Industrial production line anomaly detection and root cause tracing method based on multi-modal large model

Abstract

The application discloses a multi-modal large model-based industrial production line anomaly detection and root cause tracing method, and belongs to the technical field of industrial intelligent detection and data processing. The method comprises the following steps: preprocessing multi-modal data frames of an industrial production line to generate unified structured input data; extracting abnormal feature representations based on the unified structured input data, and constructing modal anomaly descriptors based on the abnormal feature representations; based on production process correlation, cross-process correlation organization is performed on the modal anomaly descriptors to generate correlated abnormal context; a pre-trained multi-modal anomaly reasoning model is called to jointly reason the correlated abnormal context and the modal anomaly descriptors, and abnormal type classification results, abnormal evidence weight distribution and candidate abnormal trigger nodes are obtained; reverse time sequence tracing analysis is performed from the candidate abnormal trigger nodes as the starting point, and combined with counterfactual verification, an abnormal root cause node and an abnormal propagation path are determined.

Description

Technical Field

[0001] This application relates to the field of industrial intelligent detection and data processing technology, and more specifically, to a method for anomaly detection and root cause tracing of industrial production lines based on a multimodal large model. Background Technology

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, industrial production lines are gradually evolving towards high automation and complex collaboration. In actual operation, production systems typically involve multi-source heterogeneous data such as vision, vibration, audio, temperature, and current. These different data have strong coupling and temporal correlations, making the industrial process state exhibit high-dimensionality, strong dynamics, and complex nonlinear characteristics. Traditional anomaly detection methods based on single sensors or fixed thresholds are difficult to effectively characterize the correlation between multimodal data and are prone to missed or false detections under complex operating conditions.

[0003] While existing anomaly detection methods based on machine learning or deep learning have improved detection capabilities to some extent, most still rely on single-modal modeling or simple feature concatenation, lacking a unified ability to express cross-modal semantic information and temporal dependencies. Furthermore, industrial anomalies often propagate across equipment and processes; existing methods typically only output anomaly results, failing to further characterize the propagation path and triggering sources of anomalies within the production process, thus limiting subsequent decision support capabilities.

[0004] Furthermore, in root cause analysis, existing technologies largely rely on statistical correlation analysis or empirical rule inference, lacking effective causal reasoning mechanisms. This results in insufficient stability and weak interpretability of root cause localization results. In industrial environments with multiple coupled processes, abnormal signals may be triggered by the superposition of multiple factors. Existing methods struggle to distinguish between causal and correlational relationships, thus affecting the reliability of source tracing analysis.

[0005] In recent years, although multimodal data analysis methods based on large models have been gradually introduced into the industrial field, they still have shortcomings in modeling complex process structures, analyzing cross-process anomaly correlations, and inferring causal consistency. A unified analysis framework suitable for complex industrial scenarios has not yet been formed.

[0006] In summary, improving the accuracy of anomaly identification in multi-source heterogeneous industrial data, realizing cross-process anomaly correlation modeling, and enhancing the accuracy and interpretability of anomaly root cause localization and propagation path analysis have become urgent technical problems to be solved. Summary of the Invention

[0007] To overcome a series of shortcomings in existing technologies, the purpose of this application is to provide a method for anomaly detection and root cause tracing in industrial production lines based on a multimodal large model, comprising the following steps: Preprocess the multimodal data frames from the industrial production line to generate unified structured input data; Anomaly feature representations are extracted based on unified structured input data, and modal anomaly descriptors are constructed based on these anomaly feature representations. Based on the production process relationships, modal anomaly descriptors are organized across processes to generate associated anomaly contexts. The pre-trained multimodal anomaly reasoning model is invoked to perform joint reasoning on the associated anomaly context and modal anomaly descriptors, resulting in anomaly type classification results, anomaly evidence weight distribution, and candidate anomaly triggering nodes. Starting from the candidate anomaly trigger node, perform reverse time-series source analysis and combine it with counterfactual verification to determine the root cause node and anomaly propagation path.

[0008] In some embodiments, the method for preprocessing multimodal data frames from an industrial production line is as follows: The system acquires visual data, vibration data, audio data, temperature data, and current data collected from each process node of the industrial production line, and performs timestamp synchronization processing to obtain time-aligned multimodal data. Denoising is performed on the time-aligned modal data to obtain the corresponding cleaned data; The purified data is segmented into frames based on preset frame length and frame shift parameters to construct a multimodal data frame sequence with a unified time window index. Missing or anomalous data in a multimodal data frame sequence is completed to obtain the completed multimodal data frame and the corresponding data quality confidence information; The completed multimodal data frames are associated with and stored with the corresponding data quality confidence information to generate preprocessed structured multimodal data frames.

[0009] In some embodiments, the method for extracting abnormal feature representations is as follows: Reconstruction error analysis is performed on visual data in unified structured input data to identify abnormal regions in the image and extract the corresponding visual abnormality feature vectors. Spectral deviation analysis is performed on vibration data in the unified structured input data to determine the vibration anomaly state and extract the corresponding vibration anomaly feature vector; Acoustic feature deviation analysis is performed on audio data in unified structured input data to determine audio anomalies and extract corresponding audio anomaly feature vectors. Time-series variation analysis is performed on temperature and current data in the unified structured input data, and trend change features, abrupt change features, and limit exceedance persistence features are extracted to obtain the corresponding scalar anomaly feature vectors. Cross-modal fusion processing is performed on visual anomaly feature vectors, vibration anomaly feature vectors, audio anomaly feature vectors, and scalar anomaly feature vectors to obtain fused anomaly features; Anomaly description results are constructed based on fused anomaly features, resulting in anomaly feature representations that include structured semantic information and numerical representation information.

[0010] In some embodiments, the method for generating the associated exception context is as follows: Based on the process node identifiers corresponding to each modal anomaly descriptor in the current data frame, determine the set of abnormal process nodes; Based on the set of abnormal process nodes, the corresponding upstream and downstream process nodes and process relationships are retrieved from the industrial knowledge graph; Based on the process association, obtain the modal anomaly descriptors corresponding to the upstream and downstream process nodes in the current data frame and the preset historical time window; Based on the process relationships and the corresponding modal anomaly descriptors, identify the abnormal process nodes with relationships and construct the associated anomaly group; Based on the propagation delay information corresponding to the process association, the temporal relationship of abnormal events in the associated abnormal group is verified. The abnormal association type of the associated abnormal group is determined based on the time sequence relationship verification results; Construct an associated anomaly context based on the associated anomaly group, the anomaly association type, and the corresponding process topology information.

[0011] In some embodiments, the method for performing reverse temporal source tracing analysis starting from a candidate anomaly triggering node is as follows: Obtain candidate exception triggering nodes and determine them as the current backtracking nodes; Obtain the upstream process node corresponding to the current backtracking node, and determine the process node to be backtracked based on the historical causal strength information corresponding to the process association relationship; Based on the first occurrence time of the anomaly at the current backtracking node and the propagation delay information of the corresponding process relationship, the anomaly search time interval is determined; Obtain the corresponding multimodal data frames within the anomaly search time interval and extract the corresponding anomaly features; Based on the abnormal characteristics, determine whether there are any abnormal signals in the process node to be traced that occurred earlier than the current traced node; If it exists, update the process node to be traced back to the current traceback node; If it does not exist, or if the backtracking level reaches the preset level, the current backtracking node will be identified as the root cause node to be verified. The process nodes involved in the backtracking process and the corresponding first occurrence time of the anomaly are associated and recorded to form an anomaly propagation path, and corresponding anomaly propagation path data is generated based on the anomaly propagation path.

[0012] In some embodiments, the method for determining the root cause node of anomalies by combining counterfactual verification is as follows: Obtain the root cause node to be verified and its corresponding anomaly propagation path; Counterfactual intervention hypotheses are constructed for each root cause node to be verified, and the probability of occurrence of abnormal events at each downstream node under counterfactual conditions is calculated by combining abnormal time series information and process correlation in the industrial knowledge graph. Based on the probability of occurrence, calculate the counterfactual causal inference score and causal consistency score for each root cause node to be verified. Based on the causal consistency score of each root cause node to be verified, the abnormal root cause nodes are identified. When the causal consistency scores of multiple root cause nodes to be verified meet the preset association conditions, the multiple root cause nodes to be verified are identified as joint root cause nodes, and their corresponding contribution weights are determined.

[0013] In some embodiments, the method for determining anomaly propagation paths using counterfactual verification is as follows: Obtain the set of abnormal propagation process nodes and process relationships corresponding to the abnormal root cause node; Based on the set of abnormal propagation process nodes and propagation relationships, a directed abnormal propagation path graph is constructed, and corresponding attributes are configured for each node and propagation relationship. Based on the propagation characteristics of each node during the abnormal propagation process, the corresponding node role type is determined; Based on node attributes, edge attributes, and counterfactual causal inference results corresponding to each node, the overall confidence score of the abnormal propagation path is calculated. Based on the directed anomaly propagation path diagram and the overall confidence score, anomaly propagation path data is generated and associated with anomaly event records in the industrial knowledge base.

[0014] In some embodiments, the industrial knowledge base adopts a hybrid storage architecture that combines relational databases and graph databases. The relational database is used to store structured abnormal event records, equipment ledger information, process parameter specifications, and sensor baseline statistical data, while the graph database is used to store node data and edge relationship data in the industrial knowledge graph.

[0015] In some embodiments, the industrial production line anomaly detection and root cause analysis method further includes: During the reverse time-series source tracing analysis, based on the anomaly propagation path and real-time acquired multimodal data frames, it is determined whether the preset anomaly elimination conditions are met. If not met, continue reverse time-series source analysis and real-time monitoring until the anomaly elimination conditions are met; If the conditions are met, the reverse time-series source analysis process ends, and an interpretable source report is generated based on the abnormal root cause node and abnormal propagation path. The explainable traceability report is written back to the industrial knowledge base, and the multimodal anomaly inference model is incrementally updated based on the updated industrial knowledge base.

[0016] In some embodiments, the method for generating an explainable source tracing report is as follows: Obtain information on the root cause nodes of anomalies, the directed anomaly propagation path diagram, the anomaly type classification results, and the anomaly evidence weight distribution data, and perform unified structured processing to generate report input data; The report input data is input into the report generation language model, and combined with a predefined report template to generate an anomaly event summary, root cause analysis content, and anomaly propagation path content; Based on key modal anomaly descriptors, anomaly intensity quantification values, and historical records of similar anomaly events, a multimodal evidence list and handling recommendations are generated. The generated report content is linked, integrated, and uniformly arranged to construct a structured, interpretable, and traceable report; Output the PDF and JSON format files corresponding to the structured and interpretable traceability report.

[0017] Compared with the prior art, this application has the following beneficial effects: This application achieves cross-modal accurate identification of production line anomalies, cross-process correlation analysis, and high-confidence interpretable tracing of root cause nodes and propagation paths by integrating multimodal large models with industrial process topology and counterfactual causal inference, thereby significantly improving the accuracy of industrial anomaly detection and the efficiency of root cause localization. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the industrial production line anomaly detection and root cause tracing method based on a multimodal large model disclosed in this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some embodiments of this invention, but not all embodiments.

[0020] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The embodiments and directional terms described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0022] like Figure 1 As shown, the method for anomaly detection and root cause analysis of industrial production lines based on a multimodal large model includes the following steps: Preprocess the multimodal data frames from the industrial production line to generate unified structured input data; Anomaly feature representations are extracted based on unified structured input data, and modal anomaly descriptors are constructed based on these anomaly feature representations. Based on the production process relationships, modal anomaly descriptors are organized across processes to generate associated anomaly contexts. The pre-trained multimodal anomaly reasoning model is invoked to perform joint reasoning on the associated anomaly context and modal anomaly descriptors, resulting in anomaly type classification results, anomaly evidence weight distribution, and candidate anomaly triggering nodes. Starting from the candidate anomaly triggering node, perform reverse time-series source tracing analysis, and combine counterfactual verification to determine the root cause node and anomaly propagation path; During the reverse time-series source tracing analysis, based on the anomaly propagation path and real-time acquired multimodal data frames, it is determined whether the preset anomaly elimination conditions are met. If not met, continue reverse time-series source analysis and real-time monitoring until the anomaly elimination conditions are met; If the conditions are met, the reverse time-series source analysis process ends, and an interpretable source report is generated based on the abnormal root cause node and abnormal propagation path. The explainable traceability report is written back to the industrial knowledge base, and the multimodal anomaly inference model is incrementally updated based on the updated industrial knowledge base.

[0023] The aforementioned method for anomaly detection and root cause tracing in industrial production lines based on a multimodal large model ensures the consistency and integrity of subsequent model input data through unified preprocessing and structured encapsulation of multi-source heterogeneous data. It achieves accurate representation of cross-modal anomaly information through multimodal anomaly feature extraction and modal anomaly descriptor construction. Furthermore, it effectively locates candidate anomaly triggering nodes by combining process correlations with joint reasoning of the large model. Further, it accurately determines the root cause and propagation path of anomalies by combining reverse temporal tracing and counterfactual verification mechanisms. Simultaneously, it continuously improves the model's detection accuracy and root cause tracing capabilities in the target process scenario through dynamic anomaly elimination judgment and closed-loop knowledge base updates. This significantly enhances the accuracy, interpretability, and adaptive evolution capabilities of anomaly detection in industrial production lines, meeting the high-reliability intelligent operation and maintenance requirements in complex industrial scenarios.

[0024] In some embodiments, the method for preprocessing multimodal data frames from an industrial production line is as follows: The system acquires visual data, vibration data, audio data, temperature data, and current data collected from each process node of the industrial production line, and performs timestamp synchronization processing to obtain time-aligned multimodal data. Denoising is performed on the time-aligned modal data to obtain the corresponding cleaned data; The purified data is segmented into frames based on preset frame length and frame shift parameters to construct a multimodal data frame sequence with a unified time window index. Missing or anomalous data in a multimodal data frame sequence is completed to obtain the completed multimodal data frame and the corresponding data quality confidence information; The completed multimodal data frames are associated with and stored with the corresponding data quality confidence information to generate preprocessed structured multimodal data frames.

[0025] The multimodal data frame preprocessing method described in this application eliminates temporal discrepancies in the acquisition times of each modality through timestamp synchronization processing, ensuring precise alignment of multimodal data in the time dimension; it suppresses random noise interference introduced during sensor acquisition through denoising processing, improving the data signal-to-noise ratio; it constructs a multimodal data frame sequence with a consistent time window index through unified frame segmentation processing, providing standardized input for subsequent cross-modal feature extraction; and it quantifies the credibility of each data frame while ensuring data integrity through missing and anomalous data completion and data quality confidence labeling, thereby significantly improving the quality consistency of multimodal input data and the reliability of subsequent anomaly detection.

[0026] In some embodiments, the unified structured input data is encapsulated using a unified tensor format; wherein visual data is encoded as a spatial feature tensor, vibration data and audio data are encoded as time-frequency feature tensors, temperature data and current data are encoded as time series vectors, and process identifiers, equipment numbers, sampling timestamps and data quality confidence information are appended to form corresponding modal data packets; The data packets of each modality are organized according to the topological order of the production process to construct a multimodal frame object. The multimodal frame object includes a frame-level header information area, a modal data area, and a process association description area. The process association description area is used to store the material flow direction relationship and energy transfer relationship between process nodes.

[0027] The unified structured input data encapsulation method described in this application achieves format unification while preserving the original information of each modality by employing the most suitable tensor encoding method for different modal data types, thereby reducing the processing complexity of heterogeneous data access. By adding meta-information such as process identifiers, equipment numbers, timestamps, and data quality confidence levels, it provides complete process context support for subsequent model inference. By introducing a process association description area into the multimodal frame object, it embeds material flow and energy transfer relationships into the data structure, thereby significantly improving the multimodal anomaly inference model's ability to understand process semantics and providing structured prior evidence for cross-process association analysis.

[0028] In some embodiments, the method for extracting abnormal feature representations is as follows: Reconstruction error analysis is performed on visual data in unified structured input data to identify abnormal regions in the image and extract the corresponding visual abnormality feature vectors. Spectral deviation analysis is performed on vibration data in the unified structured input data to determine the vibration anomaly state and extract the corresponding vibration anomaly feature vector; Acoustic feature deviation analysis is performed on audio data in unified structured input data to determine audio anomalies and extract corresponding audio anomaly feature vectors. Time-series variation analysis is performed on temperature and current data in the unified structured input data, and trend change features, abrupt change features, and limit exceedance persistence features are extracted to obtain the corresponding scalar anomaly feature vectors. Cross-modal fusion processing is performed on visual anomaly feature vectors, vibration anomaly feature vectors, audio anomaly feature vectors, and scalar anomaly feature vectors to obtain fused anomaly features; Anomaly description results are constructed based on fused anomaly features, resulting in anomaly feature representations that include structured semantic information and numerical representation information.

[0029] The above-mentioned abnormal feature representation extraction method of this application effectively improves the detection sensitivity of each single-modal abnormal signal by adopting anomaly analysis methods that match the physical characteristics of each modality for visual, vibration, audio and scalar data; it achieves complementary enhancement of multimodal abnormal information through cross-modal fusion processing, making up for the perception limitations of a single modality in specific abnormal scenarios; and it significantly improves the semantic understanding ability and quantitative analysis accuracy of subsequent anomaly type discrimination and root cause tracing by encoding the fused abnormal features into anomaly feature representations that simultaneously contain structured semantic information and numerical representation information.

[0030] In some embodiments, the modal anomaly descriptor is a structured data object, including an anomaly semantic description field, an anomaly intensity quantification field, an anomaly timestamp field, a process node identifier field, and a modal confidence score field. Specifically: the anomaly semantic description field is constructed based on a predefined industrial anomaly ontology terminology library and is used to describe the anomaly detected in the corresponding modality; the anomaly intensity quantification field stores the Mahalanobis distance between the corresponding anomaly feature vector and the normal operating condition baseline to characterize the degree of anomaly; the anomaly timestamp field records the time of occurrence of the anomaly event; the process node identifier field identifies the process node to which the anomaly belongs; and the modal confidence score field is calculated jointly based on the data quality confidence score and the corresponding sensor's historical failure rate, and is used to characterize the confidence level of the current modal anomaly information.

[0031] The modal anomaly descriptor substructure design described in this application achieves standardized semantic expression of anomalies by constructing anomaly semantic description fields based on an industrial anomaly ontology terminology library, eliminating ambiguity in descriptive terms under different process scenarios; it quantifies anomaly intensity by Mahalanobis distance values, effectively eliminating the influence of differences in the dimensions of various feature dimensions and improving the objectivity of anomaly degree assessment; and it calculates modal credibility scores by jointly using data quality confidence and sensor historical failure rate, enabling the inference model to adaptively reduce the influence weight of low-confidence modalities, thereby significantly improving the robustness of anomaly inference results under sensor failure or data quality degradation scenarios.

[0032] In some embodiments, the production process relationships are described using an industrial knowledge graph and stored in an industrial knowledge base; wherein, knowledge graph nodes correspond to process units and key equipment in the production line, and knowledge graph edges correspond to material flow relationships, energy flow relationships and control signal flow relationships between processes; The edge attributes of the knowledge graph include dependency type, propagation delay estimate, and historical causal strength coefficient; the historical causal strength coefficient is obtained based on the statistical analysis of upstream and downstream anomaly time-series correlations in historical anomaly events and is dynamically updated according to the anomaly tracing results; When performing cross-process correlation analysis, the corresponding K-hop neighborhood subgraph is retrieved based on the process node identification information in the modal anomaly descriptor, and the neighborhood subgraph structure and edge attribute information are input into the multimodal anomaly inference model as the process topology context.

[0033] The industrial knowledge graph construction and application method described in this application unifies the process relationships of material flow, energy flow, and control signal flow into a graph structure, fully expressing the multidimensional dependencies between processes; by introducing propagation delay estimates and historical causal strength coefficients into the graph edge attributes, it provides quantitative prior knowledge for cross-process anomaly timing verification and root cause tracing; by dynamically updating the historical causal strength coefficients, it achieves continuous self-optimization of the knowledge graph as operational data accumulates; and by injecting local process topology context into the inference model through K-hop neighborhood subgraph retrieval, it significantly improves the accuracy of causal inference in complex multi-process coupled scenarios for anomaly root cause tracing.

[0034] In some embodiments, the method for generating the associated exception context is as follows: Based on the process node identifiers corresponding to each modal anomaly descriptor in the current data frame, determine the set of abnormal process nodes; Based on the set of abnormal process nodes, the corresponding upstream and downstream process nodes and process relationships are retrieved from the industrial knowledge graph; Based on the process association, obtain the modal anomaly descriptors corresponding to the upstream and downstream process nodes in the current data frame and the preset historical time window; Based on the process relationships and the corresponding modal anomaly descriptors, identify the abnormal process nodes with relationships and construct the associated anomaly group; Based on the propagation delay information corresponding to the process association, the temporal relationship of abnormal events in the associated abnormal group is verified. The abnormal association type of the associated abnormal group is determined based on the time sequence relationship verification results; Construct an associated anomaly context based on the associated anomaly group, the anomaly association type, and the corresponding process topology information.

[0035] The aforementioned method for generating context for associated anomalies in this application expands a single-point anomaly signal into a complete set of anomaly events covering upstream and downstream processes by combining knowledge graph retrieval and historical time window expansion, effectively capturing the spatiotemporal correlation information of cross-process anomaly propagation; by introducing propagation delay information to verify the temporal relationship of anomaly events, it filters out false correlations with inconsistent temporal sequences, improving the accuracy of constructing associated anomaly groups; by classifying associated anomaly groups and integrating process topology information, it provides a structured context containing causal semantics for the inference model, thereby significantly improving the root cause discrimination capability of the multimodal anomaly inference model in scenarios of chained anomalies and concurrent anomalies.

[0036] In some embodiments, the pre-trained multimodal anomaly reasoning model uses a large-scale pre-trained language model as the backbone network, and adds a cross-modal cross-attention module on the basis of the self-attention layer of the backbone network. The cross-modal attention module uses the semantic embedding corresponding to the context of the associated anomaly as the query vector, and the result of linear mapping of the numerical feature vector corresponding to the modal anomaly descriptor as the key vector and value vector. It performs attention retrieval and weighted aggregation on the numerical feature information to realize cross-modal association modeling between semantic information and multimodal anomaly evidence. The output layer of the multimodal anomaly inference model includes an anomaly type decoding head, an evidence weight decoding head, and a root cause node decoding head. The anomaly type decoding head is used to output the category probability distribution corresponding to the anomaly type, the evidence weight decoding head is used to output the evidence weight corresponding to each modal anomaly descriptor, and the root cause node decoding head is used to output the candidate anomaly triggering nodes and their corresponding confidence scores. The pre-training data for the multimodal anomaly reasoning model is constructed from historical anomaly event records in the industrial knowledge base. The training samples include anomaly type labels, root cause process node labels, and modal evidence importance labels. After pre-training, a parameter-efficient fine-tuning method based on low-rank decomposition is used to adapt the backbone network to the domain. Specifically, low-rank adaptation parameters are configured on the query matrix and numerical matrix of each attention layer, only the low-rank adaptation parameters are updated, and the backbone network parameters are kept frozen. The fine-tuning training data uses historical abnormal event data corresponding to the target production line to achieve adaptation training for the target process scenario.

[0037] The multimodal anomaly reasoning model architecture and training method described in this application achieve deep fusion between semantic context information and multimodal numerical anomaly evidence by introducing a cross-modal cross-attention module on the basis of a pre-trained language model backbone network, thus fully leveraging the advantages of large models in semantic reasoning. By setting three independent decoding heads—anomaly type, evidence weight, and root cause node—joint optimization of classification, attribution, and localization is achieved, providing structured output support for the interpretability of the results. By introducing an efficient fine-tuning method for low-rank decomposition parameters, the computational cost of domain-adaptive training is significantly reduced while retaining the general semantic reasoning capabilities of the backbone network, thereby significantly improving the anomaly detection accuracy and cross-scene transfer adaptation efficiency of the model in the target process scenario.

[0038] In some embodiments, the anomaly type classification results are represented by the leaf node category labels in the industrial anomaly ontology; the industrial anomaly ontology is constructed according to the hierarchical structure of "anomaly category - anomaly subcategory - specific anomaly pattern", wherein the anomaly category includes equipment mechanical failure, process parameter deviation, raw material quality defect and external environmental interference. Each anomaly classification result corresponds to a category confidence score and an uncertainty estimate; the uncertainty estimate is calculated based on the category probability distribution output by the model and is used to characterize the reliability of the classification result. For pre-defined high-safety-level anomaly types, temperature scaling calibration is performed on the category probability distribution before calculating the uncertainty estimate to improve the consistency of the probability output. The multimodal anomaly reasoning model simultaneously outputs anomaly evidence weight distribution; the anomaly evidence weight distribution is represented by a weight vector that corresponds one-to-one with the input modal anomaly descriptor. After normalization, the sum of each weight value is 1, where the weight value is used to characterize the degree of contribution of the corresponding modal anomaly descriptor to the anomaly type determination result. The weight distribution of abnormal evidence is used to generate information on the importance of evidence in the results of anomaly tracing, and serves as the data basis for subsequent anomaly analysis and result interpretation. When the category confidence score is lower than the preset threshold, a manual review label is generated and stored in the industrial knowledge base along with the anomaly analysis results.

[0039] The method for representing and outputting the above-mentioned anomaly type classification results in this application achieves fine-grained and semantically standardized anomaly type descriptions by using leaf node labels of a hierarchical industrial anomaly ontology to represent the classification results, facilitating integration with industrial operation and maintenance knowledge systems. It quantifies the model's classification confidence level through uncertainty estimates, providing a basis for human-machine collaborative decision-making for high-safety-level anomalies. It improves the consistency and reliability of probability output by introducing temperature scaling calibration for high-safety-level anomaly types. It achieves interpretability and traceability of classification results through normalized anomaly evidence weight distribution output. Furthermore, it effectively avoids the safety risks of autonomous decision-making by the model in low-determinism scenarios through a manual review and labeling mechanism in low-confidence scenarios, thereby significantly improving the refinement, credibility, and production safety assurance capabilities of the anomaly classification results.

[0040] In some embodiments, the multimodal anomaly inference model outputs a set of candidate anomaly triggering nodes; the set of candidate anomaly triggering nodes includes multiple process nodes arranged in descending order of confidence score, wherein the number of candidate nodes is determined by a preset maximum number of candidate nodes; For each candidate anomaly triggering node, generate corresponding supporting evidence information; the supporting evidence information includes the modal anomaly descriptor identifier that supports the candidate anomaly triggering node and the corresponding anomaly intensity quantification information; Before performing reverse time-series sourcing analysis, process reachability verification is performed on each candidate anomaly triggering node based on the industrial knowledge graph to determine whether the corresponding process node meets the origin conditions of the current anomaly type; candidate anomaly triggering nodes that do not meet the process reachability constraints are eliminated. When the number of candidate anomaly triggering nodes that have passed the process reachability verification is lower than a preset threshold, supplementary process nodes that are related to the candidate anomaly triggering nodes are introduced according to the industrial knowledge graph to form an extended candidate node set. The verified candidate anomaly triggering nodes, their corresponding supporting evidence information, and confidence scores are structured and organized to generate candidate node data packets, which are then used as input data for the reverse time-series source tracing analysis module.

[0041] The proposed method for generating and verifying candidate anomaly triggering nodes enhances the interpretability of candidate nodes while preserving the diversity of root cause tracing results by outputting multiple candidate nodes in descending order of confidence and attaching supporting evidence information. Through process accessibility verification, the method utilizes the process semantic constraints of the industrial knowledge graph to filter candidate nodes, effectively eliminating nodes that are unlikely to be the origin of an anomaly in terms of process, thus reducing the search space and computational overhead of subsequent tracing analysis. Furthermore, the method employs a supplementary expansion mechanism when the number of candidate nodes is insufficient, ensuring the completeness of the tracing analysis even in scenarios with low model output confidence. This significantly improves the targeting, efficiency, and success rate of reverse time-series tracing analysis in root cause localization.

[0042] In some embodiments, the method for performing reverse temporal source tracing analysis starting from a candidate anomaly triggering node is as follows: Obtain candidate exception triggering nodes and determine them as the current backtracking nodes; Obtain the upstream process node corresponding to the current backtracking node, and determine the process node to be backtracked based on the historical causal strength information corresponding to the process association relationship; Based on the first occurrence time of the anomaly at the current backtracking node and the propagation delay information of the corresponding process relationship, the anomaly search time interval is determined; Obtain the corresponding multimodal data frames within the anomaly search time interval and extract the corresponding anomaly features; Based on the abnormal characteristics, determine whether there are any abnormal signals in the process node to be traced that occurred earlier than the current traced node; If it exists, update the process node to be traced back to the current traceback node; If it does not exist, or if the backtracking level reaches the preset level, the current backtracking node will be identified as the root cause node to be verified. The process nodes involved in the backtracking process and the corresponding first occurrence time of the anomaly are associated and recorded to form an anomaly propagation path, and corresponding anomaly propagation path data is generated based on the anomaly propagation path.

[0043] The reverse temporal tracing analysis method described in this application traces upstream layer by layer along the process topology, starting from the candidate anomaly trigger node, following the physical laws of anomaly propagation along material and energy flows in industrial production lines to ensure the process rationality of the tracing path; by combining historical causal strength information to prioritize upstream nodes with high causal correlations, the tracing search space is effectively compressed; by precisely limiting the anomaly search time interval based on propagation delay information, interference from timing mismatch signals is avoided; and by comparing anomaly timings layer by layer and controlling hierarchical termination, the anomaly propagation path is fully recorded while efficiently locating the root cause node to be verified, thereby significantly improving the accuracy of root cause node location and the temporal integrity of anomaly propagation path reconstruction.

[0044] In some embodiments, the method for determining the root cause node of anomalies by combining counterfactual verification is as follows: Obtain the root cause node to be verified and its corresponding anomaly propagation path; Counterfactual intervention hypotheses are constructed for each root cause node to be verified, and the probability of occurrence of abnormal events at each downstream node under counterfactual conditions is calculated by combining abnormal time series information and process correlation in the industrial knowledge graph. Based on the probability of occurrence, calculate the counterfactual causal inference score and causal consistency score for each root cause node to be verified. Based on the causal consistency score of each root cause node to be verified, the abnormal root cause nodes are identified. When the causal consistency scores of multiple root cause nodes to be verified meet the preset association conditions, the multiple root cause nodes to be verified are identified as joint root cause nodes, and their corresponding contribution weights are determined.

[0045] The aforementioned method for determining abnormal root cause nodes based on counterfactual verification in this application constructs counterfactual intervention hypotheses and calculates the changes in downstream anomaly probabilities before and after intervention. It quantifies the root cause probability of each node to be verified from the perspective of causal inference rather than simple correlation analysis, effectively avoiding root cause misjudgment caused by the co-occurrence of anomalies in multiple processes. By introducing causal consistency scores for quantitative comparison of multiple candidate nodes, the objectivity of root cause determination is improved. By combining root cause node identification and contribution weight calculation mechanisms, it effectively supports complex root cause attribution in multi-source concurrent anomaly scenarios, thereby significantly improving the causal reliability and adaptability to complex scenarios of abnormal root cause node determination results.

[0046] In some embodiments, the method for determining anomaly propagation paths using counterfactual verification is as follows: Obtain the set of abnormal propagation process nodes and process relationships corresponding to the abnormal root cause node; Based on the set of abnormal propagation process nodes and propagation relationships, a directed abnormal propagation path graph is constructed, and corresponding attributes are configured for each node and propagation relationship. Based on the propagation characteristics of each node during the abnormal propagation process, the corresponding node role type is determined; Based on node attributes, edge attributes, and counterfactual causal inference results corresponding to each node, the overall confidence score of the abnormal propagation path is calculated. Based on the directed anomaly propagation path diagram and the overall confidence score, anomaly propagation path data is generated and associated with anomaly event records in the industrial knowledge base.

[0047] The method for determining the anomaly propagation path described in this application constructs a directed anomaly propagation path graph and configures rich attributes for each node and propagation relationship, thus fully depicting the structured propagation process of anomalies spreading from the root cause node to downstream processes. By assigning role type labels to each node, it provides semantic support for maintenance personnel to quickly understand the scope of anomaly impact and intervention priorities. By combining node attributes, edge attributes, and counterfactual causal inference results to comprehensively calculate the overall confidence score, it quantifies the credibility of the propagation path. By establishing an association storage with an industrial knowledge base, it significantly improves the traceability of anomaly propagation path data and its accumulated value for subsequent incremental updates to the model.

[0048] In some embodiments, the preset anomaly elimination condition is a logical AND combination of the following three types of sub-conditions: The first type of sub-condition is the sensor signal recovery condition, which requires that all modal sensor signals corresponding to each process node in the abnormal root cause node and abnormal propagation path be continuously within the normal operating condition baseline interval for no less than a preset number of frames. Among them, for scalar sensor signals with an approximately normally distributed pattern, the normal operating condition baseline interval is determined by the mean of historical normal operating data plus or minus three times the standard deviation. For non-normally distributed modal signals, a single-class support vector machine or kernel density estimation method based on historical normal data is used to determine the normal operating condition boundary. The second type of sub-condition is the multimodal anomaly inference model re-verification condition, which requires that after the current real-time data frame is input into the multimodal anomaly inference model, the probability of no anomaly category output by the model exceeds the preset re-verification confidence threshold. The third type of sub-condition is the process parameter recovery condition, which requires that the core process parameters of the process corresponding to the abnormal root cause node be restored to the target value tolerance range corresponding to the production process specification, and the duration shall not be less than the minimum stable period of the process. The judgment results of each sub-condition are stored in the abnormal state monitoring record in the form of Boolean values ​​and are updated in each real-time data frame acquisition cycle; When all three sub-conditions are met simultaneously, the anomaly elimination condition is determined to be met, and the source tracing analysis process is terminated to avoid premature termination of the source tracing analysis due to short-term recovery of a single modal signal.

[0049] The aforementioned pre-set anomaly elimination condition design in this application constructs a multi-dimensional collaborative verification anomaly elimination judgment mechanism by combining three types of sub-conditions—sensor signal recovery conditions, model verification conditions, and process parameter recovery conditions—in a logical AND manner. This effectively prevents premature termination of the traceability process due to short-term recovery in a single dimension. By employing corresponding baseline interval determination methods for normally and non-normally distributed signals, the accuracy of recovery judgment for sensor signals with different statistical characteristics is improved. By requiring process parameters to continuously meet the minimum stable period, the process status is ensured to truly achieve stable recovery rather than transient improvement. This significantly enhances the rigor of anomaly elimination judgment and the accuracy of the timing of traceability process termination, ensuring the closed-loop integrity of anomaly handling.

[0050] In some embodiments, the method for generating an explainable source tracing report is as follows: Obtain information on the root cause nodes of anomalies, the directed anomaly propagation path diagram, the anomaly type classification results, and the anomaly evidence weight distribution data, and perform unified structured processing to generate report input data; The report input data is input into the report generation language model, and combined with a predefined report template to generate an anomaly event summary, root cause analysis content, and anomaly propagation path content; Based on key modal anomaly descriptors, anomaly intensity quantification values, and historical records of similar anomaly events, a multimodal evidence list and handling recommendations are generated. The generated report content is linked, integrated, and uniformly arranged to construct a structured, interpretable, and traceable report; Output the PDF and JSON format files corresponding to the structured and interpretable traceability report.

[0051] The aforementioned interpretable source tracing report generation method of this application unifies and structures the source tracing results, such as root cause node information, propagation path diagram, classification results, and evidence weights, and inputs them into the report generation language model. Combined with predefined report templates, it realizes the automated natural language description of source tracing analysis conclusions, reducing the time cost of manual report compilation. By introducing historical records of similar anomaly handling to generate handling suggestions, it enhances the practical guidance value of the report for on-site maintenance personnel. By simultaneously outputting PDF and structured JSON dual-format files, it meets the needs of manual reading and automated system processing, respectively, thereby significantly improving the interpretability, transmission efficiency, and industrial operation and maintenance decision support capabilities of anomaly source tracing results.

[0052] In some embodiments, the industrial knowledge base adopts a hybrid storage architecture that combines relational databases and graph databases. The relational database is used to store structured abnormal event records, equipment ledger information, process parameter specifications, and sensor baseline statistical data, while the graph database is used to store node data and edge relationship data in the industrial knowledge graph.

[0053] The hybrid storage architecture of the industrial knowledge base described in this application addresses the differentiated storage needs of structured record data and graph topology relationship data by mapping the storage advantages of relational databases and graph databases to the respective storage requirements. It fully leverages the capabilities of relational databases in structured querying and transaction consistency assurance, as well as the efficiency of graph databases in multi-hop neighborhood traversal and graph structure relationship retrieval. This significantly improves the storage efficiency and query performance of the industrial knowledge base in scenarios involving the management of massive abnormal event records and the retrieval of complex process relationship relationships, providing low-latency knowledge support for real-time traceability analysis.

[0054] In some embodiments, the method for writing an interpretable traceability report back to an industry knowledge base is as follows: Add a new abnormal event record in the relational database and write the event's unique identifier, abnormal type label, root cause node identifier, JSON serialized string corresponding to the abnormal propagation path, source tracing report file path, abnormal elimination time, and overall confidence score of this source tracing analysis. In the graph database, based on the identifier of this abnormal event, the historical anomaly frequency attributes of the root cause node and the nodes involved in the propagation path are updated, and the historical recurrence frequency attributes of each directed edge in the propagation path are accumulated. When this source tracing analysis identifies an inter-process causal dependency relationship that has not yet been established in the industrial knowledge graph, a corresponding directed edge is added to the graph database, and the historical causal strength coefficient of the directed edge is initialized to a preset minimum confidence value so that iterative updates can be continuously made based on newly added abnormal events. The write-back operation is executed atomically using a database transaction mechanism. If any sub-step fails, all executed modifications are rolled back, and the write-back failure information is appended to the exception handling log. The write-back operation is then re-executed after manual confirmation to ensure the consistency and integrity of the industrial knowledge base data.

[0055] The aforementioned traceability report write-back method in this application achieves full-element structured archiving of abnormal events by writing complete traceability analysis result metadata into a relational database, supporting subsequent retrieval of historical abnormal events based on multi-dimensional conditions; continuously enriches the statistical prior information of the knowledge graph by updating the historical anomaly frequency of nodes and the historical recurrence frequency of edges in the graph database, providing a more accurate causal strength reference for subsequent traceability analysis; achieves automated expansion and evolution of the knowledge graph by adding causal dependency edges between processes that have not yet been established; and ensures the atomicity of the write-back operation through a transaction mechanism, thereby significantly improving the consistency, integrity, and continuous self-learning accumulation capabilities of industrial knowledge base data.

[0056] In some embodiments, the method for incrementally updating the multimodal anomaly inference model based on the updated industrial knowledge base is as follows: The incremental update process of the multimodal anomaly reasoning model is determined based on preset trigger conditions. The trigger conditions include: the number of new anomaly event records added in the industrial knowledge base since the last incremental update exceeds a preset threshold, and / or the time interval since the last incremental update exceeds a preset maximum update cycle. When the triggering conditions are met, newly added abnormal event records are extracted from the updated industrial knowledge base, and the associated abnormal context and modal abnormal descriptor subsets corresponding to each abnormal event are obtained as incremental training input data. Obtain the manually confirmed anomaly type label, root cause node identifier, and modal evidence importance score corresponding to each newly added anomaly event, and construct an incremental training sample set based on the incremental training input data and supervision labels; Incremental training of the multimodal anomaly inference model is performed based on the incremental training sample set, and the model parameters are constrained and optimized using the elastic weight consolidation method to suppress the model's forgetting of historical anomaly detection capabilities during the learning of new sample distributions. The importance of each model parameter in historical tasks is estimated based on the diagonal approximation Fisher information matrix, and a regularization constraint term is constructed based on the expected value of the squared gradient of each parameter to reduce the computational complexity of the incremental training process. After incremental training is completed, regression testing is performed on the updated multimodal anomaly inference model based on the preset retention set, and it is determined whether the decrease in classification accuracy of the model on historical anomaly types is lower than the preset tolerance threshold. When the regression test results meet the preset requirements, the old model version running online is replaced with the updated multimodal anomaly inference model, and the model version number, number of incremental training samples, training time and regression test accuracy indicators are recorded in the industrial knowledge base to realize model version management and performance traceability.

[0057] The incremental update method for the multimodal anomaly inference model described in this application controls the update timing through dual triggering conditions based on quantity thresholds and time intervals. This ensures that the model evolves synchronously with the anomaly distribution on the production site while avoiding the waste of computational resources caused by frequent updates. By employing an elastic weight consolidation method combined with a diagonal approximation Fisher information matrix to construct regularization constraints, it effectively suppresses catastrophic forgetting while introducing the learning ability of new anomaly samples, maintaining the model's ability to detect anomaly types in historical operating conditions. Through regression testing and accuracy degradation threshold control before deployment, the method significantly improves the adaptive evolution capability, historical capability retention, and version management standardization of the multimodal anomaly inference model in the long-term continuous operation scenario of industrial production lines, thereby achieving continuous improvement in anomaly detection and root cause tracing capabilities.

[0058] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for anomaly detection and root cause analysis in industrial production lines based on a multimodal large model, characterized in that, Includes the following steps: Preprocess the multimodal data frames from the industrial production line to generate unified structured input data; Anomaly feature representations are extracted based on unified structured input data, and modal anomaly descriptors are constructed based on these anomaly feature representations. Based on the production process relationships, modal anomaly descriptors are organized across processes to generate associated anomaly contexts. The pre-trained multimodal anomaly reasoning model is invoked to perform joint reasoning on the associated anomaly context and modal anomaly descriptors, resulting in anomaly type classification results, anomaly evidence weight distribution, and candidate anomaly triggering nodes. Starting from the candidate anomaly trigger node, perform reverse time-series source analysis and combine it with counterfactual verification to determine the root cause node and anomaly propagation path.

2. The method for detecting and tracing the root causes of anomalies in industrial production lines according to claim 1, characterized in that, The method for preprocessing multimodal data frames from industrial production lines is as follows: The system acquires visual data, vibration data, audio data, temperature data, and current data collected from each process node of the industrial production line, and performs timestamp synchronization processing to obtain time-aligned multimodal data. Denoising is performed on the time-aligned modal data to obtain the corresponding cleaned data; The purified data is segmented into frames based on preset frame length and frame shift parameters to construct a multimodal data frame sequence with a unified time window index. Missing or anomalous data in a multimodal data frame sequence is completed to obtain the completed multimodal data frame and the corresponding data quality confidence information; The completed multimodal data frames are associated with and stored with the corresponding data quality confidence information to generate preprocessed structured multimodal data frames.

3. The method for detecting and tracing the root causes of anomalies in industrial production lines according to claim 2, characterized in that, The method for extracting abnormal features is as follows: Reconstruction error analysis is performed on visual data in unified structured input data to identify abnormal regions in the image and extract the corresponding visual abnormality feature vectors. Spectral deviation analysis is performed on vibration data in the unified structured input data to determine the vibration anomaly state and extract the corresponding vibration anomaly feature vector; Acoustic feature deviation analysis is performed on audio data in unified structured input data to determine audio anomalies and extract corresponding audio anomaly feature vectors. Time-series variation analysis is performed on temperature and current data in the unified structured input data, and trend change features, abrupt change features, and limit exceedance persistence features are extracted to obtain the corresponding scalar anomaly feature vectors. Cross-modal fusion processing is performed on visual anomaly feature vectors, vibration anomaly feature vectors, audio anomaly feature vectors, and scalar anomaly feature vectors to obtain fused anomaly features; Anomaly description results are constructed based on fused anomaly features, resulting in anomaly feature representations that include structured semantic information and numerical representation information.

4. The method for detecting and tracing the root causes of anomalies in industrial production lines according to claim 1, characterized in that, The method for generating associated exception contexts is as follows: Based on the process node identifiers corresponding to each modal anomaly descriptor in the current data frame, determine the set of abnormal process nodes; Based on the set of abnormal process nodes, the corresponding upstream and downstream process nodes and process relationships are retrieved from the industrial knowledge graph; Based on the process association, obtain the modal anomaly descriptors corresponding to the upstream and downstream process nodes in the current data frame and the preset historical time window; Based on the process relationships and the corresponding modal anomaly descriptors, identify the abnormal process nodes with relationships and construct the associated anomaly group; Based on the propagation delay information corresponding to the process association, the temporal relationship of abnormal events in the associated abnormal group is verified. The abnormal association type of the associated abnormal group is determined based on the time sequence relationship verification results; Construct an associated anomaly context based on the associated anomaly group, the anomaly association type, and the corresponding process topology information.

5. The method for detecting and tracing the root causes of anomalies in industrial production lines according to claim 1, characterized in that, The method for performing reverse time-series source tracing analysis starting from the candidate anomaly trigger node is as follows: Obtain candidate exception triggering nodes and determine them as the current backtracking nodes; Obtain the upstream process node corresponding to the current backtracking node, and determine the process node to be backtracked based on the historical causal strength information corresponding to the process association relationship; Based on the first occurrence time of the anomaly at the current backtracking node and the propagation delay information of the corresponding process relationship, the anomaly search time interval is determined; Obtain the corresponding multimodal data frames within the anomaly search time interval and extract the corresponding anomaly features; Based on the abnormal characteristics, determine whether there are any abnormal signals in the process node to be traced that occurred earlier than the current traced node; If it exists, update the process node to be traced back to the current traceback node; If it does not exist, or if the backtracking level reaches the preset level, the current backtracking node will be identified as the root cause node to be verified. The process nodes involved in the backtracking process and the corresponding first occurrence time of the anomaly are associated and recorded to form an anomaly propagation path, and corresponding anomaly propagation path data is generated based on the anomaly propagation path.

6. The method for detecting and tracing the root causes of anomalies in industrial production lines according to claim 1, characterized in that, The method for identifying the root cause of anomalies by combining counterfactual verification is as follows: Obtain the root cause node to be verified and its corresponding anomaly propagation path; Counterfactual intervention hypotheses are constructed for each root cause node to be verified, and the probability of occurrence of abnormal events at each downstream node under counterfactual conditions is calculated by combining abnormal time series information and process correlation in the industrial knowledge graph. Based on the probability of occurrence, calculate the counterfactual causal inference score and causal consistency score for each root cause node to be verified. Based on the causal consistency score of each root cause node to be verified, the abnormal root cause nodes are identified. When the causal consistency scores of multiple root cause nodes to be verified meet the preset association conditions, the multiple root cause nodes to be verified are identified as joint root cause nodes, and their corresponding contribution weights are determined.

7. The method for detecting and tracing the root causes of anomalies in industrial production lines according to claim 1, characterized in that, The method for determining the abnormal propagation path by combining counterfactual verification is as follows: Obtain the set of abnormal propagation process nodes and process relationships corresponding to the abnormal root cause node; Based on the set of abnormal propagation process nodes and propagation relationships, a directed abnormal propagation path graph is constructed, and corresponding attributes are configured for each node and propagation relationship. Based on the propagation characteristics of each node during the abnormal propagation process, the corresponding node role type is determined; Based on node attributes, edge attributes, and counterfactual causal inference results corresponding to each node, the overall confidence score of the abnormal propagation path is calculated. Based on the directed anomaly propagation path diagram and the overall confidence score, anomaly propagation path data is generated and associated with anomaly event records in the industrial knowledge base.

8. The method for detecting and tracing the root causes of anomalies in industrial production lines according to claim 7, characterized in that, The industrial knowledge base adopts a hybrid storage architecture that combines relational databases and graph databases. The relational database is used to store structured abnormal event records, equipment ledger information, process parameter specifications, and sensor baseline statistical data, while the graph database is used to store node data and edge relationship data in the industrial knowledge graph.

9. The method for detecting and tracing the root causes of anomalies in industrial production lines according to any one of claims 1-8, characterized in that, The method for detecting and tracing the root causes of anomalies in industrial production lines also includes the following steps: During the reverse time-series source tracing analysis, based on the anomaly propagation path and real-time acquired multimodal data frames, it is determined whether the preset anomaly elimination conditions are met. If the conditions are not met, reverse time-series source analysis and real-time monitoring will continue until the anomaly elimination conditions are met. If the conditions are met, the reverse time-series source analysis process ends, and an interpretable source report is generated based on the abnormal root cause node and abnormal propagation path. The explainable traceability report is written back to the industrial knowledge base, and the multimodal anomaly inference model is incrementally updated based on the updated industrial knowledge base.

10. The method for detecting and tracing the root causes of anomalies in industrial production lines according to claim 9, characterized in that, The method for generating an explainable source tracing report is as follows: Obtain information on the root cause nodes of anomalies, the directed anomaly propagation path diagram, the anomaly type classification results, and the anomaly evidence weight distribution data, and perform unified structured processing to generate report input data; The report input data is input into the report generation language model, and combined with a predefined report template to generate an anomaly event summary, root cause analysis content, and anomaly propagation path content; Based on key modal anomaly descriptors, anomaly intensity quantification values, and historical records of similar anomaly events, a multimodal evidence list and handling recommendations are generated. The generated report content is linked, integrated, and uniformly arranged to construct a structured, interpretable, and traceable report; Output the structured and interpretable traceability report in PDF and JSON formats.

Images