A lithium battery failure analysis method and system based on multi-channel RAG

Abstract

The application discloses a lithium battery failure analysis method and system based on a multi-channel RAG, and belongs to the technical field of lithium battery failure analysis. The method comprises the following steps: constructing a field knowledge vector library and a multi-level failure analysis knowledge graph; constructing a multi-channel retrieval enhancement generation framework, analyzing a user query intention, and obtaining related knowledge through a multi-channel cooperative mechanism of knowledge graph structure retrieval, knowledge graph semantic retrieval and field knowledge vector retrieval; calling a large language model to infer and generate a failure analysis result, and performing self-adaptive adjustment based on quality evaluation; and dynamically maintaining the knowledge base through internal and external knowledge updating. The application realizes unified modeling and structured expression of multi-source heterogeneous knowledge of failure analysis, improves the relevance and integrity of the retrieval result, reduces the dependence on artificial experience, and enhances the explainability and intelligent level of failure analysis.

Description

Technical Field

[0001] This invention relates to the field of lithium battery failure analysis technology, and in particular to a lithium battery failure analysis method and system based on multi-channel RAG. Background Technology

[0002] In recent years, lithium batteries have been widely used in various fields such as new energy vehicles and energy storage power stations due to their high energy density, long cycle life, and low self-discharge rate. However, with the increasing scale of their use, the failure problem of lithium batteries has become increasingly prominent. Currently, companies mainly rely on manual experience, traditional electrochemical testing methods, and disassembly analysis techniques for identifying and analyzing failed batteries. These traditional methods often suffer from low diagnostic efficiency, high subjectivity, and reliance on expert experience. Furthermore, although companies have accumulated professional knowledge documents such as fault diagnosis manuals and failure case reports, these documents have not been effectively utilized due to issues such as inconsistent standards and non-standard formats.

[0003] Currently, although existing research has applied machine learning (such as SVM, CatBoost, etc.) and deep learning models (such as ResNet, BiLSTM, etc.) to fault diagnosis and state detection of lithium batteries, these methods typically rely on large-scale, high-quality labeled data for model training, resulting in high data acquisition costs and sensitivity to label quality. Furthermore, these methods primarily target numerical data for modeling and analysis, making it difficult to effectively model and utilize professional domain knowledge that exists in textual form within enterprises. In particular, they lack the ability to express and reason about knowledge such as failure mechanisms, causal relationships, and empirical rules, leading to insufficient applicability and interpretability of existing methods in complex failure analysis scenarios.

[0004] Large language models, such as ChatGPT, Gemini, and DeepSeek, have demonstrated powerful capabilities in natural language understanding, reasoning, and generation. However, their application in lithium battery failure analysis still faces challenges. On the one hand, general-purpose large language models lack domain-specific knowledge such as lithium battery failure mechanisms, electrochemical characteristics, enterprise fault diagnosis manuals, and case studies, making them prone to "illusions" during diagnostic analysis and hindering the reliability and interpretability of results. On the other hand, existing vector retrieval-based enhanced generation techniques typically rely on semantic vector similarity to retrieve unstructured text. While this can alleviate the lack of domain knowledge in large language models to some extent, it still has significant shortcomings: single-vector retrieval mechanisms primarily rely on semantic similarity matching, making it difficult to effectively utilize the causal relationships, hierarchical structures, and logical connections involved in lithium battery failure analysis, resulting in a lack of structured support in the retrieval results. Furthermore, the retrieved knowledge fragments often exhibit discrete and fragmented characteristics, lacking the ability to integrate cross-source knowledge and failing to form a complete failure mechanism analysis chain, thus affecting the accuracy and consistency of subsequent reasoning and analysis results.

[0005] Therefore, there is a need for a method and system that can significantly reduce reliance on human experience and expertise, improve the efficiency of lithium battery failure analysis, automatically generate solutions and diagnostic reports, and effectively enhance the intelligence level of lithium battery failure analysis decision-making. Summary of the Invention

[0006] The purpose of this invention is to provide a lithium battery failure analysis method and system based on multi-channel RAG, in order to solve the problems of low utilization of professional knowledge, reliance on human experience, and fragmented search results, lack of structured support and logical correlation caused by a single search mechanism in the prior art.

[0007] To achieve the above objectives, this invention provides a lithium battery failure analysis method based on multi-channel RAG, comprising the following steps: S1 collects, cleans, and performs semantic-level preprocessing on multi-source data and domain knowledge related to lithium battery failure, and then performs vectorized modeling to build a domain knowledge vector library for subsequent failure analysis.

[0008] S2, based on the knowledge graph modeling method, automatically extracts and structures key information from lithium battery failure case reports, constructs a multi-level lithium battery failure analysis knowledge graph, and performs semantic enhancement processing on the knowledge graph.

[0009] S3 constructs a multi-channel retrieval-enhanced failure analysis framework. Based on the semantic parsing results of the user query, it acquires knowledge related to the failure analysis task through a multi-channel retrieval mechanism; it calls a large language model to perform inference to generate failure analysis results, and adaptively adjusts the retrieval strategy and inference generation process; at the end of the task, it summarizes the failure analysis task and generates a corresponding failure analysis report.

[0010] S4 dynamically maintains knowledge related to lithium battery failure through internal and external knowledge updates, in order to keep the timeliness and consistency of the knowledge on which the multi-channel RAG framework depends.

[0011] Preferably, step S1 specifically includes the following steps: S11 involves collecting multi-source data and knowledge in the field of lithium battery failure, and performing format unification, structure standardization, and noise removal to obtain a standardized text dataset. The multi-source data and knowledge related to lithium battery failure mainly include two categories: the first category is unstructured general text data, including but not limited to academic literature related to lithium-ion battery failure mechanisms and internal enterprise failure analysis manuals; the second category is semi-structured failure case data, including but not limited to historically accumulated failure analysis reports.

[0012] S12, for unstructured general text data such as professional documents and corporate manuals, constructs a semantic preprocessing workflow driven by a large language model. This workflow performs semantic segmentation, semantic filtering, and information compression on the cleaned text data to generate structured semantic text units related to lithium battery failure mechanisms, failure modes, and cause diagnosis. Specifically, semantic segmentation divides long texts into multiple semantically continuous text segments based on a preset length threshold and semantic integrity constraints.

[0013] S13. Input the semantic text unit into the pre-trained vector encoding model to generate the corresponding high-dimensional semantic vector, and establish the association between the semantic vector and the original text content, source information and semantic tags; store the semantic vector and its associated information in the vector database to build a domain knowledge vector library.

[0014] Preferably, step S2 specifically includes the following steps: S21. Construct a multi-level fusion knowledge graph ontology structure for lithium battery failure analysis. The ontology structure includes three core layers and their associated mechanisms: The first layer is the failure case entity layer: used to characterize the failure phenomenon and handling process, defining entity nodes and their attributes, including at least "failure case", "failed cell", "failure event", "failure cause", and "improvement measures". It also defines the "case semantics" node and its attributes generated by subsequent semantic enhancement processing.

[0015] The second layer is the failure mode community layer: it is used to represent a set of cases with similar failure modes, and defines the "failure mode community" node.

[0016] The third layer is the fault tree analysis logic layer: used to characterize the deep mechanism deduction of failure, defining event nodes including "top event", "intermediate event" and "bottom event", as well as logical gate relationship edges used to connect parent events and child events. Among them, the logical gate types of the logical gate relationship edges include "AND gate" and "OR gate".

[0017] Association Mechanism: Define inter-layer logical mapping rules to assign the "failure case" node in the failure case entity layer to the corresponding "failure mode community" node in the failure mode community layer. Then, associate the "failure mode community" node or "failure event" node with the "top event" node in the fault tree analysis logic layer to construct a cross-layer index path. In some implementations, when a corresponding "failure mode community" node has not yet been formed, the "failure event" node can be directly associated with the "top event" node.

[0018] S22 executes the instantiation, logical aggregation, and semantic enhancement encapsulation of the knowledge graph based on large language models and graph computing algorithms, including: First, for the collected semi-structured failure case reports, structured extraction of failure case entities and construction of the failure case entity layer are performed. Based on lithium battery failure warning word templates and failure mode dictionaries built with expert experience, a large language model is used to extract the core information of the case reports, including failure case metadata and key entity information such as failed cells, failure events, failure causes, and improvement measures. The semantic relationships and causal associations between failed cells, failure events, failure causes, and improvement measures are further identified and organized into a standardized failure causal chain containing "failed cell - failure event - failure cause - improvement measure". After terminology normalization, entity alignment, and logical verification, the extracted information is instantiated into entity nodes and relationships in a knowledge graph, completing the initial construction of the failure case entity layer.

[0019] Secondly, based on graph structure similarity and attribute semantic similarity, a graph clustering algorithm is used in the failure case entity layer to cluster failure cases with high similarity features into multiple "failure mode communities", construct the failure mode community layer, and establish its affiliation association with the failure case entity layer.

[0020] Next, a fault tree analysis structure is extracted from the failure case data using a large language model. This involves identifying and extracting the hierarchical structure of "top event - intermediate event - bottom event" and the logical relationships between events, constructing a fault tree analysis logic layer, and establishing a mapping relationship between the "top event" node and the aforementioned "failure mode community" nodes. In some implementations, the "failure event" node can also be directly associated with the "top event" node.

[0021] Finally, cross-level semantic enhancement and vectorization encapsulation are performed on the overall graph. Starting from the "failure case" node in the failure case entity layer, a path traversal operation is performed across the failure case entity layer, failure mode community layer, and fault tree analysis logic layer according to the predefined inter-layer mapping relationship to extract the full-link failure subgraph corresponding to "failure phenomenon-failure mode-failure mechanism". The large language model is called to reconstruct the contextual semantics of the subgraph to generate an enhanced semantic summary that integrates structured factual information and causal logical relationships. Subsequently, a pre-trained vector encoding model is used to transform the enhanced semantic summary into a high-dimensional semantic vector to generate the corresponding "case semantic" node. The enhanced semantic summary and the high-dimensional semantic vector are stored as attributes of the "case semantic" node, and the association relationship between the "case semantic" node and the corresponding "failure case" node is established.

[0022] Preferably, the multi-channel RAG framework in step S3 specifically includes: S31, Intent Recognition and Task Parsing: This function performs semantic understanding and intent recognition on user-input natural language queries. It utilizes a large language model to identify key lithium battery failure entity elements within the query and maps the query to predefined failure analysis task types. Key entity elements include at least one or more of the following: failure event description, failed cell identification, and operating condition information. Task types include, but are not limited to, root cause tracing, case comparison analysis, and improvement measure recommendations. Based on the task type, initial weight parameters are determined for subsequent multi-channel retrieval result fusion. Simultaneously, multiple rounds of conversations with the user are maintained to support the resolution of referential inferences and semantic completion of subsequent user queries until the current failure analysis task is completed.

[0023] S32, Multi-channel Retrieval: Used for progressive retrieval of multi-source knowledge related to failure analysis under the control of query parsing results. Multi-channel retrieval includes: The knowledge graph structure retrieval channel is configured to generate query statements that conform to graph database query specifications based on key entities identified by the intent recognition and task parsing modules and predefined relational constraints of the knowledge graph; it performs cross-level structured traversal of the knowledge graph to obtain entity nodes and their attribute information in the first-level failure case entity layer, and further retrieves failure mode communities in the second layer and the corresponding fault trees in the third layer through inter-layer mapping association.

[0024] The knowledge graph semantic retrieval channel is configured to use a query-based semantic vector representation to calculate the vector similarity between the query and the pre-stored "case semantic" nodes in the knowledge graph. This is used to recall nodes that meet the similarity criteria and their associated full-link failure subgraphs, serving as a supplement to the semantic level of graph structure retrieval.

[0025] The domain knowledge vector library retrieval channel is configured to perform vector similarity calculations on text slices in the domain knowledge vector library based on the semantic vector representation of the query, in order to recall unstructured domain background knowledge text that is highly related to the query semantics.

[0026] Step S32 employs the following progressive multi-channel retrieval strategy: First, based on the semantic parsing results of the user query, check if there are any key entities that can be mapped to the failure case knowledge graph. If a mappable entity exists, the knowledge graph structure retrieval channel is triggered first to obtain the failure case subgraph directly related to the query target.

[0027] Subsequently, based on the results of the previous step, an enhanced query vector is constructed. If the knowledge graph structure retrieval channel is triggered, the original query is fused with the structured feature information in the failure case subgraph to construct the enhanced query vector; otherwise, the query vector remains unchanged, i.e., the enhanced query vector is equal to the original query vector.

[0028] Then, enhanced query vectors are used to perform similarity retrieval on the "case semantics" node in the knowledge graph and the domain knowledge vector library, respectively, to recall structured and unstructured knowledge related to the query semantics.

[0029] Finally, redundancy removal was performed on the search results for each of the three channels, and the top N results for each channel were retained according to the preset channel weights. i The results yield a candidate set containing N candidate results. Then, a re-ranking model is introduced to score and rank the semantic relevance of the candidate results, and the top M best search results are extracted to construct an enhanced context for inference analysis, where N and M are configurable parameters of the system.

[0030] S33, Result Generation and Adaptive Adjustment: Used to generate inference results from failure analysis based on multi-channel retrieval results, and to achieve adaptive adjustment through a result evaluation and feedback mechanism. Includes: First, the user's instructions and the enhanced context obtained from multi-channel retrieval are filled into the preset prompt word template to construct inference instructions and input into the large language model for inference analysis, generating the initial structured failure analysis results.

[0031] Then, the failure analysis results are evaluated for quality using a result evaluation mechanism. When the evaluation results do not meet the preset quality requirements, the fusion weights of the multi-channel retrieval results and the constraints of the large model inference generation are adaptively adjusted based on the evaluation results, and the retrieval result fusion and failure analysis result generation process is re-executed. When the evaluation results meet the quality requirements or reach the preset number of iterations, the current optimal failure analysis result is output. Furthermore, the initial weights for the corresponding task type are updated based on the optimal weight coefficients for use in initializing the retrieval weights for subsequent similar failure analysis tasks.

[0032] S34, Task Summary: This is used to perform global backtracking and information aggregation on the multi-round interaction records and phased analysis results generated during the current failure analysis task when the current failure analysis task is terminated, and to generate a failure analysis report that is suitable for the analysis task type. The failure analysis report comprehensively reflects the core conclusions, reasoning process and key evidence of this failure analysis task, serves as the final output of this task, and serves as one of the standardized input sources for updating failure analysis knowledge in the subsequent step S4.

[0033] Preferably, step S4 specifically includes: S41, Internal Failure Analysis Knowledge Incremental Update: The failure analysis report output in step S34 is transformed into reusable domain knowledge, and updated to the domain knowledge vector library and failure analysis knowledge graph according to the knowledge form. Internal failure analysis knowledge incremental update includes structured knowledge mapping and semantic knowledge accumulation: Structured knowledge mapping: For structured information in the failure analysis report (such as failure causes, improvement measures, etc.), according to the predefined knowledge graph schema mapping rules, operations such as adding node instances, updating node attributes, and supplementing relationship instances are performed to integrate the new knowledge in the failure analysis report into the knowledge graph.

[0034] Semantic knowledge accumulation: The failure analysis report is treated as a new unstructured text. The corresponding semantic vector is generated according to the text preprocessing and semantic representation method in step S1, and stored as new domain knowledge in the domain knowledge vector library.

[0035] S42, External Multi-Source Knowledge Periodic Synchronization, is used to periodically introduce external lithium battery failure-related knowledge, including but not limited to new lithium battery industry standards, technical manuals, and academic papers. The external knowledge is processed using the text preprocessing method in step S1, generating semantic vectors which are then stored in the domain knowledge vector library. Simultaneously, the external knowledge is structurally extracted and aligned using the knowledge graph modeling method in step S2, updating it in the multi-level lithium battery failure analysis knowledge graph to supplement knowledge coverage for new cell models or novel failure modes.

[0036] This invention also provides a lithium battery failure analysis system based on multi-channel RAG, used to implement the above-described method, comprising: The multi-source heterogeneous failure analysis knowledge base construction module is configured to perform format standardization and semantic segmentation of multi-source text data through a large language model preprocessing workflow, and generate a domain knowledge vector library using a pre-trained vector encoding model. At the same time, it is configured to construct a multi-level failure analysis knowledge graph based on a preset three-layer ontology structure through knowledge extraction and logical aggregation, and generate "case semantic" nodes through cross-level path traversal to achieve the association and storage of structured facts and unstructured semantics.

[0037] The intent recognition and task parsing module is configured to use a large language model to identify key elements such as failure events and cell identifiers from user queries, map query tasks to predefined analysis task types, and initialize the fusion weights of multi-channel retrieval based on the task type. During multi-round interactions, it is also configured to perform referential resolution, context information completion, and semantic consistency verification to maintain the continuity and integrity of the failure analysis task context.

[0038] The multi-channel retrieval module is configured to perform multi-channel retrieval based on user queries, including knowledge graph structure retrieval, knowledge graph semantic retrieval, and domain knowledge vector library retrieval. It also performs recall and reordering of structured and unstructured failure analysis knowledge and constructs enhanced context for analytical reasoning.

[0039] The results generation and adaptive adjustment module is configured to generate failure analysis results based on the enhanced context call of the large language model, and to adaptively adjust the retrieval strategy and inference generation process according to the quality assessment of the generated results, so as to output a failure analysis report that meets the preset quality requirements.

[0040] The task summary module is configured to summarize and perform a global analysis of the multi-round interaction records, retrieval results at each stage, and inference generation results generated during the current failure analysis task when the current failure analysis task terminates. It generates a structured failure analysis report based on the analysis task type. The failure analysis report serves as the final output of this failure analysis task and as one of the input data for the subsequent domain knowledge dynamic update module.

[0041] The domain knowledge dynamic update module is configured to standardize the failure analysis results that have passed the quality assessment and update them to the failure analysis knowledge base to achieve incremental accumulation of failure analysis experience. In addition, it regularly scans external multi-source data, performs semantic or structural processing on new knowledge according to its data form, and updates it to the failure analysis knowledge base synchronously.

[0042] Therefore, the lithium battery failure analysis method and system based on multi-channel RAG with the above-described structure of the present invention has the following beneficial effects: (1) By constructing a multi-source heterogeneous knowledge system that integrates domain knowledge vector library and failure analysis knowledge graph, this invention realizes unified modeling and structured expression of unstructured and semi-structured knowledge scattered in academic literature, corporate manuals and historical case reports, and solves the problems of low utilization rate of professional knowledge and inconsistent data standards in the prior art.

[0043] (2) By introducing a multi-channel collaborative retrieval mechanism of knowledge graph structure retrieval, knowledge graph semantic retrieval and domain knowledge vector retrieval, the present invention enables the retrieval process to not only be based on semantic similarity, but also to combine causal relationships and hierarchical structure information in failure analysis, thereby significantly improving the relevance and completeness of the retrieval results.

[0044] (3) This invention constructs a large language model reasoning generation mechanism based on multi-channel retrieval enhancement, and inputs structured knowledge graph information and unstructured domain knowledge as enhanced context into the large language model. By utilizing its natural language understanding and causal reasoning capabilities, it realizes the automation of failure mechanism analysis, root cause tracing and improvement measures generation, thereby reducing the dependence on human experience and improving the intelligence level of lithium battery failure analysis.

[0045] (4) By constructing an internal and external knowledge dynamic update mechanism, the present invention transforms the generated failure analysis results into reusable knowledge and continuously updates it to the knowledge base. At the same time, it introduces new external knowledge for periodic supplementation, realizes the continuous evolution and self-optimization of the knowledge system, and improves the system's adaptability to new failure modes and its long-term application value.

[0046] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a lithium battery failure analysis method based on multi-channel RAG proposed in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of a multi-channel RAG lithium battery failure analysis framework proposed in Embodiment 1 of the present invention. Detailed Implementation

[0048] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0049] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0050] Example Please refer to Figure 1 As shown, this embodiment provides a lithium battery failure analysis method based on multi-channel RAG, including: S1 collects, cleans, and preprocesses multi-source data and domain knowledge related to lithium battery failure, and constructs a domain knowledge vector library for subsequent failure analysis.

[0051] Specifically, the multi-source data and knowledge related to lithium battery failure in this embodiment includes two categories: the first category is unstructured general text data, including but not limited to academic literature related to lithium-ion battery failure mechanisms (such as journal articles involving lithium plating and internal short circuit mechanisms) and internal failure analysis manuals of enterprises (such as "Battery Fault Diagnosis Manual" and "Cell Undervoltage Analysis Work Instruction"); the second category is semi-structured failure case data, including but not limited to historically accumulated "Failure Analysis Reports" (containing "Failure Phenomenon Description", "Faulty Component Analysis", "Root Cause", "Improvement Measures", etc.).

[0052] Specifically, a large-scale model workflow based on the LangChain framework is introduced to preprocess the first type of unstructured general text data to obtain a domain knowledge set and store it in a domain vector knowledge base. The specific steps are as follows: First, the document loader provided by LangChain is used to load unstructured text data such as documents and manuals. Then, regular expressions are used to remove headers and footers, table of contents numbers, garbled characters, and formatting content that is irrelevant to failure analysis, resulting in standardized text for semantic processing.

[0053] Secondly, for the cleaned and standardized text, a recursive character text segmenter is used to semantically slice the text according to preset sliding window parameters. In this embodiment, the length threshold of a single text segment is set to 500 characters, and the overlap is set to 50-100 characters, so as to ensure the semantic continuity of the text while making the segmented text segments meet the processing constraints of the large language model context window.

[0054] Then, for each text segment, under the constraints of the preset prompt word template, the large language model is invoked to perform semantic filtering and information compression on the text content to determine its relevance to failure analysis. Key semantic content related to the failure mechanism, failure mode and analysis experience of lithium batteries is extracted, and information irrelevant to the failure analysis task is removed to generate a standardized text segment for vectorized modeling.

[0055] Finally, the BGE-M3 embedding model is invoked to normalize the text fragments. Perform high-dimensional semantic vector encoding to map text fragments into high-dimensional dense vectors. Its mathematical expression is: ; and the generated semantic vector Its corresponding original text fragment The source metadata is stored together in the FAISS vector database to complete the domain vector knowledge base. The initial construction.

[0056] S2, based on the knowledge graph modeling method, automatically extracts and structures key information from semi-structured failure case data, constructs a multi-level lithium battery failure analysis knowledge graph, and performs semantic enhancement processing on the knowledge graph.

[0057] In this embodiment, a multi-level fused knowledge graph ontology structure for lithium battery failure analysis is first defined. The ontology structure includes three parts: a failure case entity layer, a failure mode community layer, and a fault tree analysis logic layer, as well as inter-layer association rules.

[0058] Specifically, the failure case entity layer is used to characterize the failure phenomenon and handling process, defining entity nodes and their attributes, including types such as "failure case," "failure cell," "failure event," "failure cause," and "improvement measures," as well as "case semantics" nodes and their attributes generated by subsequent semantic enhancement processing. The failure mode community layer is used to characterize a set of cases with similar failure modes and defines "failure mode community" nodes. The fault tree analysis logic layer is used to characterize the deep mechanism deduction of failure, defining event nodes including "top event," "intermediate event," and "bottom event," as well as logical gate relationship edges used to connect parent events and child events. The logical gate types of the logical gate relationship edges include "AND gate" and "OR gate." The inter-layer association mechanism defines the rules for assigning "failure case" nodes in the failure case entity layer to the corresponding "failure mode community" nodes in the failure mode community layer, and then associating the "failure mode community" nodes to the "top event" nodes in the fault tree analysis logic layer. In some implementations, when a corresponding "failure mode community" node has not yet been formed, the "failure event" node can be directly associated with the "top event" node.

[0059] As a preferred embodiment of the present invention, the specific definition of the multi-level knowledge graph ontology structure can be found in Table 1.

[0060] Table 1. Definition of the ontological structure of the knowledge graph for failure analysis of multi-level lithium batteries

[0061] It should be noted that the failure case entity layer, failure mode community layer, and fault tree analysis logic layer mentioned above achieve cross-level indexing and knowledge fusion through a predefined inter-layer association mechanism. Specifically, this includes: (1) Assign the “failure case” node in the failure case entity layer to the corresponding “failure mode community” node in the failure mode community layer to establish an “assignment” relationship; (2) Associate the “Failure Mode Community” node in the Failure Mode Community layer with the “Top Event” node in the Fault Tree Analysis Logic Layer to establish a “mapping” relationship; (3) In some implementations, when the corresponding “failure mode community” node has not yet been formed, the “failure event” node in the failure case entity layer can be directly associated with the “top event” node in the fault tree analysis logic layer.

[0062] Furthermore, after completing the ontology structure definition, this embodiment performs knowledge graph instantiation construction based on large language model and graph computing algorithm for semi-structured failure case data.

[0063] Specifically, including: Entity Relationship Extraction for Failure Cases: A workflow for entity extraction is built using LangChain. A large language model is invoked to parse failure case data under preset prompt word template constraints. On one hand, metadata is automatically extracted from failure cases, and the large language model performs semantic compression on the full report text to generate a summary text. This summary text is stored as the "Case Text Summary" attribute of the "Failure Case" node. On the other hand, key entity information such as failed battery cells, failure events, failure causes, and improvement measures is extracted. Based on the extraction results, alignment and disambiguation are performed on the extracted entities, and semantic relationships between entities are identified and mapped to predefined relation types in the ontology structure, forming corresponding "entity-relation-entity" triples. Furthermore, a self-reflection mechanism is introduced to guide the large language model to self-evaluate and iteratively correct the extracted triples, filtering and outputting truly valid triples. In this embodiment, a failure analysis knowledge graph is constructed with entities as nodes and relations as edges and stored in the Neo4j graph database.

[0064] Failure Mode Community Aggregation: Based on the structural association features and semantic similarity features among cases in the entity layer of failure cases, the Leiden community detection clustering algorithm is used to divide failure cases into communities and determine the set of failure cases with high similarity failure mode features; for each set of failure cases, a corresponding "failure mode community" node is generated, and the affiliation relationship between case nodes and community nodes is established.

[0065] Fault tree analysis logic layer construction: A large language model is used to extract the fault tree analysis tree structure from the failure case data. "Top event," "intermediate event," and "bottom event" nodes are instantiated, and edges with logical gate type attributes are established based on logical relationships. Subsequently, a semantic matching strategy is used to calculate the vector similarity between the "community description" and the "top event description," mapping community nodes to the corresponding "top event" nodes in the fault tree analysis logic layer.

[0066] Cross-level semantic enhancement processing: After completing the construction of a multi-level knowledge graph, this embodiment performs cross-level path traversal operations on the graph, using the "failed case" node in the failed case entity layer. Starting from the entity layer, the failure mode community layer, and the fault tree analysis logic layer, a full-link failure subgraph is extracted. The large language model is invoked to reconstruct the subgraph using natural language, generating an enhanced semantic summary that integrates failure phenomena, failure modes, and failure mechanisms. Subsequently, the enhanced semantic summary is vectorized and encoded using the BGE-M3 embedding model, the mathematical expression of which is: ; in, This indicates the use of large model inference. This represents path traversal. The enhanced semantic summary is then combined with the generated semantic vector. Store it in the "Case Semantics" node and connect it to the "Failed Cases" node.

[0067] S3 constructs a multi-channel retrieval-enhanced failure analysis framework. Based on the semantic parsing results of the user query, it obtains structured and unstructured knowledge related to the failure analysis task through a multi-channel retrieval mechanism, and calls a large language model to generate failure analysis results under structured reasoning and multi-source enhanced context constraints.

[0068] Please refer to Figure 2 , Figure 2 The diagram shown is a general overview of the multi-channel RAG lithium battery failure analysis framework in this embodiment, which specifically includes the following execution steps: S31, Intent Recognition and Task Parsing: Used to convert unstructured user natural language input into machine-executable retrieval instructions.

[0069] Entity recognition and intent classification: Receiving natural language query commands input by the user. The large language model is used as the semantic parsing and recognition unit. Guided by a small number of prompt words, it performs semantic parsing and recognition. Perform named entity recognition and intent classification.

[0070] Specifically, identify and extract the set of key entity elements contained in the query. ,in, This represents the key entities of failed battery cells extracted from user queries. Represents the key entities of failure events extracted from user queries; Analysis The semantic tendencies are analyzed to identify the core intent and the queries are mapped to predefined analysis task types. ,in, Indicates the type of failure root cause tracing task. This indicates the type of task for case comparison analysis. This indicates the type of task recommended for improvement measures.

[0071] Weight configuration: based on the mapped task type Determine the initial weight vector for fusing multi-channel retrieval results. , Represents the retrieval weight of the knowledge graph structure. Represents the semantic retrieval weight of the graph. Represents the weight of knowledge vector retrieval in the domain.

[0072] Session context maintenance: After completing entity recognition and intent classification, the system maintains the current query context based on the query command. The identified key entity set and task type The session context information for the current failure analysis task is recorded and used for referential resolution and semantic completion in subsequent multi-turn dialogues.

[0073] S32, Multi-channel retrieval: Under the constraints of the key entity set and weight vector, a progressive retrieval logic is executed.

[0074] Structured knowledge acquisition: First, determine the entity set. Does the data contain valid entities that can be mapped to a knowledge graph? If so, the knowledge graph structure retrieval channel is triggered first, generating a Cypher query statement, performing a multi-hop traversal across levels, obtaining entity and attribute information from the first-level failure case entity layer, retrieving the associated "failure mode community" in the second level through attribution relationships, and retrieving the corresponding "fault tree analysis" subgraph in the third level through mapping relationships. Let the recalled structured knowledge subgraph be denoted as... .

[0075] Dynamic query expansion: If Not empty, will return the original query and Semantically concatenate the key feature text to obtain the expanded query. and call the BGE-M3 embedding model to Vectorization yields an enhanced query vector. Otherwise, let .

[0076] Parallel generalized retrieval of graph semantics and document channels: Successfully obtained enhanced query vector Subsequently, the semantic retrieval channel of the knowledge graph and the retrieval channel of the domain knowledge vector base are triggered in parallel. Cosine similarity is used to calculate semantic relevance, and its mathematical expression is: ; In the formula, It is the semantic vector of any "case semantic" node in the semantic retrieval channel of the knowledge graph, or the semantic vector of any text fragment in the domain knowledge vector library.

[0077] Based on the above formula, calculate The cosine similarity with the "case semantics" node vector, and the recall similarity exceeding the threshold. The end-to-end failure subgraph, denoted as At the same time, calculation Cosine similarity with document fragments in the FAISS vector library; recall cosine similarity exceeding a threshold. The relevant background document fragments are denoted as .

[0078] Search result fusion and reordering: After the search is completed, the structured knowledge subgraphs retrieved by the knowledge graph structure search channel are first processed. The knowledge graph semantic retrieval channel recalls the full-link failure subgraph. And related background document fragments recalled by the domain knowledge vector base. The data is deduplicated and merged into a candidate knowledge fragment set. Separately, for each channel, the recall results are retained according to their weights, and each channel retains the top results based on preset channel weights. The results yielded a total of [number] results. A candidate set of candidate results is generated. Then, the BGE-Reranker cross-coding model is introduced to process the knowledge fragments within the candidate set. Calculate the original query With each knowledge fragment The semantic relevance score is then re-ranked, and its mathematical expression is: The highest score was ultimately achieved. These fragments construct an enhanced context for reasoning analysis. .

[0079] S33, Result Generation and Adaptive Adjustment: The structured and unstructured knowledge obtained from multi-channel retrieval is fused and reasoned, and the quality of the output results is improved through adaptive adjustment of weight parameters and optimization of large model generation configuration.

[0080] Inference result generation: based on the analysis task type Call the corresponding suggestion word template to expand the query. With enhanced context Fill in the template to construct complete inference instructions. Input the instructions into the large language model to generate initial structured failure analysis results. Specifically, the prompt template includes a pre-set mechanism to enhance the thought chain.

[0081] Evaluation Mechanism for Generation Results: In this embodiment, an evaluator based on a large language model is set up to evaluate the results from five preset dimensions. Scoring is performed to calculate the overall quality score. : ; in, Normalized scores for each dimension (knowledge completeness, semantic consistency, logical coherence, comprehensibility, and hallucination detection) These are the weighting coefficients for each dimension.

[0082] Adaptive Adjustment and Iteration: Setting Quality Thresholds (For example, 0.85). If Output As the final result. Furthermore, if the maximum number of iterations has not been reached, a feedback adjustment mechanism is triggered. This module executes corresponding optimization strategies based on the bottleneck dimension output by the evaluator, including adjusting the weight vector. Alternatively, optimize the large language model generation configuration (e.g., adjust temperature parameters) and re-trigger steps S32 and S33 until the quality requirements are met or the number of iterations is reached, ultimately outputting the analysis result with the highest score. .

[0083] S34, Task Summary: This function is used to globally backtrack and structurally aggregate the multi-round human-computer interaction records, interim analysis results, and intermediate inference information generated during the current failure analysis task when the user chooses to terminate the current task, generating a failure analysis report that matches the analysis task type. Specifically, the failure analysis report has report templates corresponding to different failure analysis tasks.

[0084] S4 uses a knowledge update mechanism to dynamically maintain knowledge related to lithium battery failure cases, ensuring the timeliness and consistency of the knowledge upon which the multi-channel RAG framework relies.

[0085] Specifically, this embodiment dynamically maintains the failure analysis knowledge base by combining incremental updates of internal failure analysis knowledge with periodic synchronization of external multi-source knowledge, including: S41, Internal Failure Analysis Knowledge Incremental Update: When the system receives the user's "Input to Database" instruction for the failure analysis report output in S34, it marks the report as a high-confidence sample. Based on predefined knowledge graph mapping rules, it performs operations such as adding node instances, updating node attributes, and supplementing relationship instances, integrating the newly added knowledge from the failure analysis report into the knowledge graph. Simultaneously, the failure analysis report is treated as new unstructured text, and its corresponding semantic vector is generated according to the text preprocessing and semantic representation methods in step S1, and stored as new domain knowledge in the domain knowledge vector library.

[0086] S42, Periodic Synchronization of External Multi-Source Knowledge: The system presets a periodic synchronization time window and automatically scans external data sources to obtain newly added or updated lithium battery failure-related documents. For the acquired incremental external document set, the text preprocessing and vectorization methods of step S1 are reused to store the generated document semantic vectors in the domain knowledge vector library. At the same time, based on the knowledge graph modeling method of step S2, the domain concepts, failure mechanisms, and technical rules identified in the external documents are structured and modeled. New concepts are introduced into the knowledge graph as incremental nodes, or integrated into existing nodes through node attribute enhancement and entity relationship updates.

[0087] This embodiment also provides a lithium battery failure analysis system based on multi-channel RAG, used to implement the above-mentioned lithium battery failure analysis method based on multi-channel RAG, as follows: The multi-source heterogeneous failure analysis knowledge base construction module is used to build the knowledge foundation to support failure analysis. Specifically, it includes: a semantic modeling unit, which uses a text preprocessing workflow based on a large language model to perform format normalization, semantic slicing, and vectorization modeling on unstructured text data, and builds a domain knowledge vector library; and a graph instantiation unit, which uses a large language model to perform structured information extraction and semantic alignment on failure case reports based on a preset three-layer knowledge graph ontology structure, builds a failure analysis knowledge graph, and generates "case semantics" nodes.

[0088] The intent recognition and task parsing module is used to realize semantic parsing and task modeling of user queries. Specifically, it includes: a query parsing unit, which receives user queries, performs intent classification and entity element extraction, and configures the initial weight vector for multi-channel retrieval based on the identified analysis task type; and a context maintenance unit, which performs referential resolution, context information completion, and semantic consistency verification during multiple rounds of interaction.

[0089] The multi-channel retrieval module is used to realize the retrieval, recall, and failure analysis reasoning of multi-source heterogeneous knowledge. Specifically, it includes: a multi-channel retrieval unit, which calls the knowledge graph structure retrieval channel, the knowledge graph semantic retrieval channel, and the domain knowledge vector library retrieval channel respectively based on the user query, and uses a progressive retrieval strategy to obtain relevant heterogeneous knowledge fragments; and a fusion and reordering unit, which performs deduplication and reordering of the knowledge fragments recalled by each retrieval channel to construct enhanced contextual information.

[0090] The results generation and adaptive adjustment module is used to perform quality assessment on the generated failure analysis results and to adaptively adjust the system based on the assessment results. Specifically, it includes: an enhanced reasoning unit, which calls the large language model and generates structured lithium battery failure analysis results by combining enhanced contextual information under the constraints of the thought chain prompt word template; a multi-dimensional evaluation unit, which performs quantitative scoring on the analysis results using a preset evaluation mechanism; and a feedback control unit, which triggers feedback control when the quality score does not reach a preset threshold, driving the system to re-execute the knowledge retrieval and analysis reasoning process by adjusting the multi-channel retrieval weight vector or optimizing the generation constraints of the large language model.

[0091] The task summary module is used to globally summarize and output the results of the current failure analysis task. Specifically, it includes: a result summary unit, which integrates and processes the search results and inference results generated during multiple rounds of interaction; and a report generation unit, which generates a structured failure analysis report based on the analysis task type.

[0092] The domain knowledge dynamic update module is used to continuously update and expand failure analysis knowledge. Specifically, it includes: an internal experience accumulation unit, which takes the summarized failure analysis report as unstructured knowledge input, performs vectorized modeling, and stores it in the domain knowledge vector library. At the same time, it extracts the structured conclusions in the report and updates them synchronously to the failure analysis knowledge graph; and an external synchronization unit, which periodically scans external knowledge sources, automatically reuses modeling methods to process incremental knowledge, and synchronously updates it to the knowledge base.

[0093] Working Principle: The lithium battery failure analysis method and system based on multi-channel RAG provided by this invention operates on the following core principles: First, by constructing a multi-source heterogeneous knowledge system comprising a domain knowledge vector library and a multi-level knowledge graph, the system unifies and structures the modeling of dispersed, unstructured failure analysis knowledge. Second, upon receiving a user query, the system understands the user's intent through the intent recognition module and utilizes the multi-channel retrieval module to collaboratively invoke the structured retrieval capabilities of the knowledge graph (capturing causality and hierarchical logic) and the semantic generalization retrieval capabilities of the vector library, efficiently and comprehensively recalling relevant knowledge. Then, the result generation module combines the retrieved enhanced context with the large language model to perform causal reasoning with thought chains, generating interpretable failure analysis results, and continuously optimizing output quality through an adaptive evaluation feedback mechanism. Finally, through internal and external knowledge update mechanisms, newly generated analysis conclusions and newly acquired external knowledge are dynamically integrated into the knowledge base, enabling the system's self-evolution.

[0094] Therefore, the present invention adopts a lithium battery failure analysis method and system based on multi-channel RAG with the above-mentioned structure. By constructing a multi-source heterogeneous knowledge system, designing a multi-channel collaborative retrieval mechanism, introducing adaptive evaluation and feedback adjustment, and establishing a dynamic knowledge update mechanism, it effectively solves the problems of low utilization rate of professional knowledge, fragmented retrieval results, and lack of structured logical support in the analysis process in the prior art, and realizes the high efficiency, intelligence and enhanced interpretability of lithium battery failure analysis.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-channel RAG-based lithium battery failure analysis method, characterized in that, Includes the following steps: S1 collects, cleans, and performs semantic-level preprocessing on multi-source data and domain knowledge related to lithium battery failure, and performs vectorized modeling to build a domain knowledge vector library for subsequent failure analysis. S2, based on the knowledge graph modeling method, automatically extracts and structures key information from lithium battery failure case reports, constructs a multi-level lithium battery failure analysis knowledge graph, and performs semantic enhancement processing on the knowledge graph; S3, construct a multi-channel retrieval-enhanced failure analysis framework, and obtain knowledge related to the failure analysis task through a multi-channel retrieval mechanism based on the semantic parsing results of the user query; The system calls upon a large language model to perform inference and generate failure analysis results, and adaptively adjusts the retrieval strategy and inference generation process. At the end of the task, the system summarizes the failure analysis task and generates a corresponding failure analysis report. S4 dynamically maintains knowledge related to lithium battery failure through internal and external knowledge updates, in order to keep the timeliness and consistency of the knowledge on which the multi-channel RAG framework depends.

2. The multi-pass RAG-based lithium battery failure analysis method of claim 1, wherein, The method for constructing the multi-level lithium battery failure analysis knowledge graph in step S2 includes: Based on a large language model, entity extraction and relation recognition are performed on semi-structured failure case reports, and a failure case entity layer including nodes such as "failure case", "failed cell", "failure event", "failure cause" and "improvement measures" is instantiated. Based on graph structure similarity and attribute semantic similarity, failure cases in the failure case entity layer are clustered to generate "failure mode community" nodes and construct the failure mode community layer. The fault tree analysis structure is extracted using a large language model, and "top event", "intermediate event", "bottom event" nodes and logical gate relationship edges are instantiated to construct the fault tree analysis logic layer; It also defines an inter-layer association mechanism that assigns "failure case" nodes to "failure mode community" nodes and associates "failure mode community" nodes or "failure event" nodes to "top event" nodes.

3. The multi-pass RAG-based lithium battery failure analysis method of claim 1, wherein, Step S3, which involves multi-channel retrieval enhancement to generate a failure analysis framework, specifically includes: S31, Intent Recognition and Task Parsing, is used to perform semantic understanding and intent recognition on the natural language query input by the user, identify the key entity elements of lithium battery failure involved in the query, map the query to a predefined failure analysis task type, and determine the initial weight parameters for subsequent multi-channel retrieval result fusion based on the task type. S32, multi-channel retrieval, is used to progressively retrieve multi-source knowledge related to failure analysis under the control of query parsing results. Multi-channel retrieval includes knowledge graph structure retrieval channel, knowledge graph semantic retrieval channel and domain knowledge vector library retrieval channel; S33, Result Generation and Adaptive Adjustment, is used to generate inference results of failure analysis based on multi-channel retrieval results, and to achieve adaptive adjustment through the generation result evaluation mechanism and feedback mechanism; S34, Task Summary, is used to perform global backtracking and information aggregation on the multi-round interaction records and phased analysis results generated during the current failure analysis task when the current failure analysis task is terminated, and to generate a failure analysis report that is compatible with the analysis task type.

4. The lithium battery failure analysis method based on multi-channel RAG according to claim 3, characterized in that, Step S32 employs a progressive multi-channel retrieval strategy, specifically including: First, based on the semantic parsing results of the user query, check whether there are key entities that can be mapped to the failure case knowledge graph. If so, the knowledge graph structure retrieval channel is triggered first to obtain the failure case subgraph directly related to the query target. Subsequently, an enhanced query vector is constructed based on the results of the previous step: if the knowledge graph structure retrieval channel is triggered, the original query is fused with the structured feature information in the failure case subgraph to construct an enhanced query vector; otherwise, the enhanced query vector is equal to the original query vector. Then, enhanced query vectors are used to perform similarity searches on the "case semantics" node in the knowledge graph semantic retrieval channel and the domain knowledge vector library retrieval channel, respectively. Finally, the search results of the three channels are respectively de-duplicated, and the first N results are respectively retained according to the preset channel weights i of each channel to obtain a candidate set containing N candidate results in total; subsequently, a reordering model is introduced to score and sort the candidate results in terms of semantic relevance, and the first M optimal search results are intercepted to construct an enhanced context for reasoning analysis.

5. The lithium battery failure analysis method based on multi-channel RAG according to claim 3, characterized in that, The specific execution of the result generation and adaptive adjustment in step S33 is as follows: The user commands and the enhanced context obtained from multi-channel retrieval are filled into the preset prompt word template to construct inference commands and input into the large language model for inference analysis, generating the initial structured failure analysis results; The failure analysis results are evaluated by a result evaluation mechanism. When the evaluation results do not meet the preset quality requirements, the fusion weights of the multi-channel retrieval results and the constraints of the large model inference generation are adaptively adjusted based on the evaluation results, and the retrieval result fusion and failure analysis result generation process is re-executed. When the evaluation results meet the quality requirements or reach the preset number of iterations, the current optimal failure analysis result is output, and the initial weights under the corresponding task type are updated according to the optimal weight coefficients.

6. The lithium battery failure analysis method based on multi-channel RAG according to claim 5, characterized in that, The generated results evaluation mechanism is used to score the failure analysis results from multiple preset evaluation dimensions, including knowledge completeness, semantic consistency, logical relevance, comprehensibility, and hallucination detection.

7. The lithium battery failure analysis method based on multi-channel RAG according to claim 1, characterized in that, Step S4 specifically includes: S41, Internal failure analysis knowledge incremental update, transforming the failure analysis report output in step S34 into reusable domain knowledge, and updating it to the domain knowledge vector library and failure analysis knowledge graph respectively according to the knowledge form; S42, External Multi-Source Knowledge Periodic Synchronization, is used to periodically introduce external lithium battery failure-related knowledge. The external knowledge is processed through the text preprocessing in step S1 and the knowledge graph modeling method in step S2, and then updated to the corresponding failure analysis knowledge base.

8. A lithium battery failure analysis system based on multi-channel RAG, used to implement the method as described in any one of claims 1-7, characterized in that, include: The multi-source heterogeneous failure analysis knowledge base construction module is configured to perform format standardization and semantic segmentation of multi-source text data through a large language model preprocessing workflow, and generate a domain knowledge vector library using a pre-trained vector encoding model; at the same time, it is configured to construct a multi-level failure analysis knowledge graph based on a preset three-layer ontology structure through knowledge extraction and logical aggregation, and generate "case semantics" nodes through cross-level path traversal. The intent recognition and task parsing module is configured to use a large language model to identify key elements from user queries, map query tasks to predefined analysis task types, and initialize the fusion weights of multi-channel retrieval based on task types. The multi-channel retrieval module is configured to perform multi-channel retrieval based on user queries, including knowledge graph structure retrieval channel, knowledge graph semantic retrieval channel, and domain knowledge vector library retrieval channel. It also reorders the recalled knowledge and constructs an enhanced context for analysis and reasoning. The results generation and adaptive adjustment module is configured to generate failure analysis results based on the enhanced context call of the large language model, and to adaptively adjust the retrieval strategy and inference generation process according to the quality assessment of the generated results. The task summary module is configured to summarize and perform a global analysis of the multi-round interaction records and reasoning results generated during the task when the current failure analysis task terminates, and generate a structured failure analysis report. The domain knowledge dynamic update module is configured to standardize the failure analysis results that have passed the quality assessment and update them to the failure analysis knowledge base. It also periodically scans external multi-source data, processes new knowledge, and updates it to the failure analysis knowledge base synchronously.

9. A lithium battery failure analysis system based on multi-channel RAG according to claim 8, characterized in that, The multi-channel retrieval module is configured to execute the following progressive retrieval strategy: First, based on the output of the intent recognition and task parsing module, check whether there are key entities that can be mapped to the knowledge graph. If so, the knowledge graph structure retrieval channel is triggered first. Subsequently, the original query is fused with the retrieved structured information to construct an enhanced query vector; Then, the enhanced query vectors are used to trigger the semantic retrieval channel of the knowledge graph and the retrieval channel of the domain knowledge vector library in parallel; Finally, the search results from each channel are weighted, merged, and reordered to construct an enhanced context.

10. A lithium battery failure analysis system based on multi-channel RAG according to claim 8, characterized in that, The domain knowledge dynamic update module includes: The internal experience accumulation unit is used to take the verified failure analysis report as unstructured knowledge input, perform vectorized modeling and store it in the domain knowledge vector library. At the same time, the structured conclusions in the report are extracted and synchronously updated to the failure analysis knowledge graph by adding nodes or updating attributes. The external synchronization unit is used to periodically scan external knowledge sources, automatically reuse the modeling methods of the multi-source heterogeneous failure analysis knowledge base construction module to process incremental knowledge, and synchronously update it to the domain knowledge vector library and failure analysis knowledge graph.

Images