKGROUND OF THE INVENTION

[0003] For several years now, civil aviation has been mobilizing to contribute to the fight against climate change.

[0004] Technological research efforts have already led to very significant improvements in the environmental performance of aircraft. This application results from the consideration of impact factors in all phases of design and development to obtain less energy-intensive, more environmentally friendly aeronautical components and products whose integration and use in civil aviation have moderate environmental impacts with the aim of improving the energy efficiency of aircraft.

[0005] In other words, the present invention results from the ongoing work to reduce the climate impact of civil aviation by using methods and operating virtuous development and manufacturing processes that minimize greenhouse gas emissions to the minimum possible to reduce the environmental footprint of its activity. As shown by the present invention, this sustained research and development work relates in particular to the development of the use of electrical technologies to provide propulsion. In this context, it is particularly important to be able to detect the formation of electrical arcs.

[0006] As a reminder, an electric arc is a self-sustaining high-current discharge. There are two main types of electric arcs that occur in aeronautical environments: breaking arcs, present in contactors at when a circuit is opened or closed and fault arcs that can appear unexpectedly on all components of the electrical chain. These can cause severe damage to equipment, systems and even the structure of the aircraft. This is therefore a point of particular vigilance.

[0007] Fault arcs can be of two types depending on their position in the electrical circuit: parallel arcs and series arcs. Series arcs generally occur at the terminals of an intermediate interface (terminal block, connector) following a connection problem during maintenance / installation, or degradation linked to natural aging or anticipated by chemical, electrical or mechanical interactions.

[0008] On an electrical distribution network, series arcs are always more complex to detect compared to parallel arcs. Indeed, from the point of view of the electrical signature, parallel arcs generate a strong current draw, which allows real-time diagnostic systems to identify them efficiently and quickly in order to give a command to protect the line by the cut-off system. On the other hand, in the case of a series arc, the current is only slightly impacted. This is explained by the fact that an electric arc is a low impedance fault. Consequently, when an arc is present in series on the electrical system, the current will only suffer a slight fluctuation and more precisely a reduction of a small percentage of its nominal value. It is for this main reason that series arcs are more complex to detect by standard control means, compared to parallel arcs.In addition, the waveform of the electrical signal of the network is also a factor that can complicate detection. Concerning AC voltage networks, in the half-cycle, the passage through a voltage of 0 V is the moment when the electric arc can no longer persist. When a series arc appears on such a type of network, this moment can then result in a brief open circuit leaving the line current at 0 A for a short time (this phenomenon gives rise to current shoulders), which can be favorably used to diagnose a series arc fault. On the contrary, when the waveform is continuous (DC network), this phenomenon no longer exists and the arc persists without an open circuit, thus eliminating one of the possibilities for diagnosing the fault.

[0009] On current networks, due to the relatively moderate distributed power, the occurrence of electric arcs is mitigated because the voltage levels distributed in aeronautics are a maximum of 230 / 400V. In addition, the distribution is in the form of an alternating wave (AC) for these voltage levels (230 / 400V maximum), allowing the self-extinction of the arc to be promoted at each half-period when the voltage passes through 0V. In addition, the distribution of a continuous waveform (DC) is limited mainly to 28V. The impact of electric arcs is also limited by the use of passive protections (choice of “arc resistant” materials, definition of segregation distances) in order to limit the consequences of fault arcs on the environment, as well as by the use of active protections (detection and opening of line) in the event of an overcurrent fault observed by current sensors. This advantageously makes it possible to detect most parallel arcs, but not series arcs for which no active protection has been deployed until now.

[0010] Current work on aircraft electrification / hybridization considers the distribution of continuous waveforms (DC) at voltage levels up to one kilovolt. This paradigm calls into question the arc fault passivation strategy just mentioned. The series electric arc must then be taken into account in the same way as the parallel arc because the damage associated with the generation of higher power faults subjected to DC voltage risks damaging the aircraft and affecting the safety of passengers and crew.

[0011] One of the ways to mitigate the series arc fault is to identify (detect) it in order to isolate the faulty line, if necessary.

[0012] The major problem raised by the detection of electrical arcs in aeronautics is based on a significant expectation of performance levels. More specifically, a first requirement is reliability, i.e. the power of a system to systematically detect dangerous electrical phenomena in all circumstances. This is its ability to achieve a true positive rate as close as possible to 100%. The second requirement is robustness, i.e. the ability of a system to be immune to any event other than the one for which it was programmed. In other words, it is its ability to achieve a false positive rate as close as possible to 0%. The system must therefore be robust against different aircraft load signatures (having behaviors that resemble the signature of arcs) as well as all the environmental parameters associated with the flight and life cycles of aircraft.

[0013] Until now, existing solutions for detecting fault arcs have systematically shown their weaknesses, either in terms of reliability or in terms of robustness. The reasons are multiple, but it is interesting to cite the imprecision of the fault arc models used to calibrate the systems, the lack of knowledge (or simplification) of the targeted electrical architecture, or the environmental constraints encountered, which are all reasons that reduce the performance of the proposed technologies.

[0014] In an attempt to meet these demanding requirements, new methods are being sought. Among them, artificial learning methods are beginning to be used more and more. They offer new alternative solutions. These methods are divided into unsupervised learning, semi-supervised learning, supervised learning, and reinforcement learning.

[0015] The most applied branch in the detection of electrical ars faults in the state of the art is that of supervised learning. This type of learning is a group of methods whose purpose is to generate a function capable of relating a set of inputs to an output using a base of examples and sophisticated mathematical methods. The model infers this function from labeled data consisting of a set of training examples. Detection methods using a model obtained by supervised learning are divided into regression and classification methods.

[0016] The objective of a classification is to predict a discrete value that allows any data to be assigned a category called a class. The decision function obtained by a classification method therefore separates the existing classes. The objective of a regression is to predict a continuous quantity. The decision function obtained by a regression will therefore be able to represent the behavior of the system studied by continuously representing the quantity of interest (and not in a discrete manner as does the classification).

[0017] In the state of the art, the detection of electric arcs (including in an aeronautical environment) is generally done from electrical measurements using supervised artificial learning, but in the state of the art, the labeling method used is not specified, or the labeling method is specified but does not identify all phenomena correctly. For example, in the state of the art, a measurement of the arc voltage can be used to complement the labeling. The paper entitled “AC Series Arc-Fault Detection With A Transformer Neural Network, A Chabert et al., June 2022” proposes a detection of series arcs in AC voltage with a supervised learning method. A fully automatic labeling algorithm is proposed using the arc voltage.

[0018] However, known solutions using artificial learning tools aimed at detecting fault arcs sometimes lack representativeness. Generally, their labeling methods (essential for this approach) are not described and / or are not precise enough in view of the erratic nature of the electric arc. However, if the labeling step is not done correctly, including all the phenomena considered for the detection phase, the learning step will be biased and will produce a detection tool with poor performance in a real environment.

[0019] In order to detect the formation of fault arcs, document EP3695476 A1 proposes detection by artificial learning using a recurrent neural network. This is a detection method using only the mathematical information proposed by artificial learning tools. In particular, in the proposed method, the learning phase does not use physical equations of arcs. The document "Physics-Informed Learning for High Impedance Faults Detection, Wenting Li, Deepjyoti Deka, 2021 IEEE Madrid PowerTech" proposes a so-called physically informed detection method for high impedance faults in microgrids (it is therefore not the same type of discharge as that produced by an electric arc). More specifically, the detection is based on a method derived from artificial learning as well as knowledge from the physics of the phenomenon to be detected.However, in this document, the physical information does not come from a complete model of knowledge of the behavior of the discharge. The physical information concerning the discharge is in fact derived from a measurement of the difference between a voltage-current phase diagram in a network, without fault and with fault.

[0020] Thus, the methods of the state of the art do not, in general, allow obtaining the reliability and robustness required for use in aeronautics. Also, there is a need for a method of detecting electric arcs allowing to significantly improve the detection capacity of existing methods, particularly for series arc type electric arcs. SUMMARY OF THE INVENTION

[0021] The invention provides a solution to the problems mentioned above, by proposing a detection method based on an approach which consists of taking advantage of both experimental data and physical knowledge for the design by learning of a semi-physical arc detection model.

[0022] For this, a first aspect of the invention relates to a computer-implemented method for detecting the presence of an electric arc on a power line using an artificial learning method, the learning being carried out using: a plurality of measurements relating to the formation of an electric arc, each measurement of the plurality of measurements comprising a time-correlated measurement of the arc voltage V arc , of the current / and the voltage of SOUrCe ^source > of a knowledge model linking the arc voltage V arc and the current / to constants depending on the knowledge model.

[0023] The method according to a first aspect of the invention is implemented using a learning model resulting from a learning phase during which, using the measurements of the plurality of measurements and the knowledge model, a step of adjusting the learning model is implemented, the learning model taking as input the current / and the source voltage V source and providing the arc voltage V as output arc and the constants of the knowledge model, learning being carried out using a cost function L given by the following relation L = L + L2 where is a first cost subfunction measuring a first difference between the arc voltage V^ c edlt predicted by the learning model and the arc voltage V arc measured:

[0025] and where L2 is a second cost subfunction measuring a second deviation between the knowledge model prediction and the measured values.

[0026] The method according to a first aspect of the invention comprises a detection phase comprising: A continuous measurement step of the source current / and voltage V source on the power line, so as to acquire input data relating to the source current / and voltage V source on a measurement window V source (t m )}^ l=1 where M is the number of measurement points in the measurement window, / (t m ) is the current at time t m and Source ( f m) is the source voltage at time A step of continuous determination of the presence of an arc, using a decision function allowing, as a function of the last M input data relating to the current / and the source voltage V source measured on the power line during the measurement step, to detect the presence of an arc, the decision function being a function of the second cost sub-function.

[0027] Thanks to the invention, it is possible to use arc physics (on the basis of knowledge models of electric arcs) in order to predict part of the behavior of a fault arc and thus contribute to the learning of the model. This makes it possible to make a more efficient decision than the solutions proposed by the state of the art, and to reduce the quantity of raw data necessary for learning (for example, by limiting the quantity of load signals from the aircraft architecture), in comparison with a learning model according to the state of the art. This result is notably obtained by the association of experimental information with a physical law. In addition, the variation of the arc length does not need to be known a priori, because it appears implicitly in the variables of the knowledge model.

[0028] In addition to the characteristics which have just been mentioned in the preceding paragraph, the method according to a first aspect of the invention may have one or more complementary characteristics among the following, considered individually or according to all technically possible combinations.

[0029] In one embodiment, the decision function is defined by the following expression:

[0030]

[0031] Where X m = {I t m ), V source (t m ')}, e is a detection threshold value, and L2(X m ') is the value of the second cost subfunction L2 determined using the learning model and calculated at point X m .

[0032] In one embodiment, the knowledge model is the Mayr model and the second cost subfunction L2 takes the following form:

[0034] where T is the arc constant and P o is the cooling power.

[0035] In one embodiment, the knowledge model is the Cassie model and the second cost subfunction L2 takes the following form:

[0037] where T is the arc constant, and 70 is une constant corresponding to a minimum arc voltage (i.e. a minimum voltage value on the power line where the arc appears).

[0038] In one embodiment, the knowledge model is the Shavemaker model and the second cost subfunction L2 takes the following form:

[0040] where T is the arc constant, P Q is the cooling power and 70 is a constant corresponding to a minimum arc voltage (i.e. a minimum voltage value on the power line where the arc appears).

[0041] In one embodiment, the electrical line is equipped with a cut-off device for cutting the current in said line, the method comprising, when an electric arc is detected during the detection step, a step of sending, to the cut-off device, an instruction to cut the current on the line in question.

[0042] In one embodiment, the learning model is performed using a formal neural network.

[0043] In one embodiment, the learning model is performed using a support vector machine.

[0044] There is therefore a need for a data labeling method that makes it possible to obtain data representative of the phenomenon of the formation of a fault arc, in particular data representative of an electric arc of the series arc type in direct voltage. There is also a need for a detection method using, during the learning phase, data labeled with the method in question.

[0045] In one embodiment, the method may provide a solution to the problems discussed above, by ensuring that data from electrical measurements of fault arcs are correctly labeled and by enabling: To associate this binary labeling (arc or non-arc) with a supervised artificial learning classification method; To ensure that an arc is detected in a robust and reliable manner; To perform automated arc detection.

[0046] Another objective of the process may be to optimize the time spent on data labeling during the model training phase, for the detection of electric arcs, and to reduce the resources required for labeling.

[0047] For this, in one embodiment, the method comprises: a design phase by artificial learning of a decision function using the plurality of measurements, the design phase comprising: ■ For each measurement of the plurality of measurements, a step of labeling the arc current measurement using the arc voltage measurement associated with it; ■ A step of extracting descriptors from the labeled measurements; ■ A learning step for determining the decision function from at least part of the descriptors extracted during the extraction step; the detection phase including: ■ A step of measuring the current flowing on the line; ■ A step of extracting the descriptors of the current measurement obtained during the previous current measurement step; ■ From the descriptors extracted from the current measurement and using the decision function determined during the design phase, a step of detecting the presence of arcs.

[0048] Furthermore, in the method according to this embodiment, the labeling step comprises, for each measurement of the plurality of measurements and for each time interval t of the measurement considered: When the measured arc voltage U arc exceeds a threshold value U for the first time s predetermined, a recording step of the interval considered as starting point t d of the electric arc; When the measured arc voltage U arc falls below the threshold value U for the first time spredetermined, a recording step of the interval preceding the interval considered as end point t e of the electric arc.

[0049] Thanks to the process, the labeling logic combines arc physics, an automated procedure and, in some cases, human expertise. Thus, it becomes possible to correctly label the vast majority, if not all, arc or nominal signals, to carry out this labeling efficiently thanks to the automated procedure, to correct the algorithmic inaccuracies undermined by the variability of arc signatures (when human expertise is required), to label the windows as precisely as possible (to the nearest sample), and to create a database thus labeled and suitable for the design of an artificial learning classifier for the detection of electric arcs.

[0050] This embodiment of the method may form a fully fledged detection method, the method being implemented by computer and using an artificial learning method, the learning being carried out using a plurality of measurements relating to the formation of electric arcs, each measurement of the plurality of measurements comprising a time-correlated measurement of the arc voltage and the arc current, the method comprising the steps of this embodiment as described previously.

[0051] In one embodiment, for each measurement, the labeling is associated with a measurement window over which the learning is carried out, said window being given by the interval P g x min((t e - t d ), T max y, P d x min((t e - t d ), T max ) where T max is a predefined maximum window duration, P g is a step to the left of point td having a normalized value for the duration of the window under study and P d is a step to the right of point t d having a normalized value for the duration of the study window given by the relation P d = 1 - P a .

[0052] In one embodiment, the method comprises, before the learning phase, a step of acquiring signals necessary for learning so as to obtain a plurality of measurements, each measurement comprising a time-correlated measurement of the arc voltage and the arc current.

[0053] In one embodiment, the design phase comprises, before the labeling step, a step of manual pre-identification by expertise, in each measurement of the plurality of measurements, of all the electric arcs satisfying predetermined criteria so as to define a position of each arc within each measurement of the plurality of measurements.

[0054] In one embodiment, the design phase comprises, at the end of the labeling step, a step of manually verifying the outputs of the labeling step by expertise so as to verify that the labeled measurements include all of the regions considered for learning.

[0055] In one embodiment, the descriptors used in the step of determining the decision function comprise the spectral entropy and / or the autocorrelation of the arc current.

[0056] In one embodiment, the design phase comprises, between the step of extracting descriptors from the labeled measurements and the step of determining the decision function, a step of ranking and selecting the descriptors so as to retain the most relevant subset of descriptors. For example, the ranking could be carried out using a Gram-Schmidt method and the selection could be carried out using an incremental method. Of course, other methods can be used for ranking or selection.

[0057] In one embodiment, the step of determining the decision function comprises a plurality of iterations of the following two sub-steps: a sub-step of designing, by learning, a classifier; and a sub-step of validating the classifier from the labeled data; these sub-steps being repeated until an optimal classifier is obtained, the decision function being determined from this optimal classifier.

[0058] A second aspect of the invention relates to a device for detecting the presence of an arc in a current line comprising a measuring means configured to measure the current flowing in the current line and means configured to implement the method according to the first aspect of the invention.

[0059] A third aspect of the invention relates to a computer program comprising instructions which, when executed by a computer, cause the device according to the second aspect of the invention to carry out the method according to the first aspect of the invention.

[0060] A fourth aspect of the invention relates to a computer-readable medium, on which the computer program according to the third aspect of the invention is recorded.

[0061] A fifth aspect of the invention relates to an aircraft comprising the device for detecting the presence of an arc as defined above.

[0062] The invention and its various applications will be better understood by reading the following description and examining the accompanying figures. BRIEF DESCRIPTION OF THE FIGURES

[0063] The figures are presented for information purposes only and in no way limit the invention.

[0064] [Fig. 1] shows a first flowchart of the method according to an exemplary embodiment of the invention.

[0065] [Fig. 2] shows a second flowchart of a method according to an example of the invention.

[0066] [Fig. 3] shows a schematic representation of a time-correlated arc voltage measurement and arc current measurement.

[0067] [Fig. 4] shows a schematic representation of a time-correlated arc voltage measurement and arc current measurement after pre-identification.

[0068] [Fig. 5] shows a flowchart of a labeling step according to an example of the invention.

[0069] [Fig. 6] shows a schematic representation of the labeling principle according to the invention from a time-correlated arc voltage measurement and an arc current measurement.

[0070] [Fig. 7] shows a schematic representation of arc voltage as a function of arc length.

[0071] [Fig. 8] shows a schematic representation of a device according to an exemplary embodiment of the invention. DETAILED DESCRIPTION

[0072] The figures are presented for information purposes only and in no way limit the invention. Method for detecting the presence of an electric arc on a line

[0073] A first aspect of the invention illustrated in [Fig. 1] relates to a computer-implemented method 100 for detecting the presence of an electric arc on a power line using an artificial learning method. More particularly, in the method according to the invention, the learning is carried out using a plurality of measurements relating to the formation of an electric arc, each measurement of the plurality of measurements comprising a time-correlated measurement of the arc voltage V arc , of the current / and of the source voltage V source .

[0074] The method according to the invention is notably original in that, in addition to a learning model, its design is also based on the use of a knowledge model making it possible to connect the arc voltage V arcand the current / to constants which depend on the choice of the knowledge model. More particularly, the principle of the present invention is based on the implementation of a learning PA phase which uses two techniques: modeling by artificial learning and knowledge modeling. This makes it possible to take advantage of the capacity of modeling by artificial learning and the expertise provided by the method derived from the physics of the fault arc mechanism.

[0075] As a reminder, machine learning modeling uses a mathematical regression tool derived from machine learning methods. For example, this tool can take the form of a formal neural network. This type of approach, sometimes called "black box" modeling, where only measurements are used, does, however, have certain limitations, notably because the database used during learning is not exhaustive and does not have all the types of arcs to be detected.

[0076] Also for the record, knowledge modeling makes it possible to describe, at least partially, the physical behavior of the arc that we are trying to detect. However, this modeling also has certain limitations. In particular, the generalization of this method can be called into question: its scope of validity and its performance can be reduced, for example, by the great variability of the parameters of the arc and its environment.

[0077] The approach of the present invention proposes a method 100 which makes it possible to correct the limitations of each of these methods by combining them. More particularly, in the method according to the invention, the knowledge model derived from the physics of the arc makes it possible to define a framework for the behavior of the fault to be detected. Thus, in the case of a scarcity of data for adjusting the model by learning, the knowledge model makes it possible to guide the decision and not to deviate from the known operation of the mechanism studied, here the fault arc.

[0078] On the other hand, an inaccuracy of the knowledge model can be adjusted thanks to the data that contain all the information on the variation of the quantities of interest. More precisely, the learning of the learning model is done by the recognition of arc signatures within the experimental data. At the same time, it is a question of guiding this learning phase by an expertise coming from one (or more) physical law(s) resulting from the knowledge of the physics of arcs.

[0079] In summary, in the present invention, arc detection is designed and adjusted using tools from artificial learning and knowledge from the physics of the system. Phase of

[0080] The method according to the invention firstly comprises a learning phase. The aim of this phase is to adjust the black box model using the knowledge model and thus obtain a decision function which, from measurements carried out on the power line, gives a Boolean (or any other equivalent representation) representative of the presence or absence of a fault arc on the power line.

[0081] In order to be able to carry out this learning, it is necessary to have a plurality of measurements, each measurement of the plurality of measurements comprising a time-correlated measurement of the current I, of the source voltage V source and the arc voltage V arc . By "time correlated" we mean that a measurement of current I, a measurement of the source voltage V source and an arc voltage measurement V arc having the same abscissa t were measured at the same time t.

[0082] If these measurements are not available, it may be necessary to collect them, for example during laboratory measurements. For this, in one embodiment, the learning phase PA comprises a step of acquiring a plurality of measurements, each measurement comprising a time-correlated measurement of the current I, of the source voltage V source and the arc voltage V arc . As already mentioned, this step is usually implemented in the laboratory where tests are carried out in order to generate a database comprising the plurality of raw measurements.

[0083] At the end of this step, a plurality of measurements is obtained, each measurement of this plurality of measurements being able to be described using the following notation:

[0085] where t m is the discretized time, / (t m ) is the current as a function of t m , V arc (t m is the arc voltage as a function of tm , Vsource tm) is the source voltage as a function of t m and M is the total number of points available in the measurement considered.

[0086] The learning phase PA also comprises, using the measurements of the plurality of measurements, a step E1 of adjusting a learning model taking as input the current / and the source voltage V source and providing the arc voltage V as output arc and the constants associated with the chosen knowledge model (for example, the arc constant T and the cooling power P o when the knowledge model is the Mayr model). During this step E1, learning is carried out using a cost function L to be minimized and given by the following relation L = + L2. As a reminder, a cost function is a parameterized function defining an error. The closer the result of the cost function is to 0, the smaller the error.

[0087] In the previous relationship, the term is defined as the comparison of the output values ​​calculated by the learning model with the measured values, this error being calculated only for the arc voltage V arc . Also, Li represents the gap between the arc voltage V^ r r c edlt predicted by the learning model and the arc voltage V arc measured and given by the following relation:

[0089] The term L2 represents the difference between the prediction of the knowledge model and the measured values. Its expression therefore depends on the knowledge model used. In any case, the knowledge model used must be able to provide, as output, the arc voltage V arc .

[0090] In an embodiment that will serve as an illustration in the remainder of the description, the knowledge model used is the Mayr model. In this embodiment, the model adjusted by learning is carried out using a formal neural network.

[0091] In an alternative embodiment, the knowledge model is a Cassie model. In this embodiment, the model fitted by learning may be realized using a support vector machine.

[0092] In an alternative embodiment, the knowledge model is a Shavemaker model. In this embodiment, the model fitted by learning may be realized using a support vector machine.

[0093] In the following, and for illustrative purposes, the Mayr model will be used. As a reminder, the Mayr model is derived from the energy balance between the amount of energy lost due to cooling and that supplied by the circuit. In particular, the Mayr model proposes the following dynamic equation:

[0095] where G arc = — = — is the conductance of the arc column (neglecting the effect of the arc feet). This equation can be written as:

[0097] where T is the arc time constant with T = T(T, P, Z) and P o is the cooling power with P o= P0(T,P, Z), T being the ambient temperature, P being the ambient pressure and Z being the arc length.

[0098] The cost function L2 can then be written as follows:

[0100] The learning stage then aims in particular to adjust a mathematical model which will determine a function:

[0102] More specifically, the learning algorithm uses the measurement values ​​to determine the output variables (y arc (t m However, only the arc voltage V arc (t m ) is available in the measurements of the plurality of measurements. Also, the determination of the couple (T, P0) (in the case of the Mayr model), which must satisfy the equation will be carried out using the chosen knowledge model(s).

[0103] Thus, at the end of this step E1, the function f introduced previously is determined (and no longer evolves during the detection phase presented below). PD detection phase

[0104] Once the learning phase is completed, the model obtained which best minimizes L is then capable of correctly predicting the outputs (y arc (tm )>T, P0) from the inputs (l(t m ,V source (t m )) from arc signals.

[0105] During the PD detection phase, the model will thus receive a quantity M of inputs which represent a signal window acquired from sensors (typically the line current / and the source-side voltage V source , but other sizes can be considered depending on the knowledge model such as temperature, pressure, etc.). Then, the learning model through the function introduced previously makes a prediction of M outputs {V arc t m ), T, P o}" =1 , which depend on the chosen knowledge model. The outputs calculated by the model make it possible to evaluate the L2 error function and thus determine whether the corresponding signal is an arc (if the prediction is good, i.e. the value of the error function is low) or not (if the prediction is bad, i.e. the value of the error function is not low).

[0106] The principle stated in the previous paragraph can be formulated in a simple way using the L2 error function introduced previously using the following expression:

[0108] The adjustment of the threshold value E will depend on the availability of so-called nominal data. This threshold value may be more or less sensitive depending on the field of application. It can be defined empirically (possibly by calibration), depending on the desired performance of the detection model. Performance measures can, for example, be evaluated as the rate of false negatives and false positives.

[0109] Also, the detection phase PD of the method according to the invention comprises a step E2 of continuous measurement of the current / and the source voltage V source on the power line. It is from these continuous measurements that it is then possible to constitute a window of M measurements noted {l(t m ~), V source (t m ')}^ l=1 (previously introduced). Of course, this is a sliding window that evolves over the course of the measurements acquired during the E2 measurement step.

[0110] It is from this sliding window that it is possible to continuously detect the presence of a fault arc.

[0111] The detection phase PD also includes a step E3 of continuous determination of the presence of an arc using a decision function allowing, as a function of the last M input data relating to the current / and the source voltage V source measured on the power line during the measurement step, to detect the presence of an arc. The decision function is, according to an example, defined by the following expression: r L oo 1121 J

[0113] Where X m = { / (t m ), source (t m )} .

[0114] In one embodiment, the power line is equipped with a current cut-off device for cutting the current in the line on which the detection is carried out, and the method 100 according to the invention comprises, when an electric arc is detected during step E3 of determining the presence of an arc, a step of sending, to the cut-off device, an instruction to cut the current on the line in question. This makes it possible to avoid or, at the very least, limit the damage associated with the formation of an electric arc on the current line in question. Design of a decision function by apprenticeship

[0115] In one embodiment, the method 100 according to the invention takes place in at least two phases, here called PC, PD: a design PC phase making it possible to determine a decision function, that is to say a function which takes as input a plurality of descriptors from a current measurement and which gives as output a Boolean (or any equivalent representation) representative of the presence or absence of an electric arc in the measurement considered; a detection PD phase, during which the current on an electric line is measured so as to extract from this measurement descriptors identified during the design PC phase and detect the presence or absence of an electric arc using the decision function determined during the design PC phase.

[0116] Optionally, when the plurality of measurements necessary for the design phase is not available, the method may also comprise an acquisition phase aimed at obtaining said plurality of measurements.

[0117] As already mentioned, it may be necessary to collect measurements if the latter are not available. For this, in one embodiment, the method 100 comprises, before the design PC phase, an acquisition phase comprising a step of acquiring a plurality of measurements, each measurement comprising a time-correlated measurement of the arc voltage and the arc current. Such a measurement is illustrated in [Fig. 3] showing, at the top, the arc current measurement, and at the bottom, the corresponding arc voltage measurement, the two measurements being time-correlated (i.e. an arc current measurement and an arc voltage measurement having the same abscissa t were measured at the same time t). The power line is, for example, an aircraft network power line.

[0118] This phase is generally implemented in a laboratory where tests are carried out in order to generate a database comprising the plurality of raw measurements. The raw measurements are also accompanied by information relating to the measurement and in particular: The sampling frequency during measurement, for example 1 MHz; The description of the measured signals including: ■ Measurement of arc current (or line current); ■ Measurement of arc voltage; The size of the database: the number of trials (and therefore the number of nominal regions and arc regions - preferably sufficiently large, for example several hundred or thousands of windows).

[0119] Furthermore, in each measurement, the arc voltage and current measurement are time correlated.

[0120] To learn information resulting in a high true positive rate (reliability aspect) and a low false positive rate (robustness aspect), two categories of signals are necessary: ​​electric arc signatures and nominal regions. Nominal regions are all signatures on which an electric arc is not present. This therefore corresponds to nominal but also abnormal regimes ("inrush", load reconfiguration, etc.) of electrical loads connected to the network and powered by the cables under surveillance. The database must therefore contain the largest possible distribution of these signals (aircraft loads, tests in severe representative environments).

[0121] After their acquisition, it may be advantageous to pre-process the measurements in order to facilitate their analysis. Also, in one embodiment, the phase The acquisition process of the method according to the invention includes a step of pre-processing the acquired measurements. This pre-processing step makes it possible in particular to standardize the acquisitions, and to eliminate offsets and defects in the measurement chain. Decision function design phase

[0122] In one embodiment, the method 100 comprises a phase PC of designing a decision function using the plurality of measurements (possibly preprocessed as described previously), this phase being intended to determine a decision function as described previously, that is to say a function which takes as input a plurality of descriptors extracted from an arc current measurement and which gives as output a Boolean (or any equivalent representation) representative of the presence or absence of an electric arc. Pre-identification stage

[0123] Preferably, the PC design phase first includes a pre-identification step by expertise involving a rough identification (without precision to the nearest sample) of all the electric arcs that satisfy predetermined criteria. For this first step, the expert roughly identifies a first time all the arcs on all the tests of the set of measurements. The identification is not precise to the nearest sample. It is simply a matter of defining the position of each arc within the acquisition data. An illustration of these data after pre-identification is given in [Fig. 4] in which two arcs (Arc1 and Arc2) are identified by dotted circles.

[0124] In one embodiment, this pre-identification is done based on specification criteria such as minimum arc duration, minimum arc voltage, etc. In one embodiment, the measurements of the plurality of measurements are stored in one or more files, and the time position and number of arcs per measurement file are stored in a summary file.

[0125] Although optional, this step helps improve the performance of the decision function obtained at the end of the PC design phase.

[0126] The learning phase comprises a labeling step E11. In the present invention, the labeling step E11 is original in that it implements a logic based on the physics of arcs: for each measurement, it uses the arc voltage measurement (only available in the laboratory) as a reference to label the arc current measurement (or line current - available in the laboratory and which will be available later during the PD detection phase) in an automated manner. During this step E11, an algorithm which will be described below scans all the measurements of the plurality of measurements so as to precisely identify the arc start and end information in each arc voltage measurement, and thus select the relevant part (in other words, the labeled part corresponding to a measurement window) in each corresponding arc current measurement.The relevant part thus selected in each arc current measurement of the plurality of measurements will then be used for training. The windows thus labeled are labeled and saved. It is thus possible to label the data on the basis of criteria chosen upstream (e.g., a threshold voltage U. s - see the detailed description of the labeling step below). Verification step

[0127] Preferably, the PC design phase includes, at the end of the labeling step, a step of manual verification of the labeling by expertise. During this step, one or more experts come to check all the windows thus saved. During this step, unlabeled windows, but nevertheless corresponding to the occurrence of an electric arc, can thus be manually labeled. In the same way, windows labeled incorrectly (because the phenomenon presents atypical characteristics) can be relabeled. As a general rule, the labeled windows are saved while the other measurements are anonymized and removed from the backup.In other words, this step allows us to verify that the labeled measurements include all the regions considered for learning, that is, a window including a part of the measurement prior to the formation of the arc, the measurement corresponding to the formation of the arc, as well as a part of the measurement after the formation of the arc.

[0128] Although optional, this step can be useful because it is possible that some arcs are not correctly identified during the labeling step E11. This optional step can therefore contribute to improving the decision function obtained at the end of the PC design phase. Manual identification step

[0129] Preferably, the PC design phase comprises, at the end of the labeling step, a step of manually identifying the remainders, i.e. the missing arcs. During this step, a check of the discrepancies between the number and position of the arcs identified during the pre-identification step (when the latter is implemented) and during the labeling verification step (when the latter is implemented) is carried out. All the arcs identified during the pre-identification step and not marked at the end of the labeling verification step are then marked manually, the objective being to label all the arcs present in the measurements of the plurality of measurements.

[0130] Although optional, this step can therefore contribute to improving the decision function obtained at the end of the PC design phase. Descriptor extraction step

[0131] The PC design phase also comprises, for each measurement of the plurality of measurements, a step E12 of extracting the descriptors of each window of the arc current measurement labeled during the labeling step E11. Indeed, to adjust a model by learning, it is necessary to extract descriptors from a data window relating to each measurement representative of the phenomenon to be detected, this window having been identified during the labeling step E11. These descriptors are informative characteristics on the analyzed measurement window and therefore on the phenomenon to be detected, here, the formation of an electric arc.

[0132] Extraction step E12 can be done in at least two different ways: Using an unsupervised learning method such as autoencoders; in this case, the descriptors have no physical interpretation; Using prior physical knowledge that may be contained in the measurement, such as: ■ Frequency content, for example after a Fourier or wavelet transform; ■ Statistical moments, such as mean, standard deviation, kurtosis, and skewness; ■ Temporal analyses.

[0133] Preferably, in the method 100 according to the invention, the descriptors are extracted using prior physical knowledge and comprise the spectral entropy and / or the autocorrelation of the arc current measured over the selected window.

[0134] In one embodiment, the method comprises, at the end of the step of extracting the descriptors, a step of normalizing the measurements of the plurality of measurements.

[0135] In one embodiment, the PC design phase also includes a descriptor ranking and selection step. Indeed, not all extracted descriptors necessarily have characteristics that can lead to the good segregation of the two existing window classes (arcs and nominals). Thus, analyzing the importance and correlation level of the descriptors with the window label can be useful and improve the performance of the decision function. Indeed, a bad descriptor can possibly disrupt the prediction, decreasing (sometimes drastically) the final performance of the decision function. Also, the selection of descriptors can generally lead to better learning performances, such as higher efficiency, lower computational cost and a more efficient automatic arc detection model.Also, this embodiment is particularly advantageous when the number of descriptors is high.

[0136] Thus, the goal of the descriptor selection step is to maximize relevance and minimize redundancy of information. In an exemplary implementation, this step is performed by a supervised filtering method.

[0137] In an exemplary embodiment, the ranking could be performed using a Gram-Schmidt method and the selection could be performed using an incremental method. Of course, other methods could be used for ranking or selection.

[0138] Although optional, this step can therefore contribute to improving the decision function obtained at the end of the PC design phase. Decision function determination step

[0139] The PC design phase also includes a step E13 for determining the decision function, noted f(X) in the following (where X represents the descriptors extracted from the arc current measurement and provided as input to the decision function).

[0140] In one embodiment, this step E13 comprises a plurality of iterations of the following two sub-steps: a sub-step of designing, by learning, a classifier; and a sub-step of validating the classifier from the labeled data.

[0141] Furthermore, these sub-steps are repeated until the best performing classifier is obtained, i.e. for example the classifier minimizing below a certain threshold the classification error made on a non-learned database (so as to ensure a generalization capacity of said classifier). The decision function f(X) is determined from this classifier.

[0142] Thus, the determination of a classifier allows, by learning (corresponding to the plurality of iterations of the sub-steps described previously), to obtain a high-performance classifier so as to make an effective decision.

[0143] In order to implement this step E13, several supervised classification methods can be used. The choice of method can be influenced in particular by the desired detection time: some methods may have a higher computational complexity, but better performance. Preferably, the classification method used during this step is a support vector machine method. This method being well known to the person skilled in the art, it will not be described further here.

[0144] In an alternative embodiment, the classification method is selected from one of the following methods: Formal neural networks (simple, convolutional, recurrent); Hidden Markov Models; K-nearest neighbors method; Decision tree; Set learning (pairing two methods from the methods previously listed).

[0145] All steps E11-E13 described so far are implemented prior to the PD detection phase and, at the end of the PC design phase, the decision function (and therefore the detection strategy) is fixed. Also, the descriptors selected during the PC design phase are the descriptors that will be used during the PD detection phase. During the PD detection phase, these will therefore be extracted from the measured signals (for example via signal processing) and used as inputs to the decision function which will assign to the signal either the “arc” category or the “nominal regime” category (i.e. absence of electric arc). Detailed description of the labeling step

[0146] In the following, in order to illustrate the labeling step E11, the following notation is used: t is the instant considered; t end is the total duration of the current trial; T maxis the maximum window duration; T min is the minimum window duration; P g is the left step; its value is normalized for the duration of the window under study and between [0,1]; P d is the right step; P d = 1 - P g , it is therefore also a normalized value for the duration of the window under study and between [0,1]; t d is the starting time of the arc; t e is the instant of extinction of the arc; tevaiuation is the time of evaluation of the percentage of points which are above U s before starting an arc; U arc is the arc voltage; l arc is the arc current; U s is the minimum voltage threshold from which the existence of an electric arc is considered true; this voltage threshold (or threshold voltage) is a function of knowledge of the physics of arcs.

[0147] As detailed previously, the labeling step E11 according to the invention is subsequent to the signal acquisition step or even to the preprocessing step of said signals. In the case of the electrical signatures of fault arcs, the only information making it possible to guarantee the presence of an electric arc is the measurement of the voltage at its terminals (the arc voltage), since this value will be zero in the absence of discharge and greater than a certain value only in the presence of an arc. As a reminder, the information of this arc voltage is not measured (because it is not available) during the PD detection phase. It is only accessible and used during the PC design phase. It is therefore from this information that the labeling strategy according to the invention is defined.In other words, the arc voltage measurement is used exclusively to label the current measurement (for training and validation data that are acquired in the laboratory) and this use is limited to the PC design phase.

[0148] Furthermore, during labeling step E11, the characteristics of the phenomenon to be detected must be well defined. They represent the specificity of the defect that we are seeking to detect.

[0149] In an exemplary embodiment, the characteristics of the arcs to be detected are as follows: an arc duration > 1 ms; an increase in the arc voltage from 0 to 10V minimum.

[0150] In an exemplary embodiment, the characteristics of the windows labeled as arcs are as follows: a window duration between 1 and 10 ms; presence of a negative time shift of a duration proportional to the length of the window; absence of end-of-arc signature (correlated to a drop in arc voltage towards 0V).

[0151] These values ​​are provided as an example but do not represent a limitation of the invention.

[0152] In the state of the art, labeling is a delicate and time-consuming operation. However, in the method 100 according to the invention, this step E11 is automated, making it possible to ensure the same result as manual implementation, while optimizing the time spent.

[0153] For this, the labeling step E11 includes, for each measurement and for each time interval t of the measurement considered, the following sub-steps: when the measured arc voltage U arc exceeds a threshold value U for the first time spredetermined (i.e. at the first occurrence of the following relation: U arc (t) > U s and U arc (t - 1) < U s where t is the interval considered), a step of recording the interval considered as the starting point of the electric arc (in other words t d = t); when the measured arc voltage U arc falls below the said threshold value U for the first time s predetermined (i.e. at the first occurrence of the following relation: U arc (t) < U s and U arc (t - 1) > U s where t is the interval considered), a step of recording the interval preceding the interval considered (i.e. t - 1) as the end point of the electric arc (i.e. t e = t - 1).

[0154] These two sub-steps are illustrated in the form of a flowchart in [Fig. 5]. In this flowchart, the measured arc voltage is compared with the threshold arc voltage. If the measured arc voltage is higher than the threshold arc voltage, the previous values ​​of the arc voltage are checked. If none of the said previous values ​​were above this threshold value, this means that the arc voltage is for the first time above the threshold value, signaling the start of the arc. Otherwise, nothing is done. Similarly, if the measured arc voltage value is below the threshold value and the threshold value was above this same threshold value, then this signals that the arc voltage is falling back below the threshold value, signaling the end of the arc.

[0155] In one embodiment, for each measurement, the labeling is associated with a measurement window over which the learning is carried out, said window being given by the interval: [P g x min (t e - t d ,T max , P d x min (t e - t d ,T max)]. It is from the data of this interval that the descriptors will be extracted. Such a measurement window is illustrated in [Fig. 6] which represents, at the top, the arc current as a function of time and, at the bottom, the arc voltage as a function of time. This is a simplified representation of electrical profiles appearing in the case of a series arc in direct voltage on an electrical system supplying a resistive load. In the example of [Fig. 6], the appearance of the arc is represented on the voltage curve by a voltage increase of a few Volts or a few tens of Volts. On the current curve, this is represented by the appearance of a current drop occurring at the same time. In addition to these drops, the labeling step seeks to retain a small part of the signal to the left of the drop (nominal region) and an even larger part of the signal to the right of the drop (established arc) in order to store a significant frequency content of the phenomenon.As already mentioned, the beginning of an arc, at time t. d , is first identified as the moment when the measured arc voltage U arc exceeds a threshold value U s determined, chosen in relation to the physics of the arcs. In one embodiment, the value of U s is given by the sum of the anodic and cathodic drop, for example a value of 20V for the sum of the two voltage drops in air.

[0156] Once the start of the arc is identified, its end (t e ) is determined in the same way, but representing the moment when the voltage returns to a value lower than the established threshold. With this, it is possible to identify the duration of the analyzed electric arc (d = t e - t d ). If this duration is greater than the maximum windowing value desired in the detection (T max), this predefined maximum value is taken as a reference. Otherwise, where the arc duration is less than that of the established windowing and greater than the minimum duration (T min ) acceptable for the arc, the arc duration is taken as the reference window. The regions to the right and left of t d are then added according to a percentage (P d + P g = 1) of this reference duration. All arcs are then labeled in this way.

[0157] In one embodiment, the threshold voltage is determined from the evolution of the voltage measured at the terminals of an electric arc as a function of the arc length, this evolution being presented in [Fig. 7]. It is possible to identify three zones in this figure. In the zones located at the beginning and end of the arc, the arc voltage evolves in a non-linear manner. These are the anodic and cathodic zones within which the arc voltage drop represents a non-negligible, unavoidable value and systematically greater than 0 V. In general, the values ​​of V c + V aare between 10 and 25V in the environmental conditions encountered in the aeronautics field. The third zone is the central zone whose arc voltage progresses linearly as a function of the distance traveled by the arc and is not of particular interest in the implementation of the method according to the invention. In other words, the appearance of an electric arc will be systematically accompanied by the production of an arc voltage whose value will be defined by a minimum threshold explained by the voltage drops at the terminals of the electrodes. Detection phase

[0158] In order to be able to detect the presence of an electric arc on a power line, the method comprises a PD detection phase subsequent to the PC design phase and during which the decision function obtained during the PC design phase will be used to determine the presence or absence of an electric arc in a current measurement. Unlike the PC design phase where the arc voltage is an available measurement, in real conditions (outside the laboratory), that is to say during the PD detection phase, only the arc current measurement may be available. It is therefore preferably from this measurement alone that the detection will take place.

[0159] More specifically, the PD detection phase firstly comprises a step E14 for measuring the current flowing on the power line. This step E14 is carried out continuously and in real time so as to enable rapid detection of the presence of an arc.

[0160] The PD detection phase also includes a step E15 for extracting descriptors from the current measurement. As already mentioned, the descriptors extracted during this step E15 are identical to those selected during the PC design phase and serve as input data for the decision function. They are also extracted continuously, as the current measurements are acquired.

[0161] The detection phase PD also includes a step E16 for detecting the presence of an arc in the measurement of the current flowing on the electrical line L1. This step uses the descriptors extracted from the current measurement and the decision function determined during the design phase PC.

[0162] In one embodiment, it is possible to cut the current in the electrical line L1 on which the detection is carried out using a current cutting device, and the method 100 according to the invention comprises, when an electric arc is detected during the detection step E16, a step of sending, to the cutting device MMC, an instruction to cut the current on the line L1 in question. This makes it possible to avoid or, at the very least, limit the damage associated with the formation of an electric arc on the current line L1 in question.

[0163] The quality of the detection depends on the quality of the learning which itself depends on the quality of the data and their labeling. One of the contributions of the method according to the invention is to automate this labeling by using a plurality of measurements (generally carried out in the laboratory), each measurement comprising a time-correlated measurement of the arc voltage and the arc current. Device for detecting the presence of an electric arc

[0164] A second aspect of the invention illustrated in [Fig. 8] relates to a DI device comprising the means configured to implement a method 100 according to the invention. More particularly, the device comprises a calculation means MC (for example a processor or an ASIC card) associated with a memory MM (for example a RAM memory and / or a hard disk), said memory MM being configured to store the data and the instructions necessary for implementing the method 100 according to the invention. The DI device according to the invention also comprises a measurement means CPT configured to measure, in real time, the current flowing through a current line L1 when the latter is connected to such a current line L1 and communicate the measurement to the calculation means MC of the DI device according to the invention.In an exemplary embodiment, the device is an RCCB (for “remote controller circuit breaker” in English) or even an SSPC (for “solid state power controller” in English).

[0165] In one embodiment, the current flowing through the current line L1 may be cut off by a controllable MCC current cut-off device. distance, and the DI device according to the invention is configured to, when the presence of an arc is detected on the current line L1, send an instruction to the current cutting device MCC so as to cut the current in the current line.

Claims

CLAIMS

1. Computer-implemented method (100) for detecting the presence of an electric arc on a power line (L1) using an artificial learning method, the learning being performed using: - a plurality of measurements relating to the formation of an electric arc, each measurement of the plurality of measurements comprising a time-correlated measurement of the arc voltage V arc , of the current / and of the source voltage V source ; - of a knowledge model linking the arc voltage V arcand the current / to constants depending on the knowledge model; the method (100) being implemented using a learning model resulting from a learning phase (PA) during which, using the measurements of the plurality of measurements and the knowledge model, a step (E1) of adjusting the learning model is implemented, the learning model taking as input the current / and the source voltage V source and providing the arc voltage V as output arc and the constants of the knowledge model, learning being carried out using a cost function L given by the following relation L = L + L2 where is a first cost subfunction measuring a first difference between the arc voltage V^ c edlt predicted by the learning model and the arc voltage V arc measured: and where L2 is a second cost sub-function measuring a second deviation between the prediction of the knowledge model and the measured values; the method comprising a detection phase (PD) comprising: - A step (E2) of continuous measurement of the source current / and voltage V source on the power line (Ll), so as to acquire input data relating to the current / and the source voltage V source on a measurement window where M is the number of measurement points in the measurement window, / (t m ) is the current at time t m and V source (t^ is the source voltage at time - A step (E3) of continuous determination of the presence of an arc, using a decision function allowing, as a function of the last M input data relating to the current / and the source voltage V sourcemeasured on the power line (Ll) during the measurement step, to detect the presence of an arc, the decision function being a function of the second cost sub-function.

2. Method according to the preceding claim in which the decision function is defined by the following expression: Where X m = {I (t m ),V source (t m )}, £ is a detection threshold value, and L2(Xm) is the value of the second cost subfunction L2 determined using the learning model calculated at point X m .

3. Method according to the preceding claim in which the knowledge model is the Mayr model and the second cost sub-function L2 takes the following form: where T is the arc constant and P o is the cooling power.

4. Method according to claim 1 in which the knowledge model is the Cassie model and the second cost sub-function L2 takes the following form: where T is the arc constant and 70 is the minimum arc voltage.

5. A method according to claim 1 wherein the knowledge model is the Shavemaker model and the second cost subfunction L2 takes the following form: where T is the arc constant, P o is the cooling power and 70 is la minimum arc voltage.

6. Method (100) according to one of the preceding claims, in which the electrical line (L1) is equipped with a cut-off device (MCC) making it possible to cut the current in said line, the method (100) comprising, when an electric arc is detected during the detection step (E3), a step of sending, to the cut-off device (MCC), an instruction to cut the current on the line (L1) in question.

7. The method (100) of claim 1, wherein the method further comprises: - a phase (PC) of designing a decision function by artificial learning, using the plurality of measurements, the design phase (PC) comprising: ■ For each measurement of the plurality of measurements, a step (E11) of labeling the arc current measurement using the arc voltage measurement associated with it; ■ A step (E12) of extracting descriptors from the labeled measurements resulting from the labeling step; ■ A step (E13) of determining the decision function, by artificial learning, from at least part of the descriptors extracted during the extraction step (E12); in which the phase (PD) of detecting the presence of electric arcs comprises: ■ A step (E14) of measuring the current flowing on the line (Ll); ■ A step (E15) of extracting the descriptors of the current measurement obtained during the previous current measurement step (E14); ■ From a part of the descriptors extracted during the previous step and using the decision function determined during of the design phase (PC), a step (E16) of detecting the presence of arcs in the current measurement; the labeling step (E11) comprising, for each measurement of the plurality of measurements and for each time interval t of the measurement considered: - When the measured arc voltage U arc exceeds a threshold value U for the first time s predetermined, a recording step of an interval considered as starting point t d of the electric arc; - When the measured arc voltage U arc falls below the threshold value U s predetermined for the first time, a recording step of the interval preceding an interval considered as end point t e of the electric arc.

8. Method according to claim 7 in which, for each measurement, the labeling is associated with a measurement window on which the learning is carried out, said window being given by the interval [P g x 7îii7î((t e t d , T max ), P d x min (t e t d T max ) where T max is a predefined maximum window duration, P g is a step to the left of t d having a normalized value for the duration of the window under study, and P d is a step to the right of t d having a normalized value for the duration of the study window given by the relation P d = 1 - P a .

9. Method according to one of claims 7 or 8, comprising, before the design phase, a step of acquiring signals necessary for learning so as to obtain a plurality of measurements, each measurement comprising a time-correlated measurement of the arc voltage and the arc current.

10. Method according to one of claims 7 to 9, in which the design phase comprises, before the labeling step, a step of manual pre-identification by expertise, in each measurement of the plurality of measurements, of all the electric arcs satisfying predetermined criteria so as to define a position of each arc within each measurement of the plurality of measurements.

11. Method according to one of claims 7 to 10, in which the design phase comprises, at the end of the labeling step (E11), a step of manual verification of the outputs of the labeling step by expertise so as to verify that the labeled measurements include all of the regions envisaged for learning.

12. Method according to one of claims 7 to 11, in which the descriptors used during the step of determining (E13) the decision function comprise the spectral entropy and / or the autocorrelation of the arc current.

13. Device (DI) for detecting the presence of an arc in a current line (L1), the device comprising a measuring means (CPT) configured to measure the current flowing in the current line and means configured to implement a method (100) according to one of claims 1 to 5.

14. A computer program comprising instructions which, when executed by a computer, cause the device according to claim 13 to execute the method according to one of claims 1 to 12.

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

16. An aircraft comprising a detection device according to claim 13.

Images