Training method of arc detection model, arc fault detection method and device

Abstract

The present disclosure provides a method for training an arc detection model, a method for detecting an arc fault and equipment, which can be applied to the fields of artificial intelligence technology and arc detection technology. The method comprises: obtaining a first sample current data set; inputting the first sample current data into a first arc detection model to output a first sample target current feature data set; training the first arc detection model by using the first sample target current feature data and labels corresponding to the first sample current data to obtain a second arc detection model; constructing a third arc detection model by taking the second arc detection model as a submodel of the third arc detection model; inputting second sample current data into the third arc detection model to output a second sample target current feature data set; and training the third arc detection model by using the second sample target current feature data and labels corresponding to the second sample current data to obtain a trained target arc detection model.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence technology and arc detection technology, and in particular to a training method for an arc detection model, a method for detecting arc faults, and a device. Background Technology

[0002] With the rapid development of artificial intelligence technology, new ideas have been provided for arc fault identification. How to identify faulty arcs in the early stages of arcing based on artificial intelligence, reduce the harm of faulty arcs, and improve electrical safety has become a key research focus in related fields both at home and abroad.

[0003] In the process of realizing the present invention, the inventors discovered that the application of neural network models in arc detection schemes has the following drawbacks: complex models require a large amount of storage space and computing resources, making them difficult to apply on embedded platforms; and arc detection has high time requirements, so models with excessive complexity do not meet the real-time requirements of arc detection. Summary of the Invention

[0004] In view of the above problems, this disclosure provides a training method for an arc detection model, a method for detecting arc faults, and a device.

[0005] According to a first aspect of this disclosure, a method for training an electric arc detection model is provided, comprising:

[0006] Obtain the first sample current dataset, wherein the first sample current data in the first sample current dataset includes the label corresponding to the first sample current data;

[0007] The first sample current data in the first sample current dataset is input into the first arc detection model, and the first sample target current feature dataset is output. The first arc detection model includes a first convolutional layer, a second convolutional layer and a third convolutional layer.

[0008] Using the first sample target current feature data and the labels corresponding to the first sample current data in the first sample target current feature dataset, a first arc detection model is trained to obtain a second arc detection model.

[0009] The second arc detection model is used as a sub-model of the third arc detection model;

[0010] A third arc detection model is constructed based on the sub-models of the third arc detection model; the second sample current data from the second sample current dataset is input into the third arc detection model, and the second sample target current feature dataset is output, including labels corresponding to the second sample current data; and

[0011] Using the second sample target current feature data and the corresponding labels in the second sample target current feature dataset, a third arc detection model is trained to obtain the trained target arc detection model.

[0012] A second aspect of this disclosure provides a method for detecting electric arc faults, comprising:

[0013] The obtained initial current dataset is input into the target arc detection model, and the predicted target text data is output. The target arc detection model is trained based on the method of any one of claims 1 to 7.

[0014] A third aspect of this disclosure provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the methods described above.

[0015] A fourth aspect of this disclosure also provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the methods described above.

[0016] According to the training method for an arc detection model, and the method and device for detecting arc faults provided in this disclosure, by acquiring a first sample current dataset, the first sample current data in the first sample current dataset can be input into a first arc detection model, outputting a first sample target current feature dataset. Thus, the first sample target current feature data and the labels corresponding to the first sample current data in the first sample target current feature dataset can be used to train the first arc detection model, obtaining a second arc detection model. Furthermore, the second arc detection model can be used as a sub-model of a third arc detection model, obtaining a third arc detection model. Then, the second sample current data in the second sample current dataset is input into the third arc detection model, outputting a second sample target current feature dataset. This allows the second sample target current feature dataset to be used for training. The second sample target current feature data and the corresponding labels in the current feature dataset are used to train the third arc detection model, further obtaining the trained target arc detection model. Since the first convolutional layer of the target arc detection model is a one-dimensional convolutional layer, it satisfies the processing of one-dimensional current data. The second convolutional layer of the arc detection model retains the depthwise separable convolution function and is suitable for one-dimensional current data. The number of convolutional kernel channels in the third convolutional layer of the arc detection model needs to be twice the number of convolutional kernel channels in the other two layers, thereby eliminating the data flow bottleneck. While satisfying the processing of one-dimensional current data, the training speed of the model is improved. At the same time, the number of parameters of the target arc detection model is greatly reduced compared with the number of parameters of the existing model, which significantly reduces the complexity of the model and improves the computational efficiency of the model. Attached Figure Description

[0017] The foregoing contents, as well as other objects, features, and advantages of this disclosure, will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0018] Figure 1 The illustration schematically shows an application scenario of the training method for the arc detection model and the arc fault detection method according to embodiments of the present disclosure;

[0019] Figure 2 A flowchart illustrating a training method for an arc detection model according to an embodiment of the present disclosure is shown schematically.

[0020] Figure 3 The schematic diagram illustrates the structure of each convolutional layer of the first arc detection model according to an embodiment of the present disclosure;

[0021] Figure 4 The schematic diagram illustrates the structure of a second pre-trained sub-model (student sub-model) according to an embodiment of the present disclosure;

[0022] Figure 5 The schematic diagram illustrates the training process of student sub-model 1 and student sub-model 2 according to embodiments of the present disclosure;

[0023] Figure 6 A schematic diagram illustrating the structure of another second pre-trained sub-model (student sub-model) according to an embodiment of the present disclosure is shown.

[0024] Figure 7 The diagram illustrates the training process of student sub-models 3 to 6 according to embodiments of the present disclosure.

[0025] Figure 8 The diagram illustrates the parameter quantities and training results of student sub-models 1 to 6 according to embodiments of the present disclosure.

[0026] Figure 9 This schematically illustrates a structural diagram of a second arc detection model comprising different numbers of modules according to an embodiment of the present disclosure;

[0027] Figure 10 The illustration shows schematic diagrams of model training set accuracy curves containing different numbers of Arc-EfficientNet modules according to embodiments of the present disclosure.

[0028] Figure 11 The illustration shows the model accuracy, number of parameters, and training time according to embodiments of the present disclosure, including different numbers of Arc-EfficientNet modules.

[0029] Figure 12This illustration schematically shows training set accuracy curves for different amounts of sample current data according to embodiments of the present disclosure.

[0030] Figure 13 The diagram illustrates the accuracy curves and training time diagrams for different amounts of first sample current data according to embodiments of the present disclosure.

[0031] Figure 14 A schematic diagram illustrating a learning rate reduction strategy and an automatic training stop strategy according to embodiments of the present disclosure is shown.

[0032] Figure 15 A schematic diagram illustrating the loss values ​​and accuracy of an Arc-EfficientNet model according to an embodiment of the present disclosure is shown.

[0033] Figure 16 The diagram illustrates the learning rate variation curve of the Arc-EfficientNet model according to an embodiment of the present disclosure;

[0034] Figure 17 The schematic diagram illustrates the model training results with different α values ​​according to embodiments of the present disclosure;

[0035] Figure 18 The schematic diagram illustrates the Softmax output at different temperatures T according to embodiments of the present disclosure;

[0036] Figure 19 The diagram illustrates training results at different temperatures T according to embodiments of the present disclosure.

[0037] Figure 20 A schematic diagram illustrating the training loss value and accuracy of the teacher sub-model according to an embodiment of the present disclosure is shown.

[0038] Figure 21 A schematic diagram illustrating the learning rate variation curve of the teacher sub-model according to an embodiment of the present disclosure is shown.

[0039] Figure 22 A schematic diagram illustrating the training loss value and accuracy of a student sub-model according to an embodiment of the present disclosure is shown.

[0040] Figure 23 The illustration shows the detection results of the student sub-model after individual training and knowledge distillation training according to embodiments of the present disclosure;

[0041] Figure 24 A schematic diagram illustrating the confusion matrix of the prediction results of the Arc-EfficientNet model according to an embodiment of the present disclosure is shown.

[0042] Figure 25A schematic diagram illustrating the confusion matrix of the prediction results of the KD_Arc_EfficientNet model according to an embodiment of the present disclosure is shown.

[0043] Figure 26 A flowchart illustrating an arc fault detection method according to an embodiment of the present disclosure is shown schematically;

[0044] Figure 27 A schematic diagram illustrating the structure of a training apparatus for an arc detection model according to an embodiment of the present disclosure; and

[0045] Figure 28 A block diagram of an electronic device suitable for implementing a training method for an arc detection model and a method for detecting arc faults, according to embodiments of the present disclosure, is shown schematically. Detailed Implementation

[0046] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0047] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0048] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0049] When using expressions such as "at least one of A, B, and C", they should generally be interpreted in accordance with the meaning that is commonly understood by a person skilled in the art (e.g., "a system having at least one of A, B, and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B, and C, etc.).

[0050] In the technical solution disclosed herein, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse.

[0051] The inventors discovered that, in related technologies, the application of neural network models in arc detection schemes has the following drawbacks: complex models require a large amount of storage space and computing resources, making them difficult to apply on embedded platforms; and arc recognition has high time requirements, with overly complex models failing to meet the real-time requirements of arc detection.

[0052] In view of this, this disclosure provides a training method for an arc detection model and a method for detecting arc faults. By acquiring a first sample current dataset, the first sample current data in the first sample current dataset can be input into a first arc detection model, outputting a first sample target current feature dataset. Thus, the first sample target current feature data and the labels corresponding to the first sample current data in the first sample target current feature dataset can be used to train the first arc detection model, obtaining a second arc detection model. Furthermore, the second arc detection model can be used as a sub-model of a third arc detection model to obtain a third arc detection model. Then, the second sample current data in the second sample current dataset is input into the third arc detection model, outputting a second sample target current feature dataset. This allows the second sample target current feature dataset to be used for training. The second sample target current feature data and the corresponding labels in the current feature dataset are used to train the third arc detection model, further obtaining the trained target arc detection model. Since the first convolutional layer of the target arc detection model is a one-dimensional convolutional layer, it satisfies the processing of one-dimensional current data. The second convolutional layer of the arc detection model retains the depthwise separable convolution function and is suitable for one-dimensional current data. The number of convolutional kernel channels in the third convolutional layer of the arc detection model needs to be twice the number of convolutional kernel channels in the other two layers, thereby eliminating the data flow bottleneck. While satisfying the processing of one-dimensional current data, the training speed of the model is improved. At the same time, the computational amount of the target arc detection model is greatly reduced compared with the computational amount of the existing model, which significantly reduces the complexity of the model and improves the computational efficiency of the model.

[0053] This disclosure provides a method for training an arc detection model, a method for detecting arc faults, and an apparatus. The method includes: acquiring a first sample current dataset, wherein the first sample current data in the first sample current dataset includes labels corresponding to the first sample current data; inputting the first sample current data in the first sample current dataset into a first arc detection model, and outputting a first sample target current feature dataset, the first arc detection model including a first convolutional layer, a second convolutional layer, and a third convolutional layer; training the first arc detection model using the first sample target current feature data in the first sample target current feature dataset and the labels corresponding to the first sample current data, to obtain a second arc detection model; using the second arc detection model as a sub-model of a third arc detection model; constructing a third arc detection model based on the sub-model of the third arc detection model; inputting the second sample current data in the second sample current dataset into the third arc detection model, and outputting a second sample target current feature dataset, the second sample current data including labels corresponding to the second sample current data; and training the third arc detection model using the second sample target current feature data in the second sample target current feature dataset and the labels corresponding to the second sample current data, to obtain a trained target arc detection model.

[0054] Figure 1 The illustration schematically shows an application scenario of the training method for the arc detection model and the arc fault detection method according to embodiments of the present disclosure.

[0055] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, a server 105, and a current acquisition device 106. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, the server 105, and the current acquisition device 106. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0056] Users can interact with server 105 via network 104 using at least one of the first terminal device 101, second terminal device 102, and third terminal device 103 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, second terminal device 102, and third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0057] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0058] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0059] It should be noted that the arc detection model training method and arc fault detection method provided in this embodiment can generally be executed by server 105. Correspondingly, the arc detection model training device and arc fault detection device provided in this embodiment can generally be located in server 105. The arc detection model training method and arc fault detection method provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103, server 105, and / or current acquisition device 106. Correspondingly, the arc detection model training device and arc fault detection device provided in this embodiment can also be located in a server or server cluster that is different from server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103, server 105, and / or current acquisition device 106.

[0060] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, current acquisition devices, and servers can be included.

[0061] Figure 2 A flowchart illustrating a training method for an arc detection model according to an embodiment of the present disclosure is shown.

[0062] like Figure 2 As shown, the training method for the arc detection model in this embodiment includes operations S210 to S270.

[0063] In operation S210, a first sample current dataset is obtained, wherein the first sample current data in the first sample current dataset includes a label corresponding to the first sample current data.

[0064] According to embodiments of this disclosure, the first sample current data can be current data acquired from a current acquisition device. The type of current acquisition device can be diverse, such as a device specifically designed for arc detection. The current acquisition device can also be a current sensor installed in the circuit, along with a data acquisition card and a computer, to acquire current data in the circuit in real time. The specific form of the current acquisition device is not limited here. The label corresponding to the first sample current data can be set according to the current type, such as current under fault conditions and current under normal conditions, or it can be classified according to resistivity, inductance, and nonlinearity.

[0065] In operation S220, the first sample current data in the first sample current dataset is input into the first arc detection model, and the first sample target current feature dataset is output. The first arc detection model includes a first convolutional layer, a second convolutional layer and a third convolutional layer.

[0066] According to embodiments of this disclosure, the first arc detection model may be an improved EfficientNet model.

[0067] In operation S230, the first sample target current feature data and the label corresponding to the first sample current data in the first sample target current feature dataset are used to train the first arc detection model and obtain the second arc detection model.

[0068] According to embodiments of this disclosure, the second arc detection model can be an Arc-EfficientNet model. The second arc detection model is obtained by training the first arc detection model using first sample target current feature data and labels corresponding to the first sample current data. The second arc detection model may include a first convolutional layer, a second convolutional layer, and a third convolutional layer.

[0069] In operation S240, the second arc detection model is used as a sub-model of the third arc detection model.

[0070] According to embodiments of this disclosure, the third arc detection model may include multiple arc detection sub-models, and the second arc detection model may be used as one of the sub-models of the third arc detection model to construct the third arc detection model.

[0071] In operation S250, the third arc detection model is constructed based on the sub-model of the third arc detection model.

[0072] According to embodiments of this disclosure, the third arc detection model can be a knowledge distillation-Arc-EfficientNet model optimized using the knowledge distillation method. Specifically, in this disclosure, the second arc detection model can be used as the teacher sub-model in the knowledge distillation-Arc-EfficientNet model. The student sub-model in the knowledge distillation-Arc-EfficientNet model can be consistent with the teacher sub-model and also built using the Arc_EffNet module.

[0073] In operation S260, the second sample current data in the second sample current dataset is input into the third arc detection model, and the second sample target current feature dataset is output. The second sample current data includes the label corresponding to the second sample current data.

[0074] In operation S270, the third arc detection model is trained using the second sample target current feature data and the labels corresponding to the second sample current data in the second sample target current feature dataset, and the trained target arc detection model is obtained.

[0075] According to embodiments of this disclosure, the target arc detection model can be a KD-Arc-EfficientNet model, which is an arc detection model obtained by training a third arc detection model using the second sample target current feature data and the labels corresponding to the second sample current data.

[0076] According to embodiments of this disclosure, by acquiring a first sample current dataset, the first sample current data in the first sample current dataset can be input into a first arc detection model, outputting a first sample target current feature dataset. This allows the first arc detection model to be trained using the first sample target current feature data and the labels corresponding to the first sample current data, resulting in a second arc detection model. The second arc detection model can then be used as a sub-model of a third arc detection model to obtain a third arc detection model. Finally, the second sample current data in the second sample current dataset is input into the third arc detection model, outputting a second sample target current feature dataset. This allows the first arc detection model to be trained using the first sample target current feature data and the labels corresponding to the first sample current data, resulting in a second arc detection model. Using the current feature data and the labels corresponding to the second sample current data, a third arc detection model is trained, further resulting in the trained target arc detection model. Since the first convolutional layer of the target arc detection model is a one-dimensional convolutional layer, it satisfies the processing of one-dimensional current data. The second convolutional layer of the target arc detection model retains the depthwise separable convolution function and is suitable for one-dimensional current data. The number of convolutional kernel channels in the third convolutional layer of the target arc detection model needs to be twice the number of convolutional kernel channels in the other two layers, thereby eliminating the data flow bottleneck. While satisfying the processing of one-dimensional current data, the training speed of the model is improved. At the same time, the computational load of the target arc detection model is greatly reduced compared to the existing models, significantly reducing the complexity of the model and improving its computational efficiency.

[0077] According to embodiments of this disclosure, inputting first sample current data from a first sample current dataset into a first arc detection model and outputting a first sample target current feature dataset includes: inputting the first sample current data from the first sample current dataset into a first convolutional layer of the first arc detection model and outputting a first sample current feature dataset, wherein the first convolutional layer includes a one-dimensional convolutional sub-layer, a first channel sub-layer, and an activation function sub-layer; inputting the first sample current feature data from the first sample current feature dataset into a second convolutional layer of the first arc detection model and outputting a second sample current feature dataset, wherein... In the first model, the second convolutional layer is a spatially separable convolutional layer, including a spatially separable sub-layer, an activation function sub-layer, a second channel sub-layer, and a pooling sub-layer. The second sample current feature data in the second sample current feature dataset is input into the third convolutional layer of the first arc detection model to obtain the first sample target current feature dataset. The third convolutional layer includes a one-dimensional convolutional sub-layer, a third channel sub-layer, and an activation function sub-layer. The ratio between the number of third channel sub-layers and any one of the number of first channel sub-layers and the number of second channel sub-layers is an integer. The ratio between the number of first channel sub-layers and the number of second channel sub-layers is also an integer.

[0078] Figure 3The schematic diagram illustrates the structure of each convolutional layer of the first arc detection model according to an embodiment of the present disclosure.

[0079] like Figure 3 As shown, the first arc detection model can contain three convolutional layers: a first convolutional layer 310, a second convolutional layer 320, and a third convolutional layer 330. The first convolutional layer 310 can be a 1×1 one-dimensional convolutional layer, including a one-dimensional convolutional sub-layer, a first channel sub-layer, and an activation function sub-layer. The number of channels in the first channel sub-layer can be half the standard number of channels, i.e., ch / 2. The standard number of channels is compared to the number of channels in the neural network models MobileNet and ShuffleNet. The second convolutional layer 320 can be a 3×1 spatially separable convolutional layer and a max-pooling layer with a stride of 2, including a spatially separable sub-layer, an activation function sub-layer, a second channel sub-layer, and a pooling sub-layer. The number of channels in the second channel sub-layer can be half the standard number of channels, i.e., ch / 2. The pooling sub-layer can be used to perform pooling operations on the input current data. The third convolutional layer 330 can be a 2×1 one-dimensional convolutional layer with a stride of 2, including a one-dimensional convolutional sub-layer, a third channel sub-layer, and an activation function sub-layer. The activation function sublayer can be an activation function sublayer based on the ReLU function. To eliminate data flow bottlenecks, the number of channels in the kernel of the third convolutional layer can be twice the number of channels in the kernels of the other two layers.

[0080] According to embodiments of this disclosure, the first convolutional layer in the first arc detection model is a one-dimensional convolutional layer, thereby satisfying the processing of one-dimensional current data. The second convolutional layer of the first arc detection model retains the depth-separable convolution function and is suitable for one-dimensional current data. The number of convolution kernel channels in the third convolutional layer of the first arc detection model needs to be twice the number of convolution kernel channels in the other two layers, thereby eliminating the data flow bottleneck and improving the training speed of the model while satisfying the processing of one-dimensional current data.

[0081] According to embodiments of this disclosure, a third arc detection model includes a first pre-trained sub-model and a second pre-trained sub-model. The first pre-trained sub-model is characterized as a second arc detection model, and the first pre-trained sub-model is associated with the second pre-trained sub-model. The process of inputting second sample current data from a second sample current dataset into the third arc detection model and outputting a second sample target current feature dataset includes: inputting second sample current data from a second sample current dataset into the first pre-trained sub-model of the third arc detection model and outputting a second sample target current feature dataset corresponding to the first pre-trained sub-model; and inputting second sample current data from a second sample current dataset into the second pre-trained sub-model of the third arc detection model and outputting a second sample target current feature dataset corresponding to the second pre-trained sub-model.

[0082] According to embodiments of this disclosure, the first pre-trained sub-model can be characterized as the second arc detection model, that is, the second arc detection model (Arc-EfficientNet model) can be used as the first pre-trained sub-model of the third arc detection model (knowledge distillation-Arc-EfficientNet model), i.e., the teacher sub-model, and the second pre-trained sub-model can be characterized as the student sub-model in the third arc detection model (knowledge distillation-Arc-EfficientNet model).

[0083] According to embodiments of this disclosure, the student sub-model is a lighter model than the teacher sub-model.

[0084] Figure 4 The schematic diagram illustrates the structure of a second pre-trained sub-model (student sub-model) according to an embodiment of the present disclosure.

[0085] According to embodiments of this disclosure, a second pre-trained sub-model (student sub-model) can be constructed by reducing the number of Arc-EfficientNet modules in the Arc-EfficientNet model. The construction method includes: building second pre-trained sub-models containing one and two Arc-EfficientNet modules respectively, such as... Figure 4 As shown, student sub-model 1 (410) and student sub-model 2 (420) contain one and two Arc-EfficientNet modules respectively, and the width of each module (i.e. the number of convolutional kernels) can be consistent with the width of the teacher sub-model (i.e. the number of convolutional kernels).

[0086] Furthermore, student sub-model 1 (410) and student sub-model 2 (420) are trained respectively.

[0087] Figure 5 The diagram illustrates the training process of student sub-model 1 and student sub-model 2 according to embodiments of the present disclosure.

[0088] like Figure 5 As shown, in the first 200 iterations, the accuracy of both the training and validation sets of Student Sub-Model 1 showed an upward trend, but the accuracy of the validation set fluctuated significantly. After 212 iterations, the accuracy of the training set reached 95.073%, still showing a slow upward trend, but the accuracy of the validation set stopped improving, indicating overfitting. This suggests that Student Sub-Model 1, which only contains one Arc-EfficientNet module, is not deep enough. After 41 iterations, Student Sub-Model 2 achieved an accuracy of 95.110%, after which the upward trend slowed down, and the accuracy of the validation set was significantly lower than that of the training set, indicating overfitting. It can be seen that compared to Student Sub-Model 1, Student Sub-Model 2 has higher accuracy, but overfitting still exists.

[0089] Figure 6 The schematic diagram illustrates the structure of another second pre-trained sub-model (student sub-model) according to an embodiment of the present disclosure.

[0090] In one feasible implementation, this can be achieved by reducing the number of convolutional kernels in the Arc-EfficientNet model. Specifically, the number of convolutional kernels in the model is halved, resulting in four different student sub-models: student sub-model 3 (610), student sub-model 4 (620), student sub-model 5 (630), and student sub-model 6 (640), as follows: Figure 6 As shown in the figure. The left side of the figure shows the model structure. It can be seen that the number of Arc-EfficientNet modules in student sub-models 3 through 6 can be the same as in the teacher sub-model, each containing 3 Arc-EfficientNet modules. However, the number of convolutional kernels can be halved sequentially from the teacher sub-model, such as... Figure 6 As shown in bold font.

[0091] Figure 7 The diagram illustrates the training process of student sub-models 3 to 6 according to an embodiment of the present disclosure.

[0092] like Figure 7 As shown, after 16 iterations, student sub-model 3 achieved an accuracy of 95.102%, demonstrating fast convergence and minimal fluctuation in validation set accuracy. After 50 iterations, the accuracy increase slowed, ultimately reaching 99.923%. Student sub-model 4 achieved 95.036% after 29 iterations, with a slightly slower convergence speed than student sub-model 3, but its accuracy was comparable. Student sub-models 5 and 6 exhibited similar accuracy curves and showed no overfitting. However, during training, their validation set accuracy fluctuated significantly, resulting in lower final accuracies of 97.781% and 95.773%, respectively.

[0093] Figure 8 The diagram illustrates the parameter quantities and training results of student sub-models 1 to 6 according to embodiments of the present disclosure.

[0094] like Figure 8As shown, considering the convergence speed, accuracy, and parameter count of each student sub-model, a suitable student sub-model can be selected to build the Knowledge Distillation-Arc-EfficientNet model, i.e., the KD_Arc-EfficientNet model. In terms of convergence speed, student sub-model 3 converges the fastest, followed by student sub-model 4. Regarding accuracy, student sub-model 3 has the highest accuracy, followed by student sub-model 4, but the validation set accuracy difference between student sub-models 3 and 4 is only 0.08%. In terms of parameter count, student sub-model 1, which contains only one module, has the largest number of parameters among all student sub-models because the final flattening layer of the model is too large. For example... Figure 4 As shown, the flat layer (fully connected layer) of student sub-model 1 contains 12,800 parameters. The number of parameters in student sub-models 2 through 6 decreases sequentially. Although student sub-model 3 performs best in terms of convergence speed and accuracy, student sub-model 4 has fewer parameters, less than 1 / 3 of student sub-model 3, and its convergence speed and accuracy are very similar to student sub-model 3. Considering the need for lightweight sub-models, student sub-model 4 is selected as the student sub-model of the KD_Arc-EfficientNet model.

[0095] According to embodiments of this disclosure, the teacher sub-model is optimized in terms of the number of Arc-EfficientNet modules and the number of convolutional kernels. Six different student sub-models are built and trained using sample datasets. Since the student sub-model can complete the detection task with fewer parameters, the student sub-model is more lightweight, has lower computational complexity, requires fewer computational resources, and is more efficient than the teacher sub-model.

[0096] According to embodiments of this disclosure, the second arc detection model is an Arc-EfficientNet model, which includes multiple Arc-EfficientNet modules, and there are i first arc detection models. The training of the first arc detection model using a first sample target current feature dataset and labels corresponding to the first sample current data includes: inputting the first sample target current feature data and labels corresponding to the first sample current data from the first sample target current feature dataset into the (i-1)th first arc detection model, and outputting the (i-1)th Arc-EfficientNet detection result, wherein the (i-1)th Arc-EfficientNet detection result contains the (i-1)th current data detection information set; and inputting the (i-1)th Arc-EfficientNet detection result into the ith first arc detection model, and outputting the ith Arc-EfficientNet detection result, wherein the ith Arc-EfficientNet detection result contains the ith current data detection information set; each of the i first arc detection models corresponds one-to-one with the length of its output data. According to embodiments of this disclosure, the Arc-EfficientNet module can be a module generated based on a combination of the first, second, and third convolutional layers in a second arc detection model.

[0097] According to embodiments of this disclosure, a specific application interface, such as Keras, can be used to build an Arc-EfficientNet module. The input data size of the module can be 800x1 sample current data. Specifically, Table 1 is used as an example in this disclosure. Table 1 shows the parameters of the Arc-EfficientNet module according to embodiments of this disclosure.

[0098] Table 1

[0099] One-dimensional convolutional layer 1x1x32 800x32 2560 Spatially separable convolutional layer 3x1 800x32 2560 Max pooling layer 400x32 1280 One-dimensional convolution 2x1x64 200x64 1280

[0100] Figure 9 The schematic diagram illustrates the structure of a second arc detection model comprising different numbers of modules according to an embodiment of the present disclosure.

[0101] According to embodiments of this disclosure, each sample in the first sample current dataset has a data length of 800. As shown in Table 1, after passing through an Arc-EfficientNet module, the length of the output feature data is 200, reduced to 1 / 4 of the input data length. Figure 9As shown, after n Arc_EffNet modules, the output feature data length is 800 / 4n. Referring to the EfficientNet model, we can select 2, 3, and 4 modules for training, denoted as 920 (Arc_EfficientNet_2 blocks), 930 (Arc_EfficientNet_3 blocks), and 940 (Arc_EfficientNet_4 blocks) respectively. Convolution calculations within the modules can use the "same" mode. After convolution, the data length is the input data length divided by the stride. Pooling calculations can use the "valid" mode; when the data length is odd, one digit is discarded. Let the number of modules be n. When n is 3, the output of the third module pooling layer has a length of 25. During the final convolution operation, "0"s are automatically padded to ensure the data length remains unchanged. After a convolution calculation with a stride of 2, the output length becomes (25+1) / 2 = 13. When n is 4, the input length of the fourth layer module is 13. During pooling, one bit is discarded, resulting in a pooled output data length of (13-1) / 2 = 6. After convolution with a stride of 2, the final output data length of the module is 3. It is evident that when n is 4, the output length of module 4 has decreased to 3. If more modules are added, the input dimension of the fifth module is already lower than the convolution kernel. During convolution, a large amount of "0" data needs to be padded to maintain the data length. However, too much "0" data can make the network difficult to train or even cause training failure. Therefore, the number of modules n should be less than or equal to 4.

[0102] In one feasible embodiment, the training accuracy can be obtained by training and comparing models with different numbers of Arc-EfficientNet modules.

[0103] Figure 10 The illustration shows schematic diagrams of the accuracy curves of a model training set containing different numbers of Arc-EfficientNet modules according to embodiments of the present disclosure.

[0104] like Figure 10As shown, Arc-EfficientNet_2blocks has relatively low accuracy, achieving only 87.883% accuracy after 150 generations of training. Arc-EfficientNet_3blocks, on the other hand, achieves an accuracy of 81.190% after only 10 generations of training, and its accuracy curve continues to rise. After 150 generations of training, the final accuracy reaches 97.214%. Arc-EfficientNet_4blocks, due to operations such as padding with "0"s in convolutions and discarding data in pooling layers, maintains relatively stable accuracy for the first 7 generations of training. After 7 generations, the detection accuracy increases rapidly. After 60 generations, the accuracy curve essentially overlaps with that of Arc-EfficientNet_3blocks, reaching 96.960% accuracy after 150 generations of training.

[0105] In one feasible implementation, models containing different numbers of Arc-EfficientNet modules exhibit differences in accuracy, number of parameters, and training time.

[0106] Figure 11 The illustration shows the model accuracy, number of parameters, and training time of different numbers of Arc-EfficientNet modules according to embodiments of the present disclosure.

[0107] like Figure 11 As shown, the second arc fault detection model can balance the number of parameters and detection accuracy. In terms of parameter count, Arc-EfficientNet_2blocks has the fewest parameters at 81.608. Arc-EfficientNet_3blocks has 74.624 more parameters than Arc-EfficientNet_2blocks, less than twice the number of Arc-EfficientNet_2blocks. Arc-EfficientNet_4blocks has the most parameters, 3.4 times that of Arc-EfficientNet_3blocks. In terms of accuracy, Arc-EfficientNet_3blocks has the highest accuracy on both the training and validation sets, at 97.214% and 96.585%, respectively. Furthermore, the training time of the Arc-EfficientNet model increases linearly with the number of modules. Arc-EfficientNet_2blocks has a small number of parameters but low accuracy, failing to meet the requirements for arc fault detection.

[0108] According to embodiments of this disclosure, by selecting an appropriate number of modules, the detection accuracy of the arc fault monitoring model is ensured while the number of model parameters and computational load are reduced, thereby improving detection efficiency.

[0109] According to embodiments of this disclosure, the third arc detection model is a KD-Arc-EfficientNet model. The KD-Arc-EfficientNet model contains multiple KD-Arc-EfficientNet modules, and there are j third arc detection models. Each third arc detection model includes third arc detection sub-models corresponding to different numbers of KD-Arc-EfficientNet modules and third arc detection sub-models corresponding to different numbers of convolutional kernels. The training of the third arc detection model using the second sample target current feature data from the second sample target current feature dataset and the labels corresponding to the second sample current data includes: [The text abruptly ends here, so the translation stops.] The labels corresponding to the second sample current data are input into the (j-1)th third arc detection model, and the (j-1)th KD-Arc-EfficientNet detection result is output, where the (j-1)th KD-Arc-EfficientNet detection result contains the (k-1)th current data detection information set; and the (j-1)th KD-Arc-EfficientNet detection result is input into the jth third arc detection model, and the jth KD-Arc-EfficientNet detection result is output, where the jth KD-Arc-EfficientNet detection result contains the jth current data detection information set; each of the j third arc detection models corresponds one-to-one with the length of its output data.

[0110] According to embodiments of this disclosure, depending on the number of Arc-EfficientNet modules and the number of model convolution kernels, the third arc detection model includes multiple third arc detection sub-models. The third arc detection sub-models can be trained using the target current feature data of the second sample to obtain the detection results of each third arc detection sub-model.

[0111] According to embodiments of this disclosure, the selection rule for the number of first sample current data includes: selecting K batches of first sample current data, wherein the number of data in each batch of first sample current data is in a multiple relationship; and using the K batches of first sample current data to train a first arc detection model to obtain the training result of the first sample current data, wherein the training result includes training time and detection accuracy; and determining the number of first sample current data based on the training result.

[0112] Figure 12The diagram illustrates the training set accuracy curves under different amounts of sample current data according to embodiments of the present disclosure.

[0113] like Figure 12 As shown, the batch size for the first sample current data can be selected from six options: 32, 64, 128, 256, 512, and 1024. It can be seen that with a batch size of 32, the gradient descent direction differs from the overall gradient descent direction of the dataset, resulting in no significant improvement in accuracy for the Arc-EfficientNet model in the first few training iterations. After 9 iterations, the accuracy begins to improve rapidly, reaching a final accuracy of 92.180% after 150 generations of training. With a batch size of 1024, the accuracy curve is similar to that with a batch size of 32; the network's accuracy does not improve significantly before the 12th generation, and the trend after the 12th generation is similar to that with a batch size of 32, ultimately reaching an accuracy of 92.128%. The accuracy curves for batch sizes of 128, 256, and 512 show no significant difference; in the early stages of training, the accuracy increase rates for all three are almost identical, and the curves are nearly parallel. The accuracy curves of the three networks overlapped during training generations 25-40, and remained similar thereafter, with final accuracies of 93.916%, 94.756%, and 94.568%, respectively. The network performed best with a batch size of 64, exhibiting the highest rate of accuracy increase and the highest final accuracy value. After 150 generations of training, the final accuracy reached 96.875%.

[0114] Figure 13 The diagram illustrates the accuracy curves and training time diagrams for different amounts of first sample current data according to embodiments of the present disclosure.

[0115] like Figure 13 As shown, under different batch sizes of the first sample current data, the training time can decrease as the batch size increases, but the difference is small. The longest training time for 150 generations is only 106 seconds longer than the shortest time. Considering the accuracy and training time of different batch sizes, a batch size of 64 is selected.

[0116] According to embodiments of this disclosure, by selecting different data quantities of first sample current data and training, the optimal data quantity of first sample current data can be obtained by balancing the relationship between training time and arc fault recognition rate.

[0117] According to embodiments of this disclosure, the first parameter value selection rule for the first arc detection model includes: selecting the first parameter value of the first arc detection model; and determining the first target parameter value of the first arc detection model based on a learning rate reduction strategy, an automatic training stop strategy, and preset rules.

[0118] According to embodiments of this disclosure, the training of the first arc detection model is a process of continuously updating parameters along the gradient direction of the loss function. By applying a learning rate descent strategy and an automatic training stop strategy to the arc detection model training process, the loss value can be used as its evaluation criterion. Specifically, Table 2 is used as an example in this disclosure. Table 2 shows the specific parameters and abbreviations of the model learning rate descent strategy and the automatic training stop strategy according to embodiments of this disclosure.

[0119] Table 2

[0120]

[0121] Figure 14 The diagram illustrates a learning rate reduction strategy and an automatic training stop strategy according to embodiments of the present disclosure.

[0122] like Figure 14As shown, the learning rate descent strategy can first set an initial learning rate *lr* and a learning rate descent criterion *lr_min_delta*. When the decrease in the Arc-EfficientNet model's loss value *loss_delta* is less than the learning rate descent criterion *lr_min_delta*, the learning rate update value *lr_num* is incremented by 1. During Arc-EfficientNet model training, the loss value fluctuates downwards. To prevent the learning rate from rapidly decreasing to its minimum value during training due to these fluctuations, a learning rate update patience value *lr_patience* is set. When *lr_num* is greater than *lr_patience*, the updated learning rate is *factor* × *lr*. In the later stages of Arc-EfficientNet model training, the loss value tends to decrease to 0, and *lr_num* will continuously increase, causing the learning rate to continuously decrease until it reaches 0, resulting in model training failure. Therefore, a minimum learning rate value *min_lr* can be set. When *factor* × *lr* is less than *min_lr*, the updated learning rate is *min_lr*, and no further updates to the learning rate are performed. Furthermore, in the first few epochs after updating the learning rate in the Arc-EfficientNet model, the loss_delta may be less than the learning rate_min_delta. To avoid continuous learning rate updates that could cause the Arc-EfficientNet model to get stuck in a local optimum during the middle of training due to an excessively low learning rate, a learning rate update cooldown can be set. This cooldown period is the time after which the algorithm restarts updating the learning rate. cool_num is the accumulated cooldown value after each learning rate update. When cool_num exceeds cooldown, the learning rate update calculation restarts.

[0123] The automatic training stop strategy and the learning rate reduction strategy can be carried out simultaneously. After each iteration, the decrease in the loss value loss_delta is compared with the stop training criterion stop_min_delta. When loss_delta is less than stop_min_delta, the stop training value stop_num is incremented by 1. When stop_num is greater than the stop training patience value stop_patience, training is stopped.

[0124] The specific parameters for the learning rate reduction strategy and the automatic training stop strategy can be the optimal parameters selected based on experimental results. During the experiment, the training parameters of the EfficientNet model can be referenced, and similar training parameters can be set. Then, fine-tuning can be performed, and the effects of different parameter values ​​on the detection accuracy and training time of the Arc_EfficientNet model can be tested. Finally, a set of optimal parameters can be determined for the detection of arc faults.

[0125] Since the learning rate update strategy and the automatic training stop strategy can be performed synchronously, to prevent the network training from stopping prematurely, the stopping criterion stop_min_delta should be less than the learning rate decrease criterion lr_min_delta, and the stopping patience value stop_patience should be greater than the learning rate update patience value lr_patience. Specifically, Table 3 is used as an example in this disclosure, showing the parameter values ​​of the Arc_EfficientNet model according to an embodiment of this disclosure.

[0126] Table 3

[0127] Batchsize 64 min_lr 0.000 01 lr 0.001 cooldown 5 lr_min_delta 0.0001 stop_min_delta 0.000 01 lr_patience 5 stop_patience 10 factor 0.1

[0128] Figure 15 The diagram illustrates the loss values ​​and accuracy of an Arc-EfficientNet model according to an embodiment of the present disclosure.

[0129] Figure 16 The diagram illustrates the learning rate variation curve of the Arc-EfficientNet model according to an embodiment of the present disclosure.

[0130] like Figure 15 As shown, after 10 training iterations, the Arc-EfficientNet model loss value decreased from the initial 1.91853 to 0.19970, achieving an accuracy of 93.433%. Between generations 10 and 35, the training set loss value steadily decreased, while the validation set loss value fluctuated significantly but still showed a downward trend. By generation 66, there were five consecutive epochs where the loss value decreased less than the learning rate reduction criterion, indicating that the Arc-EfficientNet model was converging due to oscillations at the minimum gradient point of the loss function caused by an excessively large learning rate. At this point, the learning rate could be reduced for the first time, decreasing the gradient descent step size, which further reduced the network loss value on the training set and improved the accuracy.

[0131] like Figure 16 As shown, the learning rate can be further reduced to a minimum of 0.00001 at the 80th generation. After 10 generations of training, the loss value no longer decreases, and training automatically stops to prevent overfitting. After training, the Arc-EfficientNet model achieves a recognition accuracy of 99.904% on the training set and 98.574% on the validation set. Furthermore, during the training process, the accuracy curves of the validation set and training set consistently show the same trend, without overfitting or underfitting, demonstrating the effectiveness of the Arc-EfficientNet model's learning rate reduction strategy and automatic stopping strategy.

[0132] According to embodiments of this disclosure, the selection rule for the second parameter value of the third arc detection model includes: selecting the second parameter value of the third arc detection model; and training the third arc detection model using the training set and validation set of the second parameter value to determine the second target parameter value of the third arc detection model.

[0133] According to embodiments of this disclosure, the third arc detection model may include multiple second parameter values. Specifically, in this disclosure, the second parameter values ​​may include two parameters: the α value of the third arc detection model and the distillation temperature T. The α value determines the proportion of soft-loss and hard-loss in the knowledge distillation loss function, and has a direct impact on the model's detection accuracy.

[0134] Figure 17 The diagram illustrates the model training results with different α values ​​according to embodiments of the present disclosure.

[0135] like Figure 17 As shown in Figure 17, when α is 0, the student sub-model's loss function update depends on the soft-loss part, which can lead to the student sub-model learning errors that may exist in the teacher sub-model, resulting in decreased accuracy. When α is 1, the student sub-model's loss function update depends on the hard-loss part, and the knowledge of the teacher sub-model cannot be transferred to the student sub-model. To verify the impact of different α values ​​on arc detection, five values ​​of α (0.1, 0.2, 0.3, 0.4, and 0.5) were selected for testing. As can be seen from Figure 17, the model achieves the highest accuracy on both the training and validation sets when α is 0.3. Therefore, the α value of KD_Arc_EfficientNet was set to 0.3.

[0136] According to embodiments of this disclosure, the temperature T during knowledge distillation can determine the degree of change in the output probability of the teacher sub-model Softmax.

[0137] Figure 18 The schematic diagram illustrates the Softmax output at different temperatures T according to embodiments of the present disclosure.

[0138] like Figure 18 As shown, when the original Softmax is T=1, the probability of predicting class 5 is 1. When T is greater than 1, the output probability value of Softmax tends to level off, and when T=1000, the probabilities of each predicted class are nearly equal. It is evident that the larger the value of T, the more attention the student sub-model pays to the negative label.

[0139] Figure 19 The diagram illustrates training results at different temperatures T according to embodiments of the present disclosure.

[0140] like Figure 19As shown, due to the small size and limited representational capacity of the arc detection model, a smaller distillation temperature can be selected, and some negative label information can be discarded appropriately. Experiments can be conducted on KD_Arc_EfficientNet with temperatures T of 2, 4, 6, 8, and 10. It can be seen that, from the perspective of training set accuracy, the highest accuracy (99.624%) is achieved at distillation temperature T = 2, followed by T = 6 at 99.491%. From the perspective of validation set accuracy, the highest accuracy (98.442%) is achieved at T = 6, followed by T = 10 at 98.408%. Compared to the training set, the validation set better reflects the model's generalization performance; therefore, the distillation temperature T, which has a higher validation set accuracy, should be considered first. The training set accuracy at T = 10 is lower than that at T = 6; therefore, the distillation temperature T of KD_Arc_EfficientNet can be set to 6.

[0141] According to embodiments of this disclosure, given that the α value and distillation temperature T of the third arc detection model are already determined, a knowledge-distillation-based fault arc identification model, KD_Arc_EfficientNet, can be constructed. Its teacher sub-model is an Arc_EfficientNet model, and the student sub-model can be the aforementioned student sub-model 4, containing three Arc_EfficientNet modules. The number of convolutional kernels in student sub-model 4 is 1 / 4 that of the teacher sub-model. The number of parameters in student sub-model 4 is 15,064, which is less than 1 / 10 of that in the teacher sub-model. The distillation temperature T can be set to 6, and the hard loss percentage α can be 0.3.

[0142] Figure 20 A schematic diagram illustrating the training loss value and accuracy of the teacher sub-model according to an embodiment of the present disclosure is shown.

[0143] like Figure 20As shown, the teacher sub-model is first trained by replacing the output layer of the Arc_EfficientNet model t with a Softmax layer incorporating temperature T. The same learning rate update strategy and automatic training stop strategy as described above are used. It can be seen that in the early stages of training, the model's loss curve and accuracy curve show a stable decreasing and a stable increasing trend, respectively. In iterations 10-56, the waveforms show slight fluctuations, but the overall trend remains unchanged; the loss curve continues to decrease, while the accuracy curve continues to increase. After 56 iterations, the model performance does not improve significantly, and the learning rate drops to 0.0001. At this point, the curve's change slows down. After 70 iterations, the learning rate reaches its minimum value. At this point, the training set loss value is close to 0, reaching its minimum. Although the validation set loss value is still decreasing, the decrease is very small, and the model's accuracy does not improve significantly. Training automatically stops at 77 iterations. At this point, the accuracy of the training set and validation set is 99.805% and 98.408%, respectively.

[0144] Figure 21 A schematic diagram illustrating the learning rate variation curve of the teacher sub-model according to an embodiment of the present disclosure is shown.

[0145] like Figure 21 As shown, the training learning rate update strategy for the student sub-model in the KD_Arc_EfficientNet model is consistent with the above description. Because the student sub-model is simple and less prone to overfitting, an automatic training stop strategy is not implemented. To ensure the student sub-model can fully learn from the teacher sub-model and the data labels, the number of iterations is set to 400.

[0146] Figure 22 A schematic diagram illustrating the training loss value and accuracy of a student sub-model according to an embodiment of the present disclosure is shown.

[0147] like Figure 22 As shown, the hard-loss and validation set loss values ​​fluctuated significantly in the early stages of student sub-model training, while the soft-loss showed a stable decreasing trend. This indicates that in the early stages of training, the student sub-model effectively learned the knowledge of the teacher sub-model. After 83 generations of training, the loss value began to decrease steadily. Regarding accuracy, although there were fluctuations between the training and validation sets, the overall trend was one of steady increase. The final accuracy on the training and validation sets was 99.606% and 98.110%, respectively.

[0148] Figure 23 The illustration shows a schematic diagram of the detection results of a student sub-model after individual training and knowledge distillation training according to an embodiment of the present disclosure.

[0149] like Figure 23As shown, the student sub-model can be trained separately. By comparing the detection results of the student sub-model trained separately and the student sub-model trained by knowledge distillation, it is found that the loss value of the validation set is much greater than that of the training set during the training process. Finally, the accuracy of the student sub-model on the training set and the validation set is only 93.913% and 92.546%, respectively, which is not good. It can be seen that the knowledge distillation method can significantly improve the recognition accuracy of the student sub-model by transferring knowledge from the teacher sub-model.

[0150] According to embodiments of this disclosure, the first target parameter value includes at least one of the following: learning rate update cooldown time, initial learning rate, minimum learning rate, learning rate decline criterion, learning rate update patience value, learning rate decline factor, training stop criterion, and training stop patience value.

[0151] According to embodiments of this disclosure, the arc detection model further includes: a fully connected layer and an output layer, wherein the fully connected layer is connected to the third convolutional layer and located after the third convolutional layer, and is connected to the output layer and located before the output layer.

[0152] According to embodiments of this disclosure, the one-dimensional convolutional sublayer in the first convolutional layer is a 1x1 one-dimensional convolutional sublayer; the spatially separable sublayer in the second convolutional layer is a 3×1 spatially separable convolutional sublayer, and the pooling sublayer is a max-pooling sublayer with a stride of 2; the one-dimensional convolutional sublayer in the third convolutional layer is a 2x1 one-dimensional convolutional sublayer; the number of the third channel sublayer is a multiple of the number of the first channel sublayer and the number of the second channel sublayer.

[0153] Figure 24 A schematic diagram of the confusion matrix of the prediction results of the Arc-EfficientNet model according to an embodiment of the present disclosure is shown.

[0154] like Figure 24 As shown, the performance of the Arc-EfficientNet model can be verified using validation set data. The horizontal axis of the matrix represents the categories detected by the Arc-EfficientNet model, and the vertical axis represents the true categories of the samples. The data on the diagonal represents the number of correctly detected samples. From the perspective of load classification, the Arc-EfficientNet model detected 57 incorrect samples, of which 41 were load category detection errors, such as detecting an arc on a resistive load as an arc on a motor load. These errors have little impact on arc detection and are not considered false positives or false negatives.

[0155] In one feasible embodiment, the confusion matrix can be further summarized from the perspective of arc and non-arc detection. There are a total of 16 incorrectly detected samples, of which 6 are detected as arc when there is no arc (including 3 motor-related samples and 2 power electronic samples), and 10 are detected as no arc when there is an arc (including 7 resistive loads, 1 motor-related load, and 2 power electronic loads). It can be seen that the detection errors of arc and non-arc samples are mainly concentrated on resistive loads, motor-related loads, and power electronic loads. Let the number of arc samples correctly detected by the Arc-EfficientNet model be True Positive (TP), and the number of non-arc samples correctly detected be True Negative (TN); let the number of samples where the Arc-EfficientNet model detects no arc as arc be False Positive (FP), and the number of samples where there is arc as no arc be False Negative (FN). The precision, recall, and accuracy of the Arc-EfficientNet model were calculated separately. The calculation methods for precision, recall, and accuracy are shown in Formula (1):

[0156]

[0157] Where TP is the number of arc samples correctly detected by the Arc-EfficientNet model (denoted as true), FP is the number of samples that the Arc-EfficientNet model detects as arcs when there is no arc (denoted as false positive), TN is the number of arc samples correctly detected (denoted as true negative), and FN is the number of samples that are arcs but are detected as no arc (denoted as false negative).

[0158] According to the embodiments of this disclosure, the precision, recall and accuracy of the Arc-EfficientNet model can be calculated by formula (1): precision, accuracy and recall are 99.750%, 99.584% and 99.688% respectively, all above 99.5%, which can meet the requirements of arc fault detection.

[0159] In one feasible implementation, the performance of the KD_Arc_EfficientNet model can be verified using test set data.

[0160] Figure 25 A schematic diagram of the confusion matrix of the prediction results of the KD_Arc_EfficientNet model according to an embodiment of the present disclosure is shown.

[0161] like Figure 25 As shown, the horizontal axis of the matrix represents the predicted class of the KD_Arc_EfficientNet model, and the vertical axis represents the true class of the sample. The data on the diagonal represents the number of correctly detected samples.

[0162] Specifically, Tables 4 and 5 are used as examples in this disclosure. Tables 4 and 5 show the detection results of the KD_Arc_EfficientNet model according to the embodiments of this disclosure, and a comparison of the detection results of the KD_Arc_EfficientNet model and the Arc_EfficientNet model, respectively.

[0163] Table 4

[0164]

[0165]

[0166] Table 5

[0167]

[0168] From a load classification perspective, the KD_Arc_EfficientNet model mispredicted 75 samples, 45 of which were load class detection errors, 4 more than the Arc_EfficientNet model. This shows that the load class detection capability of the KD_Arc_EfficientNet model is similar to that of the Arc_EfficientNet model. Figure 25 It can be seen that the load category detection errors of the KD_Arc_EfficientNet model are mainly concentrated on resistive loads. This is because the arc fault current characteristics of resistive loads are more typical, and the arc current characteristics of other load categories are similar to those of resistive loads, thus resulting in more category detection errors.

[0169] From the perspective of arc and non-arc detection, the KD_Arc_EfficientNet model exhibited 30 classification errors, including 17 misclassifications of non-arc as arc and 13 misclassifications of arc as non-arc. Compared to the Arc_EfficientNet model, the number of misclassified samples increased by 14, mainly concentrated in power electronic loads, with 13 detection errors. It is evident that the performance of the knowledge-distilled KD_Arc_EfficientNet model is slightly worse than the Arc_EfficientNet model, but the overall number of detection errors remains relatively low. As shown in Table 4, the detection accuracy of the KD_Arc_EfficientNet model is 99.415%, only 0.273% lower than the Arc_EfficientNet model, but the number of parameters in the KD_Arc_EfficientNet model is less than 1 / 10 of that in the Arc_EfficientNet model.

[0170] According to embodiments of this disclosure, the KD_Arc_EfficientNet model significantly reduces the number of model parameters while sacrificing only 0.273% accuracy, which is beneficial for its deployment on embedded platforms, saving computing resources and improving computational efficiency.

[0171] Figure 26 A flowchart illustrating an arc fault detection method according to an embodiment of the present disclosure is shown schematically.

[0172] like Figure 26 As shown, the arc fault detection method of this embodiment includes operation S2610.

[0173] In operation S2610, the obtained initial current dataset is input into the target arc detection model, and the predicted target text data is output. The target arc detection model is trained based on the above-mentioned arc detection model training method.

[0174] Based on the above-mentioned training method for the arc detection model, this disclosure also provides a training device for the arc detection model. The following will be combined with... Figure 14 The device is described in detail.

[0175] Figure 27 A schematic block diagram of a training apparatus for an arc detection model according to an embodiment of the present disclosure is shown.

[0176] like Figure 27 As shown, the training device 2700 for the arc detection model in this embodiment includes a first sample current dataset acquisition module 2710, a first sample target current feature dataset output module 2720, a first arc detection model training module 2730, a second arc detection model acquisition module 2740, a third arc detection model construction module 2750, a second sample target current feature dataset output module 2760, and a third arc detection model training module 2770.

[0177] The first sample current dataset acquisition module 2710 is used to acquire a first sample current dataset, wherein the first sample current data in the first sample current dataset includes a label corresponding to the first sample current data. In one embodiment, the first sample current dataset acquisition module 2710 can be used to perform the operation S210 described above, which will not be repeated here.

[0178] The first sample target current feature dataset output module 2720 is used to input the first sample current data in the first sample current dataset into the first arc detection model and output the first sample target current feature dataset. The first arc detection model includes a first convolutional layer, a second convolutional layer, and a third convolutional layer. In one embodiment, the first sample target current feature dataset output module 2720 can be used to perform the operation S220 described above, which will not be repeated here.

[0179] The first arc detection model training module 2730 is used to train a first arc detection model using the first sample target current feature data and the labels corresponding to the first sample current data in the first sample target current feature dataset, thereby obtaining a second arc detection model. In one embodiment, the first arc detection model training module 2730 can be used to perform the operation S230 described above, which will not be repeated here.

[0180] The second arc detection model, as module 2740, is used as a sub-model of the third arc detection model. In one embodiment, the second arc detection model as module 2740 can be used to perform the operation S240 described above, which will not be repeated here.

[0181] The third arc detection model construction module 2750 is used to construct a third arc detection model based on the sub-models of the third arc detection model. In one embodiment, the third arc detection model construction module 2750 can be used to perform the operation S250 described above, which will not be repeated here.

[0182] The second sample target current feature dataset output module 2760 is used to input the second sample current data from the second sample current dataset into the third arc detection model and output the second sample target current feature dataset. The second sample current data includes labels corresponding to the second sample current data. In one embodiment, the second sample target current feature dataset output module 2760 can be used to perform the operation S260 described above, which will not be repeated here.

[0183] The third arc detection model training module 2770 is used to train the third arc detection model using the second sample target current feature data and the labels corresponding to the second sample current data in the second sample target current feature dataset, thereby obtaining the trained target arc detection model. In one embodiment, the third arc detection model training module 2770 can be used to perform the operation S270 described above, which will not be repeated here.

[0184] According to embodiments of this disclosure, any multiple modules among the first sample current dataset acquisition module 2710, the first sample target current feature dataset output module 2720, the first arc detection model training module 2730, the second arc detection model construction module 2740, the third arc detection model building module 2750, the second sample target current feature dataset output module 2760, and the third arc detection model training module 2770 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this disclosure, at least one of the following modules can be implemented, at least partially, as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or any other reasonable means of integrating or packaging circuits, or as hardware or firmware, or as any one of software, hardware, and firmware implementations or any appropriate combination thereof. Alternatively, at least one of the following modules can be implemented, at least partially, as a computer program module: the first sample current dataset acquisition module 2710, the first sample target current feature dataset output module 2720, the first arc detection model training module 2730, the second arc detection model as module 2740, the third arc detection model construction module 2750, the second sample target current feature dataset output module 2760, and the third arc detection model training module 2770. When the computer program module is run, it can perform the corresponding function.

[0185] Figure 28 A block diagram of an electronic device suitable for implementing a training method for an arc detection model and a method for detecting arc faults, according to embodiments of the present disclosure, is shown schematically.

[0186] like Figure 28As shown, an electronic device 2800 according to an embodiment of this disclosure includes a processor 2801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 2802 or a program loaded from a storage portion 2808 into a random access memory (RAM) 2803. The processor 2801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 2801 may also include onboard memory for caching purposes. The processor 2801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this disclosure.

[0187] RAM 2803 stores various programs and data required for the operation of electronic device 2800. Processor 2801, ROM 2802, and RAM 2803 are interconnected via bus 2804. Processor 2801 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 2802 and / or RAM 2803. It should be noted that programs may also be stored in one or more memories other than ROM 2802 and RAM 2803. Processor 2801 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in one or more memories.

[0188] According to embodiments of this disclosure, the electronic device 2800 may further include an input / output (I / O) interface 2805, which is also connected to a bus 2804. The electronic device 2800 may also include one or more of the following components connected to the input / output (I / O) interface 2805: an input section 2806 including a keyboard, mouse, etc.; an output section 2807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 2808 including a hard disk, etc.; and a communication section 2809 including a network interface card such as a LAN card, modem, etc. The communication section 2809 performs communication processing via a network such as the Internet. A drive 2810 is also connected to the input / output (I / O) interface 2805 as needed. A removable medium 2811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 2810 as needed so that computer programs read from it can be installed into the storage section 2808 as needed.

[0189] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0190] According to embodiments of this disclosure, the computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this disclosure, the computer-readable storage medium may include ROM 2802 and / or RAM 2803 and / or one or more memories other than ROM 2802 and RAM 2803 described above.

[0191] Embodiments of this disclosure also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the training method for the arc detection model and the arc fault detection method provided in the embodiments of this disclosure.

[0192] When the computer program is executed by the processor 2801, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0193] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 28028, and / or installed from the removable medium 2811. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0194] In such an embodiment, the computer program can be downloaded and installed from a network via communication section 2809, and / or installed from removable medium 2811. When the computer program is executed by processor 2801, it performs the functions defined in the system of this disclosure embodiment. According to embodiments of this disclosure, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0195] According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can execute entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0196] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0197] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. In particular, the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0198] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The scope of this disclosure is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A training method for an electric arc detection model, comprising: Obtain a first sample current dataset, wherein the first sample current data in the first sample current dataset includes a label corresponding to the first sample current data; The first sample current data in the first sample current dataset is input into the first arc detection model, and the first sample target current feature dataset is output. The first arc detection model includes a first convolutional layer, a second convolutional layer and a third convolutional layer. Using the first sample target current feature data in the first sample target current feature dataset and the label corresponding to the first sample current data, the first arc detection model is trained to obtain the second arc detection model. The second arc detection model is used as a sub-model of the third arc detection model; The third arc detection model is constructed based on the sub-models of the third arc detection model; the second sample current data from the second sample current dataset is input into the third arc detection model, and the second sample target current feature dataset is output, wherein the second sample current data includes labels corresponding to the second sample current data; and The third arc detection model is trained using the second sample target current feature data in the second sample target current feature dataset and the labels corresponding to the second sample current data, to obtain the trained target arc detection model.

2. The method according to claim 1, wherein, The step of inputting the first sample current data from the first sample current dataset into the first arc detection model and outputting the first sample target current feature dataset includes: The first sample current data in the first sample current dataset is input into the first convolutional layer of the first arc detection model, and the first sample current feature dataset is output. The first convolutional layer includes a one-dimensional convolutional sub-layer, a first channel sub-layer, and an activation function sub-layer. The first sample current feature data in the first sample current feature dataset is input into the second convolutional layer of the first arc detection model, and the second sample current feature dataset is output. The second convolutional layer is a spatially separable convolutional layer, including a spatially separable sub-layer, an activation function sub-layer, a second channel sub-layer, and a pooling sub-layer. The second sample current feature data in the second sample current feature dataset is input into the third convolutional layer of the first arc detection model to obtain the first sample target current feature dataset. The third convolutional layer includes a one-dimensional convolutional sub-layer, a third channel sub-layer, and an activation function sub-layer. The ratio between the number of the third channel sub-layer and any one of the number of the first channel sub-layer and the number of the second channel sub-layer is an integer. The ratio between the number of the first channel sub-layer and the number of the second channel sub-layer is also an integer.

3. The method according to claim 1, wherein, The third arc detection model includes a first pre-trained sub-model and a second pre-trained sub-model. The first pre-trained sub-model is characterized as the second arc detection model, and the first pre-trained sub-model is associated with the second pre-trained model. The step of inputting the second sample current data from the second sample current dataset into the third arc detection model and outputting the second sample target current feature dataset includes: The second sample current data in the second sample current dataset is input into the first pre-trained sub-model in the third arc detection model, and the second sample target current feature dataset corresponding to the first pre-trained model is output. The second sample current data in the second sample current dataset is input into the second pre-trained sub-model in the third arc detection model, and the second sample target current feature dataset corresponding to the second pre-trained sub-model is output.

4. The method according to claim 1, wherein, The second arc detection model is an Arc-EfficientNet model, which contains multiple Arc-EfficientNet modules, and the first arc detection model has i modules; The step of training the first arc detection model using the first sample target current feature data and the labels corresponding to the first sample current data in the first sample target current feature dataset includes: The first sample target current feature data and the corresponding label from the first sample target current feature dataset are input into the (i-1)th first arc detection model, and the (i-1)th Arc-EfficientNet detection result is output, wherein the (i-1)th Arc-EfficientNet detection result contains the (i-1)th current data detection information set; and The (i-1)th Arc-EfficientNet detection result is input into the ith first arc detection model, and the ith Arc-EfficientNet detection result is output. The ith Arc-EfficientNet detection result contains the ith current data detection information set. The ith first arc detection model corresponds one-to-one with the length of its output data.

5. The method according to claim 1, wherein, The third arc detection model is a KD-Arc-EfficientNet model, which contains multiple KD-Arc-EfficientNet modules. There are j third arc detection models, which include third arc detection sub-models corresponding to different numbers of KD-Arc-EfficientNet modules and third arc detection sub-models corresponding to different numbers of convolutional kernels. The step of training the third arc detection model using the second sample target current feature data and the labels corresponding to the second sample current data from the second sample target current feature dataset includes: The second sample target current feature data and the corresponding labels from the second sample target current feature dataset are input into the (j-1)th third arc detection model, and the (j-1)th KD-Arc-EfficientNet detection result is output, wherein the (j-1)th KD-Arc-EfficientNet detection result contains the (k-1)th current data detection information set; and The (j-1)th KD-Arc-EfficientNet detection result is input into the j-th third arc detection model, and the j-th KD-Arc-EfficientNet detection result is output. The j-th KD-Arc-EfficientNet detection result contains the j-th current data detection information set. The j-th third arc detection model corresponds one-to-one with the length of its output data.

6. The method according to claim 1, wherein, The selection rules for the number of data points in the first sample current data include: Select K batches of first-sample current data, where the number of data points in each batch of first-sample current data is a multiple of each other; and The first arc detection model is trained using the K batches of first sample current data to obtain the training results of the first sample current data, wherein the training results include training time and detection accuracy. The number of data points for the first sample current data is determined based on the training results.

7. The method according to claim 4, wherein, The parameter value selection rules for the first arc detection model include: Select the first parameter value of the first arc detection model; and The first target parameter value of the first arc detection model is determined based on the learning rate reduction strategy, the automatic training stop strategy, and the preset rules.

8. The method according to claim 5, wherein, The selection rules for the second parameter value of the third arc detection model include: Select the second parameter value of the third arc detection model; and The second target parameter value of the third arc detection model is determined by training the model using the training set and validation set of the second parameter value.

9. The method according to claim 7, wherein, The first target parameter value includes at least one of the following: Learning rate update cooldown, initial learning rate, minimum learning rate, learning rate decline criteria, learning rate update patience value, learning rate decline factor, training stop criteria, and training stop patience value.

10. The method according to claim 1, wherein, The arc detection model further includes a fully connected layer and an output layer. The fully connected layer is connected to the third convolutional layer and is located after the third convolutional layer, and is connected to the output layer and is located before the output layer.

11. The method according to claim 2, wherein, The one-dimensional convolutional sublayer in the first convolutional layer is a 1x1 one-dimensional convolutional sublayer; the spatially separable sublayer in the second convolutional layer is a 3×1 spatially separable convolutional sublayer, and the pooling sublayer is a max-pooling sublayer with a stride of 2; the one-dimensional convolutional sublayer in the third convolutional layer is a 2x1 one-dimensional convolutional sublayer; the number of the third channel sublayer is a multiple of the number of the first channel sublayer and the number of the second channel sublayer.

12. A method for detecting electric arc faults, comprising: The obtained initial current dataset is input into the target arc detection model, and the predicted target text data is output. The target arc detection model is trained based on the method of any one of claims 1 to 11.

13. An electronic device, comprising: One or more processors; Storage device for storing one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors perform the method according to any one of claims 1 to 11.

14. A computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 11.

Images