Online Remote Charge and Discharge Fault Prediction Method and System for Communication Battery Sets

Abstract

The present invention provides an online remote charge-discharge fault prediction method and system for a communication battery pack, which relates to the technical field of battery pack fault prediction. It includes preprocessing time series data, extracting complexity features based on multi-scale entropy analysis, determining physical features based on a physical model, performing normalization and dimensionless processing on the complexity features and physical features to determine a primary feature set; based on the primary feature set, constructing a weighted undirected graph, expanding to generate a maximum weight connected subgraph, using the primary features in the maximum weight connected subgraph as preferred features to construct a preferred feature set; inputting the data of the battery pack to be predicted into a fault prediction network, extracting corresponding feature values based on the preferred feature set in the input layer, calculating the probability distribution of fault types through multiple hidden layers, and selecting the fault type corresponding to the highest probability as the fault prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of battery pack fault prediction, and particularly to an online remote charge-discharge fault prediction method and system for communication battery packs. Background Art

[0002] Communication power supplies usually use AC mains power supply and are equipped with communication battery packs as backup power supplies to ensure that the base stations can continue to operate when the mains power fails. However, during long-term use, communication battery packs are vulnerable to various factors, resulting in performance degradation or even failure. Battery pack failures not only affect the normal operation of base stations but may also cause safety accidents and economic losses. Therefore, timely and accurate prediction of communication battery pack failures is of great significance for ensuring the reliability and continuity of communication networks.

[0003] Traditional battery pack fault prediction methods mainly rely on manual inspections and regular maintenance, which have problems such as large workload, low efficiency, and poor accuracy. In recent years, with the development of Internet of Things and big data technologies, remote online monitoring and data analysis have gradually been applied to the field of battery pack fault prediction. Existing fault prediction methods mainly target individual batteries or small-scale battery packs, lacking consideration for large-scale distributed communication battery packs. Communication power supplies are widely distributed, and the environmental conditions and load characteristics of battery packs at different base stations vary, resulting in different fault modes and degradation laws.

[0004] In summary, how to achieve online remote charge-discharge fault prediction for large-scale communication battery packs and consider the differences and correlations between different base stations is an urgent problem to be solved, and the present invention can solve the problems in the prior art. Summary of the Invention

[0005] Embodiments of the present invention provide an online remote charge-discharge fault prediction method and system for communication battery packs, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiments of the present invention,

[0007] An online remote charge-discharge fault prediction method for communication battery packs is provided, including:

[0008] Collecting time series data of the charge-discharge process of a communication battery pack, preprocessing the time series data, extracting complexity features based on multi-scale entropy analysis, determining physical features through a parameter identification algorithm based on the physical model corresponding to the battery pack, performing normalization and dimensionless processing on the complexity features and the physical features, and determining a primary feature set of the charge-discharge process of the battery pack;

[0009] Based on the primary feature set, construct a correlation coefficient matrix. Using the correlation coefficient as the weight value, determine the weighted edges of the undirected graph. Taking the primary features as the nodes of the undirected graph, construct a weighted undirected graph; Based on the weighted undirected graph, through iterative calculation of the weighted edge connectivity of each undirected graph node, expand and generate the maximum weighted connected subgraph. Taking the primary features corresponding to each subgraph node in the maximum weighted connected subgraph as the preferred features, construct a preferred feature set;

[0010] Input the data of the battery pack to be predicted into the fault prediction network. Based on the preferred feature set in the input layer, extract the corresponding feature values. Through multiple hidden layers, abstract the feature representations layer by layer, extract high-order features, and through the activation function of the output layer, calculate the probability distribution of the fault types, and select the fault type corresponding to the highest probability as the fault prediction result.

[0011] In an alternative embodiment,

[0012] Preprocess the time series data, extract complexity features based on multi-scale entropy analysis, determine physical features through a parameter identification algorithm based on the physical model corresponding to the battery pack, perform dimensionless processing on the complexity features and the physical features, and determine that the primary feature set of the charge and discharge process of the battery pack includes:

[0013] Detect and remove outliers from the time series data, and through smoothing filtering and normalization processing, obtain standard time series data;

[0014] Perform a coarse-graining operation on the standard time series data to obtain scale subsequences at different scales. Calculate the sample entropy for each scale subsequence, determine the multi-scale entropy curve, and extract slope features, curvature features, entropy value features, and area features from the multi-scale entropy curve to construct complexity features;

[0015] Based on the electrochemical characteristics and thermodynamic behavior of the battery pack, construct a corresponding physical model. Through a parameter identification algorithm based on the recursive least squares method, evaluate the physical model parameters and determine the physical features;

[0016] Perform dimensionless processing on the complexity features and the physical features to eliminate the numerical range and dimensional differences, and determine the primary feature set.

[0017] In an alternative embodiment,

[0018] Based on the primary feature set, construct a correlation coefficient matrix. Using the correlation coefficient as the weight value, determine the undirected graph weight edges. Taking the primary features as the undirected graph nodes, construct a weighted undirected graph. Based on the weighted undirected graph, through iterative calculation of the weight edge connectivity sum of each undirected graph node, expand and generate the maximum weight connected subgraph. Taking the primary features corresponding to each connected node in the maximum weight connected subgraph as the preferred features, construct a preferred feature set including:

[0019] Based on the primary feature set, by calculating the covariance and variance between every two primary features, determine the correlation coefficient. If the absolute value of the correlation coefficient is greater than the preset correlation coefficient threshold, retain the corresponding correlation coefficient; otherwise, eliminate it. Based on the retained correlation coefficients, construct a correlation coefficient matrix.

[0020] Taking each primary feature in the primary feature set as an undirected graph node, traverse each correlation coefficient in the correlation coefficient matrix. Using the correlation coefficient as the weight value, determine the undirected graph weight edge between the corresponding two undirected graph nodes, and construct a weighted undirected graph.

[0021] Traverse each undirected graph node in the weighted undirected graph, calculate the undirected weight sum of all undirected graph weight edges connected to the undirected graph node, select the undirected graph node corresponding to the maximum value of the undirected weight sum as the first connected node, initialize the connected subgraph, and start iterative expansion of the connected subgraph:

[0022] Traverse each connected node in the connected subgraph, determine the corresponding undirected graph node from the weighted undirected graph, find all undirected graph nodes that are directly connected to the undirected graph node and are not in the connected subgraph, determine the candidate node set, calculate the weight edge connectivity sum in the weighted undirected graph between each candidate node in the candidate node set and all connected nodes in the connected subgraph, and select the candidate node corresponding to the maximum value of the weight edge connectivity sum as the new connected node and add it to the connected subgraph.

[0023] Until the candidate node set is empty, generate the maximum weight connected subgraph.

[0024] Taking the primary features corresponding to each connected node in the maximum weight connected subgraph as the preferred features, construct a preferred feature set.

[0025] In an alternative embodiment,

[0026] The construction process of the weighted undirected graph further includes:

[0027] Based on the correlation coefficient matrix and the preset correlation coefficient threshold, perform binarization processing on the correlation coefficient matrix to determine the binarized correlation matrix.

[0028] Based on the binary correlation matrix, calculate the Euclidean distance between every two features, determine the feature distance matrix, extract the maximum and minimum values in the feature distance matrix, determine the feature distance ratio, and generate the diversity metric matrix;

[0029] Taking each primary feature in the primary feature set as a node of an undirected graph, traverse each correlation coefficient in the correlation coefficient matrix and each diversity metric in the diversity metric matrix, calculate the weight value through the correlation coefficient and the diversity metric, determine the undirected graph weighted edge between the corresponding two undirected graph nodes, and construct a weighted undirected graph.

[0030] In an alternative embodiment,

[0031] Input the data of the battery pack to be predicted into the fault prediction network. Based on the preferred feature set in the input layer, extract the corresponding feature values, through multiple hidden layers, abstract the feature representations layer by layer, extract high-order features, and through the activation function of the output layer, calculate the probability distribution of the fault types, and select the fault type corresponding to the highest probability as the fault prediction result, including:

[0032] The fault prediction network is constructed based on a deep belief network, including an input layer, at least one hidden layer, and an output layer, where the number of nodes in the input layer is equal to the number of features in the preferred feature set, the hidden layer is composed of multiple stacked probabilistic neural units, and the number of nodes in the output layer corresponds one-to-one to the fault types;

[0033] Based on the unsupervised learning method, perform layer-by-layer primary training on each layer of probabilistic neural units of the fault prediction network. The output hidden layer activation probability of the previous layer of probabilistic neural units enters the input layer of the current layer of probabilistic neural units. In the hidden layer of the current layer of probabilistic neural units, by minimizing the energy function, adjust the weights and bias factors of the current layer of probabilistic neural units to generate the probability distribution corresponding to the input data of the current layer of probabilistic neural units;

[0034] When the primary training of all probabilistic neural units is completed, perform secondary adjustment on the fault prediction network based on the supervised learning method. Based on the weights and bias factors corresponding to each layer of probabilistic neural units determined by the primary training, initialize the fault prediction network;

[0035] Add a classifier to the output layer, use different fault types as labels, and through the optimization algorithm, minimize the classification loss function, and update the weights and bias factors of each layer in the fault prediction network to obtain the optimal fault prediction network;

[0036] Input the data of the battery pack to be predicted into the optimal fault prediction network, input the probability distribution of the fault types, and select the fault type with the highest probability as the fault prediction result.

[0037] In an alternative embodiment,

[0038] The output of the previous layer of probabilistic neural units is the activation probability of the hidden layer, which enters the input layer of the current layer of probabilistic neural units. In the hidden layer of the current layer of probabilistic neural units, by minimizing the energy function, the weights and bias factors of the current layer of probabilistic neural units are adjusted to generate the probability distribution corresponding to the input data of the current layer of probabilistic neural units, including:

[0039] The energy function has the following formula:

[0040]

[0041] where l represents the serial number of the hidden layer of the fault prediction network, v l represents the state of the input layer, h l represents the state of the hidden layer, E(v l , h l ) represents the energy function, W l represents the connection weight matrix between the input layer and the hidden layer, a l represents the bias vector of the input layer, b l represents the bias vector of the hidden layer, U l ij represents the interaction weight between nodes i and j in the input layer, v l i represents the state of the input layer of node i in the input layer, v l j represents the state of the input layer of node j in the input layer, V l km represents the interaction weight between nodes k and m in the hidden layer, h l k represents the state of the hidden layer of node k in the hidden layer, h l m represents the state of the hidden layer of node m in the hidden layer, and λ represents the regularization term balance parameter.

[0042] In an alternative embodiment,

[0043] Based on the fault prediction network, a distributed prediction network is constructed, including:

[0044] Taking a communication battery pack and the corresponding edge computing node as a branch network and the central server as the central network, a distributed prediction network is constructed;

[0045] The central network is constructed based on the fault prediction network and distributed to the branch networks. Independent data collection and training are performed on each branch network, and the model parameters in each trained branch network are uploaded to the central network. The central network aggregates the model parameters to obtain updated model parameters and distributes them to the branch networks.

[0046] Multiple rounds of iteration are performed in sequence until a preset iteration upper limit is reached, and the final central network and the corresponding final branch networks are determined. The final branch networks perform fault prediction.

[0047] In the second aspect of the embodiments of the present invention,

[0048] A communication battery pack online remote charge and discharge fault prediction system is provided, including:

[0049] A first unit for collecting time series data of the charging and discharging process of the communication battery pack, preprocessing the time series data, extracting complexity features based on multi-scale entropy analysis, determining physical features through a parameter identification algorithm based on the physical model corresponding to the battery pack, and performing non-dimensional processing on the complexity features and the physical features to determine the primary feature set of the charging and discharging process of the battery pack.

[0050] A second unit for constructing a correlation coefficient matrix based on the primary feature set, determining undirected graph weighted edges with the correlation coefficient as the weight value, constructing a weighted undirected graph with the primary features as undirected graph nodes; based on the weighted undirected graph, expanding and generating a maximum weight connected subgraph by iteratively calculating the weight edge connectivity of each undirected graph node, and constructing a preferred feature set with the primary features corresponding to each subgraph node in the maximum weight connected subgraph as the preferred features.

[0051] A third unit for inputting the data of the battery pack to be predicted into the fault prediction network, extracting corresponding feature values based on the preferred feature set in the input layer, abstracting feature representations layer by layer through multiple hidden layers, extracting high-order features, and calculating the probability distribution of fault types through the activation function of the output layer, and selecting the fault type corresponding to the highest probability as the fault prediction result.

[0052] In the third aspect of the embodiments of the present invention,

[0053] An electronic device is provided, including:

[0054] A processor;

[0055] A memory for storing instructions executable by the processor;

[0056] Wherein, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0057] In the fourth aspect of the embodiments of the present invention,

[0058] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0059] In the embodiments of the present invention, the multi-scale entropy analysis method is used to extract complexity features from the preprocessed time series data, capturing the non-linear and non-stationary characteristics of the signal, which helps to deeply understand the dynamic changes in the charging and discharging process of the battery pack; according to the physical model corresponding to the battery pack, an in-depth understanding of the battery state and performance can be provided; by constructing a correlation coefficient matrix and using a weighted undirected graph, the correlation between primary features can be systematically analyzed, the strongest association between features can be identified, the most useful features for fault prediction can be effectively screened out, ensuring that the selected feature set is not only representative and can retain the information content of the original data to the greatest extent; by selecting the maximum weight connected subgraph, redundant features can be effectively removed, and the correlation coefficient matrix helps to identify highly correlated features, avoiding the influence of duplicate information and improving the generalization ability and prediction accuracy of the model; constructing a branch network allows parallel processing of data for each communication battery pack, meaning that the system can process multiple data streams simultaneously, improving the overall efficiency and processing capacity of the system; a two-way communication channel is established between the central network and the branch network, enabling model parameters to be transmitted and updated between the central network and the branch network. The co-training method allows the central network to aggregate information from multiple branch networks to obtain a more comprehensive data perspective and improve the accuracy and generalization ability of the model; after the independent training of the branch network is completed, the central network will collect and aggregate the model parameters of all branch networks to generate updated global model parameters. The parameter aggregation mechanism can effectively integrate the training results of each branch network and improve the overall performance of the model; BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a schematic flowchart of the method for online remote charge and discharge fault prediction of a communication battery pack according to the embodiments of the present invention;

[0061] Figure 2 is a schematic structural diagram of the system for online remote charge and discharge fault prediction of a communication battery pack according to the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. 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.

[0063] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0064] Figure 1 It is a schematic flowchart of the online remote charge and discharge fault prediction method for the communication battery pack in the embodiments of the present invention. As Figure 1 shown, the method includes:

[0065] S101. Collect the time series data of the charge and discharge process of the communication battery pack, preprocess the time series data, extract the complexity features based on multi-scale entropy analysis, determine the physical features based on the physical model corresponding to the battery pack through the parameter identification algorithm, perform normalization and dimensionless processing on the complexity features and the physical features, and determine the primary feature set of the charge and discharge process of the battery pack;

[0066] The multi-scale entropy analysis specifically refers to a method for describing the complexity of time series data, which calculates the entropy value of the data at different time scales. By calculating the entropy value at different time scales, the multi-scale entropy analysis can provide the complexity information of the data at different time scales;

[0067] The parameter identification algorithm specifically refers to determining the parameters in the system model through calculation according to the dynamic model of the system and the measured data. In the present invention, the parameter identification algorithm can be used to determine the electrochemical parameters of the battery, such as resistance, capacitance, etc.;

[0068] The normalization and dimensionless processing specifically refers to performing dimensionless processing on the data and making the data distributed within a certain range, which can make the dimensions of different features consistent and facilitate subsequent data analysis and processing.

[0069] In this embodiment, the multi-scale entropy analysis method is used to extract complexity features from the preprocessed time series data. Multi-scale entropy is a method for measuring the complexity of time series, which can capture the non-linear and non-stationary characteristics of signals and help to deeply understand the dynamic changes in the charge and discharge process of the battery pack. According to the physical model corresponding to the battery pack, physical characteristics are determined through a parameter identification algorithm. Physical characteristics may involve the electrochemical parameters, internal resistance, voltage curve characteristics, etc. of the battery, and can provide an in-depth understanding of the battery state and performance. The extracted complexity features and physical characteristics are normalized to make them have the same dimension and range, which is convenient for subsequent feature selection and model establishment. The normalized complexity features and physical characteristics are combined to form a primary feature set for the charge and discharge process of the battery pack. The primary feature set will reflect the key features of the battery pack during the charge and discharge process and provide an important basis for subsequent tasks such as fault prediction and state assessment.

[0070] In an alternative embodiment, the time series data is preprocessed, complexity features are extracted based on multi-scale entropy analysis, physical characteristics are determined based on the physical model corresponding to the battery pack through a parameter identification algorithm, and the complexity features and the physical characteristics are normalized to dimensionless, and the primary feature set for the charge and discharge process of the battery pack is determined, including:

[0071] Outlier detection and elimination are performed on the time series data, and standard time series data is obtained through smoothing filtering processing and normalization processing;

[0072] The standard time series data is coarsened to obtain scale subsequences at different scales. Sample entropy is calculated for each scale subsequence, a multi-scale entropy curve is determined, and slope features, curvature features, entropy value features, and area features are extracted from the multi-scale entropy curve to construct complexity features;

[0073] Based on the electrochemical characteristics and thermodynamic behavior of the battery pack, a corresponding physical model is constructed, and physical model parameters are evaluated through a parameter identification algorithm based on the recursive least squares method to determine physical characteristics;

[0074] The complexity features and the physical characteristics are normalized to dimensionless to eliminate the differences in numerical range and dimension, and the primary feature set is determined.

[0075] Original time series data is collected and preliminarily inspected. The 3σ principle statistical method is used to identify outliers in the data, where the 3σ principle means that data points exceed the range of the mean plus or minus three standard deviations. The detected outliers can be processed by either directly deleting these outliers or replacing them with reasonable substitute values. Preferably, the mean or median of the previous and subsequent data points is used;

[0076] Select appropriate filtering methods, such as moving average filtering, median filtering, and Kalman filtering, to remove high-frequency noise and short-term fluctuations, and smooth the time series data; perform normalization on the smoothed time series data to scale the data to a unified range, preferably [-1, 1], to eliminate the dimensional differences between different time series. The normalization method can include min-max normalization or zero-mean unit-variance normalization; use the data after outlier removal, smoothing filtering, and normalization as the standard time series data;

[0077] Determine different scale factors to perform coarse-graining operations on the standard time series data. Use the set scale factors to divide the standard time series data into subsequences at different scales, and map each subsequence to a scale point by taking the average value, maximum value, or minimum value, etc.; calculate the sample entropy for each scale subsequence to determine the complexity and regularity of the time series at this scale; select the sample entropy values under different scale factors, plot the multi-scale entropy curve, and determine the changing trend of the complexity of the time series at different scales.

[0078] Extract complexity features from the multi-scale entropy curve. Calculate the slope of the multi-scale entropy curve in different scale intervals to obtain the rate of change of the entropy value with the scale; calculate the curvature of the multi-scale entropy curve at different scale points to obtain the non-linear degree of the entropy value with the scale; extract the maximum entropy value, minimum entropy value, and average entropy value of the multi-scale entropy curve, etc., to obtain the overall complexity level of the time series; calculate the area enclosed by the multi-scale entropy curve to obtain the cumulative effect of complexity at different scales; combine the extracted slope features, curvature features, entropy value features, and area features into a complexity feature vector;

[0079] According to the specific characteristics of the battery pack, select an equivalent circuit model, preferably an R-C circuit model. Based on the model type, set the structure and equations of the model, and define each parameter in the model, such as resistance, capacitance, thermal conductivity, etc.; collect experimental data or actual operation data, including key variables such as current, voltage, and temperature;

[0080] Initialize the model parameter values, use the recursive least squares algorithm, and continuously update the parameters to minimize the error between the predicted value and the actual observed value, and judge whether the update of the parameters converges, that is, whether the change amount of the parameters is within the pre-set tolerance range; extract the key parameters from the identified physical model and calculate the physical characteristics according to the key parameters;

[0081] Standardize the complexity features and physical features, and then through normalization, scale the numerical ranges of different features to the same interval, preferably [-1, 1], to eliminate the dimensional differences between different features, ensure that different features are within the same numerical range, which can make it easier to perform feature fusion and comparative analysis. Then fuse the normalized complexity features and physical features to construct a primary feature set.

[0082] In this embodiment, through the coarse-grained processing of different scale factors, obtain the complexity and regularity of the time series at different scales, comprehensively capture the dynamic characteristics of the time series, which helps to extract complexity features more accurately; extract features such as slope, curvature, maximum entropy value, minimum entropy value, average entropy value, and the area enclosed by the curve from the multi-scale entropy curve, comprehensively reflect the complexity and variation law of the time series, and improve the description ability and discrimination of the feature vector; by selecting an appropriate equivalent circuit model, use the recursive least squares algorithm to identify the parameters of the data, extract the key physical parameters of the battery pack, and accurately reflect the physical characteristics and operating state of the battery pack; standardize and normalize the complexity features and physical features to ensure that the numerical ranges of different features are consistent, which is convenient for feature fusion and comparative analysis. The fused primary feature set contains the dynamic complexity and physical characteristics of the battery pack, providing rich and effective feature data for subsequent fault prediction and state assessment.

[0083] S102. Based on the primary feature set, construct a correlation coefficient matrix, use the correlation coefficient as the weight value to determine the undirected graph weight edges, and use the primary features as the undirected graph nodes to construct a weighted undirected graph; based on the weighted undirected graph, through iterative calculation of the weight edge connectivity of each undirected graph node, expand and generate the maximum weight connected subgraph, and use the primary features corresponding to each subgraph node in the maximum weight connected subgraph as the preferred features to construct a preferred feature set;

[0084] The correlation coefficient matrix specifically refers to a symmetric matrix used to represent the linear correlation degree between the features in the primary feature set, and the elements in the matrix represent the correlation coefficients between pairs of features.

[0085] The weighted undirected graph specifically refers to a graph where the nodes represent the primary features, each primary feature is a node in the graph, and the edges represent the relationships between the features. If the absolute value of the correlation coefficient between two features is greater than a preset threshold, draw an edge between the two nodes, and the weight of the edge is the correlation coefficient between the features.

[0086] In this embodiment, by constructing a correlation coefficient matrix and using a weighted undirected graph, the correlation between primary features can be systematically analyzed. By iteratively calculating the weighted edge connectivity sum, the strongest associations between features can be identified, and the features most useful for fault prediction can be effectively screened out, ensuring that the selected feature set is not only representative but also can retain the information content of the original data to the greatest extent. By selecting the maximum-weight connected subgraph, redundant features can be effectively removed. The correlation coefficient matrix helps identify highly correlated features, avoiding the influence of duplicate information, and improving the generalization ability and prediction accuracy of the model. Reducing redundant features helps simplify the model and reduce the computational complexity and storage requirements.

[0087] In an alternative embodiment, based on the primary feature set, a correlation coefficient matrix is constructed. Using the correlation coefficient as the weight value, the undirected graph weighted edges are determined, and with the primary features as the undirected graph nodes, a weighted undirected graph is constructed. Based on the weighted undirected graph, by iteratively calculating the weighted edge connectivity sum of each undirected graph node, a maximum-weight connected subgraph is extended and generated. The primary features corresponding to each connected node in the maximum-weight connected subgraph are used as the preferred features to construct a preferred feature set, including:

[0088] Based on the primary feature set, by calculating the covariance and variance between every two primary features, the correlation coefficient is determined. If the absolute value of the correlation coefficient is greater than a preset correlation coefficient threshold, the corresponding correlation coefficient is retained; otherwise, it is eliminated. Based on the retained correlation coefficients, a correlation coefficient matrix is constructed.

[0089] Taking each primary feature in the primary feature set as an undirected graph node, traversing each correlation coefficient in the correlation coefficient matrix, and using the correlation coefficient as the weight value, the undirected graph weighted edges between the corresponding two undirected graph nodes are determined to construct a weighted undirected graph.

[0090] Traverse each undirected graph node in the weighted undirected graph, calculate the undirected weight sum of all undirected graph weighted edges connected to the undirected graph node, select the undirected graph node corresponding to the maximum value of the undirected weight sum as the first connected node, initialize the connected subgraph, and start iteratively expanding the connected subgraph:

[0091] Traverse each connected node of the connected subgraph, determine the corresponding undirected graph node from the weighted undirected graph, find all undirected graph nodes that are directly connected to the undirected graph node and are not in the connected subgraph, determine the candidate node set, calculate the weighted edge connectivity sum in the weighted undirected graph between each candidate node in the candidate node set and all connected nodes in the connected subgraph, and select the candidate node corresponding to the maximum value of the weighted edge connectivity sum as the new connected node and add it to the connected subgraph.

[0092] Until the candidate node set is empty, generate a maximum-weight connected subgraph;

[0093] Using the primary features corresponding to each connected node in the maximum-weight connected subgraph as the preferred features, construct a preferred feature set.

[0094] Extract all primary features from the normalized primary feature set obtained in the previous stage. For every two primary features, calculate their covariance, evaluate their linear relationship, and calculate the variance of each primary feature respectively, which represents the degree of dispersion of the feature data;

[0095] Using the covariance and variance between primary features, calculate the correlation coefficient to quantify the degree of linear correlation between features; define a correlation coefficient threshold for screening significantly correlated feature pairs. If the absolute value of the correlation coefficient is greater than the preset correlation coefficient threshold, retain the corresponding correlation coefficient; otherwise, eliminate the corresponding correlation coefficient; create a matrix where the rows and columns represent the respective primary features in the primary feature set, and fill the retained correlation coefficients into the corresponding positions in the matrix to construct a correlation coefficient matrix.

[0096] Take each primary feature in the primary feature set as a node in an undirected graph. Traverse each correlation coefficient in the correlation coefficient matrix, and use the correlation coefficient as the weight value of the edge between two nodes in the undirected graph. If the correlation coefficient exists and is retained, establish a weighted edge between the corresponding undirected graph nodes, and combine all the undirected graph nodes and undirected weighted edges to construct a complete weighted undirected graph.

[0097] Traverse each undirected graph node in the weighted undirected graph, calculate the sum of all undirected graph weight edges connected to the undirected graph node, that is, the undirected weight sum. Select the undirected graph node with the largest undirected weight sum as the first connected node, and add the first connected node to the connected subgraph to start iterative expansion:

[0098] Traverse each connected node in the current connected subgraph, find all undirected graph nodes that are directly connected to each connected node in the weighted undirected graph and are not in the connected subgraph, and form a candidate node set. For each candidate node in the candidate node set, calculate the weighted edge connection sum between the candidate node and all connected nodes in the connected subgraph, select the candidate node with the largest weighted edge connection sum, and add the selected candidate node as a new connected node to the connected subgraph;

[0099] Repeat the above steps until the candidate node set is empty.

[0100] Extract the primary features corresponding to each connected node from the maximum-weight connected subgraph, and assemble the extracted primary feature set to construct the final preferred feature set.

[0101] Optionally, assume there is a feature set F containing 5 features, and the correlation coefficient matrix R is as follows:

[0102]

[0103] And set a threshold r 0 to be 0.5;

[0104] Initialize an empty undirected graph G, which contains 5 nodes corresponding to the 5 features in the feature set F. Traverse each element in the correlation coefficient matrix R and add edges that meet the threshold condition, including:

[0105] r 12 = 0.9, > r 0 , add an edge (1, 2) with a weight of 0.9 in G;

[0106] r 15 = 0.5, = r 0 , add an edge (1, 5) with a weight of 0.5 in G;

[0107] r 24 = 0.7, > r 0 , add an edge (2, 4) with a weight of 0.7 in G;

[0108] r 25 = 0.6, > r 0 , add an edge (2, 5) with a weight of 0.6 in G;

[0109] r 34 = 0.8, > r 0 , add an edge (3, 4) with a weight of 0.8 in G;

[0110] r 45 = 0.5, = r 0 , add an edge (4, 5) with a weight of 0.5 in G.

[0111] Determine the weighted undirected graph G;

[0112] Initialize an empty connected subgraph S and an empty candidate feature set C. Calculate the sum of the weights of the edges of each node and select the node with the largest sum of weights as the initial node:

[0113] Node 1, 0.9 + 0.5 = 1.4;

[0114] Node 2, 0.9 + 0.7 + 0.6 = 2.2 (the largest);

[0115] Node 3, 0.8 = 0.8;

[0116] Node 4, 0.7 + 0.8 + 0.5 = 2.0;

[0117] Node 5, 0.5 + 0.6 + 0.5 = 1.6;

[0118] Select node 2, initialize S = {2}, C = {2}.

[0119] Initialize an empty set of candidate nodes N, traverse the nodes in the current set of candidate features C, find the nodes directly connected to it and not in S, and calculate the sum of the edge weights:

[0120] Currently, there is only node 2 in C. The nodes directly connected to node 2 and not in S are 1, 4, and 5. Calculate respectively:

[0121] Node 1, r 21 = 0.9; Node 4, r 24 = 0.7; Node 5, r 25 = 0.6;

[0122] Select the node with the largest sum of weights, that is, select node 1, add it to S and C, and get S = {2, 1}, C = {2, 1};

[0123] The nodes directly connected to node 2 and not in S are 4 and 5, and the nodes directly connected to node 1 and not in S are 5. Update N = {4, 5};

[0124] Iterate again and calculate: Node 4, r 24 = 0.7; Node 5, r 15 + r 25 = 0.5 + 0.6 = 1.1 (the largest);

[0125] Select node 5, add it to S and C. Now S = {2, 1, 5}, C = {2, 1, 5};

[0126] The nodes directly connected to nodes 2 and 5 and not in S are 4. Update N = {4};

[0127] Iterate again and calculate: Node 4, r 24 + r 45 = 0.7 + 0.5 = 1.2;

[0128] Select node 4, add it to S and C. Now S = {2, 1, 5, 4}, C = {2, 1, 5, 4};

[0129] There are no nodes directly connected to nodes 2, 5, and 4 and not in S. Update N to be empty;

[0130] When the set of candidate nodes N is empty, the iteration stops, and the final maximum-weight connected subgraph S = {2, 1, 5, 4} is obtained.

[0131] In this embodiment, by calculating the correlation coefficients between primary features to quantify the degree of linear correlation between features, and using a correlation coefficient threshold to screen out significantly correlated feature pairs, features that have a significant impact on the model can be efficiently selected, which helps reduce the number of input features of the model and improve the model training speed and efficiency; by constructing a weighted undirected graph and a maximum-weight connected subgraph to select features that have a strong correlation with the target variable, the prediction performance of the model can be improved. The features in the preferred feature set can better reflect the essential features of the data, enhancing the generalization ability and prediction effect of the model; by gradually selecting features with high correlation and constructing a maximum-weight connected subgraph, the features in the preferred feature set become more interpretable. The contribution of each feature to the final prediction result can be explained by its position and connectivity in the weighted undirected graph, which helps better understand the decision-making process of the model.

[0132] In an alternative embodiment, the process of constructing the weighted undirected graph further includes:

[0133] Based on the correlation coefficient matrix and a preset correlation coefficient threshold, perform binarization processing on the correlation coefficient matrix to determine a binarized correlation matrix;

[0134] Based on the binarized correlation matrix, calculate the Euclidean distance between every two features to determine a feature distance matrix, extract the maximum and minimum values in the feature distance matrix, determine the proportion of feature distances, and generate a diversity metric matrix;

[0135] Taking each primary feature in the primary feature set as a node of the undirected graph, traverse each correlation coefficient in the correlation coefficient matrix and each diversity metric in the diversity metric matrix, calculate the weight value through the correlation coefficient and the diversity metric, determine the undirected graph weight edge between the corresponding two undirected graph nodes, and construct a weighted undirected graph.

[0136] Extract all primary features from the primary feature set obtained in the previous steps, calculate the correlation coefficient between every two primary features to quantify the degree of linear correlation between features;

[0137] Define a correlation coefficient threshold for screening out significantly correlated feature pairs. Traverse each correlation coefficient in the correlation coefficient matrix. If the absolute value of the correlation coefficient is greater than the preset correlation coefficient threshold, set it to 1; otherwise, set it to 0. The processed result forms a binarized correlation matrix;

[0138] Create a matrix where the rows and columns represent the respective primary features in the set of primary features. For every two primary features, calculate the Euclidean distance based on their binarized correlation values, and fill the calculated Euclidean distance into the corresponding position in the feature distance matrix; extract the maximum Euclidean distance value and the minimum Euclidean distance value from the feature distance matrix, and for each value in the feature distance matrix, calculate the proportion it occupies between the maximum value and the minimum value.

[0139] Create a matrix with the same structure as the feature distance matrix, and fill the results of the feature distance ratios into the diversity metric matrix to complete the construction of the diversity metric matrix.

[0140] Take each primary feature in the set of primary features as a node in an undirected graph. Traverse each correlation coefficient in the correlation coefficient matrix and each diversity metric in the diversity metric matrix. Based on the correlation coefficient and the diversity metric, calculate the weight value of the edge between two nodes in the undirected graph. The calculation of the weight value can consider the combination of the absolute value of the correlation coefficient and the diversity metric.

[0141] For each pair of primary features, if both their correlation coefficient and diversity metric exist, establish a weighted edge between the corresponding nodes in the undirected graph, and combine all the nodes in the undirected graph and the undirected weighted edges to construct a complete weighted undirected graph.

[0142] In this embodiment, by calculating the correlation coefficients between features and performing binarization processing, it is possible to effectively screen out significantly correlated feature pairs, improve the accuracy of feature selection, ensure that the finally selected features have a strong correlation with the target variable, and improve the prediction performance of the model; the binarized correlation matrix can eliminate insignificantly correlated feature pairs, reduce feature redundancy, and reduce the problem of multicollinearity, ensuring the stability and reliability of the model and avoiding overfitting caused by redundant features; by calculating the Euclidean distance between features and constructing a diversity metric matrix, it is possible to quantify the differences between features, ensure that the selected features have higher diversity, capture more potential patterns in the data, and improve the generalization ability of the model.

[0143] S103. Input the data of the battery pack to be predicted into the fault prediction network. Based on the set of preferred features in the input layer, extract the corresponding feature values, abstract the feature representations layer by layer through multiple hidden layers, extract high-order features, and calculate the probability distribution of the fault types through the activation function in the output layer. Select the fault type corresponding to the highest probability as the fault prediction result.

[0144] The multi-layer hidden layer specifically refers to a neural network structure located between the input layer and the output layer, consisting of multiple hidden layers. Each hidden layer contains several neurons, and the neurons are connected by weights and biases. The main function of the multi-layer hidden layer is to gradually form higher-order and more abstract feature representations by abstracting and extracting the features of the input data layer by layer;

[0145] In this embodiment, the feature representation is abstracted layer by layer through the multi-layer hidden layer to extract deeper features. The high-order features can better capture the complex patterns and potential relationships in the battery pack data, improving the accuracy of fault prediction; through the activation function of the output layer, the probability distribution of the fault types is calculated, which can accurately measure the possibility of each fault type, and the fault type corresponding to the highest probability is selected as the prediction result to ensure the accuracy and reliability of the prediction; the high-order feature representation extracted by the multi-layer hidden layer can comprehensively reflect the state and potential fault information of the battery pack, improving the robustness and reliability of the system under complex and variable working conditions.

[0146] In an alternative embodiment, the data of the battery pack to be predicted is input into the fault prediction network. Based on the preferred feature set in the input layer, the corresponding feature values are extracted. Through the multi-layer hidden layer, the feature representation is abstracted layer by layer to extract high-order features, and through the activation function of the output layer, the probability distribution of the fault types is calculated. The fault type corresponding to the highest probability is selected as the fault prediction result, including:

[0147] The fault prediction network is constructed based on a deep belief network, including an input layer, at least one hidden layer, and an output layer. The number of nodes in the input layer is equal to the number of features in the preferred feature set. The hidden layer consists of multiple stacked probabilistic neural units, and the number of nodes in the output layer corresponds one-to-one to the fault types;

[0148] Based on the unsupervised learning method, the primary training of each layer of probabilistic neural units in the fault prediction network is carried out layer by layer. The output hidden layer activation probability of the previous layer of probabilistic neural units enters the input layer of the current layer of probabilistic neural units. In the hidden layer of the current layer of probabilistic neural units, by minimizing the energy function, the weights and bias factors of the current layer of probabilistic neural units are adjusted to generate the probability distribution corresponding to the input data of the current layer of probabilistic neural units;

[0149] When the primary training of all probabilistic neural units is completed, the secondary adjustment of the fault prediction network is carried out based on the supervised learning method. Based on the weights and bias factors corresponding to each layer of probabilistic neural units determined by the primary training, the fault prediction network is initialized;

[0150] A classifier is added to the output layer, with different fault types as labels. By minimizing the classification loss function through an optimization algorithm, the weights and bias factors of each layer in the fault prediction network are updated to obtain the optimal fault prediction network;

[0151] Input the data of the battery pack to be predicted into the optimal fault prediction network, input the probability distribution of the fault types, and select the fault type with the highest probability as the fault prediction result.

[0152] The energy function specifically refers to a scalar value used to describe the state of the network, reflecting the quality of the current network state, which is particularly important in probabilistic neural units. By introducing the concept of energy value, the energy function helps to understand and optimize the weights and biases of the network to minimize the energy and make the network better fit the data distribution; the energy function assigns an energy value to each state of the network, and a lower energy value corresponds to a better state, that is, the network is more in line with the probability distribution of the input data; the energy function is used to define the joint probability distribution of probabilistic neural units, and the probability of a state in the network is inversely proportional to its energy. The lower the energy, the higher the probability of the state; by minimizing the energy function, the weights and biases of the network are adjusted so that the network can more accurately learn the distribution of the input data.

[0153] Take the number of features in the preferred feature set as the number of nodes in the input layer, use multiple stacked probabilistic neural units to construct the hidden layer, and the number of units in each layer can be set according to requirements. Set the number of nodes in the output layer to be the same as the number of fault types, and each node corresponds to a fault type.

[0154] Starting from the input layer, train each layer of probabilistic neural units layer by layer. Take the output of the previous layer of probabilistic neural units as the activation probability of the hidden layer and input it into the input layer of the current layer of probabilistic neural units. In the hidden layer of the current layer of probabilistic neural units, adjust the weights and bias factors by minimizing the energy function. The energy function measures the difference between the current layer's probability distribution and the input data, and generates the probability distribution corresponding to the input data of the current layer of probabilistic neural units; repeat the training until the primary training of all hidden layers is completed.

[0155] Based on the weights and bias factors of each layer of probabilistic neural units obtained from unsupervised primary training, initialize the entire fault prediction network, add a classifier to the output layer, use different fault types as labels, and define a classification loss function to measure the difference between the prediction result and the true label. Preferably, the loss function includes the cross-entropy loss function.

[0156] Select an optimization algorithm, preferably such as the gradient descent method, to minimize the classification loss function. Update the weights and bias factors of each layer in the fault prediction network through backpropagation, and iteratively perform the optimization operation until the classification loss function converges or reaches the preset number of iterations to obtain the trained optimal fault prediction network.

[0157] Input the data of the battery pack to be predicted into the trained optimal fault prediction network, propagate the data forward, pass through the input layer, hidden layer, until the output layer, and obtain the probability distribution corresponding to each fault type at the output layer. Select the fault type with the highest probability value as the final fault prediction result.

[0158] In this embodiment, the use of the optimal feature set and the multi-layer stacked structure enable the model to more effectively extract and represent the key features of the input data, improving the prediction accuracy; the layer-by-layer unsupervised primary training can learn the essential features of the data with less dependence on labels, enhancing the robustness of the model and reducing the prediction errors caused by data noise and anomalies; through staged training and regularization means, the risk of overfitting is reduced, and the generalization ability of the model is improved, making it perform stably under different environments and data conditions; combined with the classifier and optimization algorithm, the fault prediction network can achieve automated fault diagnosis, reduce manual intervention, and improve the prediction efficiency and accuracy.

[0159] In an alternative embodiment, the output of the probability neural unit in the previous layer is the activation probability of the hidden layer, which enters the input layer of the probability neural unit in the current layer. In the hidden layer of the probability neural unit in the current layer, by minimizing the energy function, the weights and bias factors of the probability neural unit in the current layer are adjusted to generate the probability distribution corresponding to the input data of the probability neural unit in the current layer, including:

[0160] The energy function, and its formula is as follows:

[0161]

[0162] where l represents the serial number of the hidden layer of the fault prediction network, v l represents the state of the input layer, h l represents the state of the hidden layer, E(v l , h l ) represents the energy function, W l represents the connection weight matrix between the input layer and the hidden layer, a l represents the bias vector of the input layer, b l represents the bias vector of the hidden layer, U l ij represents the interaction weight between nodes i and j in the input layer, v l i represents the state of the input layer of node i in the input layer, v l j represents the state of the input layer of node j in the input layer, V l km represents the interaction weight between nodes k and m in the hidden layer, h l k represents the state of the hidden layer of node k in the hidden layer, h lm represents the hidden layer state of the hidden layer node m, and λ represents the regularization term balance parameter.

[0163] The formula of the energy function consists of three parts: a linear term, a quadratic term, and a regularization term. The linear term is used to calculate the direct contribution of the energy between the input layer and the hidden layer, as well as their respective biases, describing the direct impact of the input and hidden layer states and their interactions; the quadratic term is used to calculate the quadratic interactions between the nodes in the input layer and the hidden layer, capturing the more complex relationships between the nodes within the layer; the regularization term is used to control the model complexity and prevent overfitting. By introducing the L2 norm of the weights and biases, it ensures that the parameters do not become too large, enabling the model to have better generalization ability.

[0164] According to the formula, the energy function combines the linear term and the quadratic term between the input layer and the hidden layer. By adjusting the weights and biases, it can abstract the feature representation layer by layer, extract the high-order features of the input data, and help the network learn the complex patterns and features in the data; by introducing the non-linear activation function and the quadratic term in the hidden layer, the energy function can model the non-linear relationships in the data, enabling the network to better adapt to the complex data distribution and features; the regularization term in the energy function helps to control the model complexity, avoid overfitting the training data during the training process, and improve the generalization ability of the network; the energy function adjusts the weights and biases through the minimization process to minimize the energy of the model, generate the probability distribution corresponding to the input data, and is realized by optimization algorithms such as gradient descent, which can effectively learn the representation of the data.

[0165] In an optional embodiment, based on the fault prediction network, constructing a distributed prediction network includes:

[0166] Taking a communication battery pack and the corresponding edge computing node as a branch network, and the central server as the central network, to construct a distributed prediction network;

[0167] The central network is constructed based on the fault prediction network, and the central network is distributed to the branch networks; independent data collection and training are performed for each branch network, and the model parameters in each trained branch network are uploaded to the central network. The central network aggregates the model parameters to obtain updated model parameters and distributes them to the branch networks;

[0168] Multiple rounds of iteration are performed in sequence until the preset iteration upper limit is reached, determining the final central network and the corresponding final branch networks, and the final branch networks perform fault prediction.

[0169] Define a branch network, considering each communication battery pack and its corresponding edge computing node as a branch network. Define a central network, with the central server regarded as the central network, which is used to coordinate and aggregate the model parameters of the branch networks. Establish a communication mechanism to create a two-way communication channel between the central network and the branch networks for the transmission and update of model parameters.

[0170] Distribute the central network model constructed based on the fault prediction network to each branch network, and initialize the model parameters in the branch network, including network structure, weights, and bias factors, etc., to ensure that the initial model parameters of all branch networks are consistent with those of the central network.

[0171] In each branch network, independently perform data collection and preprocessing. Input the collected data into the model of the branch network for forward propagation calculation, calculate the loss function, evaluate the performance of the model on the current data, and use the gradient descent method to update the model parameters of the branch network, adjusting the weights and bias factors through backpropagation. ， Perform multiple rounds of iterative training on the model of the branch network until the preset number of iterations or performance metrics are reached.

[0172] After each branch network completes independent training, upload the trained model parameters to the central network. The central network receives the model parameters from all branch networks, aggregates the received model parameters, usually by taking the average or weighted average, to obtain the updated global model parameters, and distributes the updated global model parameters to all branch networks. The branch networks replace the original local model parameters with the received global model parameters.

[0173] Perform multiple rounds of iterative training. In each round of iteration, the branch networks independently train and upload the model parameters, and the central network aggregates the parameters and distributes the updates. The iterative process continues until the preset iteration upper limit is reached.

[0174] After reaching the iteration upper limit, the central network obtains the final global model parameters, distributes the final global model parameters to all branch networks, and each branch network updates its local model with the received global model parameters to obtain the final branch network model.

[0175] In this embodiment, constructing a branch network allows parallel processing of data for each communication battery pack, which means the system can process multiple data streams simultaneously, improving the overall efficiency and processing capacity of the system. A two-way communication channel is established between the central network and the branch network, enabling the transmission and update of model parameters between the central network and the branch network. The co-training method allows the central network to aggregate information from multiple branch networks to obtain a more comprehensive data perspective, improving the accuracy and generalization ability of the model. After the independent training of the branch networks is completed, the central network collects and aggregates the model parameters of all branch networks to generate updated global model parameters. The parameter aggregation mechanism can effectively integrate the training results of each branch network and improve the overall performance of the model. Through the iterative training process, the system can continuously receive new data and model parameters from the branch networks and update the global model based on this information, enabling the model to be continuously optimized as time and data grow and adapt to changing environments and data distributions. The distributed decision-making and feedback mechanism can help monitor and manage the status of communication battery packs in real time and take timely measures to handle potential faults and problems.

[0176] Figure 2 FIG. is a schematic structural diagram of an on-line remote charge and discharge fault prediction system for a communication battery pack according to an embodiment of the present invention, as Figure 2 shown, the system includes:

[0177] A first unit for collecting time series data of the charge and discharge process of a communication battery pack, preprocessing the time series data, extracting complexity features based on multi-scale entropy analysis, determining physical features through a parameter identification algorithm based on the physical model corresponding to the battery pack, performing non-dimensional processing on the complexity features and the physical features, and determining a primary feature set of the charge and discharge process of the battery pack;

[0178] A second unit for constructing a correlation coefficient matrix based on the primary feature set, determining undirected graph weight edges with the correlation coefficient as the weight value, constructing a weighted undirected graph with the primary features as undirected graph nodes; based on the weighted undirected graph, expanding and generating a maximum weight connected subgraph by iteratively calculating the weight edge connectivity of each undirected graph node, and constructing a preferred feature set with the primary features corresponding to each subgraph node in the maximum weight connected subgraph as the preferred features;

[0179] A third unit for inputting the data of the battery pack to be predicted into a fault prediction network, extracting corresponding feature values based on the preferred feature set in the input layer, abstracting feature representations layer by layer through multiple hidden layers, extracting high-order features, and calculating the probability distribution of fault types through the activation function of the output layer, and selecting the fault type corresponding to the highest probability as the fault prediction result.

[0180] In the third aspect of the embodiment of the present invention,

[0181] Provided is an electronic device, comprising:

[0182] a processor;

[0183] a memory for storing instructions executable by the processor;

[0184] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0185] In a fourth aspect of the embodiments of the present invention,

[0186] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0187] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0188] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting online remote charging and discharging faults of a communication battery pack, characterized in that: include: Collecting time series data of the charging and discharging process of the communication battery pack, preprocessing the time series data, extracting complexity features based on multi-scale entropy analysis, determining physical features through a parameter identification algorithm based on a physical model corresponding to the battery pack, normalizing the complexity features and the physical features into a dimensionless process, and determining a primary feature set of the charging and discharging process of the battery pack; Based on the primary feature set, a correlation coefficient matrix is ​​constructed, and the weighted edges of the undirected graph are determined using the correlation coefficient as the weight value. The primary features are used as the undirected graph nodes to construct a weighted undirected graph. Based on the weighted undirected graph, by iteratively calculating the weighted edge connectivity sum of each undirected graph node, a maximum weight connected subgraph is generated by expansion, and the primary features corresponding to each subgraph node in the maximum weight connected subgraph are taken as preferred features to construct a preferred feature set; The battery group data to be predicted is input into the fault prediction network. The corresponding feature values ​​are extracted based on the preferred feature set in the input layer. The feature representation is abstracted layer by layer through multiple hidden layers, and high-order features are extracted. The probability distribution of the fault type is calculated through the output layer activation function, and the fault type corresponding to the highest probability is selected as the fault prediction result.

2. The method according to claim 1, characterized in that Preprocess the time series data, extract the complexity features based on multi-scale entropy analysis, determine the physical features through parameter identification algorithm based on the physical model corresponding to the battery pack, normalize the complexity features and the physical features into dimensionless features, and determine that the primary feature set of the battery pack charging and discharging process includes: Detect and remove outliers from time series data, and obtain standard time series data through smoothing filtering and normalization. Performing a coarse-graining operation on the standard time series data to obtain scale subsequences at different scales, calculating sample entropy for each scale subsequence, determining a multi-scale entropy curve, extracting slope features, curvature features, entropy value features and area features from the multi-scale entropy curve, and constructing complexity features; Based on the electrochemical characteristics and thermodynamic behavior of the battery pack, the corresponding physical model is constructed. The physical model parameters are evaluated and the physical characteristics are determined through the parameter identification algorithm based on the recursive least squares method. The complexity characteristics and the physical characteristics are normalized and dimensionless, the numerical range and dimension differences are eliminated, and a primary feature set is determined.

3. The method according to claim 1, characterized in that Based on the primary feature set, a correlation coefficient matrix is ​​constructed, and the weighted edges of the undirected graph are determined using the correlation coefficient as the weight value. The primary features are used as the undirected graph nodes to construct a weighted undirected graph. Based on the weighted undirected graph, the weighted edge connectivity sum of each undirected graph node is iteratively calculated to expand and generate a maximum weight connected subgraph, and the primary feature corresponding to each connected node in the maximum weight connected subgraph is used as the preferred feature. Constructing the preferred feature set includes: Based on the primary feature set, the correlation coefficient is determined by calculating the covariance and variance between every two primary features. If the absolute value of the correlation coefficient is greater than a preset correlation coefficient threshold, the corresponding correlation coefficient is retained, otherwise it is removed, and a correlation coefficient matrix is ​​constructed based on the retained correlation coefficients; Taking each primary feature in the primary feature set as an undirected graph node, traversing each correlation coefficient in the correlation coefficient matrix, taking the correlation coefficient as a weight value, determining an undirected graph weight edge between two corresponding undirected graph nodes, and constructing a weighted undirected graph; Traverse each undirected graph node in the weighted undirected graph, calculate the undirected weight sum of all undirected graph weight edges connected to the undirected graph node, select the undirected graph node corresponding to the maximum value of the undirected weight sum as the first connected node, initialize the connected subgraph, and start iterative expansion of the connected subgraph: Traversing each connected node of the connected subgraph, determining a corresponding undirected graph node from the weighted undirected graph, finding all undirected graph nodes that are directly connected to the undirected graph node and are not in the connected subgraph, determining a candidate node set, calculating a weighted edge connectivity sum in the weighted undirected graph for each candidate node in the candidate node set and all connected nodes in the connected subgraph, selecting a candidate node corresponding to the maximum value of the weighted edge connectivity sum as a new connected node, and adding it to the connected subgraph; Until the candidate node set is empty, a maximum weight connected subgraph is generated; The primary feature corresponding to each connected node in the maximum weight connected subgraph is taken as the preferred feature to construct a preferred feature set.

4. The method according to claim 3, characterized in that: The construction process of the weighted undirected graph also includes: Based on the correlation coefficient matrix and a preset correlation coefficient threshold, binarizing the correlation coefficient matrix to determine a binarized correlation matrix; Based on the binary correlation matrix, the Euclidean distance between every two features is calculated, a feature distance matrix is ​​determined, the maximum value and the minimum value in the feature distance matrix are extracted, the feature distance ratio is determined, and a diversity measurement matrix is ​​generated; Each primary feature in the primary feature set is taken as an undirected graph node, each correlation coefficient in the correlation coefficient matrix and each diversity metric in the diversity metric matrix are traversed, a weight value is calculated through the correlation coefficient and the diversity metric, an undirected graph weight edge between two corresponding undirected graph nodes is determined, and a weighted undirected graph is constructed.

5. The method according to claim 1, characterized in that The battery group data to be predicted is input into the fault prediction network. Based on the preferred feature set, the corresponding feature values ​​are extracted in the input layer. The feature representation is abstracted layer by layer through multiple hidden layers, and high-order features are extracted. The fault type probability distribution is calculated through the output layer activation function, and the fault type corresponding to the highest probability is selected as the fault prediction result, including: The fault prediction network is constructed based on a deep belief network, and includes an input layer, at least one hidden layer and an output layer, wherein the number of nodes in the input layer is equal to the number of features in the preferred feature set, the hidden layer is composed of multiple layers of stacked probabilistic neural units, and the number of nodes in the output layer corresponds one-to-one to the fault type; Based on the unsupervised learning method, primary training is performed layer by layer on each layer of the probabilistic neural unit of the fault prediction network. The probabilistic neural unit of the previous layer outputs the activation probability of the hidden layer, which enters the input layer of the probabilistic neural unit of the current layer. In the hidden layer of the probabilistic neural unit of the current layer, the weight and bias factor of the probabilistic neural unit of the current layer are adjusted by minimizing the energy function to generate the probability distribution corresponding to the input data of the probabilistic neural unit of the current layer. When the primary training of all the probabilistic neural units is completed, the fault prediction network is secondary adjusted based on a supervised learning method, and the fault prediction network is initialized based on the weights and bias factors corresponding to each layer of the probabilistic neural units determined by the primary training; Adding a classifier to the output layer, taking different fault types as labels, minimizing the classification loss function through an optimization algorithm, updating the weight and bias factor of each layer in the fault prediction network, and obtaining an optimal fault prediction network; The battery group data to be predicted is input into the optimal fault prediction network, the fault type probability distribution is input, and the fault type with the highest probability is selected as the fault prediction result.

6. The method according to claim 5, characterized in that The previous layer of probabilistic neural units outputs the hidden layer activation probability and enters the input layer of the current layer of probabilistic neural units. In the hidden layer of the current layer of probabilistic neural units, the weights and bias factors of the current layer of probabilistic neural units are adjusted by minimizing the energy function to generate the probability distribution corresponding to the input data of the current layer of probabilistic neural units: The energy function has the following formula: Where l represents the hidden layer number of the fault prediction network, v l represents the input layer state, h l represents the hidden layer state, E(v l ,h l ) represents the energy function, W l represents the connection weight matrix between the input layer and the hidden layer, a l represents the input layer bias vector, b l represents the hidden layer bias vector, U l ij represents the interaction weight between node i and node j in the input layer, v l i represents the input layer state of input layer node i, v l j represents the input layer state of input layer node j, V l km represents the interaction weight between node k and node m in the hidden layer, h l k represents the hidden layer state of hidden layer node k, h l m represents the hidden layer state of the hidden layer node m, and λ represents the regularization term balance parameter.

7. The method according to claim 5, characterized in that Based on the fault prediction network, constructing a distributed prediction network includes: A distributed prediction network is constructed with a communication battery pack and corresponding edge computing nodes as branch networks and a central server as the core network; The central network is constructed based on the fault prediction network, and the central network is distributed to the branch networks; data collection and training are performed independently on each branch network, and the model parameters in each branch network after training are uploaded to the central network. The central network aggregates the model parameters to obtain updated model parameters, and distributes them to the branch networks; Multiple rounds of iterations are performed in sequence until a preset iteration upper limit is reached, and a final central network and a corresponding final branch network are determined, and the final branch network performs fault prediction.

8. A communication battery pack online remote charging and discharging fault prediction system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect time series data of the charging and discharging process of the communication battery pack, pre-process the time series data, extract complexity features based on multi-scale entropy analysis, determine physical features through parameter identification algorithm based on the physical model corresponding to the battery pack, normalize the complexity features and the physical features into dimensionless processing, and determine the primary feature set of the charging and discharging process of the battery pack; The second unit is used to construct a correlation coefficient matrix based on the primary feature set, determine the weight edge of the undirected graph with the correlation coefficient as the weight value, and construct a weighted undirected graph with the primary features as the undirected graph nodes; Based on the weighted undirected graph, by iteratively calculating the weighted edge connectivity sum of each undirected graph node, a maximum weight connected subgraph is generated by expansion, and the primary features corresponding to each subgraph node in the maximum weight connected subgraph are taken as preferred features to construct a preferred feature set; The third unit is used to input the battery group data to be predicted into the fault prediction network, extract the corresponding feature values ​​based on the preferred feature set in the input layer, abstract the feature representation layer by layer through multiple hidden layers, extract high-order features, and calculate the probability distribution of fault types through the output layer activation function, and select the fault type corresponding to the highest probability as the fault prediction result.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images