Method, device and equipment for detecting an electric arc fault

Abstract

The invention relates to a method for detecting an electric arc fault, the method comprising: - a step of calculating the RMS value of the current, prior to an artificial neural network input step, in

Description

Title of the invention: Method, device, apparatus for detecting an electric arc fault

[0001] The present invention relates to the field of methods for detecting an electric arc fault, the field of devices for detecting an electric arc fault and the field of electrical equipment for detecting an electric arc fault.

[0002] Already known from the IEEE - Research paper 2021: Y. Wang, L. Hou, KC Paul, Y. Ban, C. Chen and T. Zhao, "ArcNet: Series AC Arc Fault Detection Based on Raw Current and Convolutional Neural Network," in IEEE Transactions on Industrial Informatics, vol. 18, no. 1 , pp. 77-86, Jan. 2022, doi: 10.1109 / TII.2021 .3069849, is an arc fault detection solution called ArcNet that directly uses the measured raw current signal without preprocessing as input to a convolutional neural network. This implies that a deep and complicated model must be used in order to achieve good detection performance. Indeed, the convolutional neural network comprises at least four convolutional layers followed by three fully connected layers. One of the disadvantages of this solution is that it is expensive to implement all the necessary calculations due to the required computing unit and the energy and space consumption.Indeed, the model is implemented in a Raspberry Pi 3B nanocomputer (Quad-Core 1.2GHz Broadcom BCM2837 64bit CPU, 1 GB RAM) for classification accuracy. The proposed ArcNet solution has an average execution time of 31 milliseconds / sample of a cycle at a sampling rate of 10 kiloHertz. Furthermore, this solution requires a full cycle for detection, which implies a 20 millisecond delay for triggering and imposes a suboptimal time resolution. More precisely, the arc fault detection standard EN 62606 dictates the trigger time requirements that are partly described below. For a trigger time resolution of 20 milliseconds, in the case of a nominal current of 10 Amperes / 16 Amperes, the solution triggers rather at 240 / 140 milliseconds instead of 250 / 150 milliseconds. So, an opportunity to acquire more. of information before confirming or rejecting arc fault detection is lost. Using the raw current database, experimental results indicate that the ArcNet solution can achieve a maximum arc detection accuracy of 99.47% at a sampling rate of 10 kiloHertz.

[0003] Publication WO2021136053A1 discloses a fault arc identification method, device, and apparatus, and a storage medium. The method comprises: performing sampling on a target arc at a high frequency, and obtaining a high-frequency sampling signal S11; preprocessing the high-frequency sampling signal and obtaining a processed sampling signal S12; performing feature extraction on the processed sampling signal, and obtaining a target arc feature S13; and inputting the target arc feature into a neural network model, obtaining a target output result, and determining, according to the target output result, whether the target arc is a fault arc S14. Performing sampling on a target arc at a high frequency can obtain more arc features from the target arc.In addition, since a neural network model has favorable data classification capability, using a neural network model to make a determination with respect to the target arc can improve the accuracy and reliability of a fault arc detection result.

[0004] Finally, publication US2019199080A1 discloses an arc fault detection system that includes a high frequency detector module structured to detect a first spectral content of power having a frequency in a first frequency range, a number of arc fault circuit interrupters each having separable contacts, a low frequency detector module structured to detect a second spectral content of power having a frequency in a second frequency range, a controller structured to control the separable contacts to open, and an arc fault detection unit structured to detect an arc fault based on the detected first and second spectral contents of the power and to cause the controller to control the opening of the separable contacts in response to detection of the arc fault. A communication bus provides communication between the high frequency detector and the arc fault detection unit. frequency and arc fault circuit interrupters and the first frequency range includes frequencies that are higher than the frequencies in the second frequency range. The spectral content can be divided into high frequency spectral content either above about 1 Megahertz and low frequency spectral content either below about 1 Megahertz. This solution relies solely on the spectral content for detection and is therefore sensitive to a disturbance at the given frequency band near the chosen frequency. In this solution, the arc fault detection unit does not combine the high and low frequency spectral content, but rather analyzes the low frequency spectral content and based on the analysis of the low frequency spectral content and the indication that an arc fault is potentially present derived from the high frequency spectral content, ultimately determines whether an arc fault is present.The solution uses stacked information and sequential verification rather than a true combination strategy. The invention only works with a fixed signal acquisition platform; any hardware change could require a complete revision of the detection algorithm.

[0005] The present invention aims to overcome at least one of these drawbacks and propose a solution using a high-performance and inexpensive artificial neural network.

[0006] To this end, the invention relates to a method for detecting an electric arc fault comprising:

[0007] a reception step, during which at least one acquired current signal is received,

[0008] an extraction step, during which a set of arc fault characteristics is extracted from the arc fault characteristics from the acquired current signal,

[0009] an input step into an artificial neural network, during which a set of input data from at least the extraction step is input into an artificial neural network to determine at least one set of output data,

[0010] a decision step, in which said output data set is processed to decide whether an arc fault is present or not,

[0011] the detection method is characterized in that it further comprises:

[0012] a step of calculating the effective value of the current which is prior to the step of inputting into an artificial neural network, during which an effective value of the current is calculated from the current signal acquired over a given time interval, preferably equal to 10 milliseconds, and in that

[0013] during said step of inputting into an artificial neural network, said one set of input data is input, depending on the current range in which the effective value of the current lies, into an artificial neural network chosen at least from two artificial neural networks each associated respectively with a predetermined current range,

[0014] and in that said at least two artificial neural networks are pre-trained from arc fault characteristics included in their respective current range.

[0015] The invention further relates to a device for detecting an electric arc fault comprising at least one module for detecting an electric arc fault, said detection module is characterized in that it is configured to implement the method for detecting an electric arc fault according to the invention.

[0016] The invention further relates to an electrical apparatus for detecting an electric arc fault comprising at least:

[0017] an electrical line comprising a pair of contacts,

[0018] a current measuring device,

[0019] a device for detecting an electric arc fault,

[0020] a control module for opening the pair of contacts,

[0021] said electrical arc fault detection apparatus is characterized in that it comprises the detection device according to the invention.

[0022] The invention further relates to a computer program comprising instructions which, when the program is executed by a computer, lead the computer to implement the method for detecting an electric arc fault according to the invention.

[0023] The invention further relates to a computer-readable medium on which the computer program according to the invention is recorded.

[0024] The invention will be better understood from the following description, which relates to several preferred embodiments, given as non-limiting examples, and explained with reference to the appended schematic drawings, in which:

[0025] [Fig. 1] Figure 1 represents a schematic view of an artificial neural network according to the invention,

[0026] [Fig. 2] Figure 2 represents a schematic view of the method for detecting an electric arc fault according to the invention,

[0027] [Fig. 3] Figure 3 represents a summary table of the performance of the method for detecting an electric arc fault, and

[0028] [Fig. 4] Figure 4 represents a curve of the statistical model.

[0029] A method for detecting an electric arc fault includes:

[0030] a reception step, during which at least one acquired current signal is received,

[0031] an extraction step, during which a set of arc fault characteristics is extracted from the arc fault characteristics from the acquired current signal,

[0032] an input step into an artificial neural network, during which a set of input data (11, I2, I3, I4..., 1316) resulting at least from the extraction step is entered into an artificial neural network to determine at least one set of output data (O),

[0033] a decision step, in which said output data set (O) is processed to decide whether an arc fault is present or not.

[0034] In accordance with the invention and as illustrated in Figure 2, the detection method is characterized in that it further comprises:

[0035] a step of calculating the effective value of the current (RMS) which is prior to the step of inputting into an artificial neural network, during which an effective value of the current is calculated from the current signal acquired over a given time interval, preferably equal to 10 milliseconds, and in that

[0036] during said step of inputting into an artificial neural network, said one set of input data is input, depending on the current range in which the effective value of the current lies, into an artificial neural network chosen at least from two artificial neural networks (ANN1, ANN2, ANN3..., ANNn) each associated respectively with a predetermined current range (P1, P2, P3..., Pn),

[0037] and in that said at least two artificial neural networks (ANN1, ANN2, ANN3..., ANNn) are trained beforehand from arc fault characteristics included in their respective current range (P1, P2, P3..., Pn).

[0038] Advantageously, since the arc fault characteristics vary depending on the circuit current rate and the load, high rated current arcs are more dangerous, evolve more rapidly over time and have more distinct characteristics than low rated current arcs. Based on this observation, the present invention proposes to use several artificial neural networks (ANN1, ANN2, ANN3..., ANNn) which each correspond to a predetermined current range (P1, P2, P3..., Pn) to adapt to different arc fault characteristics depending on the current range (P1, P2, P3..., Pn). As illustrated in Figure 2, the effective value of the current (RMS) is used over a given time interval preferably equal to 10 milliseconds or half a cycle to determine which of the artificial neural networks (ANN1, ANN2, ANN3..., ANNn) should be used.A single artificial neural network (ANN1, ANN2, ANN3..., ANNn) operates during this given time interval. Due to the presence of at least two artificial neural networks (ANN1, ANN2, ANN3..., ANNn), the solution proposes a multiple classification principle instead of a monolithic classification known from the prior art. The at least two artificial neural networks (ANN1, ANN2, ANN3..., ANNn) each corresponding to a current range (P1, P2, P3..., Pn) were. trained upstream to adapt to different arc fault characteristics associated with each current range (P1, P2, P3..., Pn). Figure 3 illustrates that the performance obtained using the method according to the invention is better due to the multiple classification and therefore the presence of at least two artificial neural networks (ANN1, ANN2, ANN3..., ANNn) compared to a method using a monolithic classification. There are many ways to divide the total current range, which preferably is between 0 Amperes and 40 Amperes, into several current ranges (P1, P2, P3..., Pn), several examples are detailed below. Preferably, this division is based on the experimental results and is strongly linked to the hardware part and in particular to the current measuring device used which will be described later.The solution according to the invention uses a set of arc fault features rather than the raw current signal. For this reason, the method uses an extraction step. First, it may be necessary to select which arc fault features will be used and then a set of arc fault features are extracted in the extraction step. This way of proceeding makes the training at a later stage less prone to noisy or irrelevant inputs, helps the artificial neural networks to learn transparently and with more common sense and simplifies the architecture of the artificial neural networks (ANN1, ANN2, ANN3..., ANNn). As a result, the detection performance is improved because this method is more sensitive and more effective against false triggering.Furthermore, the computer development time is reduced compared to a human-coded algorithmic model. Development time is a few days instead of weeks. This process can also be implemented in an inexpensive processor.

[0039] Preferably, the receiving step, the extracting step, the step of calculating the effective value of the current, the step of inputting into an artificial neural network and the decision step are repeated at the given time interval preferably equal to 10 milliseconds.

[0040] The proposed solution preferably inputs data every given time interval preferably every half cycle, which sets a trigger time resolution of 10 milliseconds.

[0041] Preferably and according to a first possibility, during said one step of entering into an artificial neural network, said artificial neural network is chosen from a first artificial neural network (ANN1) associated with a first current range (P1), preferably between 0 Amps and 10 Amps, and a second artificial neural network (ANN2) associated with a second current range (P2), preferably between 10 Amps and 40 Amps, the first current range (P1) and the second current range (P2) being disjoint.

[0042] In this case and preferably, during said step of inputting into an artificial neural network, if the effective value of the current (RMS) belongs to the first current range (P1) then said set of input data (11, I2, I3, I4, 1316) is input into the first artificial neural network (ANN1) or if the effective value of the current (RMS) belongs to the second current range (P2) then said set of input data (11, I2, I3, I4, 1316) is input into the second artificial neural network (ANN2).

[0043] Advantageously, this solution uses two artificial neural networks (ANN1, ANN2) associated respectively with the first current range P1 corresponding to so-called low currents between 0 Amps and 10 Amps and with the second current range P2 corresponding to so-called high currents between 10 Amps and 40 Amps. As illustrated in Figure 3, the performances obtained are better with this solution using two artificial neural networks (ANN1, ANN2) compared to a method using a monolithic classification with a single artificial neural network (ANN), known from the prior art.

[0044] Preferably and according to a second possibility, during the step of entering an artificial neural network, said artificial neural network is chosen from a first artificial neural network (ANN1) associated with a first current range (P1), preferably between 0 Amps and 10 Amps, a second artificial neural network (ANN2) associated with a second current range (P2), preferably between 10 Amps and 20 Amps, and a third artificial neural network (ANN3) associated with a third current range (P3), preferably between 20 Amps and 40 Amps, the first current range (P1), the second current range (P2) and the third current range (P3) being disjoint.

[0045] In this case and preferably, during the step of inputting into an artificial neural network, if the effective value of the current (RMS) belongs to the first current range (P1 ) then the input data set (11 , I2, I3, I4... , 1316) is input into the first artificial neural network (ANN1 ) or if the effective value of the current (RMS) belongs to the second current range (P2) then the input data set (11 , I2, I3, I4. , 1316) is input into the second artificial neural network (ANN2) or if the effective value of the current (RMS) belongs to the third current range (P3) then the input data set (11 , I2, I3, I4... , 1316) is input into the third artificial neural network (ANN3).

[0046] Advantageously, this solution uses three artificial neural networks (ANN1, ANN2, ANN3) associated respectively with the first current range P1 corresponding to so-called low currents between 0 Amps and 10 Amps, with the second current range P2 and the third current range P3 corresponding to so-called high currents between 10 Amps and 20 Amps and between 20 Amps and 40 Amps. As illustrated in Figure 3, the performances obtained are better with this solution using three artificial neural networks (ANN1, ANN2, ANN3) compared to a method using a monolithic classification with a single artificial neural network (ANN), known from the prior art or two artificial neural networks (ANN1, ANN2) according to the invention.

[0047] Preferably, the method for detecting an electric arc fault further comprises a normalization step (N1, N2..., Nn) prior to the input step, during which said set of arc fault characteristics is normalized to form a set of normalized input data and in that during the step of inputting into an artificial neural network, the set of input data (11, I2, I3, I4..., I316) corresponds to the set of normalized input data.

[0048] The objective of the normalization shown in Figure 2 is to convert the floating point variables of the arc fault characteristics into integers between 0 and 1.

[0049] Min-max normalization is preferably used, due to the different values ​​of the arc fault characteristic in each database and each normalization step (N1, N2..., Nn) is unique.

[0050] Preferably, in the decision step, said output data set (O) is statistically processed by a sequential odds ratio (SD) test to decide whether an arc fault is present or not.

[0051] This statistical decision, which is illustrated in Figures 3 and 4, allows as much information as possible to be gathered before making a decision during the decision stage. This approach optimizes detection time and minimizes false triggers linked to the absence of dangerous sporadic arcs.

[0052] At the end of the decision step and as illustrated in Figure 2, a binary output variable is determined based on the conclusion of absence or presence of an electric arc fault. This binary output variable can constitute a control signal (SC) of an electrical apparatus according to the invention described below.

[0053] Preferably and as illustrated in Figure 1, each artificial neural network (ANN1, ANN2, ANN3..., ANNn) preferably comprises three layers (L1, L2, L3). The input data set (11, I2, I3, I4..., 1316) preferably comprises 316 variables. The artificial neural network comprises approximately 9000 coefficients and calculates approximately 9000 multiply accumulation operations over the given time interval.

[0054] A classical fully connected neural network is preferably used. The output layer (LS) preferably comprises a neuron that has a sigmoid or softmax activation function. The first three layers (L1, L2, L3) can use a sigmoid or rectified linear unit as the activation function. For the learning process, the binary cross-entropy loss is preferably used as the cost function. To minimize the cost function, the stochastic gradient of the algorithm is used, and the Classical accelerated gradient Momentum and Nesterov are preferably applied during the training process. Depending on the convergence of the neural network, regularization techniques such as early stopping, ridge regression, lasso or similar can be used.

[0055] The acquired current signal received during the receiving step can also be preprocessed before the extraction step during a preprocessing step to determine the arc fault characteristics.

[0056] Preferably, this preprocessing step comprises at least frequency analysis and waveform analysis and sampling.

[0057] Indeed, preferably only a selection of the arc fault characteristics resulting from the preprocessing step forms said set of arc fault characteristics.

[0058] The invention further relates to a device for detecting an electric arc fault comprising at least one module for detecting an electric arc fault. According to the invention, said detection module is characterized in that it is configured to implement the method for detecting an electric arc fault according to the invention and as described previously.

[0059] Advantageously, the method for detecting an electric arc fault according to the invention can be implemented in a device for detecting an electric arc fault (not shown) according to the invention and thus be presented in the form of a functional brick which can be added for example to electrical equipment (not shown).

[0060] Preferably according to an example not shown, the module for detecting an electric arc fault may comprise an electronic card comprising a processor configured to implement the method for detecting an electric arc fault according to the invention. For example, the processor consists of an ARM cortex m4-64 Megahertz processor. This type of processor has the advantage of being inexpensive and therefore capable of being integrated into an industrializable solution. When the method for detecting an electric arc fault is implemented with this type of processor, 30 percent of the processor's computing time is used.

[0061] The detection module may further comprise at least one filter for performing the frequency analysis in the preprocessing step. Preferably, the detection module comprises three bandpass filters at 20 kiloHertz, 30 kiloHertz and 50 kiloHertz.

[0062] The detection module may further comprise at least one integrator for performing the analysis of the waveform in the preprocessing step.

[0063] The invention further relates to an electrical apparatus for detecting an electric arc fault comprising at least:

[0064] an electrical line comprising a pair of separable contacts,

[0065] a current measuring device,

[0066] a device for detecting an electric arc fault,

[0067] a module for controlling the opening of the pair of separable contacts.

[0068] According to the invention, said electrical apparatus for detecting arc faults is characterized in that it comprises the detection device according to the invention and as described previously.

[0069] Advantageously and according to an example not shown, the device for detecting an electric arc fault according to the invention can be part of an electrical apparatus for detecting an electric arc fault and thus integrate an electric arc fault detection function. Thus, the current measuring device can measure the current signal of the electrical line to be protected. If an arc fault is detected by the detection module implementing the method according to the invention, a signal for controlling the opening of the contacts can in response be sent to the control module which will in response control the opening of the separable contacts.

[0070] The electrical apparatus for detecting an electric arc fault may also perform a function of protection against short-circuit faults and / or against overcurrent faults and / or against differential faults. In this case, the electrical apparatus for detecting an electric arc fault may consist of a molded case circuit breaker of the MCB-AFDD type or a molded case differential circuit breaker of the RCBO-AFDD type.

[0071] Preferably, the current measuring device is a current sensor, for example a current transformer.

[0072] The invention further relates to a computer program comprising instructions which, when the program is executed by a computer, cause the computer to implement the method for detecting an electric arc fault according to the invention and as described above.

[0073] A computer is any programmable information processing system.

[0074] The invention further relates to a computer-readable medium, on which the computer program according to the invention and as described above is recorded.

[0075] The computer-readable medium may be a chip, a hard disk, or the like. These examples are not limiting.

[0076] Of course, the invention is not limited to the embodiments described and shown in the attached drawings. Modifications remain possible, particularly from the point of view of the constitution of the various elements or by substitution of technical equivalents, without departing from the scope of protection of the invention.

Claims

1. A method for detecting an electric arc fault, comprising:- a reception step during which at least one acquired current signal is received, - an extraction step during which a set of arc-fault characteristics is extracted from the arc-fault characteristics from the acquired current signal,- an artificial neural network input step during which a set of input data (11, I2, I3, 14..., 1316) at least from the extraction step is input into an artificial neural network to determine at least one set of output data (0),- a decision step during which said set of output data (0) is processed to decide whether an arc fault is present or not, the detection method being characterized in that it further comprises:- prior to the artificial neural network input step, a step of calculating the RMS value of the current, during which an RMS value of the current is calculated from the acquired current signal over a given time interval of preferably equal to 10 milliseconds, and in thatduring said artificial neural network input step, said set of input data is input, according to the current range in which the RMS value of the current is located, into an artificial neural network chosen at least from among two artificial neural networks (ANN1, ANN2, ANN3..., ANNn) each of which is respectively associated with a predetermined current range (P1, P2, P3..., Pn),and in that said at least two artificial neural networks (ANN1, ANN2, ANN3..., ANNn) are pre-trained on the basis of arc-fault characteristics comprised in their respective current range (P1, P2, P3..., Pn).

2. The method for detecting an electric arc fault as claimed in claim 1, characterized in that the reception step, the extraction step, the step of calculating the RMS value of the current, the artificial neural network input step and the decision step are repeated at the given time interval of preferably equal to 10 milliseconds.

3. The method for detecting an electric arc fault as claimed in either one of claims 1 and 2, characterized in that during said artificial neural network input step, said artificial neural network is chosen from among a first artificial neural network (ANN1) associated with a first current range (P1), preferably comprisedbetween 0 Amperes and 10 Amperes, and a second artificial neural network (ANN2) associated with a second current range (P2), preferably comprised between 10 Amperes and 40 Amperes, the first current range (P1) and the second current range (P2) being different.

4. The method for detecting an electric arc fault as claimed in claim 3, characterized in that during said artificial neural network input step, if the RMS value of the current belongs to the first current range (P1) then said set of input data (11, I2, I3, 14..., 1316) is input into the first artificial neural network (ANN1) or if the RMS value of the current belongs to the second current range (P2) then said set of input data (11, I2, I3, 14..., 1316) is input into the second artificial neural network (ANN2).

5. The method for detecting an electric arc fault as claimed in either one of claims 1 and 2, characterized in that during the artificial neural network input step, said artificial neural network is chosen from among a first artificial neural network (ANN1) associated with a first current range (P1), preferably comprised between 0 Amperes and 10 Amperes, a second artificial neural network (ANN2) associated with a second current range (P2), preferably comprised between 10 Amperes and 20 Amperes, and a third artificial neural network (ANN3) associated with a third current range (P3), preferably comprised between 20 Amperes and 40 Amperes, the first current range (P1), the second current range (P2) and the third current range (P3) being different.

6. The method for detecting an electric arc fault as claimed in claim 5, characterized in that during the artificial neural network input step, if the RMS value of the current belongs to the first current range (P1) then the set of input data (11, I2, I3, 14..., 1316) is input into the first artificial neural network (ANN1) or if the RMS value of the current belongs to the second current range (P2) then the set of input data (11,12,13,14..., 1316) is input into the second artificial neural network (ANN2) or if the RMS value of the current belongs to the third current range (P3) then the set of input data (11, I2, I3, 14..., 1316) is input into the third artificial neural network (ANN3).

7. The method for detecting an electric arc fault as claimed in any one of claims 1 to 6, characterized in that the RMS value of the current corresponds to the root mean square of the acquired current signal over said given time interval of preferably equal to 10 milliseconds.

8. The method for detecting an electric arc fault as claimed in any one of claims 1 to 7, characterized in that it further comprises, prior to the input step, a normalization step (N1, N2..., Nn) during which said set of arc-fault characteristics is normalized to form a set of normalized input data and in that during the artificial neural network input step, the set of input data (11, I2, I3, 14..., 1316) corresponds to the set of normalized input data.

9. The method for detecting an electric arc fault as claimed in any one of claims 1 to 8, characterized in that during the decision step, said set of output data (0) is processed statistically by means of a sequential probability ratio test (SD) to decide whether an arc fault is present or not.

10. A device for detecting an electric arc fault, comprising at least one module for detecting an electric arc fault, said detection module being characterized in that it is configured to implement the method for detecting an electric arc fault as claimed in any one of claims 1 to 9.

11. An item of electrical equipment for detecting an electric arc fault, comprising at least:- an electrical line comprising a pair of contacts,- a device for measuring the current,- a device for detecting an electric arc fault,- a module for controlling the opening of the pair of contacts, said item of electrical equipment for detecting an arc fault being characterized in that it comprises the detection device as claimed in claim 10.

12. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to implement the method for detecting an electric arc fault as claimed in any one of claims 1 to 9.

13. A computer-readable medium on which the computer program as claimed in claim 12 is stored.

Images