Charging pile fault prediction method based on improved convolutional long short-term memory network

Abstract

The invention discloses a charging pile fault prediction method based on an improved convolutional long short-term memory network, and the method comprises the following steps: S1, collecting charging data, fault maintenance data and charging feedback data, cleaning the collected data, and generating a time sequence; s2, feature extraction and mode recognition are carried out on the time series data of fault prediction of the charging pile through a convolutional neural network; s3, constructing a convolutional long-short-term memory network model, and outputting probability distribution of each category; s4, optimizing the convolutional layer by using an adaptive momentum optimization algorithm; and S5, predicting fault data in the complex time sequence data of the charging pile. By improving the CNN and LSTM models and combining the Adam, the convergence speed and stability of the model are improved, and high-precision and quick-response charging pile fault prediction is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of charging pile fault diagnosis, and particularly to a charging pile fault prediction method based on an improved convolutional long short-term memory network. Background Art

[0002] With the rapid development of electric vehicles, as the infrastructure for electric vehicle charging, the number and types of charging piles are also increasing continuously. The fault problems of charging piles, especially the operation faults in high-load environments, have become important factors affecting the safety, reliability and user experience of charging piles. Charging pile faults not only cause charging interruptions for users, but may also affect the stability of the power grid, cause equipment damage, thereby increasing maintenance costs and reducing operation efficiency. Therefore, charging pile fault prediction and diagnosis have become a key issue in the management of electric vehicle charging infrastructure. The existing charging pile fault detection methods mainly rely on manual inspections and regular checks. These traditional detection means not only have problems such as slow response, limited coverage, and high labor costs, but also often cannot detect potential faults in advance, resulting in significant losses when faults occur.

[0003] In recent years, with the development of sensing technology, data such as voltage, current, and temperature generated during the operation of charging piles have been widely collected. The accumulation of this data provides a basis for data-driven fault prediction methods. By analyzing historical operation data, potential patterns of charging pile faults can be discovered. However, the existing data-based fault prediction methods still face some challenges. Traditional machine learning methods (such as decision trees, support vector machines, etc.) often show low prediction accuracy and generalization ability when dealing with large-scale, high-dimensional, and time-series operation data of charging piles. Deep learning methods, especially convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), show great potential in time-series data analysis and fault prediction, but due to their slow convergence speed, poor real-time performance, and low prediction accuracy. Summary of the Invention

[0004] The purpose of this part is to outline some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this part, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this part, the abstract, and the title, but such simplifications or omissions cannot be used to limit the scope of the present invention.

[0005] In view of the above and / or problems existing in the existing phase-shift control technology of dual active bridge circuits, the present invention is proposed.

[0006] Therefore, the object of the present invention is to provide a charging pile fault prediction method based on an improved convolutional long short-term memory network, which improves the convergence speed and stability of the model by improving the CNN and LSTM models and combining with Adam, and realizes high-precision and fast-response charging pile fault prediction.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: A charging pile fault prediction method based on an improved convolutional long short-term memory network, comprising the following steps:

[0008] S1. Collect charging data, fault repair data and charging feedback data, clean the collected data, and generate a time series;

[0009] S2. Extract features and recognize patterns from the time series data of charging pile fault prediction through a convolutional neural network;

[0010] S3. Build a convolutional long short-term memory network model and output the probability distribution of each category;

[0011] S4. Use the adaptive momentum optimization algorithm to optimize the convolutional layer;

[0012] S5. Predict the fault data in the complex time series data of the charging pile.

[0013] As a preferred solution of the charging pile fault prediction method based on the improved convolutional long short-term memory network in the present invention, wherein: the step S2 specifically includes,

[0014] S201. Given the input data Perform weighted summation on the input data X through the sliding convolutional kernel W to obtain the output feature map X';

[0015]

[0016] where B is the batch size, C in is the number of input channels, H and W are the height and width respectively, W i，m，n is the sliding convolutional kernel, X i，m，n is the input data, and the output X' of the convolution operation is obtained through the convolutional kernel and the bias where K out is the number of output channels, k H and k W are the height and width of the convolutional kernel respectively;

[0017] S202. Perform batch normalization processing on each feature map,

[0018]

[0019]

[0020] γ c is the scaling factor, β c is the offset factor, ∈ is a constant to prevent division-by-zero errors, H’ is the height, W’ is the width, X′ i，c，h，w is the variance of the c-th channel, X′ i，c，h，w is the output after the convolution operation, X″ i，c，h，w is the output after batch normalization, standardization, and scaling;

[0021] S203. Process each element of the input data through the ReLU activation function, X ReLU = max(0, X″), X ReLU is the element processed by the ReLU function;

[0022] S204. After passing through the activation layer, apply the pooling operation to each feature map. The size of the pooling window is k H ×k W , the stride is s. For each position in the input feature map, the pooling operation selects the maximum value within the window

[0023] where, X ReLU is the feature map after ReLU activation, window is the sliding window area of the pooling operation, i.e., k H ×k W ; is the output feature map after pooling; represents taking the maximum value within the window range.

[0024] As a preferred scheme of the charging pile fault prediction method based on the improved convolutional long short-term memory network in the present invention, wherein: the step S3 specifically includes,

[0025] S301. Convert the output in step S2 from a four-dimensional tensor to a three-dimensional tensor,

[0026]

[0027] X LSTM，b，t，f is the b-th sample, the t-th time step, and the f-th feature dimension of the flattened data; is the calculated row position after pooling corresponding to the time step t, and t mod W′ is the calculated column position after pooling corresponding to the time step t;

[0028] S302. Use the three-dimensional tensor as the input data of the LSTM, and sequentially calculate the output of the forget gate, the output of the input gate, and the cell state i tand the hidden state h t ;

[0029] S303. Use the three - dimensional vector output by the LSTM as the input of the fully - connected layer and perform a linear transformation, then apply the ReLU activation function to set the negative values of the input to zero and retain the positive values. Add a Dropout layer after the fully - connected layer to randomly discard the outputs of some neurons; after passing through the SoftMax function, convert the scores of the output layer into the probability values Y of each category output,c ,

[0030]

[0031] where Y output,c is the predicted probability of the c - th category, representing the probability that the sample belongs to category c; Y FC，c is the output score of the c - th category, coming from the fully - connected layer; C out is the total number of categories, is the sum of the exponential functions of the scores of all categories.

[0032] As a preferred solution of the charging pile fault prediction method based on the improved convolutional long - short - term memory network in the present invention, where: the output f of the forget gate t is, f t =σ(W f ·[h t-1 , X LSTM，b，t，f +b f ), H LSTM is the hidden layer size of the LSTM, is the weight matrix of the forget gate, is the bias term, and σ is the Sigmoid activation function with an output range of [0, 1].[[]]

[0033] As a preferred solution of the charging pile fault prediction method based on the improved convolutional long - short - term memory network in the present invention, where: the output i of the input gate t is, i t =σ(W i ·[h t-1 , X LSTM，b，t，f +b i ;

[0034] where, is the output of the input gate, is the weight matrix of the input gate, is the bias term, h t-1 is the hidden state of the previous time step, storing information from past time steps;

[0035] Generate the candidate cell state

[0036] is the weight matrix, is the bias term.

[0037] As a preferred solution of the charging pile fault prediction method based on the improved convolutional long short-term memory network in the present invention, where: the updated cell state c t is

[0038]

[0039] is the cell state at the current moment, is the cell state at the previous moment.

[0040] As a preferred solution of the charging pile fault prediction method based on the improved convolutional long short-term memory network in the present invention, where: the output o of the output gate t is o t =σ(W o ·[h t-1 , X LSTM，b，t，f +b o );

[0041] is the weight matrix of the output gate, is the bias term;

[0042] The final output h t is h t =o t ⊙tanh(c t );

[0043] where, is the LSTM output at the current moment, and tanh is the hyperbolic tangent function.

[0044] As a preferred solution of the charging pile fault prediction method based on the improved convolutional long short-term memory network in the present invention, where: step S4 is specifically

[0045] using the cross-entropy loss function to measure the difference between the model output Y output，c and the true label Y true ;

[0046] calculating the gradient of the loss function for all parameters θ of the convolutional neural network, long short-term memory network and fully connected layer;

[0047] using Adam to update the model parameters θ.

[0048] As a preferred solution of the charging pile fault prediction method based on the improved convolutional long short-term memory network in the present invention, where: the formula for the cross-entropy loss L is

[0049] Y ture，c is the one - hot encoding of the true category;

[0050] The gradient of the loss function is,

[0051]

[0052] θ ∈ {W CNN , b CNN , W h , W x , W FC , b FC};

[0053] When updating the model parameters, first is the first - order momentum update,

[0054]

[0055] β1 is the momentum decay coefficient, and m t is the first - order momentum;

[0056] Then perform the second - order moment variance update,

[0057] β2 is the variance decay coefficient, and v t is the updated second - order matrix variance;

[0058] Perform bias correction,

[0059]

[0060] Finally, perform parameter update,

[0061]

[0062] α is the learning rate, is the corrected first - order momentum, is the corrected second - order matrix variance.

[0063] Through the improved CNN - LSTM model, combining automatic feature extraction and pattern recognition, the present invention can accurately capture potential fault signals from the time - series data of charging piles, significantly improving the accuracy of fault prediction; adopting the adaptive momentum optimization algorithm (AdamW) to optimize the learning process of the convolutional layer, improving the convergence speed and training stability of the model, thus ensuring the fast response and stable operation of the fault prediction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings. Among them:

[0065] Figure 1 is the flow chart of the present invention.

[0066] Figure 2 is the confusion matrix diagram of the present invention (true is the actual label, predicted is the predicted label), showing the matching situation between the categories predicted by the model and the true categories, and observing true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN).

[0067] Figure 3 is the graph of the accuracy rate changing with the number of iterations during the training process (Training Accuracy is the accuracy rate before training, validation Accuracy is the accuracy rate after verification, and Epochs is the number of iterations).

[0068] Figure 4 The flow chart of the present invention. Detailed implementation manners

[0069] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific implementation manners of the present invention in conjunction with the accompanying drawings of the specification.

[0070] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0071] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments.

[0072] Embodiment 1

[0073] Referring to Figure 1 and Figure 4 , the embodiment of the present invention provides a charging pile fault prediction method based on an improved convolutional long short-term memory network, including the following steps:

[0074] S1. Collect charging data, fault repair data, and charging feedback data, clean the collected data, and generate a time series;

[0075] When collecting data, determine the data sources: Real-time collect data through the internal sensors and control modules of the charging pile (this is an existing technology and not an improvement point of this application), including core parameters such as voltage, current, power, and temperature. Obtain information such as repair time, fault type, and operator remarks from the fault repair records, and collect charging feedback data submitted by users, including charging interruption, payment exception, etc.;

[0076] Use the EC800K-CN Internet of Things module to upload the operation data of the charging pile to the cloud in real time through the sensor network, so as to collect real-time data. Regularly export the bill data and device operation logs from the charging pile management system to batch collect data.

[0077] During the data transmission process, data loss may occur. For missing values, methods such as mean filling, interpolation, and deleting incomplete records can be used for processing (this is an existing technology). Secondly, perform outlier detection. By setting a reasonable threshold range, identify and process abnormal data points to ensure data accuracy. For example, abnormal fluctuations in voltage or current may indicate data collection errors or equipment failures; finally, remove duplicate data, check and delete duplicate records to avoid the impact of data redundancy on the analysis results.

[0078] Preprocess the data to ensure that all data uses a unified format and unit; then perform normalization processing: Perform normalization or standardization processing on numerical data to make it meet the model input requirements and avoid biases caused by different dimensions.

[0079] The generation of the time series data set is specifically as follows. Time window division: Use the sliding window method to divide the continuous time series data into segments of a fixed length to capture the dynamic change characteristics of the data; then perform feature extraction: Extract key features from the time series, and obtain frequency domain features through Fourier transform and wavelet transform to enrich the input information of the model; finally generate labels: According to the time point when the fault occurs, label the time series data for the training of the supervised learning model.

[0080] S2. Use a convolutional neural network to perform feature extraction and pattern recognition on the time series data for charging pile fault prediction, that is, the input data is processed through multiple convolutional layers, and each convolutional layer is followed by a batch normalization layer, a ReLU activation layer, and a pooling layer, which are used to gradually extract and enhance the spatial features in the data. Specifically,

[0081] S201. Given the input data By performing weighted summation of the sliding convolutional kernel W and the input data X, the output feature map X' is obtained.

[0082]

[0083] Among them, B is the batch size, and C in is the number of input channels, H and w are the height and width respectively, and W i，m，n is the first-order moment momentum, X i，m，n is the input data, and the output X' of the convolution operation is obtained through the convolutional kernel and the bias where K out is the number of output channels, k H and k W are the height and width of the convolutional kernel respectively;

[0084] S202. Perform batch normalization processing on each feature map.

[0085]

[0086]

[0087] γ c is the scaling factor, β c is the offset factor, ∈ is a constant to prevent division by zero error, H' is the height, W' is the width, and X' i，c，h，w is, is the variance of the c-th channel, X' i，c，h，w is the output after the convolution operation, and X″ i，c，h，w is the output after batch normalization, standardization, and scaling;

[0088] S203. Process each element of the input data through the ReLU activation function, X ReLU = max(0, X″), and X ReLU is the element processed by the ReLU function. C out is the number of channels of the feature map output during batch normalization;

[0089] Input: The output X″ obtained through the convolutional layer and the batch normalization layer;

[0090] Output: The ReLU function will process each element of the input data, convert negative values to zero, and keep positive values unchanged. All values less than zero are "truncated" to zero, thus introducing non-linearity;

[0091] S204. After the input feature map X ReLU passes through the activation layer, the pooling operation applies a sliding window to each feature map, and the size of the pooling window is k H ×kW , with a stride of s, for each position in the input feature map, the pooling operation selects the maximum value within the window

[0092] where X ReLU is the feature map after ReLU activation, and window is the sliding window area of the pooling operation, i.e., k H ×k W ; is the output feature map after pooling; denotes taking the maximum value within the window range;

[0093] S3. Build a convolutional long short-term memory network model to output the probability distribution of each class. Specifically,

[0094] S301. The LSTM layer requires the input data to be three-dimensional where T is the number of time steps, and F is the feature dimension of each time step. To adapt the output X of the CNN pool to the input of the LSTM, the output in step S2 is transformed from a four-dimensional tensor to a three-dimensional tensor. For each sample b ∈ [1, B], the corresponding C out ×H′×W′ feature map is flattened into a T×F tensor. Each time step t corresponds to a feature vector after flattening, where t corresponds to the spatial position after pooling, i.e., t = h′×w′ (for each position after pooling). The formula for flattening the b-th sample is is to calculate the row position after pooling corresponding to time step t, and t mod W′ is to calculate the column position after pooling corresponding to time step t;

[0095] Through the above process, the output of the CNN is flattened into a format suitable for the input of the LSTM as the input data of the LSTM.

[0096] S302. Use the three-dimensional tensor as the input data of the LSTM, and sequentially calculate the output f of the forget gate t , the output of the input gate, the cell state c t and the output o of the output gate t . The output f of the forget gate t is, f t = σ(W f ·[h t-1 , X LSTM，b，t，f +b f ), H LSTM is the hidden layer size of the LSTM, is the weight matrix of the forget gate, is the bias term, σ is the Sigmoid activation function, and the output range is [0, 1].

[0097] The output i of the input gate t is, i t = σ(W i · [h t-1 , X LSTM，b，t，f + b i ;

[0098] Among them, is the output of the input gate, is the weight matrix of the input gate, is the bias term;

[0099] Generate a candidate cell state

[0100] is the weight matrix, is the bias term;

[0101] The updated cell state c t is,

[0102]

[0103] is the cell state at the current moment, is the cell state at the previous moment;

[0104] Determine which parts of the current cell state c t are used for output and generate the output o of the output gate at the current moment, the hidden state t is, ot = σ(W o · [h t-1 , X LSTM,b,t,f + b o );

[0105] is the weight matrix of the output gate, is the bias term;

[0106] The final output h t is, h t = o t ⊙ tanh(c t );

[0107] Among them, is the LSTM output at the current moment, tanh is the hyperbolic tangent function, which is used to perform a non-linear transformation on the cell state;

[0108] S303. Use the three-dimensional vector output by the LSTM as the input of the fully connected layer and perform a linear transformation, and then apply the ReLU activation function to set the negative values of the input to zero and retain the positive values to introduce non-linearity. To prevent overfitting, add a Dropout layer after the fully connected layer to randomly discard the outputs of some neurons; set the outputs of some neurons to zero to prevent the model from over-relying on certain features, thereby improving the generalization ability of the model. Finally, through the SoftMax function, convert the scores of the output layer into the probability values Y of each category. output,c ,

[0109]

[0110] Among them, Y output,c is the predicted probability of the c-th category, indicating the probability that the sample belongs to category c; Y FC,c is the output score of the c-th category, coming from the fully connected layer; C out is the total number of categories, is the exponential function sum of the scores of all categories, ensuring that the sum of the output probabilities is 1.

[0111] S4. Use the adaptive momentum optimization algorithm to optimize the convolutional layer. Specifically,

[0112] Use the cross-entropy loss function to measure the difference between the model output Y output,c and the true label Y true ;

[0113] Calculate the gradient of the loss function for all parameters θ of the convolutional neural network, long short-term memory network, and fully connected layer;

[0114] Use Adam to update the model parameters θ.

[0115] The formula for the cross-entropy loss L is,

[0116] Y ture，c is the one-hot encoding of the true category;

[0117] The gradient of the loss function is,

[0118]

[0119] θ ∈ {W CNN , b CNN , W h , W x , W FC , b FC};

[0120] When updating the model parameters, first perform the first-order moment momentum update,

[0121]

[0122] β1 is the momentum decay coefficient, and m t is the first-order moment momentum;

[0123] Then, the second-order moment variance is updated.

[0124] β2 is the variance decay coefficient, and v t is the updated second-order matrix variance;

[0125] Bias correction is performed.

[0126]

[0127] Finally, parameter update is performed.

[0128]

[0129] θ t is the model parameter at the current time t (i.e., the current weight), and θ t+1 is the model parameter at the next time t + 1 (the updated weight); α is the learning rate. is the corrected first-order moment momentum, is the corrected second-order matrix variance.

[0130] S5. Predict the fault data in the complex time-series data of the charging pile.

[0131] Combining step S3 and step S4 ensures more stable update of the convolutional layer parameters, accelerates the convergence of gradient descent, and thus improves the training efficiency of the entire CNN-LSTM model.

[0132] After the data preprocessing, feature extraction, time-series modeling and optimization of the present invention are completed, the last step is to use the improved CNN-LSTM model to predict the faults in the complex time-series data of the charging pile. The core objective of this step is to extract the key features of the charging pile operation data through the deep learning network, and combine the time-series information to accurately identify potential fault patterns and achieve efficient and intelligent assessment of the health status of the charging pile.

[0133] During the prediction process, the optimized CNN is responsible for extracting the local features of the charging bill data and operation and maintenance data, such as abnormal patterns of charging orders, distribution of reasons for charging end, etc., while the LSTM further models the long-term time-dependent relationship and identifies the trend of state changes of the charging pile before and after a fault occurs. Finally, using the fully connected layer and the SoftMax function, the output of the model is mapped to specific fault categories to achieve classification prediction of the charging pile state.

[0134] The present invention can effectively improve the accuracy of charging pile fault prediction, enabling the operation and maintenance system to identify potential faults in advance, reducing equipment downtime, and enhancing the stability of charging infrastructure and user experience. At the same time, the present invention can adapt to different types of charging pile data and has good generalization ability, and can be popularized and applied to the intelligent monitoring and management of large-scale charging networks.

[0135] Embodiment 2

[0136] As Figure 2 and Figure 3 This is the second embodiment of the present invention. In this embodiment, the technical effects of using the present invention for fault prediction are verified through simulation experiments.

[0137] Specifically, 1. Model training: Use the Adam optimizer to train the CNN-LSTM, optimize the convergence speed of feature extraction by the CNN, and improve the ability of the LSTM to remember temporal information. When training the data, input the charging pile temporal data (voltage, current, power, temperature, etc.), and the goal is to predict future fault types. Record the loss curve during the training process to check for overfitting or underfitting.

[0138] 2. Test set evaluation: That is, confusion matrix analysis. Use the test data set, the CNN-LSTM predicts the fault category, and calculate the confusion matrix:

[0139]

[0140] It can be seen from Figure 2 that the values on the diagonal are relatively large, and the vast majority of predictions are correct; there are certain misclassifications in category 0, where 231 are correctly classified as 0, but 19 are misclassified as 2; almost all other categories are correctly classified. From category 1 to 9 ((normal operation (1), overcurrent fault (2), overvoltage fault (3), undervoltage fault (4), insulation fault (5), temperature anomaly (6), connection anomaly (7), communication anomaly (8), relay fault (9), other faults (0)), almost all samples are correctly classified, indicating that the model has extremely strong discrimination ability in these categories.

[0141] By observing the ratios of FN (false negative) and FP (false positive), evaluate whether the model is prone to misjudgment or missed detection. The accuracy formula:

[0142]

[0143] It can be seen from Figure 3It can be seen that in the first 10 iterations, the accuracy rate increased rapidly, indicating that the model learned effective features in the early training stage. After that, the accuracy rate tended to be stable and finally approached 98%, indicating that the model achieved a high classification performance.

[0144] 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 within the scope of the claims of the present invention.

Claims

1. A charging pile fault prediction method based on an improved convolutional long short-term memory network, characterized in that: It includes the following steps: S1. Collect charging data, fault repair data, and charging feedback data, clean the collected data, and generate a time series; S2. Extract features and perform pattern recognition on the time series data for charging pile fault prediction through a convolutional neural network; S3. Construct a convolutional long short-term memory network model and output the probability distribution of each category; S4. Use the adaptive moment optimization algorithm to optimize the convolutional layer; S5. Predict the fault data in the complex time series data of the charging pile.

2. The method for predicting charging pile faults based on an improved convolutional long short-term memory network according to claim 1, wherein: The specific content of step S2 includes: S201. Given the input time series data Perform weighted summation on the time series data X through the sliding convolution kernel W to obtain the output feature map X'. Among them, B is the batch size, C in is the number of input channels, H and W are the height and width respectively, W i,m,n is the sliding convolutional kernel, X i,m,n is the input time-series data, and the output X' after the convolution operation is obtained through the convolutional kernel and the bias The number of output channels is K out is the number of output channels, k H and k W are the height and width of the convolutional kernel respectively; S202. Perform batch normalization processing on each feature map; γ c is the scaling factor, β c is the offset factor, ∈ is a constant to prevent division by zero errors, μ c is the mean value of the c-th channel calculated by the BN layer for a certain channel c; H’ is the height, W’ is the width, is the variance of the c-th channel, X′ i，c，h，w is the output after the convolution operation, X″ i，c，h，w is the output after batch normalization, standardization, and scaling; S203. Process each element of the input data through the ReLU activation function, X ReLU = max(0, X″), where X ReLU is the element processed by the ReLU function; S204. After the activation layer, the pooling operation applies a sliding window to each feature map. The size of the pooling window is k H ×k W , the stride is s. For each position in the input feature map, the pooling operation selects the maximum value within the window Among them, X ReLU is the feature map after being activated by ReLU, and window is the sliding window area of the pooling operation, that is, k H ×k W ; is the output feature map after pooling; represents taking the maximum value within the window range.

3. The method for predicting charging pile faults based on an improved convolutional long short-term memory network according to claim 1, wherein: The specific content of step S3 includes: S301. Convert the output in step S2 from a four-dimensional tensor to a three-dimensional tensor. X LSTM，b，t，f is the b-th sample, the t-th time step, and the f-th feature dimension of the flattened data; is the pooled row position corresponding to time step t, and t mod W′ is the pooled column position corresponding to time step t; S302. Use the three-dimensional tensor as the input data of the LSTM, and sequentially calculate the output of the forget gate, the output of the input gate, the cell state i t and the hidden state h t ; S303. Take the three-dimensional vector output by the LSTM as the input of the fully connected layer and perform a linear transformation. Then apply the ReLU activation function to set the negative values of the input to zero and retain the positive values. Add a Dropout layer after the fully connected layer to randomly discard the outputs of some neurons. After passing through the SoftMax function, convert the scores of the output layer into the probability values Y for each category output，c , Among them, Y output，c is the predicted probability of the c-th class, representing the probability that the sample belongs to class c; Y FC,c is the output score of the c-th class, coming from the fully connected layer; C out is the total number of classes, is the exponential function sum of the scores of all classes.

4. The method for predicting charging pile faults based on an improved convolutional long short-term memory network according to claim 3, characterized in that: The output f of the forget gate t is H LSTM the hidden layer size of the LSTM, the weight matrix of the forget gate, the bias term, and σ is the Sigmoid activation function with an output range of [0, 1].

5. The method for predicting charging pile faults based on an improved convolutional long short-term memory network according to claim 4, characterized in that: The output i of the input gate t is, i t = σ(W i · [h t-1 , X LSTM,b，t，f + b i ; Among them, is the output of the input gate, is the weight matrix of the input gate, is the bias term, h t-1 is the hidden state at the previous time step, storing information from past time steps and determining which information should be forgotten or retained; Generate candidate cell states is the weight matrix, is the bias term.

6. The method for predicting charging pile faults based on an improved convolutional long short-term memory network according to claim 5, wherein: Updated cell state c t is is the cell state at the current moment, is the cell state at the previous moment.

7. The method for predicting charging pile faults based on an improved convolutional long short-term memory network according to claim 6, characterized in that: The output o of the output gate t is, o t = σ(W o · [h t-1 , X LSTM,b，t，f + b o ); is the weight matrix of the output gate, is the bias term; The final output h t is, h t = o t ⊙hanh(c t ); Among them, is the output of the LSTM at the current moment, and tanh is the hyperbolic tangent function.

8. The method for predicting charging pile faults based on an improved convolutional long short-term memory network according to any one of claims 3 to 6, wherein: Specifically, step S4 is: Use the cross-entropy loss function to measure the difference between the model output Y output，c and the true label Y true ; Calculate the gradient of the loss function for all parameters θ of the convolutional neural network, long short-term memory network, and fully connected layer; Use Adam to update the model parameters θ.

9. The cross-entropy loss \(L\) in the charging pile fault prediction method based on the improved convolutional long short-term memory network as described in claim 8 has the formula: Y ture，c is the one-hot encoding of the true category; Gradient of the loss function is θ ∈ {W CNN , b CNN , W h , W x , W FC , b FC}; When updating the model parameters, first perform first-order moment momentum update; β1 is the momentum decay coefficient, m t is the first - order moment momentum at the current moment, m t-1 is the first - order moment momentum at the previous moment; Then perform the second-moment variance update, β2 is the variance decay coefficient, and v t is the second moment gradient variance at the current time t, and v t-1 is the second moment gradient variance at the previous time; Perform bias correction; Finally, perform parameter update. α is the learning rate, is the corrected first-order moment momentum, is the corrected second-order matrix variance.

Images