Lightweight power distribution network automation equipment fault rapid diagnosis method and device

Abstract

The invention discloses a lightweight power distribution network automation equipment fault rapid diagnosis method and device, and the method comprises the steps: obtaining the real-time environment operation data of power distribution network equipment, carrying out the preprocessing and GAF image coding, and obtaining a GAF image; inputting the GAF image into a lightweight fault diagnosis model to generate a fault type; the lightweight fault diagnosis model is obtained through the following steps: obtaining a plurality of historical environment operation data and historical fault states of the power distribution network equipment; performing preprocessing and GAF image coding on the historical environment operation data to obtain a training image; constructing a multi-stage neural network model, and inputting the training image and the historical fault state into the multi-stage neural network model for training to obtain a teacher-level fault diagnosis model; and constructing a convolutional neural network model, and inputting the soft label output by the teacher-level fault diagnosis model and the corresponding training image into the convolutional neural network model for training to obtain a lightweight fault diagnosis model.

Description

Technical Field

[0001] The present invention relates to the field of distribution network fault diagnosis, and in particular to a lightweight distribution network automation equipment fault rapid diagnosis method and device. Background Art

[0002] In the field of distribution network automation equipment fault diagnosis, traditional deep learning models usually require a large amount of terminal device data for training and use complex algorithms to ensure accurate equipment fault diagnosis. However, the amount of data generated by distribution network terminal devices is huge, which places huge demands on centralized computing resources. Due to the high complexity of the model and the large demand for computing resources, the actual deployment consumes huge computing resources and it is difficult to meet the needs of rapid iteration. At the same time, complex models are difficult to deploy on resource-constrained devices, which limits their application in distribution networks. Summary of the invention

[0003] The embodiments of the present invention provide a lightweight method and device for rapid diagnosis of faults of distribution network automation equipment, which can reduce the complexity of the model while maintaining high accuracy.

[0004] An embodiment of the present invention provides a lightweight method for rapid diagnosis of distribution network automation equipment faults, comprising:

[0005] Obtaining real-time first environment operation data of distribution network automation equipment; the first environment operation data includes: voltage, current, temperature, humidity and pressure;

[0006] Preprocessing the first environment operation data and performing GAF image encoding to obtain a real-time GAF image;

[0007] Input the real-time GAF image into the lightweight fault diagnosis model on the device side, so that the lightweight fault diagnosis model generates the corresponding fault type according to the real-time GAF image;

[0008] Among them, the lightweight fault diagnosis model is obtained by:

[0009] Obtaining several historical second environment operation data of the distribution network automation equipment and its corresponding historical fault status;

[0010] Preprocessing the second environment operation data and performing GAF image encoding to obtain a GAF training image;

[0011] A multi-stage neural network model is constructed, and several GAF training images and their corresponding historical fault states are input into the multi-stage neural network model for iterative training until the first loss function converges, thereby obtaining a teacher-level fault diagnosis model;

[0012] A convolutional neural network model is constructed, and the soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model.

[0013] Further, the first environment operation data is preprocessed and GAF ​​image encoded to obtain a real-time GAF image, including:

[0014] Performing outlier elimination, missing value supplementation, and normalization processing on the first environment operation data to obtain first preprocessed data;

[0015] Performing fully adaptive noise integrated empirical mode decomposition on the first preprocessed data to obtain second preprocessed data;

[0016] The second preprocessed data is screened for correlation by a correlation coefficient method, and normalized to obtain third preprocessed data;

[0017] A Gram matrix is ​​constructed according to the polar coordinates of the third preprocessed data, and the values ​​of the Gram matrix are mapped to obtain a GAF image.

[0018] Furthermore, outliers are eliminated, missing values ​​are supplemented, and normalization is performed on the first environment operation data to obtain first preprocessed data, including:

[0019] For each type of data in the first environment operation data, the Z value of each data is calculated by the Z-score method;

[0020] Compare the Z value of each data with the preset abnormal threshold, and remove the data corresponding to the Z value exceeding the preset abnormal threshold to obtain the data after abnormal removal;

[0021] The data after the abnormal elimination is sorted according to the time series, and the data values ​​of the previous and next time series of each missing time series in the data after the abnormal elimination are linearly supplemented to obtain the supplemented data;

[0022] The supplemented data is normalized to obtain first preprocessed data.

[0023] Further, performing fully adaptive noise integrated empirical mode decomposition on the first preprocessed data to obtain second preprocessed data includes:

[0024] According to the first preprocessed data, a data matrix is ​​constructed in time sequence; wherein the columns in the data matrix represent data of the same category, and the rows in the data matrix represent data of the same time sequence;

[0025] The cycle process is executed until the preset energy threshold is met; the cycle process includes:

[0026] Perform modal decomposition on each data in the current data matrix to obtain the intrinsic mode function matrix;

[0027] Adding white noise to each intrinsic mode function in the current intrinsic mode function matrix to generate a noise function matrix;

[0028] Calculate the average value of the noise functions with the same order in each category in the current noise function matrix to obtain the intrinsic mode function of each category;

[0029] Subtract the intrinsic mode function of the corresponding category from the current data matrix to obtain the residual signal matrix;

[0030] It is determined whether the energy of the residual signal matrix meets a preset energy threshold. If so, the intrinsic mode function of each category is used as the second preprocessed data. If not, the residual signal matrix is ​​updated to the current data matrix.

[0031] Furthermore, the multi-stage neural network model includes: a four-stage network layer, a linear projection layer, a layer normalization module, an efficient multi-head attention module, a multi-layer perception module, a feature map module, and a first output layer;

[0032] Among them, the network layer of the first stage and the network layer of the second stage both include: the first convolution layer, the first pooling layer and the activation function;

[0033] The network layer of the third stage and the network layer of the fourth stage both include: three densely connected networks.

[0034] Furthermore, the convolutional neural network model includes: two second convolutional layers, two second pooling layers, a fully connected layer and a second output layer.

[0035] Furthermore, several GAF training images and their corresponding historical fault states are input into the multi-stage neural network model for iterative training until the first loss function converges, and a teacher-level fault diagnosis model is obtained, including:

[0036] Several GAF training images and their corresponding historical fault states are divided into a training set and a test set;

[0037] The loop process is executed until the accuracy threshold is met; the loop process includes:

[0038] Each GAF training image and the corresponding historical fault state in the training set are respectively input into the current multi-stage neural network model for iterative training until the first cross entropy loss function converges, thereby obtaining a teacher-level fault diagnosis model to be tested; wherein, during each training, the multi-stage neural network model generates a first prediction probability of each type of fault state according to the current GAF training image, the preset first distillation temperature and the first softmax function; and calculates the first cross entropy loss function according to the first prediction probability of each type of fault state and the corresponding historical fault state;

[0039] Input each GAF training image and the corresponding historical fault state in the test set into the teacher-level fault diagnosis model to be tested, so that the teacher-level fault diagnosis model to be tested outputs the probability distribution result of each GAF training image in the test set;

[0040] Calculate the test accuracy based on the probability distribution results of each GAF training image in the test set and the historical fault status of the corresponding GAF training image;

[0041] Determine whether the test accuracy exceeds the preset accuracy threshold. If so, the teacher-level fault diagnosis model to be tested is used as the final teacher-level fault diagnosis model. If not, adjust the network parameters of the current multi-stage neural network model and use it as the current multi-stage neural network model in the next cycle.

[0042] Furthermore, the soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges, thereby obtaining a lightweight fault diagnosis model, including:

[0043] The first predicted probability of each type of fault state is used as the soft label output by the teacher-level fault diagnosis model;

[0044] The soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second cross entropy loss function converges to obtain a lightweight fault diagnosis model; wherein, during each training, the convolutional neural network model generates a second prediction probability for each type of fault state according to the current GAF training image, the preset second distillation temperature and the second softmax function; the distillation error is calculated according to the second prediction probability of each type of fault state and the corresponding first prediction probability of each type of fault state; the convolutional neural network model generates a hard prediction result according to the current GAF training image; the classification error is calculated according to the hard prediction result and the corresponding historical fault state; and the second cross entropy loss function is calculated according to the distillation error and the classification error.

[0045] Based on the above method embodiment, the present invention provides a corresponding device embodiment, including: a real-time data acquisition module, a data processing module and a fault diagnosis module;

[0046] The data real-time acquisition module is used to acquire the real-time first environment operation data of the distribution network automation equipment; the first environment operation data includes: voltage, current, temperature, humidity and pressure;

[0047] A data processing module, used for preprocessing the first environment operation data and encoding the GAF image to obtain a real-time GAF image;

[0048] A fault diagnosis module is used to input the real-time GAF image into the lightweight fault diagnosis model on the device side, so that the lightweight fault diagnosis model generates a corresponding fault type according to the real-time GAF image;

[0049] Among them, the lightweight fault diagnosis model is obtained by:

[0050] Obtaining several historical second environment operation data of the distribution network automation equipment and its corresponding historical fault status;

[0051] Preprocessing the second environment operation data and performing GAF image encoding to obtain a GAF training image;

[0052] A multi-stage neural network model is constructed, and several GAF training images and their corresponding historical fault states are input into the multi-stage neural network model for iterative training until the first loss function converges, thereby obtaining a teacher-level fault diagnosis model;

[0053] A convolutional neural network model is constructed, and the soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model.

[0054] Further, the data processing module includes: a data preprocessing submodule, a modal decomposition submodule, a correlation screening submodule and a GAF image mapping submodule;

[0055] The data preprocessing submodule is used to remove abnormal values, supplement missing values ​​and normalize the first environment operation data to obtain first preprocessed data;

[0056] A modal decomposition submodule, used for performing fully adaptive noise integrated empirical mode decomposition on the first preprocessed data to obtain second preprocessed data;

[0057] A correlation screening submodule, used to perform correlation screening on the second preprocessed data by using a correlation coefficient method, and perform normalization processing to obtain third preprocessed data;

[0058] The GAF image mapping submodule is used to construct a Gram matrix according to the polar coordinates of the third preprocessed data, and map the values ​​of the Gram matrix to obtain a GAF image.

[0059] Compared with the prior art, the beneficial effects of the embodiment of this solution are:

[0060] The present invention obtains the real-time first environment operation data of the distribution network automation equipment; the first environment operation data includes: voltage, current, temperature, humidity and pressure; then, the first environment operation data is preprocessed and GAF ​​image encoded to obtain a real-time GAF image; finally, the real-time GAF image is input into the lightweight fault diagnosis model of the device end, so that the lightweight fault diagnosis model generates the corresponding fault type according to the real-time GAF image; wherein, the lightweight fault diagnosis model uses the soft label output by the trained fault diagnosis model as the target learning label, and the soft label is a form of knowledge transfer of the teacher model used in knowledge distillation, which is smoother than the hard label. Therefore, the student model only needs to learn part of the information of the soft label, and does not need complex parameters to fit the details of each label, thereby reducing the number of parameters. At the same time, the fault diagnosis model belongs to a complex multi-stage neural network model, and the lightweight fault diagnosis model belongs to a simple convolutional neural network model, which is suitable for application on resource-constrained edge devices. Through the knowledge transfer of the soft label, a simple model such as the lightweight fault diagnosis model can also maintain a high accuracy.

[0061] In summary, the lightweight fault diagnosis model of the present invention obtains the knowledge of the fault diagnosis model by learning the soft labels of the complex fault diagnosis model, thereby reducing the complexity of the model while maintaining high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flowchart of a lightweight method for rapid fault diagnosis of distribution network automation equipment provided by an embodiment of the present invention;

[0063] Figure 2 It is a flowchart of a lightweight fault diagnosis model training process provided by an embodiment of the present invention;

[0064] Figure 3 is a schematic diagram of the structure of a multi-stage neural network model provided by an embodiment of the present invention;

[0065] Figure 4 It is a structural schematic diagram of a convolutional neural network model provided by an embodiment of the present invention;

[0066] Figure 5It is a structural schematic diagram of a lightweight distribution network automation equipment fault rapid diagnosis device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features.

[0069] like Figure 1 As shown, an embodiment of the present invention provides a lightweight distribution network automation equipment fault rapid diagnosis method, the method comprising at least the following steps:

[0070] Step S1: Acquire real-time first environment operation data of the distribution network automation equipment; the first environment operation data includes: voltage, current, temperature, humidity and pressure;

[0071] For step S1, the first environmental operation data of the equipment is collected from the distribution network automation equipment terminal in real time, which needs to include voltage, current, temperature, humidity and pressure, etc. These data can provide detailed information about the automation equipment to reflect the operating status and environmental conditions of the equipment.

[0072] Step S2: preprocessing the first environment operation data and performing GAF image encoding to obtain a real-time GAF image;

[0073] In a preferred embodiment, the first environment operation data is preprocessed and GAF ​​image encoded to obtain a real-time GAF image, including:

[0074] Performing outlier elimination, missing value supplementation, and normalization processing on the first environment operation data to obtain first preprocessed data;

[0075] Performing fully adaptive noise integrated empirical mode decomposition on the first preprocessed data to obtain second preprocessed data;

[0076] The second preprocessed data is screened for correlation by a correlation coefficient method, and normalized to obtain third preprocessed data;

[0077] A Gram matrix is ​​constructed according to the polar coordinates of the third preprocessed data, and the values ​​of the Gram matrix are mapped to obtain a GAF image.

[0078] In a preferred embodiment, outlier removal, missing value supplementation and normalization are performed on the first environment operation data to obtain first preprocessed data, including:

[0079] For each type of data in the first environment operation data, the Z value of each data is calculated by the Z-score method;

[0080] Compare the Z value of each data with the preset abnormal threshold, and remove the data corresponding to the Z value exceeding the preset abnormal threshold to obtain the data after abnormal removal;

[0081] The data after the abnormal elimination is sorted according to the time series, and the data values ​​of the previous and next time series of each missing time series in the data after the abnormal elimination are linearly supplemented to obtain the supplemented data;

[0082] The supplemented data is normalized to obtain first preprocessed data.

[0083] In a preferred embodiment, performing fully adaptive noise integrated empirical mode decomposition on the first preprocessed data to obtain the second preprocessed data includes:

[0084] According to the first preprocessed data, a data matrix is ​​constructed in time sequence; wherein the columns in the data matrix represent data of the same category, and the rows in the data matrix represent data of the same time sequence;

[0085] The cycle process is executed until the preset energy threshold is met; the cycle process includes:

[0086] Perform modal decomposition on each data in the current data matrix to obtain the intrinsic mode function matrix;

[0087] Adding white noise to each intrinsic mode function in the current intrinsic mode function matrix to generate a noise function matrix;

[0088] Calculate the average value of the noise functions with the same order in each category in the current noise function matrix to obtain the intrinsic mode function of each category;

[0089] Subtract the intrinsic mode function of the corresponding category from the current data matrix to obtain the residual signal matrix;

[0090] It is determined whether the energy of the residual signal matrix meets a preset energy threshold. If so, the intrinsic mode function of each category is used as the second preprocessed data. If not, the residual signal matrix is ​​updated to the current data matrix.

[0091] For step S2, a series of preprocessing steps are performed on the real-time first environment operation data acquired in step S1 to prepare for subsequent data analysis and modeling.

[0092] First, remove the outliers in the first environment operation data. Specifically, for each type of data (such as voltage, current, temperature, etc.), calculate its Z-score value using the following formula:

[0093]

[0094] Among them, Z represents the Z-score value of the data, x represents the value of the data, μ represents the mean of the data of this type, and σ represents the standard deviation of the data of this type. Z-score is a standardized method to measure the distance of data from the mean, which indicates the degree of difference between the data point and the mean.

[0095] For each data point, calculate the absolute value of its Z-score value. If its absolute value exceeds the set abnormal threshold, it will be marked as an outlier. According to the marked outliers, these data points are removed. You can choose to replace the outliers with missing values ​​(NaN). Repeat the above steps to detect and remove outliers for each type of data. By using the Z-score method for outlier removal, you can effectively identify and remove data points that are significantly different from the average value, which helps to ensure the accuracy and reliability of the data and provide a more reliable basis for subsequent data analysis and modeling.

[0096] Secondly, after removing outliers, there may be missing time series in the data. In order to supplement these missing values, the missing values ​​need to be supplemented. Specifically, the data after outliers are removed is sorted according to the time series. For each missing time series, the time series data values ​​before and after it are found, and then the replacement value of each missing time series is calculated by the following linear interpolation method:

[0097]

[0098] Among them, y represents the replacement value of the missing time series, y1 and y2 represent the data values ​​of the time series before and after the missing time series, x represents the missing time series, and x1 and x2 represent the time series before and after the missing time series. Linear interpolation is a method of estimating the value of the missing position through a straight line between known data points. Through this linear interpolation method, the information of known data points can be used to estimate and supplement the value of the missing position, which can maintain the time series continuity of the data and provide a reasonable estimate for subsequent data analysis and modeling.

[0099] Next, the supplemented data is normalized. Specifically, for each type of data, the minimum and maximum values ​​in the data after supplementing the missing values ​​are found, and min-max normalization is performed to linearly map the data to the specified range, eliminating the dimensional differences between different features so that each feature has similar importance in calculation, and finally the temperature series T is obtained. EMT 、Humidity series H T , pressure sequence P T 、Current sequence C T And the voltage sequence U T , which is used as the first preprocessed data.

[0100] The preprocessed multi-source sequence data, i.e., the first preprocessed data, is used to construct a data matrix X in chronological order n , where each column represents a data source and each row represents data at a time point;

[0101]

[0102] For the data matrix X n Perform fully adaptive noise integrated empirical mode decomposition (CEEMDAN). Specifically, first, perform empirical mode decomposition (EMD) on each data in the data matrix Xn to obtain an intrinsic mode function matrix, wherein each point in the intrinsic mode function matrix contains an IMF vector, and the IMF vector contains multiple IMF components of the EMD decomposition of the corresponding data. Then, add white noise to the intrinsic mode function matrix to generate a noise function matrix. Then, for the current noise function matrix, calculate the average value of the noise function with the same order in each category to obtain the intrinsic mode function of each category, and subtract the intrinsic mode function of the corresponding category from the current data matrix to obtain the residual signal matrix. Finally, determine whether the energy of the residual signal matrix is ​​less than the preset energy threshold. If so, it means that these intrinsic mode functions can better describe the vibration mode of the data, and the intrinsic mode function of each category is used as the second preprocessing data. If not, update the residual signal matrix to the current data matrix and repeat the above steps.

[0103] CEEMDAN is a data decomposition method that can decompose complex nonlinear and non-stationary data into multiple intrinsic mode functions (IMFs). Through CEEMDAN processing, nonlinear and non-stationary data can be effectively processed, the intrinsic mode functions can be adaptively adjusted according to the characteristics of the data, and more accurate data decomposition results can be provided.

[0104] For the second preprocessed data, that is, the intrinsic mode function of each category, the second preprocessed data is screened for correlation by the correlation coefficient method. The correlation coefficient between each mode function is calculated, and the mode function with higher correlation is screened out according to the set threshold, thereby reducing redundant information and extracting the most relevant mode function. Then, the screened data is normalized to obtain the third preprocessed data.

[0105] Finally, the third preprocessed data is converted into polar coordinates. Polar coordinates are a coordinate system that uses angles and distances to represent positions. Converting the data into polar angles and polar diameters can better capture the periodicity and directional characteristics of the data. Then, the Gram matrix is ​​obtained by calculating the inner product between the polar coordinates of the third preprocessed data. By mapping the values ​​of the Gram matrix, it can be converted into a form suitable for generating images.

[0106] Through the above steps, the acquired first environment operation data is preprocessed and GAF ​​image encoded to obtain a real-time GAF image. Through the conversion of the GAF image, data compression and dimensionality reduction are achieved, the complexity of data storage and processing is reduced, and the efficiency of data processing is improved.

[0107] Step S3: inputting the real-time GAF image into the lightweight fault diagnosis model at the device end, so that the lightweight fault diagnosis model generates a corresponding fault type according to the real-time GAF image;

[0108] Among them, the lightweight fault diagnosis model is obtained by:

[0109] Obtaining a number of historical second environment operation data of the distribution network automation equipment and its corresponding historical fault status;

[0110] Preprocessing the second environment operation data and performing GAF image encoding to obtain a GAF training image;

[0111] A multi-stage neural network model is constructed, and several GAF training images and their corresponding historical fault states are input into the multi-stage neural network model for iterative training until the first loss function converges, thereby obtaining a teacher-level fault diagnosis model;

[0112] A convolutional neural network model is constructed, and the soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model.

[0113] For step S3, the GAF image processed by step S2 is input into the lightweight fault diagnosis model deployed on the device side. The lightweight fault diagnosis model takes the GAF image as input and uses the learned knowledge and prior information to match and compare the input GAF image with the known fault mode, so as to determine the possible fault type of the current device. Finally, the predicted fault type is returned as the output result, thereby realizing fault identification.

[0114] The following is a detailed description of the training process of the lightweight fault diagnosis model:

[0115] like Figure 2 As shown in Figure 1, the training process of the lightweight fault diagnosis model includes the following steps:

[0116] Step S301: Acquire a number of historical second environment operation data of the distribution network automation equipment and its corresponding historical fault status;

[0117] For step S301, several historical second environment operation data of the distribution network automation equipment and their corresponding historical fault states are obtained from the historical database information. The second environment operation data needs to include data of parameters such as voltage, current, temperature, humidity and pressure.

[0118] Step S302: preprocessing the second environment operation data and performing GAF image encoding to obtain a GAF training image;

[0119] For step S302, the process of preprocessing the second environment operation data and GAF ​​image encoding is similar to the process of preprocessing the first environment operation data and GAF ​​image encoding. First, the second environment operation data is subjected to outlier elimination, missing value supplementation and normalization processing to obtain the first preprocessed historical data, then the first preprocessed historical data is subjected to fully adaptive noise integrated empirical mode decomposition to obtain the second preprocessed historical data, then the second preprocessed historical data is subjected to correlation screening by the correlation coefficient method, and normalization processing is performed to obtain the third preprocessed historical data; a Gram matrix is ​​constructed according to the polar coordinates of the third preprocessed historical data, and the values ​​of the Gram matrix are mapped to obtain the GAF training image.

[0120] By preprocessing the second environment operation data and processing the GAF image encoding process, a GAF training image is obtained. Such an image can better reflect the operating status and characteristics of the equipment and provide more accurate and effective input data for subsequent fault diagnosis models.

[0121] Step S303: construct a multi-stage neural network model, and input a number of second environment operation data and their corresponding historical fault states into the multi-stage neural network model for iterative training until the first loss function converges, thereby obtaining a teacher-level fault diagnosis model;

[0122] In a preferred embodiment, the multi-stage neural network model includes: a four-stage network layer, a linear projection layer, a layer normalization module, an efficient multi-head attention module, a multi-layer perception module, a feature map module and a first output layer;

[0123] Among them, the network layer of the first stage and the network layer of the second stage both include: the first convolution layer, the first pooling layer and the activation function;

[0124] The network layer of the third stage and the network layer of the fourth stage both include: three densely connected networks.

[0125] In a preferred embodiment, a plurality of second environment operation data and their corresponding historical fault states are input into a multi-stage neural network model for iterative training until the first loss function converges, thereby obtaining a teacher-level fault diagnosis model, including:

[0126] Several GAF training images and their corresponding historical fault states are divided into a training set and a test set;

[0127] The loop process is executed until the accuracy threshold is met; the loop process includes:

[0128] Each GAF training image and the corresponding historical fault state in the training set are respectively input into the current multi-stage neural network model for iterative training until the first cross entropy loss function converges, thereby obtaining a teacher-level fault diagnosis model to be tested; wherein, during each training, the multi-stage neural network model generates a first prediction probability of each type of fault state according to the current GAF training image, the preset first distillation temperature and the first softmax function; and calculates the first cross entropy loss function according to the first prediction probability of each type of fault state and the corresponding historical fault state;

[0129] Input each GAF training image and the corresponding historical fault state in the test set into the teacher-level fault diagnosis model to be tested, so that the teacher-level fault diagnosis model to be tested outputs the probability distribution result of each GAF training image in the test set;

[0130] Calculate the test accuracy based on the probability distribution results of each GAF training image in the test set and the historical fault status of the corresponding GAF training image;

[0131] Determine whether the test accuracy exceeds the preset accuracy threshold. If so, the teacher-level fault diagnosis model to be tested is used as the final teacher-level fault diagnosis model. If not, adjust the network parameters of the current multi-stage neural network model and use it as the current multi-stage neural network model in the next cycle.

[0132] For step S303, first, a complex multi-stage neural network model is constructed, such as Figure 3 As shown in the figure, this complex multi-stage neural network model includes a CNN network and a Transformer, specifically: four-stage network layers, a linear projection layer LP, a layer normalization module LN, an efficient multi-head attention module EMA, a multi-layer perception module MLP, a feature map module FM and an output layer Output. The network layers of the first stage and the second stage both include: a convolutional layer Conv, a pooling layer AveragePooling and a Relu activation function, which are used to downsample the GAF training images to help the network quickly capture the low-level features and local structural information of the image. The network layers of the third stage and the fourth stage both include: three densely connected networks DenseNet, which are used to extract features from high-resolution images.

[0133] Then, the GAF training images and their corresponding historical fault states obtained in step S302 are divided into a training set and a test set. The training set is used for model training and parameter optimization, and the test set is used to evaluate the performance and generalization ability of the model.

[0134] In the training phase, each GAF training image and the corresponding historical fault state in the training set are respectively input into the current multi-stage neural network model for iterative training, and the cross entropy loss function is used as the objective function of the model, and the model is optimized according to the value of the loss function. Specifically, during each training, the current multi-stage neural network model generates the first predicted probability of each type of fault state according to the current GAF training image, the distillation temperature T=t and the softmax function, wherein the softmax function is used to convert the prediction result of the model into a probability form, and the distillation temperature T=t is used to control the smoothness of the softmax function. Then, based on the first predicted probability of each type of fault state and the corresponding historical fault state, the first cross entropy loss function can be calculated. The loss function measures the difference between the model prediction probability and the actual fault state. The training is judged whether it is finished according to the convergence of the first cross entropy loss function. When the first cross entropy loss function converges or reaches a predetermined threshold, the model training can be considered to be completed, and the teacher-level fault diagnosis model to be tested is obtained.

[0135] In the test phase, each GAF training image in the test set and the corresponding historical fault state are respectively input into the teacher-level fault diagnosis model to be tested, and the teacher-level fault diagnosis model to be tested outputs the probability distribution result of each GAF training image in the test set. Then, the test accuracy is calculated based on the probability distribution result of each GAF training image in the test set and the corresponding historical fault state of the GAF training image.

[0136] Determine whether the test accuracy of the current teacher-level fault diagnosis model to be tested exceeds the preset accuracy threshold. If so, it means that the accuracy of the model has reached the expected requirement, and the teacher-level fault diagnosis model to be tested is used as the final teacher-level fault diagnosis model. If not, the network parameters of the current multi-stage neural network model are adjusted, and it is used as the current multi-stage neural network model in the next cycle. In each cycle, the process of the training phase and the test phase is repeated until the test accuracy reaches or exceeds the preset accuracy threshold. In the present embodiment, the preset accuracy threshold is 95%, indicating that the model test can only be passed when the accuracy of the teacher-level fault diagnosis model to be tested reaches more than 95%, otherwise the parameters are adjusted to retrain the teacher-level fault diagnosis model.

[0137] The trained teacher-level fault diagnosis model can accurately predict the fault type of distribution network automation equipment. However, since the teacher-level fault diagnosis model belongs to a large and complex network model, it cannot be directly deployed on resource-constrained distribution network automation equipment for real-time fault identification and early warning.

[0138] In order to realize real-time fault identification and early warning on resource-constrained distribution network automation equipment, the complex teacher-level fault diagnosis model is migrated to a simple lightweight fault diagnosis model through knowledge distillation technology, which can realize real-time fault identification on distribution network automation equipment while ensuring the accuracy of identification.

[0139] Step S304: construct a convolutional neural network model, and input the soft labels output by the teacher-level fault diagnosis model and the corresponding second environment operation data into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model.

[0140] In a preferred embodiment, the convolutional neural network model includes: two second convolutional layers, two second pooling layers, a fully connected layer and a second output layer.

[0141] In a preferred embodiment, the soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model, including:

[0142] The first predicted probability of each type of fault state is used as the soft label output by the teacher-level fault diagnosis model;

[0143] The soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second cross entropy loss function converges to obtain a lightweight fault diagnosis model; wherein, during each training, the convolutional neural network model generates a second prediction probability for each type of fault state according to the current GAF training image, the preset second distillation temperature and the second softmax function; the distillation error is calculated according to the second prediction probability of each type of fault state and the corresponding first prediction probability of each type of fault state; the convolutional neural network model generates a hard prediction result according to the current GAF training image; the classification error is calculated according to the hard prediction result and the corresponding historical fault state; and the second cross entropy loss function is calculated according to the distillation error and the classification error.

[0144] For step S304, first, a simple convolutional neural network model is constructed, such as Figure 4 As shown in the figure, this simple convolutional neural network model includes two convolutional layers, two pooling layers, a fully connected layer and an output layer.

[0145] After the teacher-level fault diagnosis model is trained in step S303, the probability distribution of the fault type prediction of each GAF training image in the training set by the final teacher-level fault diagnosis model can be obtained, which is also called a soft label.

[0146] Next, train the student-level fault diagnosis model:

[0147] The soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model.

[0148] Among them, during each training, the convolutional neural network model generates the predicted probabilities (soft labels) of various fault states according to the current GAF training image, distillation temperature T=t and softmax function. These predicted probabilities represent the predictions of the lightweight model for different fault types. Combined with the soft labels output by the teacher-level fault diagnosis model and the soft labels output by the student-level fault diagnosis model, the distillation error is calculated by the following formula:

[0149]

[0150] Among them, L soft represents the distillation error, and They represent the predicted probability distribution of the teacher-level fault diagnosis model and the student-level fault diagnosis model after distillation by the temperature factor T, also known as soft labels. K represents the total number of fault types. and They represent the output scores of the teacher-level fault diagnosis model and the student-level fault diagnosis model on the i-th fault category, respectively, and T represents the distillation temperature.

[0151] At the same time, the convolutional neural network model generates a hard prediction result at T = 1, that is, without distillation, based on the current GAF training image. According to the hard prediction result and the corresponding historical fault status, the classification error is calculated by the following formula:

[0152]

[0153] Among them, L hard represents the classification error, r i Indicates the corresponding historical fault status label, It represents the predicted probability distribution of the student-level fault diagnosis model at the distillation temperature T=1, also known as the hard prediction result of the student-level fault diagnosis model.

[0154] Finally, according to the preset weight coefficient, the distillation error and classification error are weighted and fused to calculate the final cross entropy loss function:

[0155]

[0156] Among them, L KD represents the final cross entropy loss function of the student-level fault diagnosis model, Represents the preset weight coefficient, which is used to balance the weight of distillation error and classification error. hard (R,P1) represents the cross entropy between the true label and the output of the student-level fault diagnosis model (the label value of a certain type of fault at T=1),

[0157] L soft (P t ,P s ) represents the cross entropy between the soft label of the teacher-level fault diagnosis model (the probability of a certain type of fault predicted under T = t) and the output of the student-level fault diagnosis model (the probability of a certain type of fault predicted under T = t).

[0158] Through the final cross entropy loss function L KD To constrain the training of the student-level fault diagnosis model and obtain the final lightweight fault diagnosis model.

[0159] After obtaining the trained lightweight fault diagnosis model, it is deployed to the edge computing terminal on site. The distribution network terminal equipment collects data from various sensors in the distribution network automation equipment, and inputs it into the lightweight fault diagnosis model after preprocessing. Through the combination of convolutional layer and pooling layer, the key features in the data are extracted, and the fully connected layer is used to integrate the extracted features to form a comprehensive judgment on the fault type. Finally, the output layer provides the final classification result, indicating the possible fault type of the distribution network automation equipment, realizing fault diagnosis, and finally uploading the result back to the distribution network cloud platform.

[0160] The present invention transfers the knowledge of large-scale teacher-level fault diagnosis models to small lightweight fault diagnosis models through the knowledge distillation method, which reduces the complexity and resource consumption of the model, enables real-time fault identification and early warning on resource-constrained devices, and improves the reliability and safety of the distribution network. In addition, due to the low complexity and resource consumption of lightweight models, they can be deployed closer to the data source, thereby reducing data transmission delays and improving data processing speed and efficiency. The advantage of this rapid deployment helps reduce operation and maintenance costs while ensuring real-time monitoring and intelligent response capabilities.

[0161] like Figure 5 As shown, based on the above method embodiment, a corresponding device embodiment is provided;

[0162] An embodiment of the present invention provides a lightweight distribution network automation equipment fault rapid diagnosis device, comprising: a real-time data acquisition module, a data processing module and a fault diagnosis module;

[0163] The data real-time acquisition module is used to acquire the real-time first environment operation data of the distribution network automation equipment; the first environment operation data includes: voltage, current, temperature, humidity and pressure;

[0164] A data processing module, used for preprocessing the first environment operation data and encoding the GAF image to obtain a real-time GAF image;

[0165] A fault diagnosis module is used to input the real-time GAF image into the lightweight fault diagnosis model on the device side, so that the lightweight fault diagnosis model generates a corresponding fault type according to the real-time GAF image;

[0166] Among them, the lightweight fault diagnosis model is obtained by:

[0167] Obtaining a number of historical second environment operation data of the distribution network automation equipment and its corresponding historical fault status;

[0168] Preprocessing the second environment operation data and performing GAF image encoding to obtain a GAF training image;

[0169] A multi-stage neural network model is constructed, and several GAF training images and their corresponding historical fault states are input into the multi-stage neural network model for iterative training until the first loss function converges, thereby obtaining a teacher-level fault diagnosis model;

[0170] A convolutional neural network model is constructed, and the soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model.

[0171] In a preferred embodiment, the data processing module includes: a data preprocessing submodule, a modal decomposition submodule, a correlation screening submodule and a GAF image mapping submodule;

[0172] The data preprocessing submodule is used to remove abnormal values, supplement missing values ​​and normalize the first environment operation data to obtain first preprocessed data;

[0173] A modal decomposition submodule, used for performing fully adaptive noise integrated empirical mode decomposition on the first preprocessed data to obtain second preprocessed data;

[0174] A correlation screening submodule, used to perform correlation screening on the second preprocessed data by using a correlation coefficient method, and perform normalization processing to obtain third preprocessed data;

[0175] The GAF image mapping submodule is used to construct a Gram matrix according to the polar coordinates of the third preprocessed data, and map the values ​​of the Gram matrix to obtain a GAF image.

[0176] It can be understood that the above-mentioned device item embodiments correspond to the method item embodiments of the present invention, which can implement the lightweight distribution network automation equipment fault rapid diagnosis method provided by any one of the above-mentioned method item embodiments of the present invention.

[0177] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art may understand and implement the present invention without creative work.

[0178] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A lightweight method for rapid fault diagnosis of distribution network automation equipment, characterized in that: include: Obtain real-time first-environment operation data of distribution network automation equipment; The first environment operation data includes: voltage, current, temperature, humidity and pressure; Preprocessing the first environment operation data and performing GAF image encoding to obtain a real-time GAF image; Inputting the real-time GAF image into the lightweight fault diagnosis model on the device side, so that the lightweight fault diagnosis model generates a corresponding fault type according to the real-time GAF image; The lightweight fault diagnosis model is obtained in the following way: Obtaining several historical second environment operation data of the distribution network automation equipment and its corresponding historical fault status; Preprocessing and GAF ​​image encoding the second environment operation data to obtain a GAF training image; Constructing a multi-stage neural network model, and inputting a number of GAF training images and their corresponding historical fault states into the multi-stage neural network model for iterative training until the first loss function converges, thereby obtaining a teacher-level fault diagnosis model; A convolutional neural network model is constructed, and the soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model.

2. The lightweight distribution network automation equipment fault rapid diagnosis method according to claim 1 is characterized in that: Preprocessing the first environment operation data and performing GAF image encoding to obtain a real-time GAF image includes: Performing outlier elimination, missing value supplementation, and normalization processing on the first environment operation data to obtain first preprocessed data; Performing fully adaptive noise integrated empirical mode decomposition on the first preprocessed data to obtain second preprocessed data; Performing correlation screening on the second preprocessed data by using a correlation coefficient method, and performing normalization processing to obtain third preprocessed data; A Gram matrix is ​​constructed according to the polar coordinates of the third preprocessed data, and the values ​​of the Gram matrix are mapped to obtain a GAF image.

3. The lightweight distribution network automation equipment fault rapid diagnosis method according to claim 2 is characterized in that: The first environment operation data is subjected to outlier removal, missing value supplementation, and normalization processing to obtain first preprocessed data, including: For each type of data in the first environment operation data, the Z value of each data is calculated by the Z-score method; Compare the Z value of each data with the preset abnormal threshold, and remove the data corresponding to the Z value exceeding the preset abnormal threshold to obtain the data after abnormal removal; The data after the abnormal elimination is sorted according to the time series, and linearly supplemented according to the data values ​​of the time series before and after each missing time series in the data after the abnormal elimination is obtained to obtain supplemented data; The supplemented data is normalized to obtain first preprocessed data.

4. The lightweight distribution network automation equipment fault rapid diagnosis method according to claim 2 is characterized in that: Performing fully adaptive noise integrated empirical mode decomposition on the first preprocessed data to obtain second preprocessed data includes: According to the first preprocessed data, a data matrix is ​​constructed in time sequence; wherein the columns in the data matrix represent data of the same category, and the rows in the data matrix represent data of the same time sequence; The cycle process is executed until a preset energy threshold is met; the cycle process includes: Perform modal decomposition on each data in the current data matrix to obtain the intrinsic mode function matrix; Adding white noise to each intrinsic mode function in the current intrinsic mode function matrix to generate a noise function matrix; Calculate the average value of the noise functions with the same order in each category in the current noise function matrix to obtain the intrinsic mode function of each category; Subtract the intrinsic mode function of the corresponding category from the current data matrix to obtain the residual signal matrix; It is determined whether the energy of the residual signal matrix meets a preset energy threshold. If so, the intrinsic mode function of each category is used as the second preprocessed data. If not, the residual signal matrix is ​​updated to the current data matrix.

5. The lightweight rapid fault diagnosis method for distribution network automation equipment according to claim 1 is characterized in that: The multi-stage neural network model includes: four-stage network layers, a linear projection layer, a layer normalization module, an efficient multi-head attention module, a multi-layer perception module, a feature map module and a first output layer; Among them, the network layer of the first stage and the network layer of the second stage both include: the first convolution layer, the first pooling layer and the activation function; The network layer of the third stage and the network layer of the fourth stage both include: three densely connected networks.

6. The lightweight rapid fault diagnosis method for distribution network automation equipment according to claim 5 is characterized in that: The convolutional neural network model includes: two second convolutional layers, two second pooling layers, a fully connected layer and a second output layer.

7. The lightweight rapid fault diagnosis method for distribution network automation equipment according to claim 6 is characterized in that: Several GAF training images and their corresponding historical fault states are input into the multi-stage neural network model for iterative training until the first loss function converges, thereby obtaining a teacher-level fault diagnosis model, including: Several GAF training images and their corresponding historical fault states are divided into a training set and a test set; The loop process is executed until the accuracy threshold is met; the loop process includes: Input each GAF training image and the corresponding historical fault state in the training set into the current multi-stage neural network model for iterative training until the first cross entropy loss function converges to obtain the teacher-level fault diagnosis model to be tested; wherein, during each training, the multi-stage neural network model generates the first prediction probability of each type of fault state according to the current GAF training image, the preset first distillation temperature and the first softmax function; calculate the first cross entropy loss function according to the first prediction probability of each type of fault state and the corresponding historical fault state; Input each GAF training image and the corresponding historical fault state in the test set into the teacher-level fault diagnosis model to be tested, so that the teacher-level fault diagnosis model to be tested outputs the probability distribution result of each GAF training image in the test set; Calculate the test accuracy based on the probability distribution results of each GAF training image in the test set and the historical fault status of the corresponding GAF training image; Determine whether the test accuracy exceeds the preset accuracy threshold. If so, the teacher-level fault diagnosis model to be tested is used as the final teacher-level fault diagnosis model. If not, adjust the network parameters of the current multi-stage neural network model and use it as the current multi-stage neural network model in the next cycle.

8. The lightweight rapid fault diagnosis method for distribution network automation equipment according to claim 7 is characterized in that: The soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model, including: Using the first predicted probability of each type of fault state as a soft label output by the teacher-level fault diagnosis model; The soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second cross entropy loss function converges to obtain a lightweight fault diagnosis model; wherein, during each training, the convolutional neural network model generates a second prediction probability of each type of fault state according to the current GAF training image, a preset second distillation temperature and a second softmax function; the distillation error is calculated according to the second prediction probability of each type of fault state and the corresponding first prediction probability of each type of fault state; the convolutional neural network model generates a hard prediction result according to the current GAF training image; the classification error is calculated according to the hard prediction result and the corresponding historical fault state; and the second cross entropy loss function is calculated according to the distillation error and the classification error.

9. A lightweight rapid fault diagnosis device for distribution network automation equipment, characterized in that: include: Real-time data acquisition module, data processing module and fault diagnosis module; The real-time data acquisition module is used to acquire the real-time first environment operation data of the distribution network automation equipment; The first environment operation data includes: voltage, current, temperature, humidity and pressure; The data processing module is used to pre-process the first environment operation data and perform GAF image encoding to obtain a real-time GAF image; The fault diagnosis module is used to input the real-time GAF image into the lightweight fault diagnosis model on the device side, so that the lightweight fault diagnosis model generates a corresponding fault type according to the real-time GAF image; The lightweight fault diagnosis model is obtained in the following way: Obtaining a number of historical second environment operation data of the distribution network automation equipment and its corresponding historical fault status; Preprocessing and GAF ​​image encoding the second environment operation data to obtain a GAF training image; Constructing a multi-stage neural network model, and inputting a number of GAF training images and their corresponding historical fault states into the multi-stage neural network model for iterative training until the first loss function converges, thereby obtaining a teacher-level fault diagnosis model; A convolutional neural network model is constructed, and the soft labels output by the teacher-level fault diagnosis model and the corresponding GAF training images are input into the convolutional neural network model for iterative training until the second loss function converges to obtain a lightweight fault diagnosis model.

10. The lightweight rapid fault diagnosis device for distribution network automation equipment according to claim 9, characterized in that: The data processing module includes: a data preprocessing submodule, a modal decomposition submodule, a correlation screening submodule and a GAF image mapping submodule; The data preprocessing submodule is used to remove abnormal values, supplement missing values ​​and normalize the first environment operation data to obtain first preprocessed data; The modal decomposition submodule is used to perform fully adaptive noise integrated empirical mode decomposition on the first preprocessed data to obtain second preprocessed data; The correlation screening submodule is used to perform correlation screening on the second pre-processed data by using a correlation coefficient method, and perform normalization processing to obtain third pre-processed data; The GAF image mapping submodule is used to construct a Gram matrix according to the polar coordinates of the third preprocessed data, and map the values ​​of the Gram matrix to obtain a GAF image.

Images