Fusion neural network-based residual life prediction method, system and device for electric energy measurement and acquisition equipment, and storage medium

Abstract

The invention discloses a method, a system and equipment for predicting residual life of electric energy metering and collecting equipment based on a fusion neural network, and a storage medium, and belongs to the technical field of health state management of the electric energy metering and collecting equipment. Comprising a data preprocessing module, a time domain convolution gating cycle unit fusion neural network construction module, a model training control module, a model test and prediction module and a fine tuning and enhancement module. The method solves the problems of insufficient prediction precision, poor generalization ability and the like of a traditional method in a complex electric power environment, realizes high-precision prediction of the residual life of the electric energy metering and collecting equipment, and realizes intelligent management from change according to age limit to change according to state; and finally, the technical targets of reducing maintenance cost, improving equipment reliability and promoting the development of a smart power grid are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of health status management of electric energy metering and acquisition equipment, and specifically relates to a method, system, device and storage medium for predicting the remaining life of electric energy metering and acquisition equipment based on a fusion neural network. Background Art

[0002] During the long-term use of electric energy metering equipment, it may be affected by factors such as environment, load fluctuations, aging, and mechanical damage, resulting in a gradual decline in its performance. To avoid sudden failures or premature scrapping, predicting the remaining life of the equipment becomes particularly important. Accurate remaining life prediction can help operators carry out effective maintenance and replacement plans, reduce failure rates and maintenance costs, and improve the reliability of the equipment and the stability of the system. Traditional methods usually rely on manual experience or rule-based judgment algorithms. When facing complex and changing power environments, these methods are prone to misdiagnosis and missed diagnosis, and cannot adapt to diverse situations such as various different devices and environmental changes.

[0003] Deep neural networks can automatically extract features and perform non-linear modeling by learning the historical operation data of the equipment, thereby achieving more accurate remaining life prediction. Especially for multi-dimensional and multi-feature data sets, deep learning models can provide higher accuracy and stability than traditional algorithms. Summary of the Invention

[0004] To solve the above technical problems, a method for predicting the remaining life of electric energy metering and acquisition equipment based on a fusion neural network is proposed, including preprocessing the acquisition data of the electric energy metering and acquisition equipment to obtain a data set for a time-domain convolutional gated recurrent unit fusion neural network model and the corresponding labels, normalizing the preprocessed data set and dividing it into three parts: a training set, a validation set, and a test set for the time-domain convolutional gated recurrent unit fusion neural network model; Constructing a combined time-domain convolutional gated recurrent unit fusion neural network of a time-domain convolutional network, a gated recurrent unit network, and a feedforward neural full connection layer, and setting hyperparameters; Specifying a tensor converter and a data loader for training the time-domain convolutional gated recurrent unit fusion neural network, a loss function, an optimizer, and a learning rate scheduler for the time-domain convolutional gated recurrent unit fusion neural network model, setting the maximum number of epochs for training the time-domain convolutional gated recurrent unit fusion neural network model, and the conditions for triggering the early termination of training the time-domain convolutional gated recurrent unit fusion neural network model; Constructing the forward propagation process, the backward propagation process, and the training loop process of the time-domain convolutional gated recurrent unit fusion neural network; Constructing a process for solving the optimal weights of the combined architecture time-domain convolutional gated recurrent unit fusion neural network; Predict the remaining life of the electric energy metering and acquisition device on the test set, and construct a fine-tuning mechanism for the time-domain convolutional gated recurrent unit fusion neural network with a combined architecture.

[0005] As a preferred solution of the method for predicting the remaining life of the electric energy metering and acquisition device based on the fusion neural network according to the present invention, wherein: the acquisition data of the electric energy metering and acquisition device is preprocessed to obtain a time-domain convolutional gated recurrent unit fusion neural network model data set and the labels corresponding to the data. The preprocessed data set is normalized and divided into three parts: a time-domain convolutional gated recurrent unit fusion neural network model training set, a validation set, and a test set, including Perform data preprocessing on the original data collected by the electric energy metering and acquisition device. The original data includes three-phase voltage, three-phase current, three-phase power, total power, power factor, daily power consumption, and the acquisition time interval is one year; and the installation date and fault record of the electric energy metering and acquisition device. Analyze and sort out the collected data, filter out invalid data, and reasonably fill in a small amount of missing data. Calculate relevant electrical indicators including voltage volatility, current volatility, current reverse polarity, power difference rate, and calculate the metering error of the electric energy metering and acquisition device. Divide the data of each electric energy metering and acquisition device into 12 samples per month and assign remaining life labels to obtain a data set. Normalize each data sample in the data set and divide it into three parts: a time-domain convolutional gated recurrent unit fusion neural network model training set, a validation set, and a test set.

[0006] As a preferred solution of the method for predicting the remaining life of the electric energy metering and acquisition device based on the fusion neural network according to the present invention, wherein: construct a combined time-domain convolutional gated recurrent unit fusion neural network of a time-domain convolutional network, a gated recurrent network, and a feedforward neural full connection layer, and set hyperparameters including Construct the input layer of the time-domain convolutional network and specify the batch size of the input samples. Construct the components of the time-domain convolutional network including the temporal convolutional unit TCU. The temporal convolutional unit TCU contains 4 component layers, and the structural order is a one-dimensional dilated causal convolutional layer, a weight parameterization layer, a Relu function activation layer, and a neuron dropout layer. Construct the main body of the time-domain convolutional network, which is successively composed of temporal convolutional units TCU1, TCU2, and TCU3 composed of dilated causal convolutional layers. Construct the skip connection structure of the time-domain convolutional network, judge whether the dimension of the input layer data matches the dimension of the output data of the temporal convolutional cloud TCU3, and output the final feature vector of the time-domain convolutional network. Construct each component of the gated recurrent unit in the gated recurrent network body. Component one is the reset gate, component two is the update gate, and they are in a parallel relationship. Component three is the candidate hidden state component, and component four is the final hidden state component; It is stipulated that the input data dimension of the gated recurrent unit network is the same as the dimension of the final feature vector of the time-domain convolutional network. The output of the last time step of the last gated recurrent unit in the gated recurrent unit network is used as the final output of the gated recurrent unit network; Construct a feedforward neural fully connected layer, which includes three feedforward fully connected layers FC1, FC2, and FC3. The number of neurons in each fully connected layer is N1, N2, and N3, where N3 is determined by the total number of categories. The first and second feedforward fully connected layers are activated with the Relu function, and the third layer is activated with the Softmax function.

[0007] As a preferred solution of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network according to the present invention, wherein: the tensor converter and data loader for training the specified time-domain convolutional gated recurrent unit fusion neural network, the loss function of the time-domain convolutional gated recurrent unit fusion neural network model, the optimizer, and the learning rate scheduler, setting the maximum number of epochs for training the time-domain convolutional gated recurrent unit fusion neural network model and the conditions for triggering the early termination of the training of the time-domain convolutional gated recurrent unit fusion neural network model include, The tensor converter and data loader for training the specified time-domain convolutional gated recurrent unit fusion neural network include converting the normalized data into a tensor matrix form and shuffling the order of the samples in the dataset and recording the data converted into a tensor matrix; The loss function of the time-domain convolutional gated recurrent unit fusion neural network model is to set the loss function of the time-domain convolutional gated recurrent unit fusion neural network as the cross-entropy loss function; The optimizer is to set the optimizer of the time-domain convolutional gated recurrent unit fusion neural network as adaptive momentum estimation; The learning rate scheduler is to set the learning rate adjustment strategy of the learning rate scheduler of the time-domain convolutional gated recurrent unit fusion neural network as follows: when the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model on the validation set has not increased after continuously completing P training epochs, the learning rate decays to: , wherein, is the learning rate before attenuation, is the learning rate after attenuation, is the attenuation coefficient; The maximum number of training epochs for setting the training of the time-domain convolutional gated recurrent unit fusion neural network model and the conditions for triggering the early termination of the training of the time-domain convolutional gated recurrent unit fusion neural network model include that when the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model on the validation set has not increased after several consecutive training epochs, the training is terminated early and the weights of the time-domain convolutional gated recurrent unit fusion neural network model with the best performance in the history of the time-domain convolutional gated recurrent unit fusion neural network are saved.

[0008] As a preferred embodiment of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network according to the present invention, wherein: the construction of the forward propagation process, the backward propagation process, and the training loop process of the time-domain convolutional gated recurrent unit fusion neural network includes, Construct the forward propagation process of a sample in the time-domain convolutional gated recurrent unit fusion neural network. The training set data is read by the data loader and becomes the input data X, which is passed into the input layer of the time-domain convolutional network TCN. The input data is sequentially passed through the time-domain convolutional network to extract the feature vector Y1; The extracted feature vector Y1 is used as the input vector and passed into the main body of the gated recurrent network to further capture the long-term dependencies of the time series, obtaining the deep feature vector Y2. The obtained deep feature vector Y2 is used as the input X3 and passed into the fully connected feedforward neural layer FC to obtain the non-linear mapping vector Y3. The non-linear mapping vector Y3 is passed into the Softmax classifier to calculate the probability of each sample in the current batch belonging to each category, and the category with the highest probability is selected as the final predicted category of the sample; Construct the backward propagation process of a sample in the time-domain convolutional gated recurrent unit fusion neural network. Calculate the cross-entropy loss between the label of the prediction result and the true label according to the loss function, calculate the weight parameter gradients of each layer of neurons in the time-domain convolutional gated recurrent unit fusion neural network, update the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model through the Adam optimizer, and clear the previously calculated gradients after each step of updating the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model; The training loop process includes passing a batch of data through the forward propagation process and the backward propagation process each time until all the training set data is passed, and then a training loop is completed.

[0009] As a preferred embodiment of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network according to the present invention, wherein: the construction of the process for solving the optimal weights of the combined architecture time-domain convolutional gated recurrent unit fusion neural network includes, After each training loop is completed, the data of the validation set is passed into the time-domain convolutional gated recurrent unit fusion neural network with the weight parameters of the current round of training to predict the life labels of the electricity meter samples on the validation set. Calculate the loss according to the cross-loss function and calculate the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model. The learning rate is improved by the learning rate scheduler but no backpropagation is performed. The expression of the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model is: , where Accuracy is the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model, represents the indicator function, which is 1 when the prediction is correct and 0 otherwise, and N is the total number of samples; Continuously repeat the training loop process until the multi-class cross-entropy loss converges, or reaches the maximum number of training loop rounds, or meets the conditions for early termination of training. Then complete the training of the time-domain convolutional gated recurrent unit fusion neural network model and save the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model with the highest classification accuracy in previous training rounds as the final weight parameters of the time-domain convolutional gated recurrent unit fusion neural network for the remaining life prediction of the electricity metering acquisition device.

[0010] As a preferred scheme of the method for predicting the remaining life of the electricity metering acquisition device based on the fusion neural network described in the present invention, wherein: predicting the remaining life of the electricity metering acquisition device on the test set includes, Using the test set data to predict the remaining life of the electricity metering acquisition device in the trained time-domain convolutional gated recurrent unit fusion neural network, and comprehensively evaluating the performance of the time-domain convolutional gated recurrent unit fusion neural network model using the cross-validation method; Divide the data set into k subsets of equal size. Each time, select one subset as the test set and one subset as the validation set. The selected test set and validation set are not repeated each time. The remaining k-2 subsets are used as the training set. Use the training set to train the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model, and use the test set to evaluate the performance of the time-domain convolutional gated recurrent unit fusion neural network model. Repeat the iteration k times, and take the average of the k evaluation results of the time-domain convolutional gated recurrent unit fusion neural network model as the final performance evaluation of the time-domain convolutional gated recurrent unit fusion neural network model; The fine-tuning mechanism for building the combined architecture of the time-domain convolutional gated recurrent unit fusion neural network includes, If the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model on the test set or validation set is less than 5 percentage points or more of the training set, first freeze the weight parameters of 3 one-dimensional dilated convolutional layers in the time-domain convolutional network, adjust the proportion of randomly inactivated neurons in the neuron dropout layer, and replace the learning rate scheduler with cosine annealing; Balance the number of samples in each category, add noise to the training set samples, insert a random array with the same dimension as the sample data into the samples of the category with a small proportion, and assign labels to the randomly generated noise samples. It is required that the number of generated noise samples is less than the number of samples in the expanded category.

[0011] Another object of the present invention is to provide a remaining life prediction system for electric energy metering and acquisition equipment based on a fusion neural network. The present invention solves the problems of non-standard data processing, incomplete feature extraction, insufficient ability of the time-domain convolutional gated recurrent unit fusion neural network model to capture time-series features, and poor model generalization performance in the life prediction of existing electric energy metering and acquisition equipment, realizes high-precision prediction of the remaining life of electric energy metering and acquisition equipment, and improves the intelligent level of equipment management and the reliability of prediction results.

[0012] As a preferred solution of the remaining life prediction system for electric energy metering and acquisition equipment based on the fusion neural network of the present invention, it is characterized in that it includes a data processing module, a fusion neural network construction module, a time-domain convolutional gated recurrent unit fusion neural network model training control module, a time-domain convolutional gated recurrent unit fusion neural network model training and evaluation module, a time-domain convolutional gated recurrent unit fusion neural network model testing and prediction module, and a fine-tuning and enhancement module; The data processing module preprocesses the original data obtained by the electric energy metering and acquisition equipment, including removing invalid data, filling in missing data, extracting electrical characteristic indicators, dividing the data into samples by month, and generating a training set, a validation set, and a test set for the time-domain convolutional gated recurrent unit fusion neural network model after normalization processing; The fusion neural network construction module constructs a fusion neural network architecture composed of a time-domain convolutional network, a gated recurrent unit network, and a feedforward fully connected layer, extracts local and global features of the time series through the time-domain convolutional gated recurrent unit fusion neural network, and completes the classification prediction of the life of the electric energy metering and acquisition equipment; The time-domain convolutional gated recurrent unit fusion neural network model training control module sets the parameter configuration for training the time-domain convolutional gated recurrent unit fusion neural network model, including a tensor converter, a data loader, a cross-entropy loss function, an Adam optimizer, a learning rate scheduling strategy, and sets the maximum number of training epochs and early termination conditions; The time-domain convolutional gated recurrent unit fusion neural network model training and evaluation module executes the forward propagation, backward propagation, and training loop processes of the time-domain convolutional gated recurrent unit fusion neural network model, calculates the weight gradient through the loss function, optimizes the parameters of the time-domain convolutional gated recurrent unit fusion neural network model, and evaluates the performance of the time-domain convolutional gated recurrent unit fusion neural network model on the validation set to determine the optimal weights of the time-domain convolutional gated recurrent unit fusion neural network model; The test and prediction module of the time-domain convolutional gated recurrent unit fusion neural network model predicts the remaining life of the device on the test set and comprehensively evaluates the performance of the time-domain convolutional gated recurrent unit fusion neural network model through k-fold cross-validation; The fine-tuning and enhancement module fine-tunes when the performance of the time-domain convolutional gated recurrent unit fusion neural network model is poor, freezes the weights of some network layers, adjusts the inactivation rate, modifies the learning rate scheduling strategy, and optimizes the generalization ability of the time-domain convolutional gated recurrent unit fusion neural network model by expanding small-sample categories and adding noise samples.

[0013] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network are implemented.

[0014] A computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, the steps of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network are implemented.

[0015] Advantages of the present invention: The present invention uses big data processing and deep learning technical solutions to solve the problem of the health status management of electric energy metering and acquisition devices, achieves the purpose of automatically detecting the health status of electric energy metering and acquisition devices, and has the effects of changing from replacing according to years to replacing according to the device status of electric energy metering and acquisition devices, saving device maintenance costs, and promoting the development of smart grids. Description of the Drawings

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for description in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0017] Figure 1 It is the overall flowchart of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network provided by an embodiment of the present invention.

[0018] Figure 2 It is the architecture diagram of the time-domain convolutional gated recurrent unit fusion neural network model TGFN of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network provided by an embodiment of the present invention.

[0019] Figure 3 It is the structural diagram of the gated recurrent unit network of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network provided by an embodiment of the present invention.

[0020] Figure 4 This is a multi - layer gated recurrent unit network diagram of the remaining life prediction system method for power metering and acquisition equipment based on a fusion neural network provided by an embodiment of the present invention.

[0021] Figure 5 This is a schematic diagram of the feed - forward neural full - connection layer of the remaining life prediction method for power metering and acquisition equipment based on a fusion neural network provided by an embodiment of the present invention. Detailed implementation manners

[0022] To make the above - mentioned objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] Example 1, referring to Figures 1 to 5 , which is the first embodiment of the present invention. This embodiment provides a remaining life prediction method for power metering and acquisition equipment based on a fusion neural network, including: Taking the historical operation and maintenance data of power metering and acquisition equipment of some users in a certain county - level area as an example to elaborate this method. The data of this example comes from the historical operation and maintenance data of power metering and acquisition equipment of low - voltage users and special - transformer users in the county - level area from June 1, 2023 to May 31, 2024. After pre - processing and normalization, the sample order in the data set is disrupted and the entire data set is divided into 12 subsets. Under the integrated development environment of python3.10.9, the time - domain convolutional gated recurrent unit fusion neural network TGFN model is built using the deep - learning framework pytorch2.4.1.

[0024] In one embodiment, it includes the following steps: S11. Perform data pre - processing on the historical power consumption data of users collected by the power metering and acquisition equipment. The original collected data includes three - phase voltage, three - phase current, three - phase power, total power, power factor, daily power consumption, and the acquisition time interval is one year; as well as the installation date and fault records of the power metering and acquisition equipment. Among them, low - voltage users only have daily power consumption data without other items. Analyze and sort out these data, filter out invalid data, reasonably fill in a small amount of missing data, calculate the relevant electrical indexes of special - transformer user power consumption data: voltage volatility, current volatility, current reverse polarity, power difference rate, and calculate the performance index of power metering and acquisition equipment for special - transformer users and low - voltage users: metering error. Divide the data of each power metering and acquisition equipment into 12 samples per month and assign remaining life labels to obtain a data set; S12. Perform max - min normalization on each data sample in the dataset described in S1. The normalization calculation formula is: , where is the value after normalization, is the value before normalization, is the minimum value of the sample data column where is the maximum value of the sample data column where

[0025] Divide the normalized dataset into 12 subsets, including 10 training sets, 1 validation set, and 1 test set for the time - domain convolutional gated recurrent unit fusion neural network model.

[0026] S21. Construct the input layer of the time - domain convolutional network as Figure 2 shown. It stipulates that the number of samples in a batch of the input data is the batch size Batch_Size = 64, and the dimensionality size of a sample data in this batch is , where t = 672 represents that a sample has a 96 - time - step sequence representing the historical operation and maintenance data of the device for a one - week time interval, and m = 21 represents that each time step has 21 feature dimensions. Then the dimensionality size of the input data X in a batch is .

[0027] S22. Construct the components of the time - domain convolutional network, as Figure 2 shown: The time - series convolutional unit TCU. This unit contains 4 component layers, and their structural order is as follows: 1. One - dimensional dilated causal convolutional layer dilated_conv, 2. Weight parameterization layer weight_norm, 3. Relu function activation layer, 4. Dropout neuron random inactivation layer. Among them, the input feature dimensions of the dilated causal one - dimensional convolutional layer dilated_conv are m=[m1,m2,m3]=[21,21,21], the output feature dimensions are h=[h1,h2,h3]=[21,21,21], the dilation factors are dilation=[d1,d2,d3]=[1,2,4], the kernel size is kernel_size = 3, and the padding size value of the one - dimensional dilated causal convolutional layer dilated_conv is: .

[0028] The expression form of the internal calculation of the convolution kernel in the one - dimensional dilated causal convolutional layer dilated_conv is as follows: , where is the (n - dk)-th element of the x sequence, is the K-th element of the convolutional kernel, d is the dilation rate, when there will be d - 1 intervals between the elements of the convolutional kernel.

[0029] The weight parameterization weight_norm rewrites the weight matrix in the form of: , where is a tensor with the same dimension as the weight matrix g is a scalar, is the direction of the weight after normalizing .

[0030] The expression of the Relu activation function is: , where A is the input of the Rule function.

[0031] The following hyperparameter configurations of the temporal convolutional unit TCU are shown in Table 1: Table 1 Hyperparameter Configuration Table , S23. Construct the main body of the temporal convolutional network TCN, as Figure 2 shown. The main body TCN is composed of temporal convolutional units TCU1, TCU2, and TCU3, which are successively composed of dilated causal convolutional layers with dilation factor parameters d1 = 1, d2 = 2, and d3 = 4.

[0032] S24. Construct the skip connection structure of the temporal convolutional network as Figure 2 shown. If the dimension of the data X in the input layer matches the dimension of the output data Y3 of the temporal convolutional unit TCU3, the finally output feature vector of the temporal convolutional network TCN is: , where Y_TCN is the finally output vector matrix of TCN, X is the incoming data matrix of the input layer of the temporal convolutional network described in S21, Y3 is the output data matrix of the temporal convolutional unit TCU3, and + is the operation of adding the numerical values of the corresponding elements of the two matrices.

[0033] If the dimension of the data X in the input layer does not match the dimension Y3 of the output data Y3 of the temporal convolutional unit TCU3, the output feature vector of the temporal convolutional network is: , where is the data X in the input layer after passing through The convolution kernel Conv_residual() is mapped to the output data dimension of the temporal convolutional unit TCU3, and then the matrix data value addition operation is performed, and the result is used as the output feature vector of the temporal convolutional network TCN.

[0034] S25. Construct each component of the gated recurrent unit in the gated recurrent unit network body GRU, such as Figure 3 shown, where Component 1 is the reset gate Reset Gate, and its expression form is as follows: , where the current time step is t, and the input at this time step is , is the output result of the reset gate Reset Gate at the current time step, is the hidden state of the previous time step, is the weight parameter matrix of the reset gate, is the Sigmoid activation function, and the reset gate controls how much information of the hidden state of the previous moment can be retained. Among them, Component 2 is the update gate Update Gate, and its expression form is as follows: , where the current time step is t, and the input at this time step is , is the output result of the update gate UpdateGate at the current time step, is the hidden state of the previous time step, is the weight parameter matrix of the reset gate, tanh is the hyperbolic tangent activation function, and the update gate controls the combination degree of the candidate hidden state at the current moment t and the hidden state of the previous moment.

[0035] The candidate hidden state of the gated recurrent unit at the current time step , and its expression form is as follows: , where the time step is t, and the input at the current time step is , is the weight matrix of the candidate state, is the output result of the reset gate Reset Gate at the current time step, is the hidden state of the previous time step, ⊙ represents element-wise multiplication, is the hyperbolic tangent activation function. Use the hidden state controlled by the reset gate to calculate the candidate hidden state , which contains the current input and the hidden state of the previous moment processed by the reset gate Combined information

[0036] The output result of the gated recurrent unit at the current time step is equal to the final hidden state at the current time step , and its expression form is as follows: , where is the output result of the update gate Update Gate at the current time step, ⊙ represents element-wise multiplication, is the hidden state at the previous time step, is the candidate hidden state at the current time step

[0037] S26. Construct a gated recurrent unit, which is composed in the following structural order, as Figure 3 shown: 1. The update gate Update Gate and the reset gate Reset Gate are juxtaposed, 2. The candidate hidden state , 3. The final hidden state ; S27. Construct the main body of the gated recurrent unit network GRU main body, as Figure 4 shown. It is stipulated that the input data dimension of the gated recurrent unit network main body GRU is the same as the feature vector Y_TCN dimension described in S24, its feature dimension is m = 21, the time step sequence length is t = 672, and the feature dimension of the output data vector of the gated recurrent unit network main body GRU is h4 = 32, the time step sequence length is t = 672. Take the output of the last time step of the last gated recurrent unit of the GRU network as the final output Y_GRU of the GRU network, and the number of layers of the gated recurrent unit network main body GRU is num_layers = 3

[0038] S28. Construct a feedforward neural fully connected layer FN, as Figure 5 shown. It contains 3 layers of feedforward fully connected layers FC1, FC2, FC3. The number of neurons in each fully connected layer is N1 = 64, N2 = 128, and N3 is determined by the total number of categories. The first and second feedforward fully connected layers are activated with the Relu function, while the third layer uses the Softmax function for activation. The internal expression form of each feedforward full connection is as follows: , where is the output of the i-th feedforward fully connected layer, is the input of the i-th feedforward neuron, is the weight matrix of the i-th fully connected layer, is the bias term of the i-th layer

[0039] The expression of the Relu function is: , where A is the input of the Rule function; The Softmax function is defined as: for the input vector , the output of Softmax is a probability distribution , where each represents the probability of the i-th class prediction, and the calculation formula is as follows: , where is the sum of the exponentiated values of all inputs.

[0040] The structure order and hyperparameters of the feedforward fully connected layer are shown in Table 2: Table 2 Structure Order and Hyperparameter Information Table of the Feedforward Fully Connected Layer , S31. Designate the data converter and data loader of the time-domain convolutional gated recurrent unit fusion neural network TGFN in S2. The data converter converts the data normalized in S12 into a tensor matrix form convenient for parallel computing , and the data loader loads the data converted into a tensor matrix by shuffling the sample order in the dataset with a sample batch size of Bach_Size = 64 ; S32. Designate the loss function of the time-domain convolutional gated recurrent unit fusion neural network TGFN combined in S2 as the cross-entropy loss function Categorical-Cross-Entropy Loss, and its definition is as follows: The true label of the sample is a class index , C is the total number of classes, and the time-domain convolutional gated recurrent unit fusion neural network model outputs the probability distribution of each class. The calculation formula of the cross-entropy loss is:

[0041] , where is the predicted probability of the time-domain convolutional gated recurrent unit fusion neural network model for each class, is the one-hot encoded vector of the true label, indicates that the sample belongs to class c, otherwise ; is the predicted probability of the time-domain convolutional gated recurrent unit fusion neural network model for class c. For each sample, the contribution of the probability terms of other classes is 0, and only the corresponding probability term of the true class will be calculated.

[0042] The optimizer of the time-domain convolutional gated recurrent unit fusion neural network TGFN is specified as adaptive momentum estimation (Adam). Its mathematical expression process is as follows: Derivative the weights and calculate the gradient of the weights. The calculation formula is as follows: , in, is the weight, is the weight at time t, is the gradient at time t; Computing Gradients The current first moment (mean) of , the calculation formula is as follows: , in, is the first-order moment at time t, is the decay rate that controls the influence of historical gradients, is the first-order moment at time t, is the first-order moment at time t-1.

[0043] Computing Gradients The second moment (uncentered variance) of , the calculation formula is as follows: , in, is the decay rate of the historical influence of the control variance, is the second-order moment at time t, is the second-order moment at time t-1.

[0044] For the current gradient The first-order moment and second-order moment are used to correct the deviation. The correction formula is as follows: , Among them, t is the current time t, , is the estimated value of the first-order moment after correction at time t and the estimated value of the second-order moment after correction at time t, They are respectively the decay rate of the control history gradient influence at time t and the decay rate of the control variance history influence at time t.

[0045] Update the weight parameters. The calculation formula for updating the weight parameters is as follows: , in, is a small constant, usually set to , used to prevent division by zero, is the updated weight parameter at time t-1.

[0046] The learning rate adjustment strategy of the learning rate scheduler is as follows: when the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model on the validation set has not increased after continuously completing P = 4 training epochs, the learning rate decays to: , where is the learning rate before decay, is the learning rate after decay, is the decay coefficient.

[0047] S33. Set the hyperparameters of the time-domain convolutional gated recurrent unit fusion neural network model: the size of the input tensor of the time-domain convolutional gated recurrent unit fusion neural network model is the dimension size of the input data of one batch described in S21 ( ), the dimension of the output data is the total number of categories h6 for task classification, the maximum number of training epochs for the time-domain convolutional gated recurrent unit fusion neural network model is max_epoches, and the condition for triggering the early termination of the training of the time-domain convolutional gated recurrent unit fusion neural network model is: when the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model on the validation set has not increased after continuously completing 4P = 16 training epochs, the training is terminated early and the model weights of the epoch with the best historical performance of the time-domain convolutional gated recurrent unit fusion neural network model are saved.

[0048] S41. Construct the forward propagation process of a sample in the time-domain convolutional gated recurrent unit fusion neural network TGFN: The first step: The training set data described in S1 is read by the data loader to become the input data X and is passed into the input layer of the time-domain convolutional network TCN described in S21, and then the input data is sequentially passed into the time-domain convolutional network to extract the feature vector Y1.

[0049] The second step: The extracted feature vector Y1 is used as the input vector X2 and passed into the gated recurrent unit network main body GRU described in S27 to further capture the long-term dependence relationship of the time series, and the deep feature vector Y2 is obtained.

[0050] The third step: The obtained deep feature vector Y2 is used as the input X3 and passed into the feedforward neural fully connected layer FC to obtain the non-linear mapping vector Y3.

[0051] The fourth step: The non-linear mapping vector Y3 is passed into the Softmax classifier to calculate the probability that each sample in the current batch belongs to each category, and the category with the largest probability is selected as the final predicted category of the sample.

[0052] S42. Construct the backpropagation process of a sample in the time-domain convolutional gated recurrent unit fusion neural network TGFN: First step, calculate the cross-entropy loss between the predicted result label and the true label according to the loss function Categorical Cross-Entropy Loss.

[0053] Second step, calculate the weight parameter gradients of each layer of neurons in the time-domain convolutional gated recurrent unit fusion neural network TGFN.

[0054] Third step, update the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model through the optimizer Adam, and clear the previously calculated gradients after each weight parameter update step of the time-domain convolutional gated recurrent unit fusion neural network model.

[0055] S43. Construct the training loop of the time-domain convolutional gated recurrent unit fusion neural network TGFN. One training loop is as follows: Each time, pass a batch of data through the forward propagation process described in S41 and the backpropagation process described in S42 until all the training set data described in S12 is passed, then a training loop is completed.

[0056] S51. Repeatedly execute the training loop described in S43 on the training set. After each training loop is completed, pass the data of the validation set into the time-domain convolutional gated recurrent unit fusion neural network TGFN with the weight parameters trained in the current round to predict the life labels of the electricity meter samples on the validation set, calculate the loss according to the cross-entropy loss function described in S32, and calculate the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model. Improve the learning rate by the learning rate scheduler described in S32 but do not perform backpropagation. The formula for calculating the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model is as follows: , where Accuracy is the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model, represents the indicator function, which is 1 when the prediction is correct and 0 otherwise, and N is the total number of samples.

[0057] S52. Specify the maximum number of epochs of the training loop as num_epoches, and continuously repeat the loop process described in S51 until the multi-class cross-entropy loss converges, or reaches the maximum number of training loop epochs, or meets the conditions for early termination of training. Then complete the training of the time-domain convolutional gated recurrent unit fusion neural network model and save the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model with the highest classification accuracy in each training epoch as the final weight parameters of the time-domain convolutional gated recurrent unit fusion neural network TGFN.

[0058] S61. Use the test set data described in S12 to predict the remaining life of the power metering acquisition device in the trained time-domain convolutional gated recurrent unit fusion neural network model TGFN described in S52, and use cross-validation to comprehensively evaluate the performance of the time-domain convolutional gated recurrent unit fusion neural network model. The steps of its cross-validation are as follows: (1) Divide the data set into k = 12 subsets of equal size.

[0059] (2) Each time, select one subset as the test set and one subset as the validation set. The selected test set and validation set are not repeated, and the remaining k - 2 = 10 subsets are used as the training set.

[0060] (3) Train the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model on the training set, and evaluate the performance of the time-domain convolutional gated recurrent unit fusion neural network model on the test set.

[0061] (4) Repeat steps (2) and (3) for a total of k = 12 times.

[0062] (5) Take the average of the evaluation results of the time-domain convolutional gated recurrent unit fusion neural network model k times as the final performance evaluation of the time-domain convolutional gated recurrent unit fusion neural network model.

[0063] S62. Build a fine-tuning mechanism for the time-domain convolutional gated recurrent unit fusion neural network model: If the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model on the test set or the validation set is much lower than that on the training set, first freeze the weight parameters of the 3 one-dimensional dilated convolutional layers dilated_conv in the time-domain convolutional network, and then adjust the proportion of randomly inactivated neurons in the Dropout neuron random inactivation layer described in S22. The adjustment amplitude each time is ; Use the L2 regularization strategy in terms of the loss function. Its mathematical expression is as follows: , where is the loss function before regularization, is the loss function after regularization, is the regularization hyperparameter, is the weight parameter of the time-domain convolutional gated recurrent unit fusion neural network model is the L2 norm; Replace the learning rate scheduler with cosine annealing. Its mathematical expression is as follows: , where t is the number of the current iteration, is the learning rate of the t-th iteration, is the minimum learning rate, is the maximum learning rate is the total number of iterations within one period.

[0064] Expand the category samples with a small proportion, that is, add noise to the samples in the training set: insert a random array with the same dimension as the sample data into the category samples with a small proportion identical random array , and for each individual column element of its m = 21 feature columns, it follows a Gaussian distribution with a mean of and a standard deviation of , where and are calculated according to the category samples with added noise, and the labels of the randomly generated noise samples are assigned, and it is required that the number of generated noise samples is less than the number of samples of the expanded category.

[0065] Embodiment 2 is the second embodiment of the present invention. This embodiment provides a remaining life prediction system for an electric energy metering and acquisition device based on a fusion neural network, including a data processing module, a fusion neural network construction module, a time-domain convolutional gated recurrent unit fusion neural network model training control module, a time-domain convolutional gated recurrent unit fusion neural network model training and evaluation module, a time-domain convolutional gated recurrent unit fusion neural network model testing and prediction module, and a fine-tuning and enhancement module; The data processing module preprocesses the original data obtained by the electric energy metering and acquisition device, including removing invalid data, filling in missing data, extracting electrical feature indicators, dividing the data into samples by month, and generating a training set, a validation set, and a test set for the time-domain convolutional gated recurrent unit fusion neural network model after normalization processing; The fusion neural network construction module constructs a fusion neural network architecture composed of a time-domain convolutional network, a gated recurrent unit network, and a feedforward fully connected layer, extracts local and global features of the time series through the time-domain convolutional gated recurrent unit fusion neural network, and completes the classification prediction of the life of the electric energy metering and acquisition device; The time-domain convolutional gated recurrent unit fusion neural network model training control module sets the parameter configuration for training the time-domain convolutional gated recurrent unit fusion neural network model, including a tensor converter, a data loader, a cross-entropy loss function, an Adam optimizer, a learning rate scheduling strategy, and sets the maximum number of training epochs and the early termination condition; The time-domain convolutional gated recurrent unit fusion neural network model training and evaluation module executes the forward propagation, backward propagation, and training loop processes of the time-domain convolutional gated recurrent unit fusion neural network model, calculates the weight gradient through the loss function, optimizes the parameters of the time-domain convolutional gated recurrent unit fusion neural network model, and evaluates the performance of the time-domain convolutional gated recurrent unit fusion neural network model on the validation set to determine the optimal weights of the time-domain convolutional gated recurrent unit fusion neural network model; The test and prediction module of the time-domain convolutional gated recurrent unit fusion neural network model predicts the remaining useful life of the device on the test set and comprehensively evaluates the performance of the time-domain convolutional gated recurrent unit fusion neural network model through k-fold cross-validation; The fine-tuning and enhancement module fine-tunes when the performance of the time-domain convolutional gated recurrent unit fusion neural network model is poor, freezes the weights of some network layers, adjusts the dropout rate, modifies the learning rate scheduling strategy, and optimizes the generalization ability of the time-domain convolutional gated recurrent unit fusion neural network model by expanding small-sample classes and adding noise samples.

[0066] Example 3 is the third example of the present invention. What is different from the previous two examples is: If the described function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0067] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or used in combination with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0068] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing it in a suitable manner if necessary, and then storing it in a computer memory.

[0069] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, the multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having suitable combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), and the like.

[0070] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network, characterized in that: including Preprocessing the collected data of the power metering and acquisition device to obtain a model data set and the labels corresponding to the data, normalizing the preprocessed data set and dividing it into three parts: a training set, a validation set, and a test set for the time-domain convolutional gated recurrent unit fusion neural network model; Constructing a combined time-domain convolutional gated recurrent unit fusion neural network of a time-domain convolutional network, a gated recurrent unit network, and a feedforward neural fully connected layer, and setting hyperparameters; Specifying a tensor converter and a data loader for training the time-domain convolutional gated recurrent unit fusion neural network, a loss function, an optimizer, and a learning rate scheduler for the time-domain convolutional gated recurrent unit fusion neural network model, setting the maximum number of epochs for training the time-domain convolutional gated recurrent unit fusion neural network model, and the conditions for triggering early termination of the training of the time-domain convolutional gated recurrent unit fusion neural network model; Constructing the forward propagation process, the backward propagation process, and the training loop process of the time-domain convolutional gated recurrent unit fusion neural network; Constructing a process for solving the optimal weights of the combined architecture time-domain convolutional gated recurrent unit fusion neural network; Predicting the remaining life of the power metering and acquisition device on the test set, and constructing a fine-tuning mechanism for the combined architecture time-domain convolutional gated recurrent unit fusion neural network.

2. The method for predicting the remaining life of the electric energy metering and acquisition device based on the fusion neural network according to claim 1, wherein: The preprocessing of the collected data of the power metering and acquisition device to obtain a time-domain convolutional gated recurrent unit fusion neural network model data set and the labels corresponding to the data, normalizing the preprocessed data set and dividing it into three parts: a training set, a validation set, and a test set for the time-domain convolutional gated recurrent unit fusion neural network model includes Performing data preprocessing on the original data collected by the power metering and acquisition device. The original data includes three-phase voltage, three-phase current, three-phase power, total power, power factor, daily power consumption, and the acquisition time interval is one year; as well as the installation date and fault records of the power metering and acquisition device; Analyzing and sorting the collected data, filtering invalid data, and reasonably filling a small amount of missing data; Calculating relevant electrical indicators including voltage volatility, current volatility, current reverse polarity, power difference rate, and calculating the metering error of the power metering and acquisition device; Dividing the data of each power metering and acquisition device into 12 samples per month and assigning remaining life labels to obtain a data set; Normalizing each data sample in the data set and dividing it into three parts: a training set, a validation set, and a test set for the time-domain convolutional gated recurrent unit fusion neural network model.

3. The method for predicting the remaining life of the electric energy metering and acquisition device based on the fusion neural network according to claim 2, wherein: The construction of a combined time-domain convolutional gated recurrent unit fusion neural network of a time-domain convolutional network, a gated recurrent network, and a feedforward neural fully connected layer, and setting hyperparameters includes Constructing the input layer of the time-domain convolutional network and specifying the batch size of the input samples; Constructing the components of the time-domain convolutional network including the temporal convolutional unit TCU; The temporal convolutional unit TCU contains 4 component layers, and the structural order is a one-dimensional dilated causal convolutional layer, a weight parameterization layer, a Relu function activation layer, and a neuron dropout layer; Constructing the main body of the time-domain convolutional network, which is successively composed of temporal convolutional units TCU1, TCU2, and TCU3 composed of dilated causal convolutional layers; Construct the skip connection structure of the time-domain convolutional network, determine whether the dimensions of the input layer data match the dimensions of the output data of the time-sequence convolutional cloud TCU3, and output the final feature vector of the time-domain convolutional network; Construct each component of the gated recurrent unit in the gated recurrent network main body. Component one is the reset gate, component two is the update gate, and they are in a parallel relationship. Component three is the candidate hidden state component, and component four is the final hidden state component; Specify that the input data dimension of the gated recurrent unit network is the same as the dimension of the final feature vector of the time-domain convolutional network, and use the output of the last time step of the last gated recurrent unit in the gated recurrent unit network as the final output of the gated recurrent unit network; Construct a feed-forward neural fully connected layer including three feed-forward fully connected layers FC1, FC2, and FC3. The number of neurons in each fully connected layer is N1, N2, and N3, where N3 is determined by the total number of categories. The first and second feed-forward fully connected layers are activated with the Relu function, and the third layer is activated with the Softmax function.

4. The method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network according to claim 3, wherein: The specified tensor converter, data loader, loss function, optimizer, and learning rate scheduler for training the time-domain convolutional gated recurrent unit fusion neural network. Set the maximum number of training epochs for training the time-domain convolutional gated recurrent unit fusion neural network model and the conditions for triggering early termination of training the time-domain convolutional gated recurrent unit fusion neural network model include, The specified tensor converter and data loader for training the time-domain convolutional gated recurrent unit fusion neural network include converting the normalized data into a tensor matrix form and shuffling the sample order in the dataset and recording the data converted into a tensor matrix; The loss function of the time-domain convolutional gated recurrent unit fusion neural network model is to set the loss function of the time-domain convolutional gated recurrent unit fusion neural network as the cross-entropy loss function; The optimizer is to set the optimizer of the time-domain convolutional gated recurrent unit fusion neural network as adaptive momentum estimation; The learning rate scheduler is to set the learning rate adjustment strategy of the learning rate scheduler of the time-domain convolutional gated recurrent unit fusion neural network as follows: when the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model on the validation set has not increased after continuously completing P training epochs, the learning rate decays to: ， Among them, is the learning rate before attenuation, is the learning rate after attenuation, is the attenuation coefficient; The settings of the maximum number of training epochs for training the time-domain convolutional gated recurrent unit fusion neural network model and the conditions for triggering early termination of training the time-domain convolutional gated recurrent unit fusion neural network model include: when the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model on the validation set has not increased after continuously completing several training epochs, terminate the training early and save the weights of the time-domain convolutional gated recurrent unit fusion neural network model with the best performance in the historical best epoch of the time-domain convolutional gated recurrent unit fusion neural network model; 5. The method for predicting the remaining life of the electric energy metering and acquisition device based on the fusion neural network according to claim 4, characterized in that: The construction of the forward propagation process, backward propagation process, and training loop process of the time-domain convolutional gated recurrent unit fusion neural network includes, Construct the forward propagation process of a sample in a time-domain convolutional gated recurrent unit fusion neural network. The training set data is read by the data loader to become the input data X and is passed into the input layer of the time-domain convolutional network (TCN). The input data is sequentially passed into the time-domain convolutional network to extract the feature vector Y1; The extracted feature vector Y1 is used as the input vector and passed into the main body of the gated recurrent network to further capture the long-term dependencies of the time series, obtaining the deep feature vector Y2. The obtained deep feature vector Y2 is used as the input X3 and passed into the fully connected feedforward neural layer FC to obtain the non-linear mapping vector Y3. The non-linear mapping vector Y3 is passed into the Softmax classifier to calculate the probability that each sample in the current batch belongs to each category, and the category with the highest probability is selected as the final predicted category of the sample; Construct the backward propagation process of a sample in a time-domain convolutional gated recurrent unit fusion neural network. Calculate the cross-entropy loss between the label of the predicted result and the true label according to the loss function, calculate the weight parameter gradients of the neurons in each layer of the time-domain convolutional gated recurrent unit fusion neural network, and update the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model through the Adam optimizer. Clear the previously calculated gradients after each step of updating the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model; The training loop process includes passing a batch of data through the forward propagation process and the backward propagation process each time until the training set data is completely passed, then a training loop is completed.

6. The method for predicting the remaining life of the power metering and acquisition device based on the fusion neural network according to claim 4, characterized in that: The process of constructing and solving the optimal weights of the combined architecture time-domain convolutional gated recurrent unit fusion neural network includes, After each training loop is completed, the data of the validation set is passed into the time-domain convolutional gated recurrent unit fusion neural network with the weight parameters trained in the current round to predict the life label of the electricity meter samples on the validation set. Calculate the loss according to the cross-loss function and calculate the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model. The learning rate is improved by the learning rate scheduler but no backward propagation is performed. The expression of the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model is: ， Among them, Accuracy is the accuracy of classification of the time-domain convolutional gated recurrent unit fusion neural network model. represents the indicator function, which is 1 when the prediction is correct and 0 otherwise, and N is the total number of samples. Continuously repeat the training loop process until the multi-class cross-entropy loss converges, or reaches the maximum number of training loop rounds, or meets the conditions for early termination of training. Then complete the training of the time-domain convolutional gated recurrent unit fusion neural network model and save the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model with the highest classification accuracy in each previous training round as the final weight parameters of the time-domain convolutional gated recurrent unit fusion neural network for the remaining life of the electricity metering acquisition device.

7. The method for predicting the remaining life of the power metering and acquisition device based on the fusion neural network according to claim 6, characterized in that: The prediction of the remaining life of the electricity metering acquisition device on the test set includes, Use the test set data to predict the remaining life of the electricity metering acquisition device in the trained time-domain convolutional gated recurrent unit fusion neural network, and comprehensively evaluate the performance of the time-domain convolutional gated recurrent unit fusion neural network model using cross-validation; Divide the dataset into k subsets of equal size. Each time, select one subset as the test set and one subset as the validation set. The selected test set and validation set are not repeated each time. The remaining k - 2 subsets are used as the training set. Use the training set to train the weight parameters of the time-domain convolutional gated recurrent unit fusion neural network model. Evaluate the performance of the time-domain convolutional gated recurrent unit fusion neural network model on the test set. Repeat the iteration k times. Take the average of the evaluation results of the time-domain convolutional gated recurrent unit fusion neural network model for k times as the final performance evaluation of the time-domain convolutional gated recurrent unit fusion neural network model; The fine-tuning mechanism of the time-domain convolutional gated recurrent unit fusion neural network for building the combined architecture includes, If the classification accuracy of the time-domain convolutional gated recurrent unit fusion neural network model on the test set or validation set is less than 5 percentage points or more than that of the training set, first freeze the weight parameters of 3 one-dimensional dilated convolutional layers in the time-domain convolutional network, then adjust the proportion of randomly inactivated neurons in the neuron dropout layer, and replace the learning rate scheduler with cosine annealing; Balance the number of samples in each category. Add noise to the training set samples. Insert random arrays with the same dimension as the sample data into the category samples with a small number ratio, and assign labels to the randomly generated noise samples. And it is required that the number of generated noise samples is less than the number of samples in the expanded category.

8. A remaining life prediction system for power metering and acquisition equipment based on a fusion neural network, which applies the method for predicting the remaining life of power metering and acquisition equipment based on a fusion neural network according to any one of claims 1 to 7, characterized in that, It includes: Data processing module, fusion neural network construction module, time-domain convolutional gated recurrent unit fusion neural network model training control module, time-domain convolutional gated recurrent unit fusion neural network model training and evaluation module, time-domain convolutional gated recurrent unit fusion neural network model testing and prediction module, fine-tuning and enhancement module; The data processing module preprocesses the raw data obtained by the power metering acquisition device, including removing invalid data, filling in missing data, extracting electrical feature indicators, and dividing the data into samples by month. After normalization, it generates the training set, validation set, and test set of the time-domain convolutional gated recurrent unit fusion neural network model; The fusion neural network construction module constructs a fusion neural network architecture composed of a time-domain convolutional network, a gated recurrent unit network, and a feedforward fully connected layer. Extracts local and global features of the time series through the time-domain convolutional gated recurrent unit fusion neural network to complete the classification prediction of the lifespan of the power metering acquisition device; The time-domain convolutional gated recurrent unit fusion neural network model training control module sets the parameter configuration for training the time-domain convolutional gated recurrent unit fusion neural network model, including tensor converter, data loader, cross-entropy loss function, Adam optimizer, learning rate scheduling strategy, and sets the maximum number of training epochs and early termination conditions; The training and evaluation module of the time-domain convolutional gated recurrent unit fusion neural network model executes the forward propagation, backpropagation, and training loop processes of the time-domain convolutional gated recurrent unit fusion neural network model, calculates the weight gradients through the loss function, optimizes the parameters of the time-domain convolutional gated recurrent unit fusion neural network model, and evaluates the performance of the time-domain convolutional gated recurrent unit fusion neural network model on the validation set to determine the optimal weights of the time-domain convolutional gated recurrent unit fusion neural network model; The testing and prediction module of the time-domain convolutional gated recurrent unit fusion neural network model predicts the remaining life of the device on the test set and comprehensively evaluates the performance of the time-domain convolutional gated recurrent unit fusion neural network model through k-fold cross-validation; The fine-tuning and enhancement module fine-tunes when the performance of the time-domain convolutional gated recurrent unit fusion neural network is poor, freezes the weights of some network layers, adjusts the inactivation rate, modifies the learning rate scheduling strategy, and optimizes the generalization ability of the time-domain convolutional gated recurrent unit fusion neural network model by expanding small sample classes and adding noise samples.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for predicting the remaining life of an electric energy metering and acquisition device based on a fusion neural network according to any one of claims 1 to 7.

Images