Electric vehicle energy storage battery fault diagnosis and predictive maintenance control method and system

Abstract

The invention provides an electric vehicle energy storage battery fault diagnosis and predictive maintenance control method and system, and relates to the technical field of fault diagnosis, and the method comprises the steps: collecting the real-time working condition data and environment parameters of an electric vehicle energy storage battery, and carrying out the data preprocessing through a cloud server; carrying out life prediction by utilizing a multi-task deep learning framework, and generating a comprehensive state evaluation report; constructing a multi-target maintenance decision optimization model based on the evaluation report, and predicting an optimal maintenance time window; historical maintenance data and fault diagnosis experience are fused, and a maintenance operation list is generated; and performing resource scheduling optimization through a mixed integer programming algorithm to form an optimal battery maintenance scheme.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a method and system for fault diagnosis and predictive maintenance control of an electric vehicle energy storage battery. Background Art

[0002] Existing methods usually only consider a single battery parameter or environmental factor, which is difficult to fully reflect the actual working status of the battery. Since battery performance is affected by a combination of factors, the analysis of a single parameter can easily lead to inaccurate diagnostic results and cannot effectively predict battery life.

[0003] Traditional battery status assessment models mostly use linear or simple nonlinear models, which are difficult to accurately capture the complex changes in battery performance over time and usage environment, resulting in a large deviation between the life prediction results and the actual situation, and cannot provide a reliable basis for maintenance decisions.

[0004] Existing maintenance decision-making methods are often based on fixed cycles or threshold triggers, and lack consideration of multi-objective optimization. These methods make it difficult to balance multiple objectives such as maintenance cost, efficiency, and battery life, and are unable to achieve optimal resource allocation or dynamically adjust maintenance strategies based on actual conditions.

[0005] Therefore, there is an urgent need for a method that can comprehensively consider multi-source data and accurately predict battery life to improve the reliability and economy of electric vehicles. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for fault diagnosis and predictive maintenance control of an electric vehicle energy storage battery, which can at least solve some of the problems existing in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for fault diagnosis and predictive maintenance control of an electric vehicle energy storage battery, comprising: Collecting real-time operating data of the tram energy storage battery and tram environmental parameters, transmitting the real-time operating data and tram environmental parameters to the on-board controller through the on-board CAN bus, uploading to the cloud server in combination with the wireless communication module, the cloud server performs data cleaning through the improved Tukey-Kramerdo multiple comparison, normalizing the cleaned data, unifying the normalized data of different dimensions into the same scale range, performing nonlinear dimensionality reduction through the improved isometric mapping algorithm, extracting key characteristic parameters from the parameters after dimensionality reduction, and obtaining pre-processed real-time operating data and pre-processed tram environmental parameters; The preprocessed real-time operating condition data and the preprocessed tram environmental parameters are added as input data to a pre-set multi-task deep learning framework, wherein the multi-task deep learning framework includes a feature fusion layer, a timing analysis layer and a prediction layer, the feature fusion layer dynamically weights the input data according to the attention mechanism to construct a fusion vector, the timing analysis layer performs timing modeling on the fusion vector according to a bidirectional long short-term memory network, extracts the performance evolution law of the energy storage battery, and based on the performance evolution law of the energy storage battery, performs life prediction through the prediction layer to generate an initial life prediction result, optimizes the parameters of the multi-task deep learning framework through a multi-objective optimization strategy based on focus loss and an online learning mechanism based on a sliding time window, performs life prediction again according to the updated multi-task deep learning framework, and obtains a comprehensive status evaluation report corresponding to the current energy storage battery; A multi-objective maintenance decision optimization model is constructed based on the comprehensive status assessment report. The life prediction results, vehicle operation plan and maintenance cost constraints are used as the status. A dual evaluation is performed through the policy network and the value network to predict the optimal maintenance time window. The pre-acquired historical maintenance data and fault diagnosis experience are deeply integrated. The maintenance knowledge features are extracted and a maintenance operation list is generated in combination with the improved graph attention mechanism. Based on the maintenance operation list, according to the improved mixed integer programming algorithm, global optimization configuration is performed with minimization of maintenance cost and maximization of maintenance efficiency as the optimization goals, and a resource scheduling plan is generated. The optimal maintenance time window, the maintenance operation list and the resource scheduling plan are integrated to form an optimal battery maintenance plan and send it to the on-board controller for visual display.

[0008] In an optional embodiment, The real-time working condition data of the tram energy storage battery and the tram environmental parameters are collected, and the real-time working condition data and the tram environmental parameters are transmitted to the on-board controller through the on-board CAN bus, and uploaded to the cloud server in combination with the wireless communication module. The cloud server performs data cleaning through the improved Tukey-Kramerdo multiple comparison, normalizes the cleaned data, unifies the normalized data of different dimensions into the same scale range, performs nonlinear dimensionality reduction through the improved isometric mapping algorithm, extracts the key characteristic parameters in the parameters after dimensionality reduction, and obtains the pre-processed real-time working condition data and pre-processed tram environmental parameters, including: The battery temperature data is collected by a graphene-based flexible temperature sensor arranged on each battery module, the battery voltage data is collected by a voltage sensor arranged at the positive and negative terminals of the battery, the battery current data is collected by a Hall effect current sensor connected in series in the main circuit of the battery, the battery strain data is collected by a piezoelectric material vibration strain sensor arranged at a key position of the battery, and the gas composition data inside the battery pack is collected by a micro-spectral sensor arranged on the top of the battery pack, and the battery temperature data, battery voltage data, battery current data, battery strain data and battery pack internal gas composition data are combined to obtain the real-time operating condition data; Ambient temperature data is collected by dual backup ambient temperature sensors arranged outside the battery compartment, ambient humidity data is collected by an ambient humidity sensor integrated in the same module as the dual backup ambient temperature sensors, altitude data is collected by a MEMS air pressure sensor, omnidirectional vibration data is collected by a nine-axis inertial measurement unit arranged at the bottom of the battery pack, and ambient temperature data, ambient humidity data, altitude data and omnidirectional vibration data are integrated to obtain the environmental parameters of the electric vehicle; The real-time operating condition data and the tram environment parameters are transmitted to the on-board controller via the on-board CAN bus, and the on-board controller uploads the real-time operating condition data and the tram environment parameters to the cloud server via a pre-set wireless communication module; The cloud server performs anomaly detection on the real-time operating condition data and the tram environmental parameters through a variational autoencoder unsupervised learning model, performs data cleaning on the data after anomaly detection according to the Tukey-Kramer multiple comparison method combined with a temporal attention mechanism, establishes a multi-scale data association analysis of data mutation analysis, trend analysis, and periodic analysis, combines an adaptive time window that dynamically adjusts the window size according to the data change rate to process the abnormal data, obtains cleaned data, and normalizes the cleaned data to unify data of different dimensions within the same scale range to obtain normalized data; An attention mechanism is used to dynamically weight the importance of different parameter features in the normalized data to obtain weighted data. The weighted data is subjected to nonlinear dimensionality reduction according to an isometric mapping algorithm enhanced by multi-task learning. The Pearson correlation coefficient between each characteristic parameter in the reduced dimensionality data and the battery capacity decay rate is calculated based on a correlation analysis method. A feature importance ranking is established and the parameters in the top 10% are selected as key characteristic parameters to obtain preprocessed real-time operating condition data and preprocessed tram environmental parameters.

[0009] In an optional embodiment, The Pearson correlation coefficient between each characteristic parameter in the data after dimensionality reduction and the battery capacity decay rate is calculated based on the correlation analysis method as shown in the following formula: ; Where R represents the Pearson correlation coefficient, x i Indicates i The eigenvalues ​​after dimensionality reduction, represents the median of the eigenvalue after dimensionality reduction, y i The rate at which the battery capacity decays i Sample values, It represents the median of the battery capacity attenuation rate, and IQR() represents the interquartile range, which is used to measure the degree of dispersion of the data. α Represents the smoothing factor.

[0010] In an optional embodiment, The preprocessed real-time operating condition data and the preprocessed tram environment parameters are added as input data to a pre-set multi-task deep learning framework, wherein the multi-task deep learning framework includes a feature fusion layer, a timing analysis layer and a prediction layer. The feature fusion layer dynamically weights the input data according to the attention mechanism to construct a fusion vector. The timing analysis layer performs timing modeling on the fusion vector according to a bidirectional long short-term memory network. Extracting the evolution law of energy storage battery performance includes: The preprocessed real-time working condition data and the preprocessed tram environment parameters are unified to a sampling frequency of 20 Hz by a linear interpolation method, a data matrix is ​​generated and added as input data to a pre-set multi-task deep learning framework, and a feature fusion layer in the multi-task deep learning framework is set with eight attention heads according to a multi-head attention mechanism, each attention head learns the correlation between different dimensions, and a corresponding attention weight is set for each dimension of the input data and weighted fusion is performed to generate a fusion vector; The fusion vector is input to the timing analysis layer, which includes a two-layer bidirectional long short-term memory network structure. The first layer of the bidirectional long short-term memory network divides the fusion vector into time segments with a length of 50, and extracts timing features for each of the time segments. The second layer of the bidirectional long short-term memory network receives the output sequence of the first layer of the bidirectional long short-term memory network. The forward long short-term memory network processes the time segments in positive time order to extract long-term dependency features, and the reverse long short-term memory network processes the time segments in reverse time order to extract short-term fluctuation features. The long-term dependency features and the short-term fluctuation features are concatenated to obtain a timing feature vector with a dimension of 512. Based on the timing feature vector, the performance evolution law of the energy storage battery is extracted by a statistical method, and the battery performance evolution law is divided into multiple data segments with the same time interval. A performance degradation threshold is set and the performance degradation amount of the energy storage battery in each data segment is determined. If the performance degradation amount is greater than the performance degradation threshold, the current data segment is marked as a faulty data segment.

[0011] In an optional embodiment, Based on the performance evolution law of the energy storage battery, life prediction is performed through the prediction layer to generate an initial life prediction result. The multi-objective optimization strategy based on focus loss and the online learning mechanism based on the sliding time window are used to optimize the parameters of the multi-task deep learning framework. Life prediction is performed again according to the updated multi-task deep learning framework to obtain a comprehensive status evaluation report corresponding to the current energy storage battery, including: The pre-acquired time series feature vector is input into the prediction layer, and the prediction layer performs initial life prediction through a three-layer fully connected network structure, wherein the first layer takes the time series feature vector and the performance evolution law of the energy storage battery as input, the input dimension is 512, and obtains a 256-dimensional output through the ReLU activation function, the second layer maps the 256-dimensional output to a 128-dimensional feature through the tanh activation function, and the third layer maps the 128-dimensional feature to a 64-dimensional feature through the sigmoid activation function, and generates the initial life prediction result corresponding to the energy storage battery according to the 64-dimensional feature, wherein the initial life prediction result includes the initial remaining cycle life prediction value, the initial capacity attenuation prediction value and the initial internal resistance growth prediction value of the energy storage battery; A multi-objective optimization strategy based on focus loss is used to calculate the first loss function. Combined with the initial life prediction result, a loss weight of 0.4 is set for the initial remaining cycle life prediction value, and a loss weight of 0.3 is set for the initial capacity decay prediction value and the initial internal resistance growth prediction value. Online learning is performed based on a sliding time window mechanism. The size of the sliding time window is 2000 data points, and adjacent windows overlap by 500 data points. If the average prediction error of any prediction value in three consecutive sliding time windows exceeds 5%, the loss weight of the corresponding prediction value is increased by 0.1. The network parameters of the deep learning framework are updated online according to the first loss function, and the life prediction is performed again using the updated multi-task deep learning framework to obtain the optimized remaining cycle life prediction value, capacity decay prediction value and internal resistance growth prediction value, which are combined to obtain a comprehensive status assessment report corresponding to the current energy storage battery.

[0012] In an optional embodiment, The network parameters of the deep learning framework are updated online according to the first loss function as shown in the following formula: ; in, i t+1 Indicates that at time step t Model parameters at +1, i t Represents the time step t The model parameters when or 0 represents the initial learning rate, exp represents the exponential function, β represents the attenuation factor, k Indicates the size of the time window, Indicates that at time step j When i j The corresponding dataset D j The loss function value on , ▽ θ represents the gradient, Indicates that at time step t When , the loss function is about the model parameters i t gradient.

[0013] In an optional embodiment, Based on the comprehensive status assessment report, a multi-objective maintenance decision optimization model is constructed. The life prediction results, vehicle operation plan and maintenance cost constraints are used as the status. A dual evaluation is performed through the policy network and the value network to predict the optimal maintenance time window. The pre-acquired historical maintenance data and fault diagnosis experience are deeply integrated. The maintenance knowledge features are extracted and a maintenance operation list is generated in combination with the improved graph attention mechanism. Based on the maintenance operation list, according to the improved mixed integer programming algorithm, global optimization configuration is performed with maintenance cost minimization and maintenance efficiency maximization as the optimization goals, and a resource scheduling plan is generated. The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form the optimal battery maintenance plan and send it to the vehicle controller for visual display, including: Acquire energy storage battery performance characteristics, operation plan characteristics and maintenance constraint characteristics, extract long-range evolution trends of the energy storage battery performance characteristics through a time series convolutional network to obtain a first coding feature, convert the operation plan characteristics into a binary sequence through one-hot encoding to obtain a second coding feature, normalize the maintenance constraint characteristics to obtain a third coding feature, and concatenate the first coding feature, the second coding feature and the third coding feature to form a state vector; The state vector is input into the policy network and the value network. The policy network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer. The first hidden layer uses a ReLU activation function. The second hidden layer fuses neuron outputs through a residual connection. The third hidden layer uses a tanh activation function. The policy network outputs an action probability distribution. The value network outputs an evaluation score. The time period with the highest evaluation score is selected as the optimal maintenance time window. Retrieve historical maintenance records, which include fault type information, maintenance measures information, and maintenance effect information, establish a maintenance knowledge graph in combination with diagnostic rules in the expert experience knowledge base, use a multi-head graph attention mechanism to process the maintenance knowledge graph, fuse adjacent node information layer by layer through multi-layer attention propagation, extract the maintenance knowledge features according to the attention weight, and generate a maintenance operation list; The maintenance operation list is converted into a resource requirement list, wherein the resource requirement list includes the number of maintenance personnel, the type of equipment, and the number of spare parts. A mixed integer programming algorithm is used to optimize the resource requirement list, with the minimum total maintenance cost and the maximum maintenance efficiency as the optimization goals. The constraints of the mixed integer programming algorithm include the number of maintenance personnel, the equipment utilization rate, and the spare parts inventory safety margin constraint, and a resource scheduling solution is obtained by solving the problem. The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form the optimal battery maintenance plan and sent to the on-board controller for visual display.

[0014] A second aspect of an embodiment of the present invention provides a fault diagnosis and predictive maintenance control system for an electric vehicle energy storage battery, comprising: The first unit is used to collect the real-time working condition data of the tram energy storage battery and the tram environmental parameters, transmit the real-time working condition data and the tram environmental parameters to the on-board controller through the on-board CAN bus, and upload them to the cloud server in combination with the wireless communication module. The cloud server performs data cleaning through the improved Tukey-Kramerdo multiple comparison, normalizes the cleaned data, unifies the normalized data of different dimensions into the same scale range, performs nonlinear dimensionality reduction through the improved isometric mapping algorithm, extracts the key characteristic parameters in the parameters after dimensionality reduction, and obtains the pre-processed real-time working condition data and the pre-processed tram environmental parameters; The second unit is used to add the preprocessed real-time operating condition data and the preprocessed tram environmental parameters as input data to a pre-set multi-task deep learning framework, wherein the multi-task deep learning framework includes a feature fusion layer, a timing analysis layer and a prediction layer, the feature fusion layer dynamically weights the input data according to the attention mechanism to construct a fusion vector, the timing analysis layer performs timing modeling on the fusion vector according to a bidirectional long short-term memory network, extracts the performance evolution law of the energy storage battery, and based on the performance evolution law of the energy storage battery, performs life prediction through the prediction layer to generate an initial life prediction result, optimizes the parameters of the multi-task deep learning framework through a multi-objective optimization strategy based on focus loss and an online learning mechanism based on a sliding time window, performs life prediction again according to the updated multi-task deep learning framework, and obtains a comprehensive status evaluation report corresponding to the current energy storage battery; The third unit is used to build a multi-objective maintenance decision optimization model based on the comprehensive status assessment report, with life prediction results, vehicle operation plan and maintenance cost constraints as states, and perform dual evaluation through policy network and value network to predict the optimal maintenance time window, deeply integrate the pre-acquired historical maintenance data and fault diagnosis experience, and extract maintenance knowledge features and generate a maintenance operation list in combination with the improved graph attention mechanism. Based on the maintenance operation list, according to the improved mixed integer programming algorithm, global optimization configuration is performed with minimization of maintenance cost and maximization of maintenance efficiency as optimization goals, and a resource scheduling plan is generated. The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form an optimal battery maintenance plan and send it to the on-board controller for visual display.

[0015] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0016] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0017] In the present invention, nonlinear dimensionality reduction is performed through normalization processing and improved isometric mapping algorithm to extract key feature parameters, reduce data redundancy, reduce computational complexity, and provide high-quality input data for subsequent deep learning models. Through multi-task learning, feature fusion, timing analysis and life prediction tasks are optimized simultaneously to improve the comprehensive performance and generalization ability of the model. The online learning mechanism enables the model to continuously learn and optimize, timely capture the degradation trend of battery performance, and generate a more accurate and reliable comprehensive status assessment report. Finally, an optimal battery maintenance plan including an optimal maintenance time window, a maintenance operation list and a resource scheduling plan is formed, and a visual display is performed through the on-board controller to provide intuitive and operational guidance for maintenance decisions. In summary, this method realizes the intelligent and precise life prediction and maintenance decision-making of electric vehicle energy storage batteries, has strong practicality and promotion value, and is of great significance for promoting the health management and intelligent maintenance of electric vehicle energy storage batteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic flow chart of a method for fault diagnosis and predictive maintenance control of an electric vehicle energy storage battery according to an embodiment of the present invention; Figure 2 The present invention is a schematic structural diagram of a fault diagnosis and predictive maintenance control system for an electric vehicle energy storage battery according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0021] Figure 1FIG. 1 is a flow chart of a method for fault diagnosis and predictive maintenance control of an electric vehicle energy storage battery according to an embodiment of the present invention. Figure 1 As shown, the method includes: S1. Collect the real-time working condition data of the tram energy storage battery and the tram environmental parameters, transmit the real-time working condition data and the tram environmental parameters to the on-board controller through the on-board CAN bus, and upload them to the cloud server in combination with the wireless communication module. The cloud server performs data cleaning through the improved Tukey-Kramerdo multiple comparison, normalizes the cleaned data, unifies the normalized data of different dimensions into the same scale range, performs nonlinear dimensionality reduction through the improved isometric mapping algorithm, extracts the key characteristic parameters in the parameters after dimensionality reduction, and obtains the pre-processed real-time working condition data and the pre-processed tram environmental parameters; The Tukey-Kramer multiple comparison is a statistical method that combines graph structure with the Kramer test method, aiming to optimize the multiple hypothesis testing process to reduce false positives. The isometric mapping algorithm is a nonlinear dimensionality reduction method that reduces dimensionality by maintaining the geodesic distance between high-dimensional data points, thereby retaining the global structure of the data.

[0022] In an optional embodiment, The real-time working condition data of the tram energy storage battery and the tram environmental parameters are collected, and the real-time working condition data and the tram environmental parameters are transmitted to the on-board controller through the on-board CAN bus, and uploaded to the cloud server in combination with the wireless communication module. The cloud server performs data cleaning through the improved Tukey-Kramerdo multiple comparison, normalizes the cleaned data, unifies the normalized data of different dimensions into the same scale range, performs nonlinear dimensionality reduction through the improved isometric mapping algorithm, extracts the key characteristic parameters in the parameters after dimensionality reduction, and obtains the pre-processed real-time working condition data and pre-processed tram environmental parameters, including: The battery temperature data is collected by a graphene-based flexible temperature sensor arranged on each battery module, the battery voltage data is collected by a voltage sensor arranged at the positive and negative terminals of the battery, the battery current data is collected by a Hall effect current sensor connected in series in the main circuit of the battery, the battery strain data is collected by a piezoelectric material vibration strain sensor arranged at a key position of the battery, and the gas composition data inside the battery pack is collected by a micro-spectral sensor arranged on the top of the battery pack, and the battery temperature data, battery voltage data, battery current data, battery strain data and battery pack internal gas composition data are combined to obtain the real-time operating condition data; Ambient temperature data is collected by dual backup ambient temperature sensors arranged outside the battery compartment, ambient humidity data is collected by an ambient humidity sensor integrated in the same module as the dual backup ambient temperature sensors, altitude data is collected by a MEMS air pressure sensor, omnidirectional vibration data is collected by a nine-axis inertial measurement unit arranged at the bottom of the battery pack, and ambient temperature data, ambient humidity data, altitude data and omnidirectional vibration data are integrated to obtain the environmental parameters of the electric vehicle; The real-time operating condition data and the tram environment parameters are transmitted to the on-board controller via the on-board CAN bus, and the on-board controller uploads the real-time operating condition data and the tram environment parameters to the cloud server via a pre-set wireless communication module; The cloud server performs anomaly detection on the real-time operating condition data and the tram environmental parameters through a variational autoencoder unsupervised learning model, performs data cleaning on the data after anomaly detection according to the Tukey-Kramer multiple comparison method combined with a temporal attention mechanism, establishes a multi-scale data association analysis of data mutation analysis, trend analysis, and periodic analysis, combines an adaptive time window that dynamically adjusts the window size according to the data change rate to process the abnormal data, obtains cleaned data, and normalizes the cleaned data to unify data of different dimensions within the same scale range to obtain normalized data; An attention mechanism is used to dynamically weight the importance of different parameter features in the normalized data to obtain weighted data. The weighted data is subjected to nonlinear dimensionality reduction according to an isometric mapping algorithm enhanced by multi-task learning. The Pearson correlation coefficient between each characteristic parameter in the reduced dimensionality data and the battery capacity decay rate is calculated based on a correlation analysis method. A feature importance ranking is established and the parameters in the top 10% are selected as key characteristic parameters to obtain preprocessed real-time operating condition data and preprocessed tram environmental parameters.

[0023] The nine-axis inertial measurement unit (9-axis IMU) is a sensor that includes an accelerometer, a gyroscope, and a magnetometer, and is used to provide information about the motion, posture, and direction of an object. The temporal attention mechanism is a neural network method for processing time series data. By calculating the attention weights at different time points, it helps the model focus on key time steps in the sequence. The Pearson correlation coefficient is a statistic that measures the linear correlation between two variables, with values ​​ranging from -1 to 1, indicating complete negative correlation, no correlation, or complete positive correlation.

[0024] Various sensors are deployed in the battery system of the electric vehicle to collect real-time operating data and environmental parameters. A flexible temperature sensor based on graphene material is installed on each battery module to collect battery temperature data. The flexible temperature sensor has good flexibility and high sensitivity, and can fit the battery surface to achieve accurate temperature measurement. Voltage sensors are installed at the positive and negative terminals of the battery to collect battery voltage data. A high-precision voltage sensor is used with a range of 0-5V and an accuracy of up to 0.1%. A Hall effect current sensor is connected in series in the main circuit of the battery to collect battery current data. A closed-loop Hall current sensor is selected with a range of ±500A and an accuracy of 0.5%. Piezoelectric material vibration strain sensors are installed at key positions of the battery to collect battery strain data. Piezoelectric material vibration strain sensors have high sensitivity and can detect tiny deformations with a range of 0-1000με. A micro-spectral sensor is installed on the top of the battery pack to collect gas composition data inside the battery pack. The sensor can detect multiple gases at the same time, including H2, CO, CH4, etc. A dual backup ambient temperature sensor is installed on the outside of the battery compartment to collect ambient temperature data. Two independent temperature sensors are used to back up each other to improve reliability. The ambient humidity sensor is integrated in the same module as the temperature sensor to collect ambient humidity data. A MEMS air pressure sensor is installed to collect altitude data. A nine-axis inertial measurement unit is installed at the bottom of the battery pack to collect all-round vibration data. The unit integrates a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer. The above real-time working condition data and environmental parameters are transmitted to the vehicle controller through the vehicle CAN bus. The CAN bus adopts differential signal transmission and has strong anti-interference ability. After receiving the data, the vehicle controller uploads the data to the cloud server through the pre-set 4G / 5G wireless communication module. The wireless module adopts high-speed data transmission technology to ensure real-time data upload; The cloud server uses a variational autoencoder unsupervised learning model to detect anomalies in real-time operating data and tram environmental parameters. The variational autoencoder identifies anomalies through reconstruction errors, and sets the threshold to 3 times the mean of the reconstruction error. The data is cleaned based on the Tukey-Kramer multiple comparison method combined with the temporal attention mechanism. The temporal attention mechanism assigns different weights according to the time correlation of the data, highlights the data features at important moments, and performs multi-scale data association analysis, including data mutation analysis, trend analysis, and periodicity analysis. The mutation threshold is set to 5 times the standard deviation of the mean for data mutation analysis. The sliding window method is used for trend analysis, and the window size is 100 data points. The periodicity analysis uses Fourier transform to detect periodic components. The window size is dynamically adjusted according to the data change rate to achieve adaptive time window processing. When the data change rate exceeds 20%, the window is reduced to 50 data points. When the change rate is less than 5%, the window is increased to 200 data points. The cleaned data is normalized and the maximum and minimum value normalization method is used to unify the data of different dimensions into the range of 0-1. The attention mechanism is used to dynamically weight the importance of different parameter features in the normalized data. The attention weight is dynamically adjusted according to the correlation between the parameter and the target variable. The higher the correlation, the greater the weight. The improved isometric mapping algorithm is used to perform nonlinear dimensionality reduction on the weighted data. The improvement lies in the introduction of multi-task learning and the simultaneous optimization of dimensionality reduction and feature selection. After dimensionality reduction, the correlation between each characteristic parameter and the battery capacity attenuation rate is calculated based on the Pearson correlation coefficient, and the top 10% parameters with the largest absolute value of the correlation coefficient are selected as key characteristic parameters.

[0025] For example, for electric car A, one hour of data is collected, including 20 parameters such as battery temperature, voltage, and current, with one data point per second, for a total of 3,600 data points. After anomaly detection, 50 anomalies are identified, and 3,550 valid data points are retained after data cleaning. After normalization, all parameter values ​​are mapped to the range of 0-1. Through the attention mechanism, parameters such as battery temperature and current are given higher weights, and nonlinear dimensionality reduction reduces the 20-dimensional data to 8 dimensions. Finally, two key feature parameters such as battery temperature, current, and ambient temperature are selected.

[0026] In this embodiment, real-time operating data such as battery temperature, voltage, current, strain and internal gas composition are comprehensively collected to provide rich data support for subsequent battery status analysis. The collected real-time operating data and tram environmental parameters are quickly transmitted to the on-board controller and the cloud server to achieve efficient data transmission and centralized storage, which is convenient for subsequent data processing and analysis. Multi-scale data association analysis is established, and an adaptive time window is used to process abnormal data to effectively improve data quality, laying the foundation for subsequent analysis. The Pearson correlation coefficient of each characteristic parameter and the battery capacity decay rate is calculated through correlation analysis, and a feature importance ranking is established to provide high-quality feature data for subsequent battery health status assessment and remaining service life prediction. In summary, this embodiment can obtain high-quality, low-redundancy, and information-rich battery operation data, provide reliable data support for subsequent tasks such as battery status monitoring, fault diagnosis, and life prediction, and improve the performance and efficiency of the entire battery management system.

[0027] In an optional embodiment, The Pearson correlation coefficient between each characteristic parameter in the data after dimensionality reduction and the battery capacity decay rate is calculated based on the correlation analysis method as shown in the following formula: ; Where R represents the Pearson correlation coefficient, x i Indicates i The eigenvalues ​​after dimensionality reduction, represents the median of the eigenvalue after dimensionality reduction, y i The rate at which the battery capacity decays i Sample values, It represents the median of the battery capacity attenuation rate, and IQR() represents the interquartile range, which is used to measure the degree of dispersion of the data. α Represents the smoothing factor.

[0028] In this embodiment, by calculating the Pearson correlation coefficient between the characteristic parameters after dimensionality reduction and the battery capacity decay rate, the degree of linear correlation can be quantified. The introduction of the interquartile range as a measure of the degree of data discreteness can reflect the degree of data concentration. The adaptive adjustment strategy can improve the adaptability and accuracy of the correlation calculation. The selection of key characteristic parameters can effectively screen out the key factors that have the greatest impact on the battery capacity decay, reduce redundant information, and improve the efficiency of subsequent modeling and prediction. In summary, this embodiment can effectively quantify the correlation between the characteristic parameters and the battery capacity decay, improve the robustness and adaptability of the correlation calculation, screen out the key factors that affect the battery health status, provide high-quality feature input for subsequent analysis and prediction tasks, and improve the performance and reliability of the entire battery health management system.

[0029] S2. Add the preprocessed real-time operating condition data and the preprocessed tram environmental parameters as input data to a pre-set multi-task deep learning framework, wherein the multi-task deep learning framework includes a feature fusion layer, a timing analysis layer and a prediction layer. The feature fusion layer dynamically weights the input data according to the attention mechanism to construct a fusion vector. The timing analysis layer performs timing modeling on the fusion vector according to a bidirectional long short-term memory network to extract the performance evolution law of the energy storage battery. Based on the performance evolution law of the energy storage battery, the prediction layer performs life prediction to generate an initial life prediction result. The multi-objective optimization strategy based on focus loss and the online learning mechanism based on the sliding time window are used to optimize the parameters of the multi-task deep learning framework. The life prediction is performed again according to the updated multi-task deep learning framework to obtain a comprehensive status evaluation report corresponding to the current energy storage battery. The multi-objective optimization strategy based on focal loss is an optimization method. By introducing the focal loss function, it emphasizes the processing of targets that are difficult to predict or have large errors in multi-objective optimization to improve the overall prediction performance. It is particularly suitable for tasks with unbalanced categories. The online learning mechanism based on sliding time windows is a dynamic learning strategy. The model is updated by sliding a fixed-size window at each time step, so that it can process time series data and adjust model parameters in real time. The performance evolution law of the energy storage battery refers to the fact that during the charge and discharge cycle of the energy storage battery, as the use time goes by, its performance (such as capacity, efficiency, etc.) gradually declines and shows a certain regularity, which is usually affected by factors such as temperature and load.

[0030] In an optional embodiment, The preprocessed real-time operating condition data and the preprocessed tram environment parameters are added as input data to a pre-set multi-task deep learning framework, wherein the multi-task deep learning framework includes a feature fusion layer, a timing analysis layer and a prediction layer. The feature fusion layer dynamically weights the input data according to the attention mechanism to construct a fusion vector. The timing analysis layer performs timing modeling on the fusion vector according to a bidirectional long short-term memory network. Extracting the evolution law of energy storage battery performance includes: The preprocessed real-time working condition data and the preprocessed tram environment parameters are unified to a sampling frequency of 20 Hz by a linear interpolation method, a data matrix is ​​generated and added as input data to a pre-set multi-task deep learning framework, and a feature fusion layer in the multi-task deep learning framework is set with eight attention heads according to a multi-head attention mechanism, each attention head learns the correlation between different dimensions, and a corresponding attention weight is set for each dimension of the input data and weighted fusion is performed to generate a fusion vector; The fusion vector is input to the timing analysis layer, which includes a two-layer bidirectional long short-term memory network structure. The first layer of the bidirectional long short-term memory network divides the fusion vector into time segments with a length of 50, and extracts timing features for each of the time segments. The second layer of the bidirectional long short-term memory network receives the output sequence of the first layer of the bidirectional long short-term memory network. The forward long short-term memory network processes the time segments in positive time order to extract long-term dependency features, and the reverse long short-term memory network processes the time segments in reverse time order to extract short-term fluctuation features. The long-term dependency features and the short-term fluctuation features are concatenated to obtain a timing feature vector with a dimension of 512. Based on the timing feature vector, the performance evolution law of the energy storage battery is extracted by a statistical method, and the battery performance evolution law is divided into multiple data segments with the same time interval. A performance degradation threshold is set and the performance degradation amount of the energy storage battery in each data segment is determined. If the performance degradation amount is greater than the performance degradation threshold, the current data segment is marked as a faulty data segment.

[0031] The short-term fluctuation characteristics refer to the rapid changes and fluctuations of the system or signal within a shorter time scale, which usually reflects the dynamic information of real-time changes. The long-term dependence characteristics refer to the changing trends of data or systems on a longer time scale, which usually contain historical information and have an impact on future states. They are commonly seen in tasks such as time series prediction. The performance degradation threshold refers to the energy storage device or system. When the performance drops to a certain critical point during use, a large degree of efficiency loss or functional failure begins to occur.

[0032] Preprocessing the real-time operating data and tram environmental parameters. The real-time operating data includes parameters such as battery voltage, current, and temperature. The tram environmental parameters include ambient temperature and humidity. The preprocessing steps include removing outliers and normalizing. The preprocessed data is unified to a sampling frequency of 20 Hz through the linear interpolation method. For example, the original voltage data is [3.6 V, 3.8 V, 3.7 V] with a sampling frequency of 10 Hz. It is expanded to [3.6 V, 3.7 V, 3.8 V, 3.75 V] through linear interpolation, and the sampling frequency becomes 20 Hz. All parameters are processed similarly to generate a data matrix with a uniform frequency.

[0033] The data matrix is ​​input into a pre-set multi-task deep learning framework, which includes three main parts: feature fusion layer, timing analysis layer and prediction layer. The feature fusion layer adopts a multi-head attention mechanism and sets 8 attention heads. Each attention head independently learns the association between different dimensions. For example, the first attention head may pay more attention to the relationship between voltage and temperature, while the second one focuses on the association between current and humidity. In specific implementation, the query vector, key vector and value vector are calculated for each dimension of the input data, and the attention score is obtained by dot product operation, and then the attention weight is obtained by softmax normalization. The outputs of the 8 attention heads are spliced, and the final fusion vector is obtained by linear transformation.

[0034] The time series analysis layer contains two layers of bidirectional long short-term memory networks. The first layer of bidirectional long short-term memory network divides the fusion vector into time segments of length 50. For each time segment, the forward long short-term memory network processes the sequence from left to right, and the reverse long short-term memory network processes it from right to left, respectively extracting forward and reverse time series features. The second layer of bidirectional long short-term memory network receives the output sequence of the first layer and further extracts long-term dependency features and short-term fluctuation features. The forward long short-term memory network mainly captures long-term trends, while the reverse long short-term memory network pays more attention to recent fluctuations. The outputs of the two directions are spliced ​​to obtain a 512-dimensional time series feature vector.

[0035] Based on the time series feature vector, a statistical method is used to extract the evolution law of energy storage battery performance. The specific steps include: calculating the statistical characteristics (mean, variance, etc.) of each time point; fitting the trend curve, extracting periodic characteristics, etc., dividing the obtained evolution law into multiple equal-length data segments, such as each segment corresponding to 1 day of data, setting a performance degradation threshold, such as a 5% capacity drop. Calculate the performance degradation amount for each data segment, such as the percentage of capacity change. If the degradation amount of a segment exceeds the threshold, it will be marked as a faulty data segment.

[0036] The prediction layer trains a multi-task learning model based on the extracted features and labeled results to simultaneously predict the remaining life and fault type. It adopts a structure of shared underlying network and task-specific output layer to balance the learning of general features and task-specific features.

[0037] In this embodiment, a unified data matrix is ​​generated as the input of the multi-task deep learning framework to provide a standardized data format for subsequent feature extraction and pattern recognition. The multi-head attention mechanism can extract the correlation patterns between features from different subspace perspectives, enhance the richness and diversity of feature representation, and extract the performance evolution law based on time series features and statistical methods, which can effectively capture the performance change trend and degradation pattern of the energy storage battery during long-term use, and provide an important basis for subsequent fault diagnosis and life prediction. The marking mechanism of the fault data segment can help identify the key time points and abnormal conditions in the performance degradation process of the energy storage battery, and provide a reference for subsequent fault analysis, cause diagnosis and maintenance decisions. In summary, this embodiment can effectively mine and utilize the spatiotemporal characteristics and performance degradation patterns of the energy storage battery during operation, provide key data support and decision-making basis for subsequent fault diagnosis, life prediction and health management, and improve the reliability and safety of the energy storage battery system.

[0038] In an optional embodiment, Based on the performance evolution law of the energy storage battery, life prediction is performed through the prediction layer to generate an initial life prediction result. The multi-objective optimization strategy based on focus loss and the online learning mechanism based on the sliding time window are used to optimize the parameters of the multi-task deep learning framework. Life prediction is performed again according to the updated multi-task deep learning framework to obtain a comprehensive status evaluation report corresponding to the current energy storage battery, including: The pre-acquired time series feature vector is input into the prediction layer, and the prediction layer performs initial life prediction through a three-layer fully connected network structure, wherein the first layer takes the time series feature vector and the performance evolution law of the energy storage battery as input, the input dimension is 512, and obtains a 256-dimensional output through the ReLU activation function, the second layer maps the 256-dimensional output to a 128-dimensional feature through the tanh activation function, and the third layer maps the 128-dimensional feature to a 64-dimensional feature through the sigmoid activation function, and generates the initial life prediction result corresponding to the energy storage battery according to the 64-dimensional feature, wherein the initial life prediction result includes the initial remaining cycle life prediction value, the initial capacity attenuation prediction value and the initial internal resistance growth prediction value of the energy storage battery; A multi-objective optimization strategy based on focus loss is used to calculate the first loss function. Combined with the initial life prediction result, a loss weight of 0.4 is set for the initial remaining cycle life prediction value, and a loss weight of 0.3 is set for the initial capacity decay prediction value and the initial internal resistance growth prediction value. Online learning is performed based on a sliding time window mechanism. The size of the sliding time window is 2000 data points, and adjacent windows overlap by 500 data points. If the average prediction error of any prediction value in three consecutive sliding time windows exceeds 5%, the loss weight of the corresponding prediction value is increased by 0.1. The network parameters of the deep learning framework are updated online according to the first loss function, and the life prediction is performed again using the updated multi-task deep learning framework to obtain the optimized remaining cycle life prediction value, capacity decay prediction value and internal resistance growth prediction value, which are combined to obtain a comprehensive status assessment report corresponding to the current energy storage battery.

[0039] The tanh activation function is a commonly used activation function, which is usually used in neural networks to help the model introduce nonlinear features so that it can better fit complex patterns.

[0040] The pre-acquired time series feature vector is input into the prediction layer. The prediction layer adopts a three-layer fully connected network structure for initial life prediction. The first layer takes the 512-dimensional time series feature vector and the performance evolution law of the energy storage battery as input, and obtains a 256-dimensional output through the ReLU activation function. The second layer maps the 256-dimensional output to a 128-dimensional feature through the tanh activation function. The third layer maps the 128-dimensional feature to a 64-dimensional feature through the sigmoid activation function. The initial life prediction result of the energy storage battery is generated according to the 64-dimensional feature, including the initial remaining cycle life prediction value, the initial capacity attenuation prediction value and the initial internal resistance growth prediction value.

[0041] For example, for the initial life prediction of a certain energy storage battery, a 512-dimensional time series feature vector [0.1, 0.2, ..., 0.5] and a performance evolution law parameter [0.01, 0.02, 0.03] are input, the first layer outputs 256-dimensional features [0.15, 0.25, ..., 0.45], the second layer outputs 128-dimensional features [0.2, 0.3, ..., 0.4], and the third layer outputs 64-dimensional features [0.25, 0.35, ..., 0.55]. Finally, the initial remaining cycle life prediction value is 800 times, the initial capacity attenuation prediction value is 5%, and the initial internal resistance growth prediction value is 10%.

[0042] A multi-objective optimization strategy based on focus loss is adopted to calculate the first loss function. A loss weight of 0.4 is set for the initial remaining cycle life prediction value, and a loss weight of 0.3 is set for the initial capacity attenuation prediction value and the initial internal resistance growth prediction value. Online learning is performed based on the sliding time window mechanism. The sliding time window size is 2000 data points, and adjacent windows overlap by 500 data points.

[0043] The prediction error is monitored within the sliding window. If the average prediction error of any prediction value in three consecutive sliding time windows exceeds 5%, the loss weight of the corresponding prediction value will be increased by 0.1. For example, the average errors of the remaining cycle life prediction value in three consecutive windows are 6%, 7%, and 8%, respectively, all exceeding the 5% threshold, then its loss weight will be increased from 0.4 to 0.5.

[0044] The network parameters of the deep learning framework are updated online according to the first loss function. For example, the original network parameters are [0.1, 0.2, 0.3], the loss function calculates the gradient [0.01, 0.02, 0.03], and the learning rate is 0.1. The updated network parameters are [0.099, 0.198, 0.297].

[0045] The updated multi-task deep learning framework is used to perform life prediction again to obtain optimized remaining cycle life prediction values, capacity attenuation prediction values, and internal resistance growth prediction values. For example, the optimized prediction results are: remaining cycle life 780 times, capacity attenuation 6%, and internal resistance growth 12%.

[0046] The optimized prediction results are combined to form a comprehensive status assessment report of the current energy storage battery, which includes a remaining cycle life prediction value of 780 times, a capacity attenuation prediction value of 6%, an internal resistance growth prediction value of 12%, and corresponding confidence intervals and uncertainty analysis information.

[0047] In this embodiment, the hierarchical fully connected network structure can gradually extract and abstract high-level features related to life prediction, and increase nonlinear expression capabilities through different activation functions to improve the accuracy of life prediction. The focal loss function can effectively alleviate the sample imbalance problem, pay more attention to difficult-to-predict samples, and improve the generalization ability of the model. The adaptive adjustment strategy can dynamically respond to changes in prediction errors, pay more attention to indicators that are difficult to predict, and improve the adaptability and robustness of the model. The online learning mechanism also enables the model to adapt to the long-term change trend of energy storage battery performance and maintain the effectiveness of the prediction. In summary, this embodiment can continuously track the health status of energy storage batteries by updating network parameters online and generating comprehensive status evaluation reports, provide real-time and accurate decision-making basis for the battery management system, and improve the operating efficiency and reliability of the energy storage system.

[0048] In an optional embodiment, The network parameters of the deep learning framework are updated online according to the first loss function as shown in the following formula: ; in, i t+1 Indicates that at time step t Model parameters at +1, i t Represents the time step t The model parameters when or 0 represents the initial learning rate, exp represents the exponential function, β represents the attenuation factor, k Indicates the size of the time window, Indicates that at time step j When i j Corresponding dataset D j The loss function value on , ▽ θ represents the gradient, Indicates that at time step t When , the loss function is about the model parameters i t gradient.

[0049] In this embodiment, the adaptive learning rate adjustment mechanism can automatically find a suitable learning rate during the training process, accelerate the convergence speed, and improve the training efficiency of the model. By considering the historical loss information, the changing trend of the loss function can be better captured, and the learning rate can be adjusted in time to make the model more adaptable to the distribution characteristics of the data. By controlling the size of the learning rate, the amplitude of the gradient update can be adjusted to avoid excessive or too small parameter updates, thereby improving the convergence and generalization capabilities of the model. The online learning mechanism enables the model to maintain good performance in a constantly changing data environment and adapt to the long-term evolution trend of the energy storage battery performance. In summary, this embodiment realizes the online update of the network parameters of the deep learning framework, which is of great significance for the energy storage battery life prediction task. It can enable the model to capture the degradation trend of the battery performance in a timely manner and provide reliable decision support for battery health management.

[0050] S3. Construct a multi-objective maintenance decision optimization model based on the comprehensive status assessment report, take the life prediction results, vehicle operation plan and maintenance cost constraints as the status, perform dual evaluation through the policy network and the value network, predict the optimal maintenance time window, deeply integrate the pre-acquired historical maintenance data and fault diagnosis experience, combine the improved graph attention mechanism to extract maintenance knowledge features and generate a maintenance operation list, based on the maintenance operation list, according to the improved mixed integer programming algorithm, take minimizing maintenance cost and maximizing maintenance efficiency as the optimization goals to perform global optimization configuration, generate a resource scheduling plan, integrate the optimal maintenance time window, maintenance operation list and resource scheduling plan, form the optimal battery maintenance plan and send it to the on-board controller for visual display.

[0051] The improved graph attention mechanism is a variant of the attention mechanism introduced in the graph neural network. By dynamically calculating the weights between nodes in the graph, the model can automatically focus on the most important neighbor nodes, thereby improving the learning effect of graph structured data. The improved mixed integer programming algorithm is an optimization based on the traditional mixed integer programming method. It usually improves the solution efficiency and expands the applicability in practical applications by introducing heuristic algorithms, relaxation techniques or branch-and-bound strategies. The policy network is a neural network used to select actions in reinforcement learning. It outputs a probability distribution of an action based on the current state to guide the decision-making process of the intelligent agent in the environment. The value network is another neural network in reinforcement learning, which is used to estimate the value of the current state or state-action pair to help the intelligent agent evaluate the potential rewards in different states.

[0052] In an optional embodiment, Based on the comprehensive status assessment report, a multi-objective maintenance decision optimization model is constructed. The life prediction results, vehicle operation plan and maintenance cost constraints are used as the status. A dual evaluation is performed through the policy network and the value network to predict the optimal maintenance time window. The pre-acquired historical maintenance data and fault diagnosis experience are deeply integrated. The maintenance knowledge features are extracted and a maintenance operation list is generated in combination with the improved graph attention mechanism. Based on the maintenance operation list, according to the improved mixed integer programming algorithm, global optimization configuration is performed with maintenance cost minimization and maintenance efficiency maximization as the optimization goals, and a resource scheduling plan is generated. The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form the optimal battery maintenance plan and send it to the vehicle controller for visual display, including: Acquire energy storage battery performance characteristics, operation plan characteristics and maintenance constraint characteristics, extract long-range evolution trends of the energy storage battery performance characteristics through a time series convolutional network to obtain a first coding feature, convert the operation plan characteristics into a binary sequence through one-hot encoding to obtain a second coding feature, normalize the maintenance constraint characteristics to obtain a third coding feature, and concatenate the first coding feature, the second coding feature and the third coding feature to form a state vector; The state vector is input into the policy network and the value network. The policy network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer. The first hidden layer uses a ReLU activation function. The second hidden layer fuses neuron outputs through a residual connection. The third hidden layer uses a tanh activation function. The policy network outputs an action probability distribution. The value network outputs an evaluation score. The time period with the highest evaluation score is selected as the optimal maintenance time window. Retrieve historical maintenance records, which include fault type information, maintenance measures information, and maintenance effect information, establish a maintenance knowledge graph in combination with diagnostic rules in the expert experience knowledge base, use a multi-head graph attention mechanism to process the maintenance knowledge graph, fuse adjacent node information layer by layer through multi-layer attention propagation, extract the maintenance knowledge features according to the attention weight, and generate a maintenance operation list; The maintenance operation list is converted into a resource requirement list, wherein the resource requirement list includes the number of maintenance personnel, the type of equipment, and the number of spare parts. A mixed integer programming algorithm is used to optimize the resource requirement list, with the minimum total maintenance cost and the maximum maintenance efficiency as the optimization goals. The constraints of the mixed integer programming algorithm include the number of maintenance personnel, the equipment utilization rate, and the spare parts inventory safety margin constraint, and a resource scheduling solution is obtained by solving the problem. The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form the optimal battery maintenance plan and sent to the on-board controller for visual display.

[0053] The time series convolutional network is a structure based on convolutional neural network (CNN), which is specially used to process time series data. It extracts local time features in time series data through convolution operations and can capture short-term and long-term dynamic changes. The one-hot encoding is an encoding method that converts categorical variables into binary vectors, in which each category value is mapped to a unique index position, and the value of its corresponding position is set to 1, and other positions are set to 0, which is used to process classified data.

[0054] Obtain the performance characteristics, operation plan characteristics and maintenance constraint characteristics of energy storage batteries. The performance characteristics of energy storage batteries include the change data of parameters such as battery capacity, internal resistance, and self-discharge rate over time. The operation plan characteristics include the predicted information such as mileage, charging times, and ambient temperature in the future. The maintenance constraint characteristics include the maintenance cost, maintenance time, spare parts inventory and other restrictions. For the performance characteristics of energy storage batteries, a time series convolutional network is used to extract long-range evolution trends. Specifically, multiple layers of one-dimensional convolutional layers are used to extract features from time series data, and the convolution kernel sizes are set to 3, 5, and 7 to capture time dependencies at different scales. A fixed-dimensional feature vector is obtained through the global maximum pooling layer as the first encoding feature. For example, for the capacity decay curve, key features such as decay rate and inflection point position can be extracted.

[0055] For the operation plan feature, one-hot encoding is used to convert the discrete plan information into a binary sequence. For example, the mileage interval for the next 7 days is encoded as [0, 1, 0, 0, 0, 0], which means that the mileage on the second day is between 100-200 km. The binary sequence obtained after encoding is used as the second encoding feature.

[0056] For the maintenance constraint features, normalization is performed to unify the dimensions of different indicators. The minimum-maximum normalization method is used to map each constraint indicator to the interval [0, 1]. The normalized feature vector is used as the third encoding feature.

[0057] The above three encoding features are concatenated to form a state vector, which is used as the input of the subsequent decision model. The state vector dimension is the sum of the three feature vector dimensions, which contains comprehensive information about the battery status, operation requirements and maintenance constraints.

[0058] The state vector is input into the policy network and the value network for evaluation. The policy network adopts a multi-layer feedforward neural network structure, including an input layer, three hidden layers, and an output layer. The number of neurons in the input layer is the same as the dimension of the state vector. The first hidden layer uses the ReLU activation function to introduce nonlinearity, and the number of neurons is twice that of the input layer. The second hidden layer fuses the output of the previous layer through residual connections to alleviate the gradient vanishing problem. The third hidden layer uses the tanh activation function to limit the output to the range of [-1, 1]. The output layer uses the softmax function to obtain the maintenance probability distribution of each time period.

[0059] The structure of the value network is similar to that of the policy network, but the output layer has only one neuron, which is used to evaluate the value score of the current state. The two networks share the parameters of the first few layers to improve training efficiency. In the inference phase, the policy network outputs the action probability distribution and the value network outputs the evaluation score. The time period with the highest evaluation score is selected as the optimal maintenance time window. For example, if the value network scores every day in the next 30 days, the day with the highest score is the optimal maintenance time.

[0060] Retrieve historical maintenance records and build a maintenance knowledge graph based on expert experience knowledge. Historical maintenance records include information such as fault type, maintenance measures, and maintenance effects. The nodes in the knowledge graph represent entities such as fault type, symptoms, causes, and solutions, and the edges represent the relationships between entities. For example, there is a "may cause" relationship edge between the "capacity decay" node and the "electrolyte leakage" node.

[0061] A multi-head graph attention mechanism is used to process the maintenance knowledge graph. Each node in the graph is initialized as a low-dimensional dense vector, and the node representation is gradually updated through multi-layer attention propagation. In each layer, for the target node, the attention weights of all adjacent nodes are calculated. The weights are determined by the similarity of the node features. The information of adjacent nodes is weightedly aggregated according to the attention weights, and the representation of the target node is updated. Through the multi-head mechanism, each head independently calculates attention and aggregates information, and the results of multiple heads are spliced. After multi-layer propagation, the node representation incorporates multi-hop neighborhood information, extracts maintenance knowledge features according to the attention weights, and generates a maintenance operation list for the current fault.

[0062] The maintenance operation list is converted into a specific resource requirement list, including the number of maintenance personnel, equipment types, and spare parts required. A mixed integer programming algorithm is used to optimize the resource requirements. The optimization goal is to minimize the total maintenance cost and maximize the maintenance efficiency. The constraints include the upper limit of the number of available maintenance personnel, the equipment utilization rate not exceeding 80%, and the spare parts inventory safety margin not less than 20%. By solving the optimization problem, the optimal resource scheduling solution that meets the constraints is obtained.

[0063] The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form a complete optimal battery maintenance plan. The plan is visualized in the form of charts, text, etc. through the on-board controller, which is easy for maintenance personnel to understand and execute.

[0064] In this embodiment, by integrating multi-dimensional information such as battery performance, operation requirements and maintenance constraints, a comprehensive consideration of maintenance decisions is achieved, the limitations that may be brought about by single-factor decisions are avoided, the scientificity and reliability of maintenance plans are improved, and a deep reinforcement learning method is used to dynamically predict the optimal maintenance time. The knowledge graph and graph attention mechanism are combined to extract maintenance knowledge features, thereby realizing intelligent and knowledge-driven maintenance decisions, which can adapt to the complex and changeable actual operation environment and improve the flexibility and adaptability of maintenance decisions. The maintenance resources are globally optimized and configured through a mixed integer programming algorithm. Under the premise of meeting various constraints, multi-objective optimization of minimizing maintenance costs and maximizing efficiency is achieved, which improves resource utilization efficiency, reduces maintenance costs, and improves maintenance effects. In summary, this embodiment realizes intelligent maintenance decisions based on battery status and operation plans, and achieves efficient configuration of maintenance resources while ensuring maintenance effects.

[0065] Figure 2 FIG. 1 is a schematic diagram of a structure of a fault diagnosis and predictive maintenance control system for an electric vehicle energy storage battery according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to collect the real-time working condition data of the tram energy storage battery and the tram environmental parameters, transmit the real-time working condition data and the tram environmental parameters to the on-board controller through the on-board CAN bus, and upload them to the cloud server in combination with the wireless communication module. The cloud server performs data cleaning through the improved Tukey-Kramerdo multiple comparison, normalizes the cleaned data, unifies the normalized data of different dimensions into the same scale range, performs nonlinear dimensionality reduction through the improved isometric mapping algorithm, extracts the key characteristic parameters in the parameters after dimensionality reduction, and obtains the pre-processed real-time working condition data and the pre-processed tram environmental parameters; The second unit is used to add the preprocessed real-time operating condition data and the preprocessed tram environmental parameters as input data to a pre-set multi-task deep learning framework, wherein the multi-task deep learning framework includes a feature fusion layer, a timing analysis layer and a prediction layer, the feature fusion layer dynamically weights the input data according to the attention mechanism to construct a fusion vector, the timing analysis layer performs timing modeling on the fusion vector according to a bidirectional long short-term memory network, extracts the performance evolution law of the energy storage battery, and based on the performance evolution law of the energy storage battery, performs life prediction through the prediction layer to generate an initial life prediction result, optimizes the parameters of the multi-task deep learning framework through a multi-objective optimization strategy based on focus loss and an online learning mechanism based on a sliding time window, performs life prediction again according to the updated multi-task deep learning framework, and obtains a comprehensive status evaluation report corresponding to the current energy storage battery; The third unit is used to build a multi-objective maintenance decision optimization model based on the comprehensive status assessment report, with life prediction results, vehicle operation plan and maintenance cost constraints as states, and perform dual evaluation through policy network and value network to predict the optimal maintenance time window, deeply integrate the pre-acquired historical maintenance data and fault diagnosis experience, and extract maintenance knowledge features and generate a maintenance operation list in combination with the improved graph attention mechanism. Based on the maintenance operation list, according to the improved mixed integer programming algorithm, global optimization configuration is performed with minimization of maintenance cost and maximization of maintenance efficiency as optimization goals, and a resource scheduling plan is generated. The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form an optimal battery maintenance plan and send it to the on-board controller for visual display.

[0066] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0067] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0068] 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 carrying computer-readable program instructions for executing various aspects of the present invention.

[0069] 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 it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fault diagnosis and predictive maintenance control of electric vehicle energy storage batteries, characterized in that: include: Collecting real-time operating data of the tram energy storage battery and tram environmental parameters, transmitting the real-time operating data and tram environmental parameters to the on-board controller through the on-board CAN bus, uploading to the cloud server in combination with the wireless communication module, the cloud server performs data cleaning through the improved Tukey-Kramerdo multiple comparison, normalizing the cleaned data, unifying the normalized data of different dimensions into the same scale range, performing nonlinear dimensionality reduction through the improved isometric mapping algorithm, extracting key characteristic parameters from the parameters after dimensionality reduction, and obtaining pre-processed real-time operating data and pre-processed tram environmental parameters; The preprocessed real-time operating condition data and the preprocessed tram environmental parameters are added as input data to a pre-set multi-task deep learning framework, wherein the multi-task deep learning framework includes a feature fusion layer, a timing analysis layer and a prediction layer, the feature fusion layer dynamically weights the input data according to the attention mechanism to construct a fusion vector, the timing analysis layer performs timing modeling on the fusion vector according to a bidirectional long short-term memory network, extracts the performance evolution law of the energy storage battery, and based on the performance evolution law of the energy storage battery, performs life prediction through the prediction layer to generate an initial life prediction result, optimizes the parameters of the multi-task deep learning framework through a multi-objective optimization strategy based on focus loss and an online learning mechanism based on a sliding time window, performs life prediction again according to the updated multi-task deep learning framework, and obtains a comprehensive status evaluation report corresponding to the current energy storage battery; A multi-objective maintenance decision optimization model is constructed based on the comprehensive status assessment report. The life prediction results, vehicle operation plan and maintenance cost constraints are used as the status. A dual evaluation is performed through the policy network and the value network to predict the optimal maintenance time window. The pre-acquired historical maintenance data and fault diagnosis experience are deeply integrated. The maintenance knowledge features are extracted and a maintenance operation list is generated in combination with the improved graph attention mechanism. Based on the maintenance operation list, according to the improved mixed integer programming algorithm, global optimization configuration is performed with minimization of maintenance cost and maximization of maintenance efficiency as the optimization goals, and a resource scheduling plan is generated. The optimal maintenance time window, the maintenance operation list and the resource scheduling plan are integrated to form an optimal battery maintenance plan and send it to the on-board controller for visual display.

2. The method according to claim 1, characterized in that The real-time working condition data of the tram energy storage battery and the tram environmental parameters are collected, and the real-time working condition data and the tram environmental parameters are transmitted to the on-board controller through the on-board CAN bus, and uploaded to the cloud server in combination with the wireless communication module. The cloud server performs data cleaning through the improved Tukey-Kramerdo multiple comparison, normalizes the cleaned data, unifies the normalized data of different dimensions into the same scale range, performs nonlinear dimensionality reduction through the improved isometric mapping algorithm, extracts the key characteristic parameters in the parameters after dimensionality reduction, and obtains the pre-processed real-time working condition data and pre-processed tram environmental parameters, including: The battery temperature data is collected by a graphene-based flexible temperature sensor arranged on each battery module, the battery voltage data is collected by a voltage sensor arranged at the positive and negative terminals of the battery, the battery current data is collected by a Hall effect current sensor connected in series in the main circuit of the battery, the battery strain data is collected by a piezoelectric material vibration strain sensor arranged at a key position of the battery, and the gas composition data inside the battery pack is collected by a micro-spectral sensor arranged on the top of the battery pack, and the battery temperature data, battery voltage data, battery current data, battery strain data and battery pack internal gas composition data are combined to obtain the real-time operating condition data; Ambient temperature data is collected by dual backup ambient temperature sensors arranged outside the battery compartment, ambient humidity data is collected by an ambient humidity sensor integrated in the same module as the dual backup ambient temperature sensors, altitude data is collected by a MEMS air pressure sensor, omnidirectional vibration data is collected by a nine-axis inertial measurement unit arranged at the bottom of the battery pack, and ambient temperature data, ambient humidity data, altitude data and omnidirectional vibration data are integrated to obtain the environmental parameters of the electric vehicle; The real-time operating condition data and the tram environment parameters are transmitted to the on-board controller via the on-board CAN bus, and the on-board controller uploads the real-time operating condition data and the tram environment parameters to the cloud server via a pre-set wireless communication module; The cloud server performs anomaly detection on the real-time operating condition data and the tram environmental parameters through a variational autoencoder unsupervised learning model, performs data cleaning on the data after anomaly detection according to the Tukey-Kramer multiple comparison method combined with a temporal attention mechanism, establishes a multi-scale data association analysis of data mutation analysis, trend analysis, and periodic analysis, combines an adaptive time window that dynamically adjusts the window size according to the data change rate to process the abnormal data, obtains cleaned data, and normalizes the cleaned data to unify data of different dimensions within the same scale range to obtain normalized data; An attention mechanism is used to dynamically weight the importance of different parameter features in the normalized data to obtain weighted data. The weighted data is subjected to nonlinear dimensionality reduction according to an isometric mapping algorithm enhanced by multi-task learning. The Pearson correlation coefficient between each characteristic parameter in the reduced dimensionality data and the battery capacity decay rate is calculated based on a correlation analysis method. A feature importance ranking is established and the parameters in the top 10% are selected as key characteristic parameters to obtain preprocessed real-time operating condition data and preprocessed tram environmental parameters.

3. The method according to claim 2, characterized in that The Pearson correlation coefficient between each characteristic parameter in the data after dimensionality reduction and the battery capacity decay rate is calculated based on the correlation analysis method as shown in the following formula: ； Where R represents the Pearson correlation coefficient, x i Indicates i The eigenvalues ​​after dimensionality reduction, represents the median of the eigenvalue after dimensionality reduction, y i The rate at which the battery capacity decays i Sample values, It represents the median of the battery capacity attenuation rate, and IQR() represents the interquartile range, which is used to measure the degree of dispersion of the data. α Represents the smoothing factor.

4. The method according to claim 1, characterized in that The preprocessed real-time operating condition data and the preprocessed tram environment parameters are added as input data to a pre-set multi-task deep learning framework, wherein the multi-task deep learning framework includes a feature fusion layer, a timing analysis layer and a prediction layer. The feature fusion layer dynamically weights the input data according to the attention mechanism to construct a fusion vector. The timing analysis layer performs timing modeling on the fusion vector according to a bidirectional long short-term memory network. Extracting the evolution law of energy storage battery performance includes: The preprocessed real-time working condition data and the preprocessed tram environment parameters are unified to a sampling frequency of 20 Hz by a linear interpolation method, a data matrix is ​​generated and added as input data to a pre-set multi-task deep learning framework, and a feature fusion layer in the multi-task deep learning framework is set with eight attention heads according to a multi-head attention mechanism, each attention head learns the correlation between different dimensions, and a corresponding attention weight is set for each dimension of the input data and weighted fusion is performed to generate a fusion vector; The fusion vector is input to the timing analysis layer, which includes a two-layer bidirectional long short-term memory network structure. The first layer of the bidirectional long short-term memory network divides the fusion vector into time segments with a length of 50, and extracts timing features for each of the time segments. The second layer of the bidirectional long short-term memory network receives the output sequence of the first layer of the bidirectional long short-term memory network. The forward long short-term memory network processes the time segments in positive time order to extract long-term dependency features, and the reverse long short-term memory network processes the time segments in reverse time order to extract short-term fluctuation features. The long-term dependency features and the short-term fluctuation features are concatenated to obtain a timing feature vector with a dimension of 512. Based on the timing feature vector, the performance evolution law of the energy storage battery is extracted by a statistical method, and the battery performance evolution law is divided into multiple data segments with the same time interval. A performance degradation threshold is set and the performance degradation amount of the energy storage battery in each data segment is determined. If the performance degradation amount is greater than the performance degradation threshold, the current data segment is marked as a faulty data segment.

5. The method according to claim 1, characterized in that Based on the performance evolution law of the energy storage battery, life prediction is performed through the prediction layer to generate an initial life prediction result. The multi-objective optimization strategy based on focus loss and the online learning mechanism based on the sliding time window are used to optimize the parameters of the multi-task deep learning framework. Life prediction is performed again according to the updated multi-task deep learning framework to obtain a comprehensive status evaluation report corresponding to the current energy storage battery, including: The pre-acquired time series feature vector is input into the prediction layer, and the prediction layer performs initial life prediction through a three-layer fully connected network structure, wherein the first layer takes the time series feature vector and the performance evolution law of the energy storage battery as input, the input dimension is 512, and obtains a 256-dimensional output through the ReLU activation function, the second layer maps the 256-dimensional output to a 128-dimensional feature through the tanh activation function, and the third layer maps the 128-dimensional feature to a 64-dimensional feature through the sigmoid activation function, and generates the initial life prediction result corresponding to the energy storage battery according to the 64-dimensional feature, wherein the initial life prediction result includes the initial remaining cycle life prediction value, the initial capacity attenuation prediction value and the initial internal resistance growth prediction value of the energy storage battery; A multi-objective optimization strategy based on focus loss is used to calculate the first loss function. Combined with the initial life prediction result, a loss weight of 0.4 is set for the initial remaining cycle life prediction value, and a loss weight of 0.3 is set for the initial capacity decay prediction value and the initial internal resistance growth prediction value. Online learning is performed based on a sliding time window mechanism. The size of the sliding time window is 2000 data points, and adjacent windows overlap by 500 data points. If the average prediction error of any prediction value in three consecutive sliding time windows exceeds 5%, the loss weight of the corresponding prediction value is increased by 0.

1. The network parameters of the deep learning framework are updated online according to the first loss function, and the life prediction is performed again using the updated multi-task deep learning framework to obtain the optimized remaining cycle life prediction value, capacity decay prediction value and internal resistance growth prediction value, which are combined to obtain a comprehensive status assessment report corresponding to the current energy storage battery.

6. The method according to claim 5, characterized in that The network parameters of the deep learning framework are updated online according to the first loss function as shown in the following formula: ； in, θ t+1 Indicates that at time step t Model parameters at +1, θ t Represents the time step t The model parameters when η 0 represents the initial learning rate, exp represents the exponential function, β represents the attenuation factor, k Indicates the size of the time window, Indicates that at time step j When θ j The corresponding dataset D j The loss function value on , ▽ θ represents the gradient, Indicates that at time step t When , the loss function is about the model parameters θ t gradient.

7. The method according to claim 1, characterized in that Based on the comprehensive status assessment report, a multi-objective maintenance decision optimization model is constructed. The life prediction results, vehicle operation plan and maintenance cost constraints are used as the status. A dual evaluation is performed through the policy network and the value network to predict the optimal maintenance time window. The pre-acquired historical maintenance data and fault diagnosis experience are deeply integrated. The maintenance knowledge features are extracted and a maintenance operation list is generated in combination with the improved graph attention mechanism. Based on the maintenance operation list, according to the improved mixed integer programming algorithm, global optimization configuration is performed with maintenance cost minimization and maintenance efficiency maximization as the optimization goals, and a resource scheduling plan is generated. The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form the optimal battery maintenance plan and send it to the vehicle controller for visual display, including: Acquire energy storage battery performance characteristics, operation plan characteristics and maintenance constraint characteristics, extract long-range evolution trends of the energy storage battery performance characteristics through a time series convolutional network to obtain a first coding feature, convert the operation plan characteristics into a binary sequence through one-hot encoding to obtain a second coding feature, normalize the maintenance constraint characteristics to obtain a third coding feature, and concatenate the first coding feature, the second coding feature and the third coding feature to form a state vector; The state vector is input into the policy network and the value network. The policy network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer. The first hidden layer uses a ReLU activation function. The second hidden layer fuses neuron outputs through a residual connection. The third hidden layer uses a tanh activation function. The policy network outputs an action probability distribution. The value network outputs an evaluation score. The time period with the highest evaluation score is selected as the optimal maintenance time window. Retrieve historical maintenance records, which include fault type information, maintenance measures information, and maintenance effect information, establish a maintenance knowledge graph in combination with diagnostic rules in the expert experience knowledge base, use a multi-head graph attention mechanism to process the maintenance knowledge graph, fuse adjacent node information layer by layer through multi-layer attention propagation, extract the maintenance knowledge features according to the attention weight, and generate a maintenance operation list; The maintenance operation list is converted into a resource requirement list, wherein the resource requirement list includes the number of maintenance personnel, the type of equipment, and the number of spare parts. A mixed integer programming algorithm is used to optimize the resource requirement list, with the minimum total maintenance cost and the maximum maintenance efficiency as the optimization goals. The constraints of the mixed integer programming algorithm include the number of maintenance personnel, the equipment utilization rate, and the spare parts inventory safety margin constraint, and a resource scheduling solution is obtained by solving the problem. The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form the optimal battery maintenance plan and sent to the on-board controller for visual display.

8. A fault diagnosis and predictive maintenance control system for electric vehicle energy storage batteries, 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 the real-time working condition data of the tram energy storage battery and the tram environmental parameters, transmit the real-time working condition data and the tram environmental parameters to the on-board controller through the on-board CAN bus, and upload them to the cloud server in combination with the wireless communication module. The cloud server performs data cleaning through the improved Tukey-Kramerdo multiple comparison, normalizes the cleaned data, unifies the normalized data of different dimensions into the same scale range, performs nonlinear dimensionality reduction through the improved isometric mapping algorithm, extracts the key characteristic parameters in the parameters after dimensionality reduction, and obtains the pre-processed real-time working condition data and the pre-processed tram environmental parameters; The second unit is used to add the preprocessed real-time operating condition data and the preprocessed tram environmental parameters as input data to a pre-set multi-task deep learning framework, wherein the multi-task deep learning framework includes a feature fusion layer, a timing analysis layer and a prediction layer, the feature fusion layer dynamically weights the input data according to the attention mechanism to construct a fusion vector, the timing analysis layer performs timing modeling on the fusion vector according to a bidirectional long short-term memory network, extracts the performance evolution law of the energy storage battery, and based on the performance evolution law of the energy storage battery, performs life prediction through the prediction layer to generate an initial life prediction result, optimizes the parameters of the multi-task deep learning framework through a multi-objective optimization strategy based on focus loss and an online learning mechanism based on a sliding time window, performs life prediction again according to the updated multi-task deep learning framework, and obtains a comprehensive status evaluation report corresponding to the current energy storage battery; The third unit is used to build a multi-objective maintenance decision optimization model based on the comprehensive status assessment report, with life prediction results, vehicle operation plan and maintenance cost constraints as states, and perform dual evaluation through policy network and value network to predict the optimal maintenance time window, deeply integrate the pre-acquired historical maintenance data and fault diagnosis experience, and extract maintenance knowledge features and generate a maintenance operation list in combination with the improved graph attention mechanism. Based on the maintenance operation list, according to the improved mixed integer programming algorithm, global optimization configuration is performed with minimization of maintenance cost and maximization of maintenance efficiency as optimization goals, and a resource scheduling plan is generated. The optimal maintenance time window, maintenance operation list and resource scheduling plan are integrated to form an optimal battery maintenance plan and send it to the on-board controller for visual display.

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