A method and system for early warning of energy storage battery failure based on multi-source operation data

Abstract

The application belongs to the technical field of energy storage battery early warning monitoring, and discloses a kind of energy storage battery fault early warning method and system based on multi-source operation data, the method comprises the following steps: according to the topology structure of energy storage battery system and the multi-source operation data after preprocessing, construct space-time correlation graph, respectively establish physical twin model and data-driven benchmark model;Filter out battery abnormal nodes, and based on the bidirectional attention weight of space-time correlation graph, locate the fault source node according to the bidirectional attention weight distribution;Extract multi-dimensional abnormal feature vector from the fault source node, and calculate the marginal contribution degree of each dimension feature to the fault probability through counterfactual deduction;Identify the dominant root cause of the fault, and input the dominant root cause of the fault into the dynamic Bayesian network, combine the space-time correlation graph, predict the fault evolution propagation path and time window, and output the fault early warning result.The application realizes the accurate distinction of fault source node and propagation affected node, and provides spatial positioning basis for targeted disposal.

Description

Technical Field

[0001] This invention relates to the field of early warning and monitoring technology for energy storage batteries, and in particular to a method and system for early warning of energy storage battery faults based on multi-source operating data. Background Technology

[0002] As the core energy carrier of electrochemical energy storage power stations, electric vehicles, and distributed energy systems, the safe operation of energy storage battery systems is directly related to grid stability and the safety of people and property. During actual operation, the battery management system continuously collects multi-source operational data such as voltage, current, temperature, and insulation resistance. This data contains rich information about the battery's internal electrochemical state and external thermal state. With the expansion of energy storage system scale, the surge in the number of individual cells, and the increasing complexity of operating conditions, the quality of multi-source data varies, and fault mechanisms and propagation paths are intertwined, placing higher demands on the accuracy, interpretability, and foresight of fault early warning systems.

[0003] Existing early warning methods for energy storage batteries, while capable of initially identifying abnormal battery states by setting thresholds or based on single data-driven models, suffer from high false alarm and false negative rates due to the lack of cross-validation mechanisms between physical mechanisms and statistical data. This lack of cross-validation makes it difficult to distinguish between sensor communication anomalies and actual battery faults, impacting the reliability of the early warning system. Furthermore, existing methods typically treat each battery cell as an independent monitoring object, failing to fully utilize the electrical connections and thermal coupling topology of the energy storage system. This results in insufficient fault location accuracy and difficulty in tracing the propagation direction and evolution path of faults within the system.

[0004] Meanwhile, the state transition reasoning process of existing early warning methods lacks spatial constraints on the physical topology of the energy storage system, which leads to a disconnect between multi-node fault propagation prediction and the actual system structure, reducing the spatial accuracy of large-scale energy storage power station-level fault evolution prediction.

[0005] Therefore, how to provide a method and system for early warning of energy storage battery faults based on multi-source operating data is an urgent problem to be solved. Summary of the Invention

[0006] This invention provides a method and system for early warning of energy storage battery faults based on multi-source operating data, in order to solve the above-mentioned technical problems in the prior art.

[0007] According to a first aspect of the present invention, a method for early warning of energy storage battery faults based on multi-source operating data is provided.

[0008] In one embodiment, the energy storage battery fault early warning method based on multi-source operational data includes:

[0009] Acquire multi-source operational data of the energy storage battery system and perform preprocessing;

[0010] Based on the topology of the energy storage battery system and the preprocessed multi-source operating data, a spatiotemporal correlation graph is constructed, and a physical twin model and a data-driven benchmark model are established based on each node in the spatiotemporal correlation graph.

[0011] The preprocessed multi-source operational data is input into the physical twin model and the data-driven baseline model respectively to obtain the physical prediction residual and the data prediction residual;

[0012] The physical prediction residual and the data prediction residual are cross-validated to screen out abnormal battery nodes. The bidirectional attention weight between the abnormal battery node and its neighboring nodes is calculated based on the spatiotemporal correlation graph. The fault source node is located according to the bidirectional attention weight distribution.

[0013] Extract multi-dimensional abnormal feature vectors from the fault source node, and calculate the marginal contribution of each feature to the fault probability through counterfactual inference.

[0014] Based on the marginal contribution, the dominant root cause of the failure is identified and input into a dynamic Bayesian network. Combined with the spatiotemporal correlation graph, the failure evolution and propagation path and time window are predicted.

[0015] Based on the root cause of the fault, the fault evolution and propagation path, and the time window, output the fault warning result.

[0016] According to a second aspect of the present invention, a fault early warning system for energy storage batteries based on multi-source operating data is provided.

[0017] In one embodiment, the energy storage battery fault early warning system based on multi-source operational data includes:

[0018] The multi-source data acquisition module is used to acquire and preprocess multi-source operating data of the energy storage battery system;

[0019] The spatiotemporal correlation graph construction module is used to construct a spatiotemporal correlation graph based on the topology of the energy storage battery system and the preprocessed multi-source operating data, and to establish a physical twin model and a data-driven benchmark model based on each node in the spatiotemporal correlation graph.

[0020] The dual residual generation module is used to input the preprocessed multi-source running data into the physical twin model and the data-driven baseline model respectively to obtain the physical prediction residual and the data prediction residual;

[0021] The residual cross-validation module is used to cross-validate the physical prediction residual and the data prediction residual, filter out abnormal battery nodes, and calculate the bidirectional attention weight between the abnormal battery node and its neighboring nodes based on the spatiotemporal correlation graph, and locate the fault source node according to the bidirectional attention weight distribution.

[0022] The fault source localization module is used to extract multi-dimensional abnormal feature vectors from fault source nodes and calculate the marginal contribution of each feature dimension to the fault probability through counterfactual inference.

[0023] The root cause identification and path prediction module is used to identify the dominant root cause of the failure based on the marginal contribution, and input the dominant root cause of the failure into a dynamic Bayesian network, combined with the spatiotemporal correlation graph, to predict the failure evolution and propagation path and time window.

[0024] The early warning output module is used to output fault early warning results based on the dominant root cause of the fault, the fault evolution and propagation path, and the time window.

[0025] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0026] (1) This invention sets up a dual residual cross-verification mechanism of physical twin model and data-driven benchmark model to cross-verify electrochemical mechanism law and historical statistical law, and distinguishes sensor communication abnormalities and battery abnormalities online, so as to avoid false alarms caused by aging drift and improve the reliability and robustness of the early warning system.

[0027] (2) Based on the electrical connection and thermal coupling relationship of the energy storage system, the present invention constructs a spatiotemporal correlation diagram and uses bidirectional attention weight distribution to identify the asymmetric propagation direction of abnormal features, thereby improving fault location from single-unit anomaly detection to system-level fault source tracking, realizing accurate differentiation between fault source nodes and nodes affected by propagation, and providing spatial location basis for targeted treatment.

[0028] (3) This invention maps the dominant root cause of the fault to observational evidence of a dynamic Bayesian network and performs forward probabilistic reasoning under the spatial constraints of the spatiotemporal correlation graph to predict the evolution and propagation path of the fault along the electrical-thermal topology and the time window for reaching the preset danger threshold, thereby realizing the leap from single-moment state warning to full-time and spatiotemporal evolution warning and gaining advance response time for safety protection.

[0029] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0030] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0031] Figure 1 This is a flowchart illustrating an energy storage battery fault early warning method based on multi-source operating data, according to an exemplary embodiment.

[0032] Figure 2 This is a schematic diagram of a fault early warning system for energy storage batteries based on multi-source operating data, according to an exemplary embodiment. Detailed Implementation

[0033] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0034] Each module in the apparatus or system of this application can be implemented in whole or in part by software, hardware, or a combination thereof.

[0035] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0036] Figure 1 An embodiment of a method for early warning of energy storage battery faults based on multi-source operating data according to the present invention is shown.

[0037] In this optional embodiment, the energy storage battery fault early warning method based on multi-source operational data includes:

[0038] Step S101: Obtain multi-source operating data of the energy storage battery system and perform preprocessing;

[0039] Specifically, the battery management system uses voltage acquisition circuits, Hall current sensors, distributed thermocouple temperature sensors, insulation monitoring modules, and communication interfaces to collect multi-source operating data of the energy storage battery system in a multi-rate asynchronous sampling manner. The electrical data includes the terminal voltage, charging and discharging current, and state of charge of each battery cell; the thermal data includes the cell temperature, module temperature difference, and ambient temperature; and the system status data includes insulation resistance value, leakage current, sampling packet loss rate, and communication delay. The voltage and current sampling frequency is set to 1Hz to 10Hz, the temperature sampling frequency is set to 0.1Hz to 1Hz, and the insulation status sampling frequency is set to 0.01Hz to 0.1Hz.

[0040] Due to differences in sampling frequency and communication delay among various data sources, a linear interpolation method is used to align multi-source asynchronous data to a unified time base. The timestamp resolution is set to 1 second. Missing value detection is performed on the aligned data. For data segments with a continuous missing duration of no more than 3 sampling points, linear interpolation is used to fill in the missing data. For data segments with a continuous missing duration of more than 3 sampling points, the missing data is marked as missing and the confidence level of the data in that period is reduced. A sliding window outlier detection method based on the 3σ criterion is used to remove outliers. The sliding window length is set to 100 sampling points and the window step size is set to 1 sampling point. The removal condition is that the data point exceeds the range of the mean plus or minus three standard deviations within the window. Finally, Min-Max normalization is performed on the cleaned data, mapping voltage, current, temperature, and insulation resistance to the [0,1] interval. The normalization parameters are determined based on the statistical extreme values ​​of historical healthy operation data to obtain the preprocessed multi-source operation data.

[0041] Step S102: Based on the topology of the energy storage battery system and the preprocessed multi-source operation data, construct a spatiotemporal correlation diagram, and based on each node in the spatiotemporal correlation diagram, establish a physical twin model and a data-driven benchmark model respectively.

[0042] In this optional embodiment, the step of constructing a spatiotemporal correlation graph based on the topology of the energy storage battery system and preprocessed multi-source operating data, and establishing a physical twin model and a data-driven benchmark model based on each node in the spatiotemporal correlation graph, includes:

[0043] Based on the electrical connection relationships and thermal coupling conduction relationships between individual battery cells or modules in the energy storage battery system, the nodes and edges of the spatiotemporal correlation graph are determined; wherein, the edges corresponding to the electrical connection relationships are electrical edges, and the edges corresponding to the thermal coupling conduction relationships are thermal edges;

[0044] Based on the real-time electrical parameters in the preprocessed multi-source operating data, calculate the edge weights of the electrical edges; based on the real-time thermal parameters in the preprocessed multi-source operating data, calculate the edge weights of the thermal edges; and construct a spatiotemporal correlation graph based on the node set, edge set, and dynamically updated edge weights.

[0045] For each node of the spatiotemporal correlation graph, a physical twin model based on an equivalent circuit coupled thermal resistance network is established according to the electrochemical mechanism parameters;

[0046] For each node in the spatiotemporal correlation diagram, a data-driven benchmark model based on load condition binning regression is established using historical health operation data.

[0047] Specifically, based on the series-parallel topology configuration information stored in the energy storage battery management system, each battery cell or module is identified as a node in the spatiotemporal correlation graph, forming a node set V; simultaneously, an electrical edge set is determined based on electrical connection relationships. Electrical edges are established between adjacent cells connected in series, and between cells at the same level in parallel branches. The set of thermal edges is determined based on the heat conduction paths between adjacent cells or modules in the battery pack structure layout. The thermal edge characterizes the thermal coupling relationship between spatially adjacent nodes; the electrical edge weight is dynamically calculated based on the real-time electrical parameters in the preprocessed multi-source operating data. Specifically, for the nodes connected by the electrical edge... With nodes Extract real-time current and the voltage difference between nodes ;

[0048] Based on the real-time electrical parameters in the preprocessed multi-source operating data, the edge weights of the electrical edges are calculated. Specifically, the edge weights are dynamically calculated based on the offline calibrated connection impedance reference values ​​of the busbar impedance and contact resistance, combined with the real-time current and the voltage difference between nodes. The larger the voltage difference between nodes and the smaller the real-time current, the smaller the electrical edge weight. When the real-time current is lower than the preset current lower limit, the electrical edge weight is determined only by the connection impedance reference value to avoid division by zero abnormality.

[0049] For hot edges, the weight of the hot edge is dynamically calculated based on the preprocessed real-time thermal parameters, and the nodes connected to the hot edge are... With nodes Extract the real-time temperature difference between nodes ;

[0050] The thermal edge weight is dynamically calculated based on the thermal resistance reference value and the real-time temperature difference between nodes. The larger the temperature difference between nodes, the smaller the thermal edge weight.

[0051] Based on the aforementioned node set V, edge set E (including electrical and thermal edges), and dynamically updated edge weight set W, a spatiotemporal correlation graph G=(V,E,W) is constructed. For each node in the spatiotemporal correlation graph G, a physical twin model is established. This model adopts a structure of a first-order RC equivalent circuit coupled to a lumped-parameter thermal resistance network, and the electrical part includes an open-circuit voltage source. Ohmic internal resistance Polarization internal resistance and polarization capacitor The thermal component includes the thermal resistance from the battery core to the surface. and surface thermal resistance to the environment Electro-thermal coupling via Joule thermal power The entropy-change heat power is calculated based on the real-time current, the absolute temperature value of the surface at the previous moment, and the rate of change of the open-circuit voltage with temperature. The surface temperature needs to be converted to the Kelvin scale before being included in the calculation. The model parameters... , , Obtained offline through standard HPPC pulse test data. The thermal resistance parameters are determined by looking up a pre-calibrated SOC-OCV mapping table. and Offline calibration was performed using a steady-state charge-discharge temperature rise experiment.

[0052] For each node, a data-driven benchmark model is established simultaneously. Historical operating data of the node under healthy conditions is collected. The health condition is determined by a capacity retention rate greater than 95% and an internal resistance growth rate less than 10%. The historical data is divided into operating condition bins according to the SOC range [0,20), [20,40), [40,60), [60,80), [80,100]%, the temperature range [-10,10), [10,25), [25,40), [40,55]℃, and the charge / discharge rate range [0,0.5), [0.5,1), [1,2), [2,3]C. A regression sub-model based on the XGBoost algorithm is trained in each operating condition bin. The input features are SOC, temperature, and charge / discharge rate, and the output targets are the voltage benchmark value and the temperature benchmark value. The hyperparameters of each sub-model are determined by 5-fold cross-validation, thus obtaining the data-driven benchmark model of each node under different operating condition-aging coupling conditions.

[0053] Step S103: Input the preprocessed multi-source running data into the physical twin model and the data-driven baseline model respectively to obtain the physical prediction residual and the data prediction residual;

[0054] In this optional embodiment, the step of inputting the preprocessed multi-source runtime data into the physical twin model and the data-driven baseline model respectively to obtain the physical prediction residual and the data prediction residual includes:

[0055] The electrical and thermal data from the preprocessed multi-source operating data are input into the physical twin model to obtain the physical prediction values ​​of voltage and temperature. The difference between the measured voltage and the physical prediction value, and the difference between the measured temperature and the physical prediction value are calculated separately and then fused to obtain the physical prediction residual.

[0056] The operating parameters in the preprocessed multi-source operating data are input into the data-driven benchmark model to obtain the voltage data benchmark value and the temperature data benchmark value. The difference between the measured voltage and the voltage data benchmark value and the difference between the measured temperature and the temperature data benchmark value are calculated respectively, and the data prediction residual is obtained by fusion.

[0057] Specifically, the electrical quantity data (including charging and discharging current) at the current moment from the preprocessed multi-source operating data will be... and the state of charge updated by ampere-hour integration ) and thermal data (including ambient temperature) The input is sent to the physical twin model corresponding to each node. For the electrical part, based on the current... Open circuit voltage can be obtained by looking up a table. And solve the polarization voltage differential equation. The polarization voltage was obtained by discrete iterative calculation using the forward Euler method with a step size of 1 second. Then calculate the physical prediction value of the terminal voltage. For the thermal part, according to Joule heat power Entropy change heat power To calculate the total heat generation, for the thermal part, the total heat generation power of the battery core is obtained by summing the Joule heat power and the entropy change heat power. The heat dissipation power from the core to the surface is calculated based on the temperature difference between the battery core and the surface, and the thermal resistance from the core to the surface. The rate of change of the core temperature is calculated based on the difference between the total heat generation power and the heat dissipation power, and the core heat capacity. This process is then iteratively solved to obtain the physical prediction value of the core temperature. For the surface temperature, the heat dissipation power from the core to the surface, the temperature difference between the surface and the ambient temperature, and the thermal resistance from the surface to the environment are used to calculate the heat dissipation power from the surface to the environment. The rate of change of the surface temperature is calculated based on the difference between the heat dissipation power from the core to the surface and the heat dissipation power from the surface to the environment, and the surface heat capacity. This process is then iteratively solved to obtain the physical prediction value of the surface temperature. .

[0058] Calculate voltage physical prediction residual and temperature physics prediction residual , and according to The residuals are merged into physical predictions, where and These are the normalized weighting coefficients for voltage and temperature residuals, determined based on the reciprocals of the standard deviations of voltage and temperature from historical health data statistics; simultaneously, the operating condition parameters (including current) from the preprocessed multi-source operating data are... Measured temperature and charge / discharge rate ,in (Rated capacity) is input to the data-driven baseline model corresponding to each node, based on... , and The input condition sub-model calls the corresponding XGBoost sub-model to output voltage data reference value. and temperature data baseline value Calculate voltage data prediction residuals and temperature data prediction residuals , and according to The data is merged into prediction residuals, thus obtaining the physical prediction residuals of each node at the current time. data prediction residual .

[0059] Step S104: Perform cross-validation between physical prediction residuals and data prediction residuals, filter out abnormal battery nodes, calculate the bidirectional attention weights between abnormal battery nodes and neighboring nodes based on the spatiotemporal correlation graph, and locate the fault source node based on the bidirectional attention weight distribution.

[0060] In this optional embodiment, the step of cross-validating the physical prediction residual and the data prediction residual to filter out abnormal battery nodes, calculating the bidirectional attention weights between the abnormal battery nodes and their neighboring nodes based on the spatiotemporal correlation graph, and locating the fault source node according to the bidirectional attention weight distribution includes:

[0061] When both the physical prediction residual and the data prediction residual exceed their respective preset residual thresholds, the node is determined to be a battery abnormal node and is included in the abnormal feature propagation set.

[0062] When the data prediction residual exceeds the preset residual threshold, but the physical prediction residual does not exceed the threshold, it is determined that there is a sensor communication anomaly in the corresponding data source and the confidence of the data source is reduced, and the node is removed from the abnormal feature propagation set.

[0063] When the physical prediction residual exceeds the preset residual threshold, while the data prediction residual does not exceed the threshold, the physical twin model's parameter adaptive update is triggered.

[0064] Based on the edge weights of electrical edges and thermal edges in the spatiotemporal correlation graph, calculate the bidirectional attention weights between each battery anomaly node and its neighboring nodes in the anomaly feature propagation set.

[0065] Based on the bidirectional attention weight, when the attention weight of the abnormal feature propagation set in a certain direction is greater than the attention weight of the opposite direction, the starting node in that direction is determined to be the fault source node and the ending node is the node affected by the fault propagation.

[0066] Specifically, a preset residual threshold is determined based on historical health operation data statistics, wherein the physical prediction residual threshold is... The threshold for data prediction residuals is the mean plus three standard deviations of the physical prediction residual sequence under healthy conditions. Take the mean of the predicted residual sequence of data under healthy conditions plus three times the standard deviation; calculate the physical prediction residual for each node at the current time. data prediction residual Perform mutual verification when Greater than and Greater than When this node is identified as a battery anomaly node, it is included in the anomaly feature propagation set. Greater than and Not greater than When an anomaly is detected in the corresponding data source, the confidence coefficient of that node's data source is reduced to 0.5 times its original value, and a sensor self-check procedure is triggered. Simultaneously, if the node already exists in the anomaly feature propagation set, it is removed. Greater than and Not greater than When the parameters of the physical twin model are determined to have drifted due to battery aging, an adaptive update of the parameters is triggered. Specifically, a recursive least squares method with a forgetting factor is used to update the ohmic internal resistance in the physical twin model online. and polarization resistance The forgetting factor ranges from 0.95 to 0.99.

[0067] It should be noted that the above-mentioned physical twin model parameter adaptive update mechanism is applicable to progressive aging drift scenarios. To avoid real aging faults being misjudged as model parameter drift, after triggering the parameter adaptive update, the aging drift confidence timing window is started simultaneously. If the physical prediction residual after parameter update continues to exceed the preset residual threshold within a preset time period (e.g., no less than two complete charge-discharge cycles), or the data prediction residual shows a synchronous trend of increasing, the parameter adaptive update is terminated, and the node is re-included in the abnormal feature propagation set for fault determination.

[0068] After mutual verification is completed, for each battery anomaly node in the anomaly feature propagation set... and its neighboring nodes in the spatiotemporal relationship graph Extract nodes fused residual feature vector With nodes fused residual feature vector and extract nodes and Electrical edge weights between and hot edge weight Calculate the attention coefficient ,in It is a linear transformation matrix. For attention vectors, The vector concatenation is represented by a LeakyReLU negative slope of 0.2, which is then normalized using softmax to obtain the bidirectional attention weights. In the formula, and Representing nodes respectively and The set of neighboring nodes.

[0069] Based on the edge weights of electrical edges and thermal edges in the spatiotemporal correlation graph, the bidirectional attention weights between each battery abnormal node and its neighboring nodes in the abnormal feature propagation set are calculated, and the first attention weight from the battery abnormal node to its neighboring node and the second attention weight from the neighboring node to the battery abnormal node are obtained respectively.

[0070] Based on the bidirectional attention weights, when the first attention weight is greater than a preset multiple of the second attention weight, it is determined that the abnormal feature is propagating from the abnormal battery node to the neighboring node, the abnormal battery node is the fault source node, and the neighboring node is the node affected by the fault propagation; when the second attention weight is greater than a preset multiple of the first attention weight, it is determined that the abnormal feature is propagating from the neighboring node to the abnormal battery node, the neighboring node is the fault source node, and the abnormal battery node is the node affected by the fault propagation.

[0071] The preset multiplier is an empirical directional discrimination threshold determined by offline data calibration based on the energy storage system topology, the statistical distribution benchmark of bidirectional attention weights under healthy conditions, and the acceptable risk of false alarms / false alarms.

[0072] Step S105: Extract multi-dimensional abnormal feature vectors from the fault source node, and calculate the marginal contribution of each dimension feature to the fault probability through counterfactual inference.

[0073] In this optional embodiment, the step of extracting multi-dimensional anomaly feature vectors from the fault source node and calculating the marginal contribution of each dimension feature to the fault probability through counterfactual inference includes:

[0074] The target dimension features in the multi-dimensional abnormal feature vector are replaced with the corresponding normal baseline values ​​through counterfactual inference to obtain the counterfactual feature vector;

[0075] Based on multi-dimensional anomaly feature vectors and counterfactual feature vectors, the first fault probability and the second fault probability are calculated respectively, and the marginal contribution is output.

[0076] In this optional embodiment, the calculation of the first fault probability and the second fault probability based on the multi-dimensional anomaly feature vector and the counterfactual feature vector, and the output of the marginal contribution, includes:

[0077] Obtain the normal baseline value of the fault source node under the current operating condition and health status from the pre-configured operating condition-aging coupled baseline library;

[0078] Replace the target dimension features in the multi-dimensional abnormal feature vector with normal baseline values ​​one by one, while keeping the other dimensions unchanged, to generate counterfactual feature vectors.

[0079] Using a preset fault probability prediction function, the fault probabilities of the original feature vector and the counterfactual feature vector are calculated respectively, and the difference between the two is taken as the marginal contribution of the target dimension feature.

[0080] Specifically, a multi-dimensional abnormal feature vector is extracted from the fault source node. This vector includes voltage deviation features, temperature rise rate features, internal resistance increment features, and voltage difference anomaly features. The voltage deviation feature is the difference between the current measured voltage of the node and the voltage data benchmark value output by the data-driven benchmark model. The temperature rise rate feature is the difference between the surface temperature of the node at the current time and the previous time divided by the sampling interval. The internal resistance increment feature is the difference between the current ohmic internal resistance after adaptive update of the physical twin model and the factory-calibrated internal resistance divided by the factory-calibrated internal resistance. The voltage difference anomaly feature is the absolute value of the difference between the voltage of the node and the average voltage of the module.

[0081] A pre-configured operating condition-aging coupled baseline library is provided. This baseline library is divided into three dimensions: charge state range [0,20), [20,40), [40,60), [60,80), [80,100]%, temperature range [-10,10), [10,25), [25,40), [40,55]℃, and health state range [90,100), [80,90), [70,80)%. Each bin stores the statistical baseline values ​​of each dimension feature in historical health operation data. The statistical baseline value is the median of the corresponding feature in that bin.

[0082] Based on the current state of charge, temperature, and health status determined by capacity retention of the fault source node, the corresponding bins in the operating condition-aging coupled baseline library are matched to obtain the normal baseline values ​​for each dimension of features, thus obtaining the normal baseline vector. ;

[0083] Let the original multidimensional anomaly feature vector be . By calculating the marginal contribution of each dimension of features one by one through counterfactual deduction, the marginal contribution of the first feature is determined. Each target dimension feature is replaced with its corresponding normal baseline value. while keeping the rest With each dimension of feature unchanged, a counterfactual feature vector is generated. .

[0084] It should be noted that the above counterfactual inference method, which calculates the marginal contribution by replacing each dimension independently, is based on the following premise: within the current operating condition and health status bins defined by the operating condition-aging coupling baseline library, the features of each dimension are approximately decoupled; for feature dimensions with strong physical coupling (such as internal resistance increment and temperature rise rate), their marginal contribution characterizes the independent explanatory power of the dimension under the existing operating condition boundary conditions. In actual root cause identification, the joint contribution of coupled feature clusters can be combined for auxiliary verification to avoid distortion of the marginal contribution caused by feature coupling.

[0085] The original feature vectors are calculated using a preset fault probability prediction function. Failure probability counterfactual feature vector Failure probability The fault probability prediction function uses a multilayer perceptron model trained with historical fault samples, where the number of input layer nodes equals the number of feature dimensions. The hidden layer consists of two fully connected layers, each with 64 neurons and using the ReLU activation function. The output layer uses the Sigmoid activation function to output the fault probability. During model training, the ratio of positive to negative samples is balanced using SMOTE oversampling; the difference between the two is used as the marginal contribution of the feature in that target dimension. traverse all The marginal contribution vector is obtained after considering each dimension. .

[0086] Step S106: Based on the marginal contribution, identify the dominant root cause of the failure and input the dominant root cause of the failure into the dynamic Bayesian network. Combined with the spatiotemporal correlation graph, predict the failure evolution and propagation path and time window.

[0087] In this optional embodiment, the step of identifying the dominant root cause of the failure based on the marginal contribution, and inputting the dominant root cause of the failure into a dynamic Bayesian network, combined with a spatiotemporal correlation graph, to predict the failure evolution and propagation path and time window includes:

[0088] The marginal contribution of each dimension is sorted in descending order, and the dimension feature that ranks first and exceeds the preset root cause threshold is determined as the dominant root cause of the failure.

[0089] Root cause confidence weights are generated based on marginal contribution, and the evidence confidence of the observed nodes mapped to the dominant root cause of the fault is corrected based on the root cause confidence weights to obtain a dynamic Bayesian network.

[0090] The primary cause of the fault is input as initial evidence of anomaly into a dynamic Bayesian network. By combining the spatiotemporal correlation graph with forward probabilistic reasoning, the fault evolution and propagation path and the time window for each node to reach the preset danger threshold are obtained.

[0091] In this optional embodiment, the step of inputting the dominant root cause of the fault as initial anomalous evidence into a dynamic Bayesian network, and performing forward probabilistic inference in conjunction with a spatiotemporal correlation graph to obtain the fault evolution propagation path and the time windows for each node to reach a preset danger threshold includes:

[0092] A dynamic Bayesian network is constructed with battery state variables as nodes and state transition relations as edges. A mapping relationship between each dimension of the multi-dimensional anomaly feature vector and the battery state variables is established, and the nodes corresponding to the mapping relationship are used as observation nodes.

[0093] Based on the topological connections of the spatiotemporal correlation graph, determine the spatial constraints of state transitions for each node in the dynamic Bayesian network.

[0094] The primary cause of the failure is used as initial anomalous evidence and input into the observation node. Forward probabilistic reasoning is performed under spatial constraints, and the failure evolution probability of each node is calculated through forward probabilistic reasoning.

[0095] Determine the fault evolution propagation path based on the propagation direction of the fault evolution probability on the spatiotemporal correlation graph;

[0096] The time window is predicted based on the probability distribution of the cumulative failure evolution probability reaching a preset danger threshold.

[0097] In this optional embodiment, the step of inputting the dominant root cause of the failure as initial anomalous evidence into the observation node, performing forward probabilistic inference under spatial constraints, and calculating the failure evolution probability of each node through forward probabilistic inference includes:

[0098] Based on the magnitude of the marginal contribution, the dominant root cause of the fault is quantified into the soft evidence distribution of the corresponding observation nodes in the dynamic Bayesian network, thus obtaining the initial anomaly evidence configuration.

[0099] Extract the attention weights and edge weights from the fault source node to each neighbor node in the spatiotemporal correlation graph, and construct a spatial coupling transition matrix;

[0100] Based on the posterior probability distribution of each node in the previous time slice, the initial abnormal evidence configuration, and the spatial coupling transition matrix, the state transition prior between nodes of the dynamic Bayesian network is corrected, and the posterior probability distribution of each node's state is calculated iteratively on a time slice basis.

[0101] Based on the increment of the failure state probability in the posterior probability distribution of adjacent time slices, calculate the failure evolution probability of each node transitioning from a healthy state to a failure state.

[0102] Specifically, regarding the marginal contribution vector Sort in descending order and set a preset root cause threshold. The 95th percentile of the statistical distribution of the marginal contribution of each dimension in the historical fault samples will be ranked first and satisfy the following conditions: Greater than The dimensional features are determined to be the dominant root cause of the failure. If the marginal contribution of the first-ranked factor does not exceed [the specified value], [further details are needed]. It is then determined to be a compound root cause and the first two digits are taken.

[0103] Based on the marginal contribution of the dominant root cause of the failure Generate root cause confidence weights The evidence confidence of the observed node mapped by the dominant root cause of the fault is corrected based on the root cause confidence weight. Specifically, the soft evidence distribution of the observed node is corrected according to the root cause confidence weight, so that the larger the root cause confidence weight, the stronger the probability of the fault state of the observed node and the weaker the probability of the healthy state, and the sum of the probabilities of each state after correction remains 1.

[0104] A dynamic Bayesian network is constructed with the battery state variables of each battery cell as nodes and the state transition relations as edges. The state variables include the state of charge, core temperature, internal resistance, cell voltage difference and insulation resistance. Each state variable is discretized into three state levels: healthy, abnormal and fault. A mapping relationship between each dimension of the multi-dimensional abnormal feature vector and the above battery state variables is established, and the nodes corresponding to the mapping relationship are used as observation nodes.

[0105] Based on the topological connections of the spatiotemporal correlation graph, construct the spatial coupling transition matrix. , of which elements Representation Nodes and The spatial coupling strength between nodes; for node pairs in the spatiotemporal correlation graph that are not directly connected by electrical or thermal edges, their spatial coupling coefficient is set to zero.

[0106] The primary cause of the fault is used as initial anomalous evidence input to the corresponding observation node. A forward algorithm is used to iteratively calculate the posterior probability distribution of each node's state on a time-slice basis. During the inference process of each time slice, the prior probability of each node's fault state is corrected according to the spatial coupling transition matrix. Specifically, for any node, if the posterior probability of its neighboring node's fault state in the previous time slice is higher than the node's baseline prior probability, and the spatial coupling strength between the node and its corresponding neighboring node is greater, then the prior probability of the node's fault state in the current time slice is increased; otherwise, it is weakened. The corrected prior probabilities are truncated and normalized to ensure that they are within the probability interval [0,1].

[0107] Based on the increment of the failure state probability in the posterior probability distribution of adjacent time slices, calculate the failure evolution probability of each node transitioning from a healthy state to a failure state.

[0108] Step S107: Output the fault warning result based on the dominant root cause of the fault, the fault evolution and propagation path and the time window.

[0109] Specifically, the system receives the root cause of the fault, the path of fault evolution and propagation, and the time window for each node to reach the preset danger threshold output in step S106, and determines the warning level according to the mapping rule between the cumulative value of fault evolution probability and the warning level.

[0110] When the probability of a fault source node's current fault state is below 0.3, a Level 3 alert is output; when it is between 0.3 and 0.7, a Level 2 warning is output; and when it is above 0.7, a Level 1 emergency warning is output. Simultaneously, the warning level is dynamically adjusted based on the time window length. If the time window for a node to reach the danger threshold is less than 5 minutes, the warning level for that node is increased by one level. Based on the pre-set handling strategy library for matching the dominant root cause of the fault, a fault warning result is generated, including root cause location information, evolutionary propagation path, the time window for each node to reach the danger state, and quantitative intervention suggestions. The quantitative intervention suggestions are automatically generated according to the root cause type. For example, when the root cause is voltage deviation, it is recommended to reduce the charging rate or suspend charging; when the root cause is temperature rise rate, it is recommended to start the cooling system or reduce operating power; and when the root cause is internal resistance increment, it is recommended to limit the depth of discharge and arrange maintenance.

[0111] The fault warning results are output to the battery management system and operation and maintenance display terminal in the form of structured data packets. The structured data packets include the warning level identifier, fault source node number, root cause characteristic description, propagation path sequence, upper and lower limits of time windows for each node, recommended handling measures and evidence chain index. At the same time, the corresponding safety control strategies are triggered, including sending a power reduction command to the battery management system, disconnecting the fault module contactor or starting the fire emergency procedure. All intermediate calculation results and decision basis of this warning are stored in the historical database to support subsequent model optimization and accident tracing.

[0112] It should be noted that the above-mentioned mechanism for dynamically adjusting the warning level based on the length of the time window has the highest level of emergency warning as the upper limit. If the current warning level of a node is already the level of emergency warning, then even if the time window for it to reach the danger threshold is less than 5 minutes, the level of emergency warning will be maintained and no further adjustment will be performed.

[0113] Figure 2 An embodiment of an energy storage battery fault early warning system based on multi-source operating data according to the present invention is shown.

[0114] In this optional embodiment, the energy storage battery fault early warning system based on multi-source operational data includes:

[0115] The multi-source data acquisition module 201 is used to acquire and preprocess multi-source operating data of the energy storage battery system.

[0116] The spatiotemporal correlation graph construction module 202 is used to construct a spatiotemporal correlation graph based on the topology of the energy storage battery system and the preprocessed multi-source operation data, and to establish a physical twin model and a data-driven benchmark model based on each node in the spatiotemporal correlation graph.

[0117] The dual residual generation module 203 is used to input the preprocessed multi-source running data into the physical twin model and the data-driven baseline model respectively to obtain the physical prediction residual and the data prediction residual;

[0118] The residual cross-validation module 204 is used to cross-validate the physical prediction residual and the data prediction residual, filter out abnormal battery nodes, calculate the bidirectional attention weight between the abnormal battery node and its neighboring nodes based on the spatiotemporal correlation graph, and locate the fault source node according to the bidirectional attention weight distribution.

[0119] The fault source localization module 205 is used to extract multi-dimensional abnormal feature vectors from the fault source node and calculate the marginal contribution of each dimension feature to the fault probability through counterfactual inference.

[0120] The root cause identification and path prediction module 206 is used to identify the dominant root cause of the fault based on the marginal contribution, and input the dominant root cause of the fault into a dynamic Bayesian network, combined with the spatiotemporal correlation graph, to predict the fault evolution and propagation path and time window.

[0121] The early warning output module 207 is used to output the fault early warning result based on the dominant root cause of the fault, the fault evolution and propagation path and the time window.

[0122] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle, as follows:

[0123] Multi-source operational data, including voltage, current, and temperature, of the energy storage battery system are collected and preprocessed. A spatiotemporal correlation graph is constructed based on the system's electrical connections and thermal coupling topology. Simultaneously, a physical twin model based on electrochemical mechanisms and a data-driven benchmark model based on historical health data are established. The preprocessed data is input into the dual models to obtain physical prediction residuals and data prediction residuals. A cross-validation mechanism is used to distinguish between real battery anomalies and sensor anomalies. After identifying real anomaly nodes, a graph attention network is used to propagate anomaly features on the spatiotemporal correlation graph. The fault source node is located based on the asymmetric distribution of attention weights. Multi-dimensional anomaly feature vectors are extracted from the fault source node. The marginal contribution of each dimension feature to the fault probability is calculated one by one through counterfactual inference, thereby identifying the dominant root cause of the fault. The dominant root cause is then used as initial anomaly evidence and input into a dynamic Bayesian network. Forward probabilistic inference is performed under the spatial constraints of the spatiotemporal correlation graph to predict the evolution and propagation path of the fault along the electrical-thermal topology and the time window for each node to reach the danger threshold. Finally, a fault warning result containing warning levels and quantitative intervention suggestions is generated based on the root cause, path, and time window.

[0124] The following example illustrates the application scenario of an energy storage module consisting of four lithium iron phosphate cells (B1, B2, B3, and B4) connected in series under a 1C charging condition:

[0125] The battery management system (BMS) collects the voltage, current, and surface temperature of each cell at a frequency of 1Hz. The measured voltage of B1 is 3.25V and temperature is 32℃; the measured voltage of B2 is 3.30V and temperature is 26℃; the measured voltage of B3 is 3.29V and temperature is 25℃; and the measured voltage of B4 is 3.28V and temperature is 25℃. The system current is 5A. Preprocessed data is obtained after time alignment, 3σ criterion outlier removal, and Min-Max normalization. A spatiotemporal correlation graph is constructed based on the four-cell series topology and the heat conduction paths between adjacent cells within the module. The electrical edge weights are calculated based on the busbar impedance and real-time current. The electrical edge weight between B1 and B2 is 0.80, and the thermal edge weight is 0.30 (calculated based on real-time temperature difference and structural thermal resistance). The electrical edge weight between B2 and B3 is 0.75, and the thermal edge weight is 0.2. 5. The electrical edge weight between B3 and B4 is 0.78, and the thermal edge weight is 0.28. There are no edge connections between non-adjacent units. A physical twin model based on a first-order RC equivalent circuit coupled thermal resistance network is established for each node. The model parameters are calibrated by offline HPPC testing. The factory internal resistance of B1 is 1.5mΩ and the polarization internal resistance is 3.0mΩ. At the same time, a data-driven benchmark model based on XGBoost is established. With SOC, temperature, and rate as input, the voltage and temperature benchmark values ​​are obtained through sub-bin training under operating conditions. The preprocessed data is input into the dual model. The physical twin model outputs the physical prediction value of B1 voltage as 3.30V and the physical prediction value of temperature as 26℃. The data-driven benchmark model outputs the data benchmark value of B1 voltage as 3.29V and the data benchmark value of temperature as 25.5℃. The physical prediction residual of B1 is calculated. =0.12, data prediction residual =0.11, preset residual threshold =0.08、 =0.10, both B1 residuals exceed the threshold and are therefore included in the abnormal feature propagation set. The physical twin model outputs a physical prediction value of 3.38V for voltage and 22℃ for temperature for B2. The data-driven baseline model outputs a baseline value of 3.39V for voltage and 23℃ for temperature for B2. The calculated physical prediction residuals for B2 are... =0.09, data prediction residual =0.11, both residuals of B2 exceed the threshold, so they are included in the abnormal feature propagation set; both residuals of B3 and B4 are normal; the bidirectional attention weights between B1 and its neighbor B2 are calculated based on the spatiotemporal correlation graph, and the attention weights in the B1→B2 direction are obtained. B2→B1 direction ,satisfy >1.5× Therefore, B1 is determined to be the fault source node, and B2 is the node affected by the propagation. A multi-dimensional abnormal feature vector is extracted from B1, including voltage deviation of -0.05V, temperature rise rate of 6℃ / min, internal resistance increment of 15%, and differential voltage abnormality of 0.03V (i.e., the absolute value of the difference between the measured voltage of this node (3.25V) and the average voltage of the module (3.28V)). Corresponding normal baseline values ​​of 0V, 0.5℃ / min, 5%, and 0.02V are obtained from the operating condition-aging coupled baseline library. Through counterfactual deduction: replacing the voltage deviation with the baseline value of 0V reduces the fault probability from 0.85 to 0.50, with a marginal contribution of 0.35; replacing the temperature rise rate with 0.5℃ / min reduces the fault probability to 0.73, with a marginal contribution of 0.12; replacing the internal resistance increment with 5% reduces it to 0.77, with a marginal contribution of 0.08; and replacing the differential voltage with 0.02V reduces it to 0.80, with a marginal contribution of 0.05. A preset root cause threshold is used. Therefore, the voltage deviation was determined to be the dominant root cause of the fault, and the root cause confidence weight was determined accordingly. A dynamic Bayesian network is constructed, with state variables including internal resistance, temperature, and pressure difference. Each variable is discretized into three levels: healthy, abnormal, and fault. A mapping relationship between voltage deviation and internal resistance state nodes is established, and these nodes are set as observation nodes. Spatial constraints are determined based on the spatiotemporal correlation graph topology. The state transition between B1 and B2 is subject to the spatial coupling coefficient. Constraints are applied such that there is no direct connection between B1 and B3, therefore the transition probability constraint is zero. A spatial coupling transition matrix is ​​constructed based on the spatial coupling coefficient, and the state transition probabilities between the corrected nodes are normalized row by row to ensure that the sum of the state probability distributions in each time slice is 1. At t=1, the posterior probability of the B2 fault state increases to 0.28, and the fault evolution probability ΔP=0.23; at t=2, it increases to 0.48; at t=3, it increases to 0.68, corresponding to the first time the cumulative probability exceeds 0.5; at t=4, it reaches 0.92, which is significant. Since the cumulative probability first exceeds 0.9 and surpasses the preset danger threshold of 0.70, the lower limit of the time window for B2 to reach a dangerous state is 3 minutes and the upper limit is 4 minutes; the failure probability of B3 is 0.19 at t=2, 0.42 at t=4, 0.48 at t=5, reaches 0.71 at t=6 (the cumulative probability first exceeds 0.5), and reaches 0.93 at t=7 (the cumulative probability first exceeds 0.9), with a lower limit of 6 minutes and an upper limit of 7 minutes; the failure probability of B4 is 0.15 at t=3, t= At 6:00, the probability was 0.38; at t=8:00, it was 0.48; at t=9:00, it reached 0.72 (the cumulative probability exceeded 0.5 for the first time); at t=10:00, it reached 0.91 (the cumulative probability exceeded 0.9 for the first time). The lower limit of the time window was 9 minutes, and the upper limit was 10 minutes. Based on the propagation direction of the probability increment of the fault evolution in each time slice, the probability increment of B1→B2 (0.23) was greater than that of the reverse direction (0.04); the increment of B2→B3 (0.19) was greater than that of the reverse direction (0.03); and the increment of B3→B4 (0.16) was greater than that of the reverse direction (0.02). The fault propagation path is B1→B2→B3→B4; the final output fault warning result is: the warning level is Level 1 emergency warning, the main root cause of the fault is the voltage deviation of B1, the fault propagation path is B1→B2→B3→B4, B2 is expected to reach a dangerous state in 3 to 4 minutes, B3 is expected to reach a dangerous state in 6 to 7 minutes, and B4 is expected to reach a dangerous state in 9 to 10 minutes. The handling recommendation is to immediately reduce the charging rate from 1C to 0.5C and disconnect the contactor of module B1, and at the same time start the module cooling system, as shown in Table 1.

[0126] Table 1 Fault Early Warning Judgment Results for Each Node

[0127] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A method for early warning of energy storage battery faults based on multi-source operational data, characterized in that, include: Acquire multi-source operational data of the energy storage battery system and perform preprocessing; Based on the topology of the energy storage battery system and the preprocessed multi-source operating data, a spatiotemporal correlation graph is constructed, and a physical twin model and a data-driven benchmark model are established based on each node in the spatiotemporal correlation graph. The preprocessed multi-source operational data is input into the physical twin model and the data-driven baseline model respectively to obtain the physical prediction residual and the data prediction residual; The physical prediction residual and the data prediction residual are cross-validated to screen out abnormal battery nodes. The bidirectional attention weight between the abnormal battery node and its neighboring nodes is calculated based on the spatiotemporal correlation graph. The fault source node is located according to the bidirectional attention weight distribution. Extract multi-dimensional abnormal feature vectors from the fault source node, and calculate the marginal contribution of each feature to the fault probability through counterfactual inference. Based on the marginal contribution, the dominant root cause of the failure is identified and input into a dynamic Bayesian network. Combined with the spatiotemporal correlation graph, the failure evolution and propagation path and time window are predicted. Based on the root cause of the fault, the fault evolution and propagation path, and the time window, output the fault warning result.

2. The method for early warning of energy storage battery faults based on multi-source operational data according to claim 1, characterized in that, The process of constructing a spatiotemporal correlation graph based on the topology of the energy storage battery system and preprocessed multi-source operational data, and establishing a physical twin model and a data-driven baseline model based on each node in the spatiotemporal correlation graph, includes: Based on the electrical connection relationships and thermal coupling conduction relationships between individual battery cells or modules in the energy storage battery system, the nodes and edges of the spatiotemporal correlation graph are determined; wherein, the edges corresponding to the electrical connection relationships are electrical edges, and the edges corresponding to the thermal coupling conduction relationships are thermal edges; Based on the real-time electrical parameters in the preprocessed multi-source operating data, calculate the edge weights of the electrical edges; based on the real-time thermal parameters in the preprocessed multi-source operating data, calculate the edge weights of the thermal edges; and construct a spatiotemporal correlation graph based on the node set, edge set, and dynamically updated edge weights. For each node of the spatiotemporal correlation graph, a physical twin model based on an equivalent circuit coupled thermal resistance network is established according to the electrochemical mechanism parameters; For each node in the spatiotemporal correlation diagram, a data-driven benchmark model based on load condition binning regression is established using historical health operation data.

3. The method for early warning of energy storage battery faults based on multi-source operational data according to claim 1, characterized in that, The step of inputting the preprocessed multi-source operational data into the physical twin model and the data-driven baseline model respectively to obtain the physical prediction residuals and data prediction residuals includes: The electrical and thermal data from the preprocessed multi-source operating data are input into the physical twin model to obtain the physical prediction values ​​of voltage and temperature. The difference between the measured voltage and the physical prediction value, and the difference between the measured temperature and the physical prediction value are calculated separately and then fused to obtain the physical prediction residual. The operating parameters in the preprocessed multi-source operating data are input into the data-driven benchmark model to obtain the voltage data benchmark value and the temperature data benchmark value. The difference between the measured voltage and the voltage data benchmark value and the difference between the measured temperature and the temperature data benchmark value are calculated respectively, and the data prediction residual is obtained by fusion.

4. The method for early warning of energy storage battery faults based on multi-source operational data according to claim 1, characterized in that, The process of cross-validating the physical prediction residuals and the data prediction residuals to screen out abnormal battery nodes, and calculating the bidirectional attention weights between abnormal battery nodes and their neighboring nodes based on the spatiotemporal correlation graph, and locating the fault source node based on the bidirectional attention weight distribution includes: When both the physical prediction residual and the data prediction residual exceed their respective preset residual thresholds, the node is determined to be a battery abnormal node and is included in the abnormal feature propagation set. When the data prediction residual exceeds the preset residual threshold, but the physical prediction residual does not exceed the threshold, it is determined that there is a sensor communication anomaly in the corresponding data source and the confidence of the data source is reduced, and the node is removed from the abnormal feature propagation set. When the physical prediction residual exceeds the preset residual threshold, while the data prediction residual does not exceed the threshold, the physical twin model's parameter adaptive update is triggered. Based on the edge weights of electrical edges and thermal edges in the spatiotemporal correlation graph, calculate the bidirectional attention weights between each battery anomaly node and its neighboring nodes in the anomaly feature propagation set. Based on the bidirectional attention weight, when the attention weight of the abnormal feature propagation set in a certain direction is greater than the attention weight of the opposite direction, the starting node in that direction is determined to be the fault source node and the ending node is the node affected by the fault propagation.

5. The method for early warning of energy storage battery faults based on multi-source operational data according to claim 1, characterized in that, The step of extracting multi-dimensional abnormal feature vectors from the fault source node and calculating the marginal contribution of each feature dimension to the fault probability through counterfactual inference includes: The target dimension features in the multi-dimensional abnormal feature vector are replaced with the corresponding normal baseline values ​​through counterfactual inference to obtain the counterfactual feature vector; Based on multi-dimensional anomaly feature vectors and counterfactual feature vectors, the first fault probability and the second fault probability are calculated respectively, and the marginal contribution is output.

6. The method for early warning of energy storage battery faults based on multi-source operational data according to claim 5, characterized in that, The first fault probability and the second fault probability are calculated based on the multi-dimensional anomaly feature vector and the counterfactual feature vector, respectively, and the marginal contribution is output including: Obtain the normal baseline value of the fault source node under the current operating condition and health status from the pre-configured operating condition-aging coupled baseline library; Replace the target dimension features in the multi-dimensional abnormal feature vector with normal baseline values ​​one by one, while keeping the other dimensions unchanged, to generate counterfactual feature vectors. Using a preset fault probability prediction function, the fault probabilities of the original feature vector and the counterfactual feature vector are calculated respectively, and the difference between the two is taken as the marginal contribution of the target dimension feature.

7. The method for early warning of energy storage battery faults based on multi-source operational data according to claim 1, characterized in that, The process of identifying the dominant root cause of the failure based on marginal contribution, inputting the dominant root cause into a dynamic Bayesian network, and combining it with a spatiotemporal correlation graph to predict the failure evolution and propagation path and time window includes: The marginal contribution of each dimension is sorted in descending order, and the dimension feature that ranks first and exceeds the preset root cause threshold is determined as the dominant root cause of the failure. Root cause confidence weights are generated based on marginal contribution, and the evidence confidence of the observed nodes mapped to the dominant root cause of the fault is corrected based on the root cause confidence weights to obtain a dynamic Bayesian network. The primary cause of the fault is input as initial evidence of anomaly into a dynamic Bayesian network. By combining the spatiotemporal correlation graph with forward probabilistic reasoning, the fault evolution and propagation path and the time window for each node to reach the preset danger threshold are obtained.

8. The method for early warning of energy storage battery faults based on multi-source operational data according to claim 7, characterized in that, The step of inputting the dominant cause of the fault as initial anomalous evidence into a dynamic Bayesian network, and combining it with a spatiotemporal correlation graph for forward probabilistic inference to obtain the fault evolution propagation path and the time windows for each node to reach the preset danger threshold includes: A dynamic Bayesian network is constructed with battery state variables as nodes and state transition relations as edges. A mapping relationship between each dimension of the multi-dimensional anomaly feature vector and the battery state variables is established, and the nodes corresponding to the mapping relationship are used as observation nodes. Based on the topological connections of the spatiotemporal correlation graph, determine the spatial constraints of state transitions for each node in the dynamic Bayesian network. The primary cause of the failure is used as initial anomalous evidence and input into the observation node. Forward probabilistic reasoning is performed under spatial constraints, and the failure evolution probability of each node is calculated through forward probabilistic reasoning. Determine the fault evolution propagation path based on the propagation direction of the fault evolution probability on the spatiotemporal correlation graph; The time window is predicted based on the probability distribution of the cumulative failure evolution probability reaching a preset danger threshold.

9. A method for early warning of energy storage battery faults based on multi-source operational data according to claim 8, characterized in that, The process of inputting the dominant cause of the failure as initial evidence of anomaly into the observation nodes, performing forward probabilistic inference under spatial constraints, and calculating the failure evolution probability of each node through forward probabilistic inference includes: Based on the magnitude of the marginal contribution, the dominant root cause of the fault is quantified into the soft evidence distribution of the corresponding observation nodes in the dynamic Bayesian network, thus obtaining the initial anomaly evidence configuration. Extract the attention weights and edge weights from the fault source node to each neighbor node in the spatiotemporal correlation graph, and construct a spatial coupling transition matrix; Based on the posterior probability distribution of each node in the previous time slice, the initial abnormal evidence configuration, and the spatial coupling transition matrix, the state transition prior between nodes of the dynamic Bayesian network is corrected, and the posterior probability distribution of each node's state is calculated iteratively on a time slice basis. Based on the increment of the failure state probability in the posterior probability distribution of adjacent time slices, calculate the failure evolution probability of each node transitioning from a healthy state to a failure state.

10. A fault early warning system for energy storage batteries based on multi-source operational data, characterized in that, include: The multi-source data acquisition module is used to acquire and preprocess multi-source operating data of the energy storage battery system; The spatiotemporal correlation graph construction module is used to construct a spatiotemporal correlation graph based on the topology of the energy storage battery system and the preprocessed multi-source operating data, and to establish a physical twin model and a data-driven benchmark model based on each node in the spatiotemporal correlation graph. The dual residual generation module is used to input the preprocessed multi-source running data into the physical twin model and the data-driven baseline model respectively to obtain the physical prediction residual and the data prediction residual; The residual cross-validation module is used to cross-validate the physical prediction residual and the data prediction residual, filter out abnormal battery nodes, and calculate the bidirectional attention weight between the abnormal battery node and its neighboring nodes based on the spatiotemporal correlation graph, and locate the fault source node according to the bidirectional attention weight distribution. The fault source localization module is used to extract multi-dimensional abnormal feature vectors from fault source nodes and calculate the marginal contribution of each feature dimension to the fault probability through counterfactual inference. The root cause identification and path prediction module is used to identify the dominant root cause of the failure based on the marginal contribution, and input the dominant root cause of the failure into a dynamic Bayesian network, combined with the spatiotemporal correlation graph, to predict the failure evolution and propagation path and time window. The early warning output module is used to output fault early warning results based on the dominant root cause of the fault, the fault evolution and propagation path, and the time window.

Images