Power battery abnormity monitoring method and system suitable for new energy automobile

Abstract

The invention relates to the technical field of new energy vehicles, and particularly discloses a power battery abnormity monitoring method and system suitable for a new energy vehicle. The power battery abnormity monitoring method suitable for the new energy automobile comprises the following steps: constructing a structural causal model of a power battery system, and fusing expert domain knowledge and historical abnormal cases into a causal knowledge graph; a causal structure learning algorithm based on continuous optimization is adopted, and a causal relationship is automatically found from battery time sequence data; performing multi-level causal reasoning by using an anti-factual reasoning technology, and distinguishing a direct causal relationship and an indirect causal relationship; generating a causal relationship explanation, and mapping the discovered causal relationship to a battery physical mechanism and a fault mode; through an interactive causal exploration interface, an expert is supported to correct and perfect a causal graph. According to the invention, by fusing the data driving capability of the neural network and the logicality of symbol reasoning, the abnormal root cause of the battery is accurately identified, professional explanation is provided, and continuous optimization of causal knowledge of man-machine collaboration is realized.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle technology, particularly to the field of power battery health management and fault diagnosis technology, and more specifically, to a method and system for monitoring abnormalities in power batteries for new energy vehicles. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the safety and reliability of power batteries, as core components, have become a focus of industry attention. Existing power battery monitoring systems mainly suffer from the following problems: First, anomaly detection and root cause diagnosis are separated. Most existing systems can only detect whether the battery status is abnormal, but cannot effectively identify and locate the root cause of the anomaly, resulting in low maintenance efficiency and frequent recurrence of faults. Second, they use fixed causal models, which are difficult to adapt to the development of new battery materials and technologies, as well as the complexity of batteries under different operating conditions. Third, data-driven monitoring systems typically use black-box models such as deep learning, and the detection results lack professional domain interpretation, making it difficult to effectively combine with the professional knowledge of battery engineers. Fourth, the causal chain is incomplete, failing to fully explain complex faults caused by the combined effects of multiple factors. Fifth, there is a lack of effective knowledge update mechanisms, making it impossible to automatically optimize the causal knowledge base based on new usage data and fault cases.

[0003] Therefore, there is an urgent need for a power battery anomaly monitoring method that can automatically discover causal relationships, accurately diagnose the root causes of anomalies, provide professional explanations, and support knowledge updates, in order to improve the safety, reliability, and service life of power batteries for new energy vehicles. Summary of the Invention

[0004] This invention provides a method and system for monitoring abnormalities in power batteries of new energy vehicles, solving the technical problems of existing power battery monitoring systems that are difficult to identify the root causes of abnormalities, lack interpretability and adaptability.

[0005] The first aspect of this invention discloses a method for abnormal monitoring of power batteries in new energy vehicles, comprising the following steps:

[0006] Construct a structural causal model of the power battery system, and integrate expert domain knowledge and historical anomaly cases into a causal knowledge graph;

[0007] Based on causal knowledge graphs, a causal structure learning algorithm based on continuous optimization is adopted to represent causal graph discovery as a differential constraint optimization problem, and causal relationships are automatically discovered from battery time series data.

[0008] Based on automatically discovered causal relationships, counterfactual reasoning techniques are used to perform multi-level causal reasoning, distinguish between direct and indirect causal relationships, and construct a multi-level causal abstraction mechanism.

[0009] Based on a multi-level causal abstraction mechanism, causal relationship explanations are generated, and the discovered causal relationships are automatically mapped to battery physical mechanisms and failure modes.

[0010] Based on the causal relationships mapped to battery physical mechanisms and failure modes, an interactive causal exploration interface allows domain experts to revise and improve the causal graph.

[0011] Furthermore, the step of fusing expert domain knowledge and historical anomaly cases into a causal knowledge graph includes:

[0012] Encode battery expert knowledge into differentiable probabilistic logic expressions to establish an initial correlation between battery parameter changes and failure modes;

[0013] Time-series data and diagnostic results were extracted from historical battery failure cases to construct a training dataset;

[0014] Based on domain knowledge and historical cases, an initial causal knowledge graph is constructed.

[0015] The causal knowledge graph is converted into a differentiable parameterized representation, which supports subsequent optimization.

[0016] Furthermore, the causal structure learning algorithm further includes:

[0017] Construct a scoring function and a complexity regularization term to form the optimization objective function;

[0018] To ensure that the learned causal graph satisfies the property of an acyclic directed graph, matrix constraints are introduced;

[0019] Transform the constrained optimization problem into an augmented Lagrangian form;

[0020] The gradient is calculated by backpropagation automatic differentiation, and the weight matrix is ​​iteratively optimized using gradient descent.

[0021] Furthermore, the steps of the multi-level causal reasoning include:

[0022] Construct counterfactual reasoning models to verify the causal relationships of the assumptions;

[0023] By comparing the effects of direct intervention and the cumulative effects of indirect pathways, we can distinguish between direct and indirect causal relationships.

[0024] The causal relationships are organized according to the level of abstraction to form a multi-level causal structure including the battery physical parameter layer, the battery state layer, and the fault mode layer.

[0025] Implement a hierarchical causal reasoning algorithm that supports reasoning propagation between different levels of abstraction;

[0026] The confidence level of the inference results is assessed to provide a reliability index for the inference results.

[0027] Furthermore, the step of generating a causal explanation includes:

[0028] Extract the critical path that has the greatest impact on the target failure from the complete cause-effect graph;

[0029] Map the identified causal paths to the battery's physical mechanisms;

[0030] Generate natural language explanations based on templates and specific parameter values;

[0031] Integrate the confidence level of the reasoning process into the explanation;

[0032] Generate multi-level explanations for users with different levels of expertise.

[0033] Furthermore, the implementation of the interactive causal exploration interface includes:

[0034] The learned causal structure is presented in an interactive visualization.

[0035] Establish an expert feedback collection mechanism to allow domain experts to evaluate and modify the causal structure;

[0036] Develop feedback integration algorithms to transform expert feedback into constraints or priors for causal learning;

[0037] Implement an incremental learning mechanism to integrate new data and feedback on the basis of existing causal structures;

[0038] A knowledge persistence mechanism is constructed to store the optimized causal structure and domain knowledge in a distributed database.

[0039] Furthermore, it also includes the implementation of a neural symbolic hybrid model, which includes:

[0040] A neural network component consisting of a graph neural network is used to learn the relationship representation between nodes;

[0041] Symbolic logic reasoning components implemented through differentiable logic rules;

[0042] Neural symbol fusion is achieved by using neural networks to assist symbolic reasoning and symbolic rules to guide neural learning.

[0043] Furthermore, the symbolic logic reasoning component of the neural symbolic hybrid model converts expert rules into first-order logic forms and performs continuous processing of logical operations through t-norm or Łukasiewicz logic.

[0044] Furthermore, the battery parameters of the power battery system include at least one of voltage, current, temperature, and internal resistance; the fault types include at least one of capacity decay, internal short circuit, and lithium plating.

[0045] The second aspect of this invention discloses a power battery anomaly monitoring system suitable for new energy vehicles, used to execute the aforementioned power battery anomaly monitoring method suitable for new energy vehicles, the system comprising:

[0046] The causal knowledge representation building module is used to integrate expert domain knowledge and historical anomaly cases into a causal knowledge graph;

[0047] Differentiable causal structure learning module, used to automatically discover causal relationships from battery time-series data;

[0048] A multi-level causal reasoning module is used to distinguish between direct and indirect causal relationships and to construct a multi-level causal abstraction mechanism;

[0049] The causal relationship interpretation and generation module is used to map the discovered causal relationships to battery physical mechanisms and failure modes;

[0050] An interactive causal knowledge optimization module is used to support domain experts in revising and improving causal graphs;

[0051] Through the coordinated operation of the above modules, the system achieves accurate monitoring, root cause diagnosis, and knowledge optimization of power battery anomalies.

[0052] The beneficial effects of this invention are as follows: Through a multi-level causal reasoning mechanism and counterfactual analysis method, it accurately distinguishes between fault phenomena and root causes, enabling maintenance personnel to quickly locate the source of problems and significantly reduce recurring failure rates and maintenance costs. Based on a differentiable causal structure learning algorithm, it automatically discovers new causal relationships and fault modes from continuously accumulated battery monitoring data, eliminating the need for manual pre-setting of all causal models. It can identify novel fault modes not included in traditional expert systems and automatically generate corresponding diagnostic rules, adapting to different types of batteries and changing usage environments. Through a causal relationship explanation generator, it transforms complex data analysis results into natural language explanations consistent with battery field expertise, bridging the gap between data analysis and professional knowledge. This is significantly superior to traditional machine learning methods, enabling maintenance personnel to understand the principles of fault formation and guiding precise maintenance. By constructing a complete causal chain from underlying battery parameters to high-level fault phenomena, it achieves multi-level anomaly reasoning, providing comprehensive knowledge support for fault prevention. Through an interactive causal exploration interface, it integrates expert feedback into the causal structure optimization process, achieving continuous knowledge updates and improvements. Through an incremental learning mechanism, it maintains high accuracy while adapting to characteristic changes during battery aging, significantly reducing model maintenance costs and extending the effective service life of the system. Attached Figure Description

[0053] Figure 1 This is a flowchart of a power battery anomaly monitoring method applicable to new energy vehicles according to the present invention. Detailed Implementation

[0054] Before providing a detailed description of this application, the terms used in this application will be explained below to aid in understanding the technical solutions:

[0055] Power battery: refers to the energy storage device that provides power to new energy vehicles, usually a lithium-ion battery pack or a lithium iron phosphate battery, etc.

[0056] Causal knowledge representation: refers to the data structure used to express causal relationships between things. It is usually represented by a graph structure, where nodes represent events or states and edges represent causal relationships.

[0057] Differentiable causal structure: refers to a causal relationship representation structure that can be learned and updated through gradient optimization methods.

[0058] Neural-symbolic hybrid model: refers to a hybrid computing model that combines the data-driven capabilities of neural networks with the logical expression capabilities of symbolic reasoning.

[0059] Counterfactual reasoning: refers to the method of reasoning by simulating hypothetical situations of "what if...", and is used to explore causal relationships between events.

[0060] Causal abstraction: refers to the process of abstracting underlying specific causal relationships into higher-level conceptual relationships, which helps to simplify the understanding and analysis of complex systems.

[0061] Differentiable probabilistic logic refers to a probabilistic logic representation that can be optimized using gradient methods, supporting the application of continuous optimization techniques on logical rules.

[0062] It should be understood that the method of this embodiment is applicable to various new energy vehicle power battery systems, including but not limited to lithium-ion batteries, lithium iron phosphate batteries, and solid-state batteries. Furthermore, when implemented in an in-vehicle system, this method can run a lightweight version on the vehicle's built-in computing unit, or it can transmit data to a cloud server via vehicle-to-everything (V2X) communication for more detailed analysis.

[0063] At least one embodiment of the present invention provides a method for abnormal monitoring of power batteries suitable for new energy vehicles, such as... Figure 1 As shown, it includes the following steps:

[0064] Step 1: Construct a structural causal model of the power battery system, and integrate expert domain knowledge and historical anomaly cases into a causal knowledge graph.

[0065] This step specifically includes:

[0066] Step 1.1: Domain Knowledge Encoding: Encode battery expert knowledge into differentiable probabilistic logical expressions to establish an initial association between battery parameter changes and failure modes. The specific implementation method is as follows:

[0067] First, we collect and organize knowledge rules provided by battery experts, including the correspondence between abnormal battery parameters and fault types, such as "a continuous increase in battery internal resistance may lead to a decrease in battery capacity" and "excessively high charging temperature may lead to lithium plating in the battery." It should be noted that these rules are formally represented using differentiable probabilistic logic expressions.

[0068] ;

[0069] in, Represents a random variable representing battery parameters. This indicates the specific values ​​of the observed battery parameters. Indicates the possible fault types, This indicates the value of an intermediate logical variable (such as a change in the internal state of a battery). Represents a given intermediate state The following fault types The conditional probability, Indicates the value of the given parameter The value of the lower intermediate state is The conditional probability. Indicates the value of the given parameter The following fault types The conditional probabilities are the target probabilities that ultimately need to be inferred. These conditional probabilities can all be learned from data. Furthermore, expert rules are transformed into prior constraints on these conditional probabilities.

[0070] In some implementations, different formal representation methods can be selected; for example, fuzzy logic representation can be optionally employed, using membership functions. Indicates parameter value For fuzzy sets The degree of attribution is determined to address the uncertainty in the relationship between parameters and faults. In this case, the rule can be expressed as:

[0071] ;

[0072] in, Indicates the first One rule, This indicates battery parameter values ​​(such as temperature, voltage, etc.). Indicates the fault type or status. and These are fuzzy sets of input parameters and output fault types, respectively. express For sets The membership degree (within the range [0,1]). When the rule When triggered, its trigger strength is determined by The degree of certainty of the output result is determined by... This type of representation is particularly suitable for handling scenarios where battery parameters have fuzzy boundaries, such as when the boundary between "high temperature" and "ultra-high temperature" of a battery is unclear.

[0073] Step 1.2: Historical Case Integration: Extract time-series data and diagnostic results from historical battery failure cases to construct a training dataset. For each failure case, record the observed battery parameter sequence. and corresponding fault types ,in , , These represent different time points 1, 2, and 3. Collected battery parameters, This indicates the total length of the time series.

[0074] Optionally, when the amount of data is insufficient, this application may also employ data augmentation techniques to expand the training samples. For example, more diverse training data can be generated by adding random noise that conforms to the physical characteristics of the battery, applying time window sliding, and adjusting parameter scales. In some embodiments, simulated data based on physical models may also be introduced, especially for fault types that are highly dangerous but have few actual samples (such as thermal runaway), by simulating parameter sequences of typical fault evolution processes using electrochemical models.

[0075] Step 1.3: Causal Knowledge Graph Generation: Based on domain knowledge and historical cases, construct an initial causal knowledge graph:

[0076] ;

[0077] in, A causal knowledge graph representing the initial state. This represents a set of nodes (including battery parameters and fault types). Represents a set of edges (representing causal relationships). This represents the set of edge weights (representing causal strength). The initial graph is constructed as follows:

[0078] Extract nodes and edges directly from expert rules;

[0079] Using Bayesian network architecture to learn algorithms to learn potential causal relationships from historical data;

[0080] By integrating the relationships between the two sources mentioned above, a comprehensive causal knowledge graph is constructed.

[0081] In some embodiments, this application can also construct a multi-level causal graph based on the battery's physical structure, linking the battery's internal electrochemical reactions, thermodynamic characteristics, and external observable parameters. For example, for lithium-ion batteries, a hierarchy including physical components such as electrode materials, electrolytes, separators, and casings can be established to form a more complete causal relationship network. Optionally, dedicated sub-graphs for various battery types can also be integrated, such as specialized causal knowledge sub-graphs for lithium iron phosphate batteries and ternary lithium batteries, automatically selecting and fusing the corresponding knowledge structures according to the specific battery type during operation.

[0082] Step 1.4: Differentiable Representation Transformation: Transform the causal knowledge graph into a differentiable parameterized representation to support subsequent optimization. Specifically, each causal edge... Represented as a parameterized function ,in For learnable parameters, Indicates the starting node (dependent variable). This represents the terminating node (the result variable). The entire graph structure is represented as follows:

[0083] ;

[0084] in It is the set of parameters for all edges. Denotes the set of edges. This represents the parameterized causal knowledge graph. Indicates the starting node (dependent variable). Indicates the terminating node (fruit variable). Indicates from node To the node Parameterized causal functions, This represents the learnable parameters of the corresponding edges. As can be seen, this representation allows the causal structure to be optimized using gradient descent.

[0085] Alternatively, in some implementations, different parameterization forms may be used, such as representing causal edges as neural networks:

[0086] ;

[0087] in, Indicates from node To the node Parameterized causal functions This represents the learnable parameters of the corresponding edge. It has parameters Small neural networks can capture more complex nonlinear causal relationships. This representation is particularly suitable for modeling the nonlinear behavior of batteries under extreme conditions, such as the nonlinear relationship between battery internal resistance and temperature in low-temperature environments. Here, Indicates the starting node (dependent variable). Representing the terminating node (effect variable), in a neural network. Learning from multi-layer nonlinear transformation arrive The mapping relationship, its parameters The network's weights and biases are optimized using the backpropagation algorithm. Compared to simple linear parameterization, neural network representations can automatically discover and model complex nonlinear causal mechanisms, especially when there are complex relationships between battery parameters such as threshold effects, saturation effects, or synergistic effects, which gives them a significant advantage.

[0088] Step 2: Based on the causal knowledge graph, a causal structure learning algorithm based on continuous optimization is adopted to represent the causal graph discovery as a differential constraint optimization problem, and automatically discover causal relationships from battery time series data.

[0089] This step specifically includes:

[0090] Step 2.1: Optimize the objective function construction: The causal structure learning problem is formulated as the following optimization objective:

[0091] ;

[0092] in, It is a cause-effect graph structure. For observation data, A scoring function to measure the degree of matching between graph structure and data. Graph structure Static regularization terms for complexity For balancing parameters, Representing the structure of a cause-effect graph Minimize operation.

[0093] Specifically, the scoring function used in this application is the marginal likelihood:

[0094] ;

[0095] in, In the diagram Medium variables The set of parent nodes, Indicates the first Variables in a sample The value, Indicates the total number of nodes. For the sample size, Representing a given graph structure The following observation data The likelihood probability, Represents a given variable Under the condition of parent node value The conditional probability, A scoring function to measure the degree of matching between graph structure and data. This represents the natural logarithm function.

[0096] Furthermore, the complexity regularization term is defined as:

[0097] ;

[0098] in, Indicates from node To the node edge weights, Indicates the total number of nodes. Represents the absolute value of the edge weight. Representing graph structure The complexity metric is calculated by summing the absolute values ​​of the weights of all possible edges.

[0099] Step 2.2: Continuous Structure Representation: To apply the gradient optimization method, this application transforms the discrete graph structure into a continuous representation. Specifically, each potential causal edge... From the weight matrix elements It means that, among them The size of the edge indicates its strength, and the sign indicates the causal direction. Indicates the total number of nodes. express 3D real matrix space:

[0100] Indicates from node To the node There is a positive causal relationship;

[0101] Indicates from node To the node There is a negative causal relationship;

[0102] Represents a node and nodes There is no direct causal relationship between them;

[0103] Step 2.3: Constraint Design: To ensure that the learned causal graph satisfies the properties of an acyclic directed graph (DAG), this application introduces the following matrix constraints:

[0104] ;

[0105] in, Representation matrix The square of the elements (Hadamard product, i.e., element-wise product). This indicates matrix exponential operations (i.e., performing exponential function transformations on matrices). Represents the trace of a matrix (i.e., the sum of the elements on the main diagonal). The total number of nodes. It is a constraint function. If and only if This constraint equation holds when representing a DAG. Based on the property of matrix exponents, this constraint ensures that there are no cycles in the graph structure, thus guaranteeing the directed acyclic nature of causal relationships.

[0106] Step 2.4: Gradient Optimization Solution: This application transforms the constrained optimization problem into an augmented Lagrangian form:

[0107] ;

[0108] in, It is a Lagrange multiplier. It is the penalty coefficient. Represented by the weight matrix A defined graph structure For observation data, This indicates the degree of matching between the graph structure and the data. Graph structure determined by weight matrix W The dynamic regularization term, To ensure the graph structure is acyclic, the constraint function, To augment the Lagrange function, To balance the parameters, the gradient is calculated using backpropagation automatic differentiation, and then iterative optimization is performed using gradient descent. :

[0109] ;

[0110] in, For learning rate, Indicates the number of iterations. Indicates the first The weight matrix at the next iteration Indicates the first The weight matrix at the next iteration Indicates the first The penalty coefficient at the next iteration Indicates the first Lagrange multipliers in the next iteration Indicates about The gradient.

[0111] Step 2.5: Temporal Causal Discovery: Based on the characteristics of battery time-series data, this application extends the above algorithm to capture temporal dependencies.

[0112] Specifically, for time series variables and To examine causal relationships with different time lags, an extended variable set is constructed using the sliding time window method. Then, the above learning algorithm is applied to discover the temporal causal structure, where, , , Representing variables respectively At the present moment The moment before , The value before a time step, , , Representing variables respectively At the present moment The moment before , The value before the specified time step.

[0113] In some implementations, an adaptive time window size selection strategy can be employed to automatically adjust the optimal window size based on the frequency of change of different battery parameters. For example, a larger window can be used for parameters that change slowly (such as capacity decay), while a smaller window can be used for parameters that change rapidly (such as instantaneous current). Optionally, multi-scale time analysis can also be introduced to discover causal relationships at different time scales (minutes, hours, days), and then the multi-scale results can be integrated through a specific fusion algorithm to capture causal patterns at different time scales.

[0114] Step 2.6: Specific Implementation of the Causal Structure Learning Algorithm: According to the embodiments of this application, the specific implementation of the differentiable causal structure learning algorithm includes the following aspects:

[0115] First, for each battery parameter time-series dataset, data preprocessing is performed, including missing value imputation, noise filtering, and normalization. Moving median filtering is applied to outliers for each variable. Calculate the median within its sliding window (width 5). and standard deviation ,when At that time, Replace with ,in, Indicates the first Measured values ​​of individual battery parameters This indicates the median value of the parameter within the sliding window. This represents the standard deviation of the parameter within the sliding window. This represents the absolute difference between the measured value and the median. This represents a threshold of three standard deviations, used for outlier detection.

[0116] Optionally, when processing battery data, anomaly detection methods aided by domain-specific knowledge can be employed. For example, outliers can be identified by combining the battery's physical characteristics (such as coulombic efficiency and internal resistance variation patterns) rather than relying solely on statistical indicators. In some implementations, seasonal decomposition preprocessing can also be used to decompose the battery parameter time-series data into trend, seasonal, and random fluctuation components, model causal relationships separately, and then integrate them to more accurately capture the periodic causal patterns in the battery's cyclic charging and discharging process.

[0117] Next, initialize the weight matrix. The matrix is ​​zero, and then the constrained optimization problem is solved using the Alternating Direction Multiplier Method (ADMM). Specifically, the following steps are performed in each iteration:

[0118] fixed Update Lagrange multipliers :

[0119] ;

[0120] in, Indicates the first Lagrange multipliers in the next iteration This represents the updated Lagrange multipliers. Indicates the first The penalty coefficient at the next iteration Indicates the first The constraint function value at the next iteration. This represents the updated Lagrange multipliers;

[0121] fixed ,renew Using the Adam optimizer:

[0122] Calculate the gradient: ;

[0123] Calculate the first moment estimate: ;

[0124] Calculate the second moment estimate: ;

[0125] Bias correction: ;

[0126] ;

[0127] Update parameters: ;

[0128] in, Indicates the first The gradient at the next iteration Indicates about gradient, Indicates the first The penalty coefficient at the next iteration This represents the updated Lagrange multipliers. Indicates the first The weight matrix at the next iteration and Let these represent the first-order moment estimates of the current and previous iterations, respectively. and Let these represent the second-order moment estimates of the current and previous iterations, respectively. and These are the first and second momentum decay rates. and This represents the first and second moment estimates after bias correction. Indicates the number of iterations. and These represent the weight matrices before and after the update, respectively. Indicates the learning rate. It is a small constant added to prevent division by zero errors. The element-wise square of the gradient is represented. and They represent and of Power of;

[0129] in, ; ; ;

[0130] For the learning rate, from Initially, a cosine annealing strategy was used to gradually reduce the temperature.

[0131] Update penalty coefficient :

[0132] ;

[0133] in, This represents the updated penalty coefficient. This represents the penalty coefficient for the current iteration. For growth rate, This is the upper limit of the penalty coefficient. This indicates that the smaller of the two parameters is taken.

[0134] In some implementations, an adaptive penalty parameter update strategy can be used to dynamically adjust the growth rate based on the degree of constraint violation. For example, the growth rate can be reduced when constraints are close to being met to avoid numerical instability; and the growth rate can be increased when constraints are far from being met to accelerate convergence. Optionally, preprocessing techniques can be applied to accelerate the optimization process, such as identifying a subset of variables that may have causal relationships through domain knowledge or preliminary statistical analysis, thereby reducing the search space.

[0135] The iteration terminates when any of the following conditions are met:

[0136] Degree of constraint satisfaction: ,in ;

[0137] The relative change is small: ,in The relative change threshold, The Frobenius norm represents the difference between the weight matrices in two consecutive iterations. This represents the Frobenius norm of the current weight matrix;

[0138] Maximum number of iterations reached: ,in The maximum allowed number of iterations, This indicates the current iteration number.

[0139] Finally, to ensure that the learned causal graph is an acyclic directed graph (DAG), the final weight matrix is... Post-processing: Sort the edges according to their absolute weights and add them in descending order. If adding an edge would cause a cycle in the graph, skip that edge.

[0140] Optionally, in the post-processing stage, the learned causal graph can be corrected by incorporating battery physics knowledge. For example, according to battery physics principles, certain parameters may or may not have a causal relationship (e.g., current always affects temperature, but temperature does not directly affect voltage). This prior knowledge can be applied as hard constraints in the post-processing. In some implementations, ensemble learning methods can also be used to learn multiple causal graph structures by using different initializations or different random subsamples multiple times, and then the results can be fused through voting or weighted averaging to improve the stability and reliability of the causal structure.

[0141] When applying this algorithm to power battery parameters in the embodiments of this application, it specifically targets 20 common parameters of the battery, such as voltage, current, temperature, and internal resistance, using a 15-minute sampling interval to construct a sliding window of up to 4 hours (i.e., This algorithm captures short- to medium-term causal relationships. In practical applications, it can identify typical causal relationships such as "sustained low temperature environment leads to increased battery internal resistance" and "rapid charging and discharging leads to local overheating".

[0142] Step 3: Based on the automatically discovered causal relationships, use counterfactual reasoning techniques to perform multi-level causal reasoning, distinguish between direct and indirect causal relationships, and construct a multi-level causal abstraction mechanism;

[0143] According to an embodiment of this application, step 3 utilizes counterfactual reasoning techniques to perform a "hypothesis-verification" analysis, distinguishing between direct and indirect causal relationships, constructing a multi-level causal abstraction mechanism, and realizing a complete reasoning path from changes in underlying battery parameters to top-level fault phenomena. This step specifically includes:

[0144] Step 3.1: Counterfactual Reasoning Model Construction: This application constructs a counterfactual reasoning model to verify the causal relationship of the hypothesis. Specifically, for variables... and The potential causal relationship between them can be assessed by evaluating the intervention effect in the following ways:

[0145] ;

[0146] in, Indicates the variable Intervene (set its value to) ) after variable The probability distribution, Represents the variable Intervene and set its value to , express and The set of common cause variables. Here Represents a random variable representing battery parameters. This indicates the specific values ​​of the observed battery parameters; Indicates the fault type variable; Represents the set of intermediate state variables. This indicates the specific values ​​that the variables in the set can take; This indicates that the battery parameter values ​​are... And the intermediate state is Fault types under conditions The conditional probability; This indicates that the intermediate state takes the value of The probability. It should be noted that this intervention calculation method achieves control over confounding factors.

[0147] For battery time-series data, this application extends the counterfactual model to handle the time dimension:

[0148] ;

[0149] in, Indicating in the Battery parameters at any time Perform human intervention (set to a specific value) )back, Time-based fault types The probability distribution, Indicates time delay. express Battery parameter variables at time t, This indicates the specific values ​​of the observed battery parameters. express The fault type variable at any given time. express The set of intermediate state variables at time 1. This indicates the specific values ​​that the variables in the set can take. Indicates in The battery parameter values ​​at any given time are: And the intermediate state is under conditions Time-based fault types The conditional probability, express The intermediate state value at time step is The probability, Indicates to Variables at time Intervene and set its value to This formula evaluates battery parameters in The value at time corresponds to the fault type. The impact of time.

[0150] Step 3.2: Distinguishing between direct and indirect causality: By comparing the effects of direct intervention and the cumulative effects of indirect pathways, this application distinguishes between direct and indirect causal relationships. For variables... , and potential mediator variables Calculate the following quantities:

[0151] Direct effects: ;

[0152] Indirect effects:

[0153] ;

[0154] Total effect: ;

[0155] in, express The baseline value, express The observed values, This represents the direct effect value. This represents the indirect effect value. This represents the total effect value. Represents the variable Set the intervention value hour The conditional probability, Represents the variable Intervention was set as the baseline value. hour The conditional probability, Indicates simultaneous application of variables and Interventions were performed and set to values ​​respectively. and hour The conditional probability, Indicates simultaneous application of variables and Interventions were conducted and each was set as a baseline value. and observed values hour The conditional probability. It should be noted that if... Then it is believed right The influence is mainly through mediating variables Implementation, in which This indicates that the absolute value of the direct effect is much smaller than the absolute value of the indirect effect.

[0156] Step 3.3: Multi-level Causal Abstraction Construction: This application organizes causal relationships according to abstraction levels to form a multi-level causal structure, specifically including:

[0157] Bottom layer: Battery physical parameter layer (original observed parameters such as voltage, current, temperature, and internal resistance);

[0158] Middle layer: Battery state layer (state of charge, state of health, internal electrochemical reaction state, etc.);

[0159] High layer: Fault mode layer (fault types such as capacity decay, internal short circuit, lithium plating, etc.).

[0160] Therefore, this multi-level structure connects changes in lower-level parameters with higher-level fault phenomena through causal chains, forming a complete reasoning path.

[0161] Step 3.4: Implementation of Hierarchical Inference Algorithm: This application implements a hierarchical causal inference algorithm, supporting inference propagation between different levels of abstraction. Specifically, given the underlying observation parameters... The high-level failure mode is inferred through the following recursive process. :

[0162] ;

[0163] in, Represents a random variable representing battery parameters. Indicates the fault type variable. This represents the set of intermediate state variables. This indicates that the value of the middle-level state variable is... High-level failure modes under certain conditions The conditional probability, This indicates that the value of the observed parameter at the bottom layer is... Mid-level state variables under certain conditions The conditional probability, This indicates that the value of the observed parameter at the bottom layer is... High-level failure modes under certain conditions The conditional probability. Furthermore, this decomposition allows the reasoning process to proceed in layers, utilizing the expertise of each layer.

[0164] Step 3.5: Inference Confidence Assessment: According to the embodiments of this application, the confidence of the inference result is assessed to provide a reliability index for the inference result. This applies to the inferred causal relationship. The confidence level is calculated as follows:

[0165] ;

[0166] in, Represents a random variable representing battery parameters. The variable represents the fault type, indicating the origin. arrive causal relationship This indicates the confidence level of the causal relationship. Indicating intervention back conditional entropy, express entropy, express The maximum possible entropy value, Represents the variable Intervention measures should be taken. Furthermore, a higher confidence level indicates a more certain causal relationship.

[0167] Step 3.6: Implementation Details of the Neural Symbolic Hybrid Model: According to embodiments of this application, the proposed neural symbolic hybrid model consists of the following components, realizing the fusion of a data-driven neural network part and a logic rule-driven symbolic reasoning part:

[0168] First, the neural network component employs a graph neural network (GNN) structure to learn the relationships between nodes (battery parameters and states). Specifically, this GNN consists of the following layers:

[0169] Node feature embedding layer: embeds the original features of each node. Mapping to hidden layer representation :

[0170] ;

[0171] in, and For learnable parameters, This is the ReLU activation function.

[0172] Graph convolutional layers: Aggregate information about nodes and their neighbors through multiple graph convolutional operations.

[0173] ;

[0174] in, Represents a node In the Hidden state representation of layers, Represents the ReLU activation function. For nodes The neighborhood group, This represents the normalization coefficient between node i and node j. and They represent the first The weight matrix of the layer and the weight matrix of the first layer are as follows: Layer bias vector, Represents a node In the Hidden state representation of layers, Represents a node The neighboring nodes. This application uses 3 layers of graph convolution with a hidden layer dimension of 64.

[0175] Output layer: Transforms the final node representation into the desired output.

[0176] ;

[0177] in, Represents a node The output results, For a specific task, the output function (such as the probability distribution for fault classification). Represents a node On the last floor The hidden state representation.

[0178] In some implementations, different types of graph neural network architectures may be optionally employed. For example, for capturing long temporal dependencies, a temporal graph neural network (TGNN) can be used, adding temporal dimension processing capabilities on top of graph convolutions.

[0179] ;

[0180] in, Indicates time node In the The hidden state of the layer, This represents the activation function. Indicates the first The weight matrix of the layer, Represents a node The neighborhood group, Represents a node and nodes The normalization coefficients between them Indicates time node In the The hidden state of the layer, The parameter matrix is ​​used to process time-series information. Indicates time node In the The hidden state of the layer, Indicates the first The layer's bias vector. This variant is particularly suitable for analyzing causal relationships in the time-series evolution of battery parameters.

[0181] Secondly, the symbolic logic reasoning component is implemented through differentiable logic rules, with each rule represented as a differentiable logical formula. and assign weights The specific implementation involves the following steps:

[0182] Transforming expert rules into first-order logic form, such as "If the battery temperature is too high and the discharge rate is too high, it may lead to thermal runaway," can be represented as:

[0183] ;

[0184] in, Indicates battery, This represents the first logical rule. A predicate indicating that the battery temperature is too high. A predicate indicating that the battery discharge rate is too high. A predicate indicating that a battery has a risk of thermal runaway. This represents the logical "AND" operator. This represents the logical "implication" operator.

[0185] Using t-norm to continue logic operations:

[0186] Connect (AND): ;

[0187] Disjunction (OR): ;

[0188] Negation (NOT): ;

[0189] Implied (IMPLIES): ;

[0190] in, and Denotes the truth degree of two propositions, with a range of values ​​of 1 to 2. , This represents the result of an AND operation on two propositional logic statements. This represents the result of an OR operation on two propositional logic statements. Expressing a proposition The result of the logical NOT operation, Indicates from the proposition To the proposition The logic "implies" the result of the operation.

[0191] The satisfaction level of a rule is treated as a continuous value, and the "truthfulness" value of the rule is calculated through logical operations.

[0192] Finally, neural symbol fusion is achieved through the following two mechanisms:

[0193] Neural network-assisted symbolic reasoning: The output of a neural network provides truth estimates for basic propositions in symbolic rules, such as... The probability of;

[0194] Symbolic rule-guided neural learning: By incorporating rule satisfaction as an additional loss function component, the neural network is guided to learn representations that conform to knowledge of the battery domain.

[0195] ;

[0196] in, For the rule-based loss function, This indicates all rules Summation, Representation rules The weight, Representation rules In the input Satisfaction on Input data representing battery parameters;

[0197] The total loss function is:

[0198] ;

[0199] in, This represents the total loss function of the model. For supervised loss based on labeled data, These are the balancing parameters.

[0200] The aforementioned neural symbolic hybrid model integrates data-driven learning with expert knowledge. Under abnormal battery conditions, it can simultaneously utilize historical data patterns and battery physical laws for reasoning, thereby improving the generalization ability to unseen fault patterns.

[0201] In some implementations, this application may also employ an attention mechanism to enhance the interaction between the neural network and the symbolic rules, dynamically weighting the applicability of different rules:

[0202] ;

[0203] in, Representation rules Dynamic weights, The current state query vector, For rules The key vector, For vector dimensions, Represents the query vector transpose, The normalized exponential function ensures that the sum of all rule weights is 1. This attention-weighted approach enables the system to adaptively select the most relevant rules based on specific battery conditions and parameter patterns, improving the relevance and accuracy of the inference process. Optionally, a hierarchical adaptive rule selection mechanism can be introduced, first applying high-level rules to determine the approximate fault type, and then selecting applicable fine-grained rules based on the preliminary results, forming a decision tree-like inference process that improves computational efficiency while maintaining inference accuracy.

[0204] Step 4: Based on a multi-level causal abstraction mechanism, generate causal relationship explanations and automatically map the discovered causal relationships to battery physical mechanisms and failure modes. This step specifically includes:

[0205] Step 4.1: Critical Cause-and-Effect Path Extraction: Extract the critical path that has the greatest impact on the target fault from the complete cause-and-effect graph. Given the fault type... and observation parameter set Key causal pathways can be identified using the following methods:

[0206] Calculate each parameter For the fault Total effect ;

[0207] Parameters are ranked according to their total effect size, and the parameters with the largest effects are selected. One parameter;

[0208] Extract these parameters to the fault The complete causal path, including intermediate nodes.

[0209] Formally, the set of critical paths is defined as:

[0210] ;

[0211] in, Indication and Fault The relevant set of key causal paths Indicates from parameter To the fault causal path, Indicates the fault The most affected front One parameter, This represents a parameter in the battery system. Indicates a specific type of fault. Indicates the number of top parameters selected. The symbol represents a set construction symbol, indicating a set consisting of all elements that satisfy a given condition. It indicates the direction of a causal relationship.

[0212] Step 4.2: Physical Mechanism Mapping: This application maps the identified causal paths to battery physical mechanisms. A template library is constructed. ,in, , , They represent the 1st, 2nd, and 3rd respectively. One template, This represents the total number of templates in the template library, with each template associated with a specific causal pattern and its corresponding physical interpretation. For observed causal paths... Find the best matching template through similarity matching:

[0213] ;

[0214] in, This represents the best matching template. This represents the template that maximizes the similarity function. Measure path With template similarity, Indicates the observed causal path, This represents the t-th specific template in the template library. In this system, a template refers to a predefined causal relationship explanation structure, which includes a standard explanation framework, key terms, and logical structure for a specific fault type. It is used to transform complex causal relationships into standardized and easily understood natural language explanations. Furthermore, similarity calculation considers factors such as node type, edge pattern, and parameter change trends.

[0215] Step 4.3: Explanation Generation Algorithm: According to embodiments of this application, a natural language explanation is generated based on the template and specific parameter values. The explanation generation algorithm includes:

[0216] Template filling: Enter the specific parameter name, value, and change mode into the selected template;

[0217] Context integration: Supplement relevant contextual information based on the battery's historical state and current operating conditions;

[0218] Causal chain description: Organize statements in order from cause to effect to describe a complete causal chain.

[0219] In some implementations, a knowledge graph-based explanation generation method can be optionally employed to retrieve relevant background knowledge from a battery-related knowledge base, enriching the explanation content. For example, when explaining the cause of lithium plating failure in lithium batteries, relevant electrochemical principles can be automatically retrieved and introduced, making the explanation both intuitive and scientifically grounded. Furthermore, the level of detail and focus of the explanation can be optionally adjusted based on the user's historical interaction records to achieve personalized explanation generation. For example, for users who frequently focus on temperature management, the explanation can elaborate on the temperature-related causal mechanisms in more detail.

[0220] It should be understood that the structure of the explanation follows the following pattern:

[0221] "The [abnormal parameter] exhibits a [specific abnormal pattern], leading to a [change in intermediate state], ultimately causing a [fault type]. This may be due to a [physical mechanism explanation]. [Repair recommendations] are suggested."

[0222] Step 4.4: Confidence Expression: This application integrates the confidence level of the reasoning process into the explanation, enabling users to assess the reliability of the results. Different expression methods are used for different confidence level intervals:

[0223] High confidence level ( Use definitive expressions, such as "cause" or "lead to".

[0224] Medium confidence level ( Use more probable expressions, such as "likely to lead to" or "usually to cause";

[0225] Low confidence level ( ): Use expressions of possibility, such as "may be related to" or "there is a possibility of...".

[0226] Step 4.5: Multi-level interpretation generation: According to the embodiments of this application, multi-level interpretations are generated for users with different levels of expertise:

[0227] Technical explanation: geared towards engineers and technical experts, including detailed parameter relationships and physical mechanisms;

[0228] Maintenance-level explanation: geared towards maintenance personnel, focusing on the cause of the fault, its location, and maintenance recommendations;

[0229] User-level explanation: For vehicle users, it provides simplified problem descriptions and urgency level prompts.

[0230] Therefore, each level of explanation uses different levels of lexical complexity, depth of technical detail, and expression.

[0231] Step 4.6: Implementation details of the causal explanation generator: According to embodiments of this application, the causal explanation generator is implemented based on a sequence-to-sequence model with an attention mechanism, and its architecture includes:

[0232] Encoder: Processes causal paths and related parameter information.

[0233] Path encoding: Each path is represented as a sequence of nodes and edges. ;

[0234] Among them, , , , , These represent the starting node of the path, the first edge, the second node, the last edge in the path, and the ending node of the path, respectively. This indicates the total number of nodes in the path.

[0235] Feature extraction: For each node Extract its type, value, and trend of change; for each edge Extract its causal strength, confidence level, and other features;

[0236] Self-attention layer: Apply a multi-head self-attention mechanism to process path sequences and capture long-distance dependencies between nodes;

[0237] Encoding representation: The vector representation of the final output path ;

[0238] Interpretation template selector: Selects the most suitable interpretation template based on the encoded representation;

[0239] Template library: Builds a collection of templates for different fault types. ;

[0240] in, , , They represent the 1st, 2nd, and 3rd respectively. There are 1 template, among which This represents the total number of templates in the template library, with each template associated with a specific causal pattern and its corresponding physical interpretation.

[0241] Similarity calculation: Calculate path encoding With each template Cosine similarity of feature vectors;

[0242] Template selection: Choose the template with the highest similarity as the base;

[0243] Decoder: Generates natural language interpretation based on selected templates and path encoding;

[0244] Conditional generation: A GRU decoder using an attention mechanism, where conditions include path encoding. and template structure;

[0245] Technical terminology insertion: Select appropriate terminology based on the user's level of expertise;

[0246] Confidence adjustment: Adjust the degree of certainty of the statement based on the confidence level of the inference;

[0247] This application optimizes the explanation generator for battery anomaly monitoring scenarios, incorporating a vocabulary of over 200 battery-related terms and explanation templates for more than 50 typical battery faults. Compared to general explanation generators, this dedicated explanation generator improves the accuracy and professionalism of battery fault explanations and supports automatic generation of multi-granularity explanations, from expert-level to user-level.

[0248] Optionally, in some implementations, the explanation generator can integrate multimodal output capabilities, generating not only text explanations but also visual aids such as time-series charts of fault parameters, causal relationship diagrams, or schematic diagrams of fault locations. For example, for battery thermal runaway faults, the system not only provides text explanations but also generates heat maps of abnormal temperature points and temperature rise curves, enabling maintenance personnel to understand the problem more intuitively. Furthermore, interactive explanations can optionally be implemented, allowing users to ask in-depth questions based on the initial explanation. The system can then provide targeted, detailed explanations of relevant parts, creating a question-and-answer style explanation experience.

[0249] Step 5: Based on the causal relationships mapped to battery physical mechanisms and failure modes, domain experts can revise and improve the causal graph through an interactive causal exploration interface;

[0250] According to an embodiment of this application, step 5 involves constructing an interactive causal exploration interface, supporting domain experts in revising and improving the causal graph through feedback, integrating user feedback into the causal structure optimization process, and achieving human-machine collaborative causal knowledge optimization. This step specifically includes:

[0251] Step 5.1: Visual Representation of Causality: This application presents the learned causal structure in an interactive visual format to facilitate user understanding and operation. Specific implementation includes:

[0252] Multi-level cause-effect graph display: Displays cause-effect relationships in layers according to abstract levels, supporting expansion and collapse between layers;

[0253] Causal strength coding: using visual elements such as edge thickness and color to encode the strength and confidence of causal relationships;

[0254] Dynamic time-series display: Supports sliding time windows to show the dynamic changes of causal relationships over time.

[0255] It should be noted that the visual interface is implemented through web technology and supports access on different devices, including in-vehicle displays, service station computers, and mobile terminals.

[0256] Step 5.2: Expert Feedback Collection: According to embodiments of this application, an expert feedback collection mechanism is constructed to allow domain experts to evaluate and modify the causal structure. The types of feedback collected include:

[0257] Causal edge confirmation / negation: Experts can mark a causal edge as correct or incorrect;

[0258] Add causal edges: Experts can add causal edges that are not yet discovered by the system;

[0259] Causal edge modification: Experts can adjust the direction, strength, or conditions of causal relationships;

[0260] Interpretation and Evaluation: Experts can evaluate and modify the interpretations generated by the system.

[0261] Therefore, this feedback is collected through structured forms and interactive graphical editing interfaces to ensure data consistency and integrity.

[0262] Step 5.3: Feedback Integration Algorithm: This application develops a feedback integration algorithm that transforms expert feedback into constraints or priors for causal learning. Formally, for the set of feedback provided by experts... This is transformed into constraints for the optimization objective:

[0263] ;

[0264] in, , , These represent the 1st, 2nd, and nth feedback messages provided by the experts, respectively, where n represents the total number of nodes. Represents a cause-effect graph structure. Represents observation data, This represents the fit score between the graph structure and the data. Indicate the complexity of the graph structure. This is the complexity penalty coefficient. metric graph structure Feedback from experts The degree of inconsistency, This serves as a balancing parameter for feedback inconsistencies. For example, for causal edges confirmed by experts. , indicating from node To the node The causal relationship can be constrained. ,in Representing an edge The weight, A positive threshold; for causal edges that are denied by experts, constraints can be added. This indicates that the weight of the edge is zero, meaning that the causal relationship does not exist.

[0265] Step 5.4: Incremental Learning Mechanism: According to the embodiments of this application, an incremental learning mechanism is implemented to integrate new data and feedback based on the existing causal structure, avoiding relearning from scratch. For new datasets... and cumulative feedback The update and optimization goals are:

[0266] ;

[0267] in, Represents a cause-effect graph structure. This represents a newly collected dataset. Represents historical datasets, This represents the cumulative set of expert feedback. This represents the fit score between the graph structure and the data. Indicate the complexity of the graph structure. This indicates the degree of inconsistency between the diagram structure and expert feedback. This represents the weighting balance coefficient for controlling the new and old data (the value ranges from 0 to 1). This represents the complexity penalty coefficient. This represents a balancing parameter indicating feedback inconsistency. It should be understood that this approach ensures the system can smoothly adapt to new data while retaining validated knowledge.

[0268] In some implementations, an incremental learning strategy based on importance sampling can be optionally employed. This involves weighting historical data by importance, retaining samples that remain important for the current task, and gradually eliminating outdated data that is no longer relevant. This approach is particularly suitable for situations where performance characteristics change gradually during battery aging, allowing the system to adjust the causal model according to the degree of battery aging. Furthermore, a progressive learning framework can be implemented, maintaining dedicated model libraries for different battery types or usage scenarios. When encountering a new battery type, it can quickly adapt through transfer learning, rather than learning a completely new model from scratch.

[0269] Step 5.5: Knowledge Persistence: This application constructs a knowledge persistence mechanism to store the optimized causal structure and domain knowledge in a distributed database for sharing among multiple vehicles and repair stations. The stored content includes:

[0270] Cause-effect graph structure: includes nodes, edges and their weights;

[0271] Expert feedback: All expert feedback collected and its source information;

[0272] Case library: Typical failure cases and their causal analysis results;

[0273] Explanation Template: Explanation templates for common faults.

[0274] Furthermore, through regular synchronization and conflict resolution mechanisms, we ensure that knowledge from different sources can be effectively integrated to form a unified knowledge base.

[0275] Step 5.6: Implementation of the Interactive Causal Exploration Interface: According to the embodiments of this application, the specific implementation of the interactive causal exploration interface includes the following components:

[0276] Visualization module:

[0277] Multi-view integration: simultaneously provides a global causal network view, a hierarchical view, and a time evolution view;

[0278] Force-directed layout: Uses a force-directed algorithm based on physics simulation to lay out cause-effect graphs, visually displaying the relationships between nodes;

[0279] Node and edge encoding: node size maps to variable importance, color maps to variable type; edge thickness maps to causal strength, color maps to confidence.

[0280] Interactive technology: Supports zooming, panning, filtering, expanding / collapse, and other operations;

[0281] Interactive operation module:

[0282] Causal hypothesis testing: allows experts to create "hypothetical scenarios" and predict intervention effects in real time;

[0283] Causal edge modification: Supports adding, deleting, or modifying causal edges via drag-and-drop operations;

[0284] Parameter threshold adjustment: Adjust the displayed causal relationship strength and confidence threshold using the slider control;

[0285] Annotation feature: Allows experts to add professional explanations or notes for specific causal relationships;

[0286] Expert feedback collection module:

[0287] Structured feedback forms: Design standardized templates to collect expert evaluations and modification suggestions;

[0288] Comparison view: Showing the original causal structure and the expert-modified structure side by side;

[0289] Conflict flagging: Automatically identifies and flags conflicting feedback from different experts;

[0290] Trust rating: Expert feedback is assigned different weights based on the expert's professional background and the accuracy of historical feedback.

[0291] Knowledge synchronization module:

[0292] Incremental update: Only transmits the changed knowledge part, saving bandwidth;

[0293] Conflict resolution algorithm: Automatically resolves conflict feedback based on majority voting and expert trust.

[0294] Version control: Maintains historical versions of causal knowledge and supports rollback operations;

[0295] Access control: Different modification permissions are set according to the expert role.

[0296] This application's interactive causal exploration interface utilizes the React framework for front-end interaction and a microservice architecture for the back-end. Causal graph visualization employs a custom D3.js component. The interface supports responsive layouts from high-end workstations to mobile devices, ensuring a consistent user experience across different devices. User research shows that this interface improves the efficiency (average 43% reduction in operation time) and accuracy (52% reduction in error rate) of expert correction of causal structures compared to traditional static displays.

[0297] In some implementations, collaborative workflow functionality can be optionally integrated, supporting multiple experts to simultaneously edit and discuss the same causal structure online, similar to the working method of a code collaboration platform. For example, a change proposal and review mechanism can be implemented, allowing experts to suggest modifications to the causal structure, which are then reviewed by other experts or automatically verified by the system before being merged. Furthermore, an augmented reality (AR) based interactive interface can be implemented, allowing causal diagnostic results to be directly overlaid onto the actual battery assembly using AR glasses at the repair site, helping repair personnel intuitively understand the location and cause of the fault and improving repair efficiency. In other implementations, voice interaction capabilities can also be integrated, supporting the exploration and modification of causal structures through natural language dialogue, freeing experts from relying on complex graphical interfaces, which is particularly suitable for scenarios such as on-site repairs.

[0298] In summary, the method of this application, by integrating the powerful representational capabilities of neural networks with the logic and interpretability of symbolic reasoning, constructs a power battery monitoring system that can automatically discover causal relationships, accurately diagnose the root causes of anomalies, provide professional explanations, and support human-machine collaborative optimization, thus effectively solving the core technical challenges in the current field of power battery anomaly monitoring.

Claims

1. A method for abnormal monitoring of power batteries in new energy vehicles, characterized in that, Includes the following steps: Construct a structural causal model of the power battery system, and integrate expert domain knowledge and historical anomaly cases into a causal knowledge graph; Based on causal knowledge graphs, a causal structure learning algorithm based on continuous optimization is adopted to represent causal graph discovery as a differential constraint optimization problem, and causal relationships are automatically discovered from battery time series data. Based on automatically discovered causal relationships, counterfactual reasoning techniques are used to perform multi-level causal reasoning, distinguish between direct and indirect causal relationships, and construct a multi-level causal abstraction mechanism. Based on a multi-level causal abstraction mechanism, causal relationship explanations are generated, and the discovered causal relationships are automatically mapped to battery physical mechanisms and failure modes. Based on the causal relationships mapped to battery physical mechanisms and failure modes, an interactive causal exploration interface allows domain experts to revise and improve the causal graph.

2. The method for abnormal monitoring of power batteries in new energy vehicles according to claim 1, characterized in that, The steps for integrating expert domain knowledge and historical anomaly cases into a causal knowledge graph include: Encode battery expert knowledge into differentiable probabilistic logic expressions to establish an initial correlation between battery parameter changes and failure modes; Time-series data and diagnostic results were extracted from historical battery failure cases to construct a training dataset; Based on domain knowledge and historical cases, an initial causal knowledge graph is constructed. The causal knowledge graph is converted into a differentiable parameterized representation, which supports subsequent optimization.

3. The method for abnormal monitoring of power batteries in new energy vehicles according to claim 1, characterized in that, The causal structure learning algorithm further includes: Construct a scoring function and a complexity regularization term to form the optimization objective function; To ensure that the learned causal graph satisfies the property of an acyclic directed graph, matrix constraints are introduced; Transform the constrained optimization problem into an augmented Lagrangian form; The gradient is calculated by backpropagation automatic differentiation, and the weight matrix is ​​iteratively optimized using gradient descent.

4. The method for abnormal monitoring of power batteries in new energy vehicles according to claim 1, characterized in that, The steps of the multi-level causal reasoning include: Construct counterfactual reasoning models to verify the causal relationships of the assumptions; By comparing the effects of direct intervention and the cumulative effects of indirect pathways, we can distinguish between direct and indirect causal relationships. The causal relationships are organized according to the level of abstraction to form a multi-level causal structure including the battery physical parameter layer, the battery state layer, and the fault mode layer. Implement a hierarchical causal reasoning algorithm that supports reasoning propagation between different levels of abstraction; The confidence level of the inference results is assessed to provide a reliability index for the inference results.

5. The method for abnormal monitoring of power batteries in new energy vehicles according to claim 1, characterized in that, The steps for generating causal explanations include: Extract the critical path that has the greatest impact on the target failure from the complete cause-effect graph; Map the identified causal paths to the battery's physical mechanisms; Generate natural language explanations based on templates and specific parameter values; Integrate the confidence level of the reasoning process into the explanation; Generate multi-level explanations for users with different levels of expertise.

6. The method for abnormal monitoring of power batteries in new energy vehicles according to claim 1, characterized in that, The implementation of the interactive causal exploration interface includes: The learned causal structure is presented in an interactive visualization. Establish an expert feedback collection mechanism to allow domain experts to evaluate and modify the causal structure; Develop feedback integration algorithms to transform expert feedback into constraints or priors for causal learning; Implement an incremental learning mechanism to integrate new data and feedback on the basis of existing causal structures; A knowledge persistence mechanism is constructed to store the optimized causal structure and domain knowledge in a distributed database.

7. The method for abnormal monitoring of power batteries in new energy vehicles according to claim 1, characterized in that, It also includes the implementation of a neural symbolic hybrid model, which includes: A neural network component consisting of a graph neural network is used to learn the relationship representation between nodes; Symbolic logic reasoning components implemented through differentiable logic rules; Neural symbol fusion is achieved by using neural networks to assist symbolic reasoning and symbolic rules to guide neural learning.

8. The method for abnormal monitoring of power batteries in new energy vehicles according to claim 7, characterized in that, The symbolic logic reasoning component of the neural symbolic hybrid model converts expert rules into first-order logic forms and performs continuous processing of logical operations through t-norm or Lukasiewicz logic.

9. A method for abnormal monitoring of power batteries in new energy vehicles according to any one of claims 1-8, characterized in that, The battery parameters of the power battery system include at least one of voltage, current, temperature, and internal resistance; the fault types include at least one of capacity decay, internal short circuit, and lithium plating.

10. A power battery anomaly monitoring system suitable for new energy vehicles, characterized in that, For executing the power battery anomaly monitoring method for new energy vehicles as described in claim 9, the system includes: The causal knowledge representation building module is used to integrate expert domain knowledge and historical anomaly cases into a causal knowledge graph; Differentiable causal structure learning module, used to automatically discover causal relationships from battery time-series data; A multi-level causal reasoning module is used to distinguish between direct and indirect causal relationships and to construct a multi-level causal abstraction mechanism; The causal relationship interpretation and generation module is used to map the discovered causal relationships to battery physical mechanisms and failure modes; An interactive causal knowledge optimization module is used to support domain experts in revising and improving causal graphs; Through the coordinated operation of the above modules, the system achieves accurate monitoring, root cause diagnosis, and knowledge optimization of power battery anomalies.

Images