Power distribution network small current single-phase ground fault identification method and device based on cloud edge cooperation

Abstract

The embodiment of the application provides a cloud-edge cooperative power distribution network small-current single-phase ground fault identification method and device, and relates to the technical field of power distribution Internet of Things. The method comprises the following steps: obtaining a teacher model based on a neural network for small-current single-phase ground fault identification which is trained; using a multi-granularity feature distillation algorithm to perform knowledge distillation transmission of the teacher model to a student model; deploying the student model to an edge side of a cloud-edge system, and the student model outputs a fault identification result based on data obtained by the edge side. The small-current single-phase ground fault identification method provided by the embodiment has the advantages of accurate identification and fast response.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution Internet of Things, in particular to a cloud-edge collaborative power distribution network small current single-phase ground fault identification method, a cloud-edge collaborative power distribution network small current single-phase ground fault identification device, an electronic device, a storage medium and a computer program product. BACKGROUND

[0002] The medium and low voltage power distribution network widely adopts a small current grounding system, the probability of single-phase ground fault is high, although the power supply can be maintained for a short time, but the existence of the fault will cause the voltage of the non-fault phase to rise, which threatens the insulation of the equipment and may cause serious accidents such as phase-to-phase short circuit; the overvoltage generated by arc grounding will also damage power equipment and affect the safe and stable operation of the power grid. In order to reduce the accident loss and improve the power supply reliability, it is necessary to accurately and quickly identify the fault and handle it. At present, in the field of small current single-phase ground fault identification, deep learning models have shown high fault identification accuracy due to their powerful feature extraction and pattern recognition capabilities. However, in the face of complex fault scenarios and massive data, the generalization ability and computational efficiency still need to be improved, and complex models are difficult to deploy on resource-limited power distribution network edge devices, so there is an urgent need to research an accurate, fast and adaptive fault identification method for power distribution network edge devices.

[0003] The existing small current single-phase ground fault identification method of the power distribution network mainly includes centralized cloud processing mode and pure edge computing mode. The traditional centralized cloud processing mode collects power distribution network fault data and uploads it to the cloud, and uses deep learning or expert system for fault analysis, which can achieve high-precision diagnosis, but it relies on stable communication and large data volume, resulting in high delay (usually > 500ms), which is difficult to meet the real-time requirements; the pure edge computing mode deploys lightweight algorithms on terminal devices to achieve millisecond-level response (< 50ms), but is limited by local computing power and cannot handle complex faults (such as intermittent arc faults), and lacks global topology analysis, resulting in high false alarm rate. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a cloud-edge collaborative power distribution network small current single-phase ground fault identification method and device, which trains and deploys teacher and student models for fault identification in a cloud-edge collaborative manner, meets the requirements of accurate, fast and adaptive power distribution network edge device computing resources for power distribution network small current single-phase ground fault identification, and at least solves some of the problems in the background art.

[0005] In order to achieve the above-mentioned purpose, a cloud-edge collaborative power distribution network small current single-phase ground fault identification method is provided in the application, comprising: a cloud-edge collaborative power distribution network small current single-phase ground fault identification method, the method comprises: obtaining a trained neural network-based teacher model for small current single-phase ground fault identification; using a multi-granularity feature distillation algorithm to perform knowledge distillation on the teacher model and transfer it to a student model; deploy the student model to the edge side of the cloud-edge system, and the student model outputs a fault identification result based on the data obtained by the edge side.

[0006] Optionally, the trained neural network-based teacher model for small current single-phase ground fault identification is deployed in the cloud end of the cloud-edge system.

[0007] Optionally, the multi-granularity features include: raw signal features, abstract semantic features, and decision features; using a multi-granularity feature distillation algorithm to perform knowledge distillation on the teacher model and transfer it to a student model comprises: extracting multi-granularity features from the processing process of input data of the teacher model and the student model; constructing a distillation loss function according to the multi-granularity features; updating the parameters of the student model according to the distillation loss function in the training process.

[0008] Optionally, extracting multi-granularity features from the processing process of input data of the teacher model and the student model comprises: extracting raw signal features from the input layer of the teacher model and the input layer of the student model respectively; extracting abstract semantic features from the hidden layer of the teacher model and the hidden layer of the student model respectively; extracting decision features from the output layer of the teacher model and the output layer of the student model respectively.

[0009] Optionally, constructing a distillation loss function according to the multi-granularity features comprises: constructing an input layer network loss function according to the morphological consistency of the raw signal features extracted from the teacher model and the raw signal features extracted from the student model; constructing a hidden layer network loss function according to the similarity measure of the abstract semantic features extracted from the teacher model and the abstract semantic features extracted from the student model; constructing an output layer network loss function according to the probability classification loss and the probability distribution difference of the decision features extracted from the teacher model and the decision features extracted from the student model; and synthesizing the input layer network loss function, the hidden layer network loss function, and the output layer network loss function to obtain the distillation loss function.

[0010] Optionally, the input layer network loss function is constructed according to the shape consistency of the original signal features extracted from the teacher model and the original signal features extracted from the student model, including: measuring the shape consistency of the original signal features by using mean square error; and obtaining the input layer network loss function according to the sum of squares of differences of corresponding elements of input layer feature maps of the teacher model and the student model.

[0011] Optionally, the hidden layer network loss function is constructed according to the similarity measurement of the abstract semantic features extracted from the teacher model and the abstract semantic features extracted from the student model, including: measuring the similarity of the abstract semantic features by using cosine similarity; and obtaining the input layer network loss function according to the cosine value of the angle between vectors formed by corresponding elements of hidden layer feature maps of the teacher model and the student model.

[0012] Optionally, the output layer network loss function is constructed according to the probability classification loss and the probability distribution difference of the decision features extracted from the teacher model and the decision features extracted from the student model, including: using the cross-entropy loss of the classification result as the probability classification loss; using the KL divergence loss of the classification result as the probability distribution difference; and using the weighted sum of the probability classification loss and the probability distribution difference as the output layer network loss function.

[0013] Optionally, the distillation loss function is obtained by integrating the input layer network loss function, the hidden layer network loss function and the output layer network loss function, including: using a weighted sum method to obtain the distillation loss function according to the respective weights of the input layer network loss function, the hidden layer network loss function and the output layer network loss function.

[0014] Optionally, the parameters of the student model are updated according to the distillation loss function during the training process, including: using the distillation loss function as the loss function of the student model; and updating the parameters of the student model by using the distillation loss function and the back propagation algorithm during the training process.

[0015] Optionally, after the student model outputs the fault identification result based on the data obtained on the edge side, the method further includes: uploading the fault identification result to the cloud; obtaining a response instruction of the cloud based on the fault identification result; and conveying a control instruction to the terminal based on the response instruction.

[0016] Optionally, after the student model outputs the fault identification result based on the data obtained on the edge side, the method further includes: caching the data obtained on the edge side, and periodically uploading the cached data to the cloud as new training samples.

[0017] Optionally, the method further comprises: training and updating a neural network-based teacher model for small-current single-phase ground fault identification by using the new training sample, to obtain an updated teacher model; performing knowledge distillation from the updated teacher model to a student model, to obtain an updated student model; and deploying the updated student model to an edge side of a cloud-edge system.

[0018] The application also provides a cloud-edge collaborative power distribution network small-current single-phase ground fault identification device, comprising: a model acquisition module, configured to acquire a trained neural network-based teacher model for small-current single-phase ground fault identification; a knowledge distillation module, configured to perform knowledge distillation from the teacher model to a student model by using a multi-granularity feature distillation algorithm; and a model deployment module, configured to deploy the student model to an edge side of a cloud-edge system, wherein the student model outputs a fault identification result based on data acquired by the edge side.

[0019] The application also provides an electronic device, comprising: at least one processor; a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the at least one processor implements the cloud-edge collaborative power distribution network small-current single-phase ground fault identification method by executing the instructions stored in the memory.

[0020] The application also provides a machine-readable storage medium, which stores instructions executable by a processor, and the instructions configure the processor to implement the cloud-edge collaborative power distribution network small-current single-phase ground fault identification method when executed by the processor.

[0021] The application also provides a computer program product, comprising a computer program, which implements the cloud-edge collaborative power distribution network small-current single-phase ground fault identification method when executed by a processor.

[0022] The above technical solutions have the following beneficial effects:

[0023] (1) The cloud-edge collaborative small-current single-phase ground fault identification model training and deployment method is innovatively proposed, which reduces the interaction delay of fault identification, improves the mass data transmission efficiency, and realizes the fault identification in seconds and global analysis optimization.

[0024] (2) The multi-granularity feature distillation algorithm is innovatively proposed, which improves the learning ability of the student model and ensures the recognition accuracy of the edge side model.

[0025] Other features and advantages of the embodiments of the application will be described in detail in the following specific implementation part. BRIEF DESCRIPTION OF DRAWINGS

[0026] The accompanying drawings are included to provide a further understanding of embodiments of the application, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application, and together with the description serve to explain embodiments of the application, but do not limit the application. In the drawings:

[0027] Figure 1 A schematic diagram of steps of a power distribution network small current single-phase ground fault identification method according to cloud-edge collaboration in an embodiment of the application is shown schematically.

[0028] Figure 2 An implementation schematic diagram of a power distribution network small current single-phase ground fault identification method according to cloud-edge collaboration in an embodiment of the application is shown schematically.

[0029] Figure 3 A structural schematic diagram of a power distribution network small current single-phase ground fault identification device according to cloud-edge collaboration in an embodiment of the application is shown schematically.

[0030] Figure 4 An internal structure diagram of an electronic device according to an embodiment of the application is shown schematically. DETAILED DESCRIPTION

[0031] The specific embodiments of the application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described here are only used to illustrate and explain the application, and are not used to limit the application.

[0032] Figure 1 A schematic diagram of steps of a power distribution network small current single-phase ground fault identification method according to cloud-edge collaboration in an embodiment of the application is shown schematically. As shown in Figure 1 a cloud-edge collaborative power distribution network small current single-phase ground fault identification method, the method comprises:

[0033] S01, obtaining a trained neural network-based teacher model for small current single-phase ground fault identification;

[0034] S02, using a multi-granularity feature distillation algorithm to perform knowledge distillation transmission of the teacher model to a student model;

[0035] S03, deploying the student model to the edge side of the cloud-edge system, and the student model outputs a fault identification result based on data obtained by the edge side.

[0036] Through the above implementation, first, a deep neural network teacher model with good performance but complexity is trained; then, a large deep neural network model is distilled by using a multi-granularity feature distillation algorithm to generate a student model with simple structure and small amount of computation, and ensure that the recognition effect of the distilled small model is infinitely close to that of the large model; finally, the distilled model is distributed and deployed on the edge side of the power distribution Internet of Things with limited computing resources, and data processing and recognition are directly performed on the edge side, improving the massive data processing efficiency of the power distribution network small current single-phase ground fault recognition and reducing the interaction delay.

[0037] In some embodiments of the present application, the trained neural network-based teacher model for small current single-phase ground fault recognition is preferably deployed on the cloud side of the cloud-edge system. The deployment of the student model to the edge side of the cloud-edge system in step S03 includes: distributing the cloud-trained student model to each intelligent feeder terminal on the edge side by using wireless communication or other communication methods, and completing the model deployment on the edge side.

[0038] In some embodiments of the present application, the multi-granularity features include: original signal features, abstract semantic features, and decision features; the multi-granularity feature distillation algorithm is used to transfer the knowledge of the teacher model to the student model, including: extracting multi-granularity features from the processing process of input data of the teacher model and the student model; constructing a distillation loss function according to the multi-granularity features; updating the parameters of the student model according to the distillation loss function during the training process. Small current single-phase ground fault data contains rich time domain, frequency domain and time-frequency domain features. In order to fully utilize these information, a multi-granularity feature distillation module is designed. In different network layers of the teacher model and the student model, original signal features of the input layer, abstract semantic features of the hidden layer, and decision features of the output layer are extracted respectively. This embodiment realizes the accurate migration of features at different levels by constructing multiple sets of distillation loss functions.

[0039] In some embodiments of the present application, multi-granularity features are extracted from the processing of input data by the teacher model and the student model, including: extracting original signal features from the input layers of the teacher model and the student model, respectively; extracting abstract semantic features from the hidden layers of the teacher model and the student model, respectively; and extracting decision features from the output layers of the teacher model and the student model, respectively. Here, extraction includes classification according to the model function layers of the teacher model and the student model, and the output of a certain functional layer is taken as the corresponding original signal feature, abstract semantic feature and decision feature. Specifically, input layer features refer to the basic information retained by the original signal feature maps extracted by the input convolution layers of the teacher model and the student model after processing the input fault data (pre-processed zero sequence voltage signals), which helps the student model learn the basic morphology of fault signals, such as amplitude variation, rising and falling trend, etc. Let the feature maps output by the shallow network of the teacher model be , and the feature maps output by the shallow network of the student model be . Hidden layer features are the preliminary abstract semantic features obtained after further processing of data by the middle network of the teacher model and the student model, which can reflect the essential attributes of faults, such as specific frequency components and waveform distortion patterns. Let the feature maps output by the middle network of the teacher model be , and the feature maps output by the middle network of the student model be . Output layer features refer to the deep features related to decision extracted by the fully connected layers of the teacher model and the student model, which are closely related to the final decision of the model and guide the student model to imitate the decision logic of the teacher model. Let the feature representation output by the deep network of the teacher model be , and the feature representation output by the deep network of the student model be .

[0040] In some embodiments of the present application, an input layer network loss function is constructed according to the morphology consistency of the original signal features extracted from the teacher model and the original signal features extracted from the student model. Specifically, it includes: using mean square error to measure the morphology consistency of the original signal features; and obtaining the input layer network loss function according to the sum of squares of the differences between the corresponding elements of the input layer feature maps of the teacher model and the student model. For example, for shallow features, since the main concern is the basic morphology consistency of the original signal, a mean square error (MSE) loss function is used, and the formula of the input layer network loss function is:

[0041]

[0042] wherein N is the number of elements in the feature map, and Let be the values ​​of the i-th elements in the input layer feature maps of the teacher and student models, respectively. This loss function minimizes the sum of squared differences between corresponding elements in the input feature maps of the teacher and student models, thereby enabling the student model to learn the basic waveform information of the original signal in the teacher model.

[0043] On the other hand, a hidden layer network loss function is constructed based on the similarity measure between the abstract semantic features extracted from the teacher model and the abstract semantic features extracted from the student model. Specifically, this includes: using cosine similarity to measure the similarity of the abstract semantic features; and obtaining the input layer network loss function based on the cosine of the angle between the vectors formed by corresponding elements of the hidden layer feature maps of the teacher and student models. An example of a hidden layer network loss function is as follows: Hidden layer features emphasize the similarity of abstract semantic features, using a cosine similarity loss function. The formula is:

[0044]

[0045] Where N is the number of elements in the feature map. and These represent the values ​​of the i-th element in the hidden layer feature maps of the teacher and student models, respectively. Cosine similarity measures the cosine of the angle between two vectors, ranging from -1 to 1; the closer the value is to 1, the more similar the two vectors are. This loss function guides the student model to learn the abstract semantic representation of the features in the teacher model, enabling the student model to capture the essential attributes of fault features.

[0046] On the other hand, the output layer network loss function is constructed based on the probability classification loss and probability distribution difference between the decision features extracted from the teacher model and the decision features extracted from the student model. Specifically, this includes: using the cross-entropy loss of the classification results as the probability classification loss; and using the KL divergence loss of the classification results as the probability distribution difference. KL divergence (Kullback-Leibler divergence) is an indicator used to measure the difference between two probability distributions, also known as relative entropy or information divergence. Its theoretical significance is that a larger KL divergence indicates a greater degree of difference between the two distributions; while a KL divergence of 0 indicates that the two distributions are completely identical. KL divergence is commonly used in machine learning and statistics to help measure how much information is lost when choosing an approximate distribution. The weighted sum of the probability classification loss and the probability distribution difference is used as the output layer network loss function. Output layer network loss function: Output layer features are closely related to decisions and require a combination of cross-entropy loss and KL divergence loss.

[0047] The cross-entropy loss function is formulated as follows:

[0048]

[0049] Where C represents the number of fault types. This represents the probability of type c faults in the true labels (usually 0 or 1). Predict the probability of type c fault for the model.

[0050] The formula for the KL divergence loss function is:

[0051]

[0052] in, To predict the probability of type c faults for the teacher model, Predict the probability of type c faults for the student model.

[0053] Output layer network loss function The formula is:

[0054]

[0055] in, and These are weighting coefficients used to adjust the proportions of cross-entropy loss and KL divergence loss in the total loss. This loss function encourages the student model to mimic the decision-making logic of the teacher model, thereby improving the accuracy of fault identification.

[0056] In some embodiments of this application, the distillation loss function is obtained by combining the input layer network loss function, the hidden layer network loss function, and the output layer network loss function. This may include: obtaining the distillation loss function by weighted summation based on the respective weights of the input layer network loss function, the hidden layer network loss function, and the output layer network loss function. for:

[0057]

[0058] in, , and These are the weight coefficients for the input, hidden, and output network losses, respectively. Using the backpropagation algorithm, based on the total distillation loss... The parameters of the student model are updated to gradually approximate the teacher model in terms of multi-level features, achieving effective knowledge distillation and transfer. This is based on the total distillation loss. The parameters of the student model can be updated by using the distillation loss function as the loss function for the student model. During the training process of the student model, the parameters are updated using the distillation loss function and the backpropagation algorithm.

[0059] In the aforementioned implementations, the architectures of the teacher and student models for fault identification can be the same or different. However, generally, the teacher model has a more complex architecture, while the student model has a simpler structure and lower computational cost. For example, both the teacher and student models may use ResNet residual networks, but the teacher model may use ResNet18 while the student model may use ResNet8.

[0060] In some embodiments of this application, after the student model outputs a fault identification result based on the data acquired at the edge, the method further includes: uploading the fault identification result to the cloud; obtaining a response instruction from the cloud based on the fault identification result; and transmitting a control instruction to the terminal based on the response instruction. This embodiment provides a subsequent processing flow for the fault identification result. By uploading the fault identification result to the cloud through a cloud-edge collaboration mechanism, it avoids storing a large amount of fault handling knowledge and measures at the edge, allowing the edge to focus more on fault identification.

[0061] In some embodiments of this application, after the student model outputs fault identification results based on the data acquired at the edge, the method further includes: caching the data acquired at the edge and periodically uploading the cached data to the cloud as new training samples. This embodiment provides a data collection function at the edge, providing data samples for the updating and iteration of the teacher model and the student model.

[0062] In some embodiments of this application, the new training samples are used to train and update a neural network-based teacher model for identifying low-current single-phase grounding faults, resulting in an updated teacher model. Knowledge distillation is then performed on the updated teacher model and transferred to a student model, resulting in an updated student model. Finally, the updated student model is deployed to the edge of the cloud-edge system. This embodiment provides an iterative update mechanism for the teacher and student models, improving the accuracy of online deployment.

[0063] Figure 2 The illustration shows a schematic diagram of an implementation of a cloud-edge collaborative method for identifying small-current single-phase grounding faults in a distribution network according to an embodiment of this application. For example... Figure 2 As shown, the specific operational process for deploying a cloud-edge collaborative low-current single-phase grounding fault identification model mainly includes three parts: data storage and uploading, model training and distribution, and task identification and response. The results include fault identification results, the data includes historical fault data, the model includes student models, and the instructions include instructions issued based on the fault identification results. Detailed steps are as follows:

[0064] (1) The cloud platform provides training samples for the original teacher and student models in the cloud based on the historical data collection of small current single-phase grounding faults.

[0065] (2) Convolutional neural networks are constructed on the cloud platform, including the selected teacher network and student network. For example, the teacher model for fault identification uses ResNet18, and the student model uses ResNet8. The generated training samples are input into the teacher model, and the teacher model is trained first. The model training process mainly includes: the problem of identifying single-phase grounding faults with small current can be modeled as a deep learning classification problem. The fault types include stable faults, arc faults, high-resistance arc faults, high-resistance discharge faults, and gap discharge faults. The historical data of zero-sequence voltage transient signals collected by the distribution network at the time of the fault are normalized, filtered and denoised in the edge-side smart feeder terminal and then uploaded to the cloud. The data is then divided into training set, validation set and test set. The teacher model ResNet18 is trained first, and then the multi-granularity feature distillation algorithm is used according to the training process parameters and results of the teacher model. The loss function guides the training of the student model ResNet8, resulting in a concise student model with low computational cost and recognition performance close to that of the teacher model. After training, the model is saved to a cloud server.

[0066] (3) The student model trained in the cloud is sent to each smart feeder terminal on the edge side via wireless communication to complete the model deployment at the edge.

[0067] (4) After receiving terminal data, the intelligent feeder terminal at the edge node performs simple data processing, adjusts it into a data format that can be input into the model, inputs it into the edge side model for fault identification, temporarily stores the data, and periodically transmits pre-processed historical fault data to the cloud.

[0068] (5) Based on the model judgment, the operation and maintenance response is carried out. The response process is as follows: the fault identification result is uploaded to the cloud, then the cloud issues the corresponding instructions, and then the edge side transmits the instructions to the terminal to control the device status.

[0069] (6) Based on the established teacher model, the teacher model and student model are trained and updated using regularly updated historical fault data, and the updated student model is sent to the edge-side intelligent feeder terminal.

[0070] As can be seen from the above implementation methods, this implementation proposes a multi-granularity feature distillation algorithm, which improves the learning ability of the student model and ensures the recognition accuracy of the edge-side model. Simultaneously, based on this cloud-edge collaborative method for training and deploying a low-current single-phase grounding fault identification model, the interaction latency of fault identification is reduced, the efficiency of massive data transmission is improved, and second-level fault identification and global analysis optimization are achieved.

[0071] Based on the same inventive concept, this application also provides a cloud-edge collaborative distribution network low-current single-phase grounding fault identification device. Figure 3 A schematic diagram of the structure of a cloud-edge collaborative distribution network low-current single-phase grounding fault identification device according to an embodiment of this application is shown. Figure 3 As shown, the device includes: a model acquisition module for acquiring a trained neural network-based teacher model for identifying low-current single-phase grounding faults; a knowledge distillation module for performing knowledge distillation on the teacher model using a multi-granularity feature distillation algorithm and passing it to a student model; and a model deployment module for deploying the student model to the edge side of the cloud-edge system, wherein the student model outputs fault identification results based on the data acquired from the edge side.

[0072] In some alternative implementations, the trained neural network-based teacher model for identifying low-current single-phase grounding faults is deployed in the cloud of the cloud-edge system.

[0073] In some optional implementations, the multi-granularity features include: original signal features, abstract semantic features, and decision features; the knowledge distillation of the teacher model and its transfer to the student model is performed using a multi-granularity feature distillation algorithm, including: extracting multi-granularity features from the processing of input data by the teacher model and the student model; constructing a distillation loss function based on the multi-granularity features; and updating the parameters of the student model during the training process based on the distillation loss function.

[0074] In some optional implementations, multi-granularity features are extracted from the processing of input data by the teacher model and the student model, including: extracting raw signal features from the input layer of the teacher model and the input layer of the student model respectively; extracting abstract semantic features from the hidden layer of the teacher model and the hidden layer of the student model respectively; and extracting decision features from the output layer of the teacher model and the output layer of the student model respectively.

[0075] In some optional implementations, constructing a distillation loss function based on the multi-granularity features includes: constructing an input layer network loss function based on the morphological consistency between the original signal features extracted from the teacher model and the original signal features extracted from the student model; constructing a hidden layer network loss function based on the similarity measure between the abstract semantic features extracted from the teacher model and the abstract semantic features extracted from the student model; constructing an output layer network loss function based on the probability classification loss and probability distribution difference between the decision features extracted from the teacher model and the decision features extracted from the student model; and combining the input layer network loss function, the hidden layer network loss function, and the output layer network loss function to obtain the distillation loss function.

[0076] In some alternative implementations, the input layer network loss function is constructed based on the morphological consistency between the original signal features extracted from the teacher model and the original signal features extracted from the student model, including: using mean squared error to measure the morphological consistency of the original signal features; and obtaining the input layer network loss function based on the sum of squared differences between corresponding elements of the input layer feature maps of the teacher model and the student model.

[0077] In some optional implementations, a hidden layer network loss function is constructed based on a similarity measure between the abstract semantic features extracted from the teacher model and the abstract semantic features extracted from the student model, including: using cosine similarity to measure the similarity of the abstract semantic features; and obtaining the input layer network loss function based on the cosine of the angle between the vectors formed by corresponding elements of the hidden layer feature maps of the teacher model and the student model.

[0078] In some optional implementations, the output layer network loss function is constructed based on the probabilistic classification loss and probability distribution difference between the decision features extracted from the teacher model and the decision features extracted from the student model, including: using the cross-entropy loss of the classification results as the probabilistic classification loss; using the KL divergence loss of the classification results as the probability distribution difference; and using the weighted sum of the probabilistic classification loss and the probability distribution difference as the output layer network loss function.

[0079] In some optional implementations, the distillation loss function is obtained by combining the input layer network loss function, the hidden layer network loss function, and the output layer network loss function, including: obtaining the distillation loss function by weighted summation based on the respective weights of the input layer network loss function, the hidden layer network loss function, and the output layer network loss function.

[0080] In some alternative implementations, updating the parameters of the student model during training based on the distillation loss function includes: using the distillation loss function as the loss function of the student model; and updating the parameters of the student model during training using the distillation loss function and the backpropagation algorithm.

[0081] In some optional implementations, after the student model outputs a fault identification result based on the data obtained from the edge side, the method further includes: uploading the fault identification result to the cloud; obtaining a response instruction from the cloud based on the fault identification result; and transmitting a control instruction to the terminal based on the response instruction.

[0082] In some alternative implementations, after the student model outputs the fault identification result based on the data obtained from the edge side, the method further includes: caching the data obtained from the edge side and periodically uploading the cached data to the cloud as new training samples.

[0083] In some alternative implementations, the method further includes: training and updating a neural network-based teacher model for identifying low-current single-phase grounding faults using the new training samples to obtain an updated teacher model; performing knowledge distillation based on the updated teacher model and transferring it to a student model to obtain an updated student model; and deploying the updated student model to the edge of the cloud-edge system.

[0084] The specific limitations of each functional module in the aforementioned cloud-edge collaborative distribution network low-current single-phase grounding fault identification device can be found in the limitations of the cloud-edge collaborative distribution network low-current single-phase grounding fault identification method described above, and will not be repeated here. Each module in the above system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module. This also achieves the beneficial effects of reducing the interaction latency of fault identification, improving the efficiency of massive data transmission, and realizing second-level fault identification and global analysis and optimization.

[0085] In some embodiments of this application, an electronic device is also provided, comprising: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which executes the aforementioned cloud-edge collaborative method for identifying small-current single-phase grounding faults in power distribution networks. Its internal structure diagram can be shown as follows. Figure 4 As shown. Figure 4 The diagram schematically illustrates the internal structure of an electronic device according to an embodiment of this application. The electronic device includes a processor A01, a network interface A02, a memory (not shown), and a database (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory includes internal memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01, a computer program B02, and a database (not shown). The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A04. The network interface A02 is used for communication with external terminals via a network connection. When the computer program B02 is executed by the processor A01, it implements a cloud-edge collaborative method for identifying small-current single-phase grounding faults in a power distribution network.

[0086] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0087] In one embodiment provided in this application, a machine-readable storage medium is provided, on which instructions are stored, which, when executed by a processor, cause the processor to be configured to perform the aforementioned cloud-edge collaborative method for identifying small-current single-phase grounding faults in power distribution networks.

[0088] In one embodiment provided in this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the aforementioned cloud-edge collaborative method for identifying single-phase grounding faults in distribution networks with low current.

[0089] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0094] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0095] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0096] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0097] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A cloud-edge collaborative method for identifying low-current single-phase grounding faults in distribution networks, characterized in that, The method includes: Obtain a trained neural network-based teacher model for identifying low-current single-phase grounding faults; A multi-granularity feature distillation algorithm is used to perform knowledge distillation on the teacher model and transfer it to the student model. The student model is deployed to the edge of the cloud-edge system, and the student model outputs fault identification results based on the data obtained from the edge. The multi-granularity features include: original signal features, abstract semantic features, and decision features; The teacher model is distilled using a multi-granularity feature distillation algorithm, and the knowledge is then transferred to the student model, including: Extract multi-granularity features from the processing of input data by the teacher model and the student model; Constructing a distillation loss function based on the multi-granularity features includes: constructing an input layer network loss function based on the morphological consistency between the original signal features extracted from the teacher model and the original signal features extracted from the student model; constructing a hidden layer network loss function based on the similarity measure between the abstract semantic features extracted from the teacher model and the abstract semantic features extracted from the student model; constructing an output layer network loss function based on the probability classification loss and probability distribution difference between the decision features extracted from the teacher model and the decision features extracted from the student model; and combining the input layer network loss function, the hidden layer network loss function, and the output layer network loss function to obtain the distillation loss function. The parameters of the student model are updated during the training process based on the distillation loss function.

2. The method according to claim 1, characterized in that, The trained neural network-based teacher model for identifying low-current single-phase grounding faults is deployed in the cloud of the cloud-edge system.

3. The method according to claim 1, characterized in that, Multi-granularity features are extracted from the processing of input data by the teacher model and the student model, including: Extract the original signal features from the input layer of the teacher model and the input layer of the student model respectively; Abstract semantic features are extracted from the hidden layers of the teacher model and the student model, respectively. Decision features are extracted from the output layers of the teacher model and the student model, respectively.

4. The method according to claim 1, characterized in that, The input layer network loss function is constructed based on the morphological consistency between the original signal features extracted from the teacher model and the original signal features extracted from the student model, including: The mean square error is used to measure the morphological consistency of the original signal features; The input layer network loss function is obtained by summing the squared differences between corresponding elements of the input layer feature maps of the teacher model and the student model.

5. The method according to claim 1, characterized in that, The hidden layer network loss function is constructed based on the similarity measure between the abstract semantic features extracted from the teacher model and the abstract semantic features extracted from the student model, including: Cosine similarity is used to measure the similarity of the abstract semantic features; The input layer network loss function is obtained by taking the cosine of the angle between the vectors formed by the corresponding elements of the hidden layer feature maps of the teacher model and the student model.

6. The method according to claim 1, characterized in that, The output layer network loss function is constructed based on the probability classification loss and probability distribution difference between the decision features extracted from the teacher model and the decision features extracted from the student model, including: The cross-entropy loss of the classification result is used as the probability classification loss; The KL divergence loss of the classification results is used as the probability distribution difference; The weighted sum of the probability classification loss and the probability distribution difference is used as the loss function of the output layer network.

7. The method according to claim 1, characterized in that, The distillation loss function is obtained by combining the input layer network loss function, the hidden layer network loss function, and the output layer network loss function, including: The distillation loss function is obtained by weighting the input layer network loss function, the hidden layer network loss function, and the output layer network loss function.

8. The method according to claim 1, characterized in that, The parameters of the student model are updated during training based on the distillation loss function, including: The distillation loss function is used as the loss function for the student model; During training, the parameters of the student model are updated using the distillation loss function and the backpropagation algorithm.

9. The method according to claim 1, characterized in that, After the student model outputs the fault identification result based on the data obtained from the edge side, the method further includes: Upload the fault identification results to the cloud; Obtain a response command from the cloud based on the fault identification results; Based on the response instruction, control instructions are conveyed to the terminal.

10. The method according to claim 1, characterized in that, After the student model outputs the fault identification result based on the data obtained from the edge side, the method further includes: The data acquired at the edge is cached, and the cached data is periodically uploaded to the cloud as new training samples.

11. The method according to claim 10, characterized in that, The method further includes: using the new training samples to train and update the neural network-based teacher model for identifying small current single-phase grounding faults, so as to obtain the updated teacher model. Based on the updated teacher model, knowledge distillation is performed and transferred to the student model to obtain the updated student model. The updated student model is deployed to the edge of the cloud-edge system.

12. A cloud-edge collaborative distribution network low-current single-phase grounding fault identification device, characterized in that, The device includes: The model acquisition module is used to acquire a trained teacher model based on neural networks for identifying low-current single-phase grounding faults. The knowledge distillation module is used to perform knowledge distillation on the teacher model using a multi-granularity feature distillation algorithm and then transfer it to the student model; and The model deployment module is used to deploy the student model to the edge side of the cloud-edge system, and the student model outputs fault identification results based on the data obtained from the edge side; The multi-granularity features include: original signal features, abstract semantic features, and decision features; The teacher model is distilled using a multi-granularity feature distillation algorithm, and the knowledge is then transferred to the student model, including: Extract multi-granularity features from the processing of input data by the teacher model and the student model; Constructing a distillation loss function based on the multi-granularity features includes: constructing an input layer network loss function based on the morphological consistency between the original signal features extracted from the teacher model and the original signal features extracted from the student model; constructing a hidden layer network loss function based on the similarity measure between the abstract semantic features extracted from the teacher model and the abstract semantic features extracted from the student model; constructing an output layer network loss function based on the probability classification loss and probability distribution difference between the decision features extracted from the teacher model and the decision features extracted from the student model; and combining the input layer network loss function, the hidden layer network loss function, and the output layer network loss function to obtain the distillation loss function. The parameters of the student model are updated during the training process based on the distillation loss function.

13. An electronic device, characterized in that, include: At least one processor; A memory connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the at least one processor implements the cloud-edge collaborative distribution network low-current single-phase grounding fault identification method according to any one of claims 1 to 11 by executing the instructions stored in the memory.

14. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the cloud-edge collaborative method for identifying single-phase grounding faults in distribution networks with low current as described in any one of claims 1 to 11.

15. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the cloud-edge collaborative method for identifying single-phase grounding faults in distribution networks with low current as described in any one of claims 1 to 11.

Images