Knowledge graph-based industrial data analysis and mining system

Abstract

The application discloses an industrial data analysis and mining system based on a knowledge graph, relates to the technical field of data mining, and is used for solving the problems of difficult parameter correlation across processes and low abnormality traceability accuracy. The system calculates the correlation strength of the coating process oven temperature and the formed part capacity direct current internal resistance of the forming process by using a time delay mutual information algorithm, generates an influence relationship edge with time-varying weight in combination with the temperature-internal resistance negative correlation process rule, and constructs an industrial knowledge graph. Taking the direct current internal resistance abnormal batch as a starting point, the knowledge graph is traversed in reverse along the process path, a root cause parameter list is generated by fusing the parameter Z-score deviation and the dynamic weight, quality root cause positioning is realized, an abnormal parameter is instantiated as an event node, star and chain propagation topologies are identified by subgraph isomorphism detection, a standardized fault mode is extracted, a real cause and effect structure is identified based on an FCI cause and effect discovery algorithm, the knowledge graph weight is dynamically adjusted, the graph adaptive evolution is realized, and the accuracy and intelligent level of complex process quality analysis are improved.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, specifically to an industrial data analysis and mining system based on knowledge graphs. Background Technology

[0002] In recent years, with the rapid development of intelligent manufacturing and digital factories, key processes in production lines such as coating, rolling, slitting, and formulation have generated massive amounts of multidimensional data. Various sensors, MES systems, and quality inspection equipment continuously record parameters such as temperature, pressure, and electrical performance, providing a data foundation for process optimization and quality prediction. However, the high coupling of the process flow leads to complex time-delay correlations and nonlinear influence relationships between parameters of different processes. Traditional data methods, which mainly focus on single-process analysis, are insufficient to reveal the potential quality evolution patterns across stages and cannot support the intelligent transformation towards refined quality control and root cause analysis.

[0003] Most existing industrial data analysis methods rely on statistical regression, principal component analysis, or black-box models based on machine learning for feature extraction and anomaly detection. These methods suffer from the following substantial drawbacks: First, the models typically assume fixed or linear correlations between process variables, ignoring the differences in process timing between different production batches, leading to unstable cross-process causal inference results. Second, traditional methods focus on index fitting rather than mechanism explanation, failing to pinpoint specific upstream process parameters and their transmission paths when complex quality issues such as high product internal resistance or capacity decay occur. Third, due to the lack of structured knowledge representation, the correlation knowledge between different processes, equipment, and batches is difficult to accumulate and reuse, resulting in data silos and knowledge gaps. Finally, existing root cause analysis systems often use fixed weights or manual experience for impact assessment, lacking the ability to adaptively adjust to dynamic changes in the process environment, making it difficult to support continuous model evolution and failure mode accumulation.

[0004] Therefore, there is an urgent need for an industrial data analysis and mining system that can realize data association modeling, causal reasoning and dynamic knowledge evolution at the cross-process level, so as to improve the interpretability, traceability and self-learning ability of the manufacturing process. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an industrial data analysis and mining system based on knowledge graphs, which solves the problems mentioned in the background.

[0006] To achieve the above objectives, this invention employs the following technical solution: an industrial data analysis and mining system based on a knowledge graph, comprising the following modules: an industrial knowledge graph construction module, used to calculate the correlation strength between the oven temperature in the coating process and the DC internal resistance of the battery cell in the formation and capacity testing process using a time-delay mutual information algorithm, and to verify the correlation direction based on predefined temperature-internal resistance negative correlation process rules, generating influence relationship edges with time-varying weight attributes; and a cross-process quality root cause localization module, used to start with a batch of abnormal DC internal resistance, and to traverse the coating-rolling-slitting process path in the knowledge graph in reverse along the material flow direction, locate all upstream process parameters affecting DC internal resistance, and calculate the Z-score deviation of coating speed and areal density parameters and their corresponding influence in the knowledge graph. The dynamic weights of the relationships are weighted and fused to generate a root cause parameter list sorted by comprehensive suspicion. The fault propagation pattern mining module is used to instantiate the abnormal parameters in the root cause parameter list into abnormal event nodes, construct an abnormal event graph connecting the coating and rolling processes, and identify the recurring star and chain propagation topologies in the graph through a subgraph isomorphism detection algorithm to form a matching standardized fault mode. The causal discovery and graph evolution module is used to execute the FCI causal discovery algorithm on the coating and rolling parameter sets involved in the fault mode, construct a partially directed acyclic graph, compare the causal arrows identified in the graph with the corresponding influence relationship edges in the knowledge graph for directional consistency, and generate weight adjustment instructions based on the comparison results to drive the time-varying weight attributes of the influence relationship edges to be dynamically updated.

[0007] Furthermore, the specific process of calculating the correlation strength between the oven temperature in the coating process and the DC internal resistance of the battery cell in the formation and capacity testing process using the time-delay mutual information algorithm is as follows: Based on the production cycle between the coating and formation and capacity testing processes, a batch-aligned time-series data mapping relationship is established. Within a preset time delay range, the oven temperature sequence and DC internal resistance sequence under different time delays are scanned at a fixed sampling interval. For each time delay point, the normalized mutual information value of the oven temperature sequence and the DC internal resistance sequence is calculated to form a time-delay-mutual information strength relationship sequence. From the time-delay-mutual information strength relationship sequence, the time delay value and correlation strength corresponding to the mutual information peak are extracted as the optimal time-delay correlation feature between parameters.

[0008] Furthermore, based on predefined temperature-internal resistance negative correlation process rules, the correlation direction is verified, and the generation logic of the influence relationship edge with time-varying weight attributes is as follows: The correlation direction in the optimal time-delay correlation feature is matched and verified with the predefined temperature-internal resistance negative correlation rules. When the correlation direction is found to be consistent with the negative correlation rule, the mutual information peak intensity is used as the basic weight, and a process consistency enhancement coefficient is added. When the correlation direction is found to conflict with the negative correlation rule, a process constraint penalty coefficient is applied to the basic weight to generate the weighted correlation weight. Based on the time window sliding mechanism, the above verification process is repeated to assign a time-varying weight attribute that evolves with time to the influence relationship edge, and a process credibility label is marked for each weight.

[0009] Furthermore, starting with the batch of abnormal DC internal resistance, the specific process of traversing the coating-rolling-slitting process path in the knowledge graph in reverse along the material flow direction to locate all upstream process parameters affecting DC internal resistance is as follows: Starting with the abnormal batch as the starting node, a reverse breadth-first traversal is performed along the material flow edge in the knowledge graph. During the traversal, all process nodes and corresponding equipment nodes are recorded. For each process node, all process parameters that are directly or indirectly related to DC internal resistance are retrieved through the parameter influence relationship edge, and a parameter-process-equipment mapping relationship table is established.

[0010] Furthermore, the specific process of generating a root cause parameter list by weighting and fusing the Z-score deviation of coating speed and areal density parameters with the dynamic weights of the corresponding influence relationships in the knowledge graph, and sorting them according to comprehensive suspicion, is as follows: For each process parameter in the parameter-process-equipment mapping relationship table, obtain its feature value sequence during the current abnormal batch production period, and calculate the Z-score statistical deviation of each parameter feature value sequence relative to the recent golden batch benchmark distribution; read the current dynamic weight value of the influence relationship edge between each parameter and DC internal resistance from the knowledge graph, and use a multi-factor fusion algorithm to weight and calculate the Z-score deviation, dynamic weight value, and process distance attenuation factor to generate a comprehensive suspicion score for each parameter; sort the parameters in descending order according to the comprehensive suspicion score to generate a root cause parameter list with priority, and label the associated process and equipment information for each parameter item.

[0011] Furthermore, the abnormal parameters in the root cause parameter list are instantiated as abnormal event nodes. The construction logic of the abnormal event graph connecting the coating and rolling processes is as follows: Extract the attribute information of each abnormal parameter from the root cause parameter list, including parameter type, deviation degree and the process equipment to which it belongs. Use the parameter deviation degree as the node weight, instantiate each abnormal parameter as an abnormal event node, and mark the node with a timestamp and process context. Based on the process flow timing relationship between the coating and rolling processes, establish directed connection edges between consecutively occurring abnormal event nodes. Align the time of cross-process abnormal event nodes through time windows to construct a complete coating-rolling abnormal event propagation graph.

[0012] Furthermore, the specific process of identifying recurring star-shaped and chain-shaped propagation topologies in the graph using the subgraph isomorphism detection algorithm to form matching standardized fault modes is as follows: Candidate propagation subgraphs are extracted from the historical anomaly event graph, a candidate subgraph pattern library is established, and topological structure analysis is performed on each candidate subgraph to identify star-shaped radiation structures and chain-shaped propagation structures; the structural similarity between the candidate subgraphs and known fault modes is calculated using the subgraph isomorphism detection algorithm, and the detected recurring propagation topologies are normalized and encoded to form standardized fault modes containing topological structure, propagation paths, and typical parameter features; a fault mode matching index is established, and each standardized fault mode is assigned a unique pattern identifier and matching threshold.

[0013] Furthermore, the specific process of constructing a partial directed acyclic graph (DAG) by applying the FCI causal discovery algorithm to the coating and rolling parameter sets involved in the failure modes is as follows: Key parameter sets for the coating and rolling processes are extracted from the standardized failure modes, a multivariate time series dataset is constructed, and conditional independence tests are performed on the parameter sets to determine the dependencies between variables; the FCI causal discovery algorithm is executed to explore the causal structure between parameters in the presence of potential confounding factors; based on the separation set and collision node identification results, a partial DAG is constructed to clarify the direct causal, indirect causal, and pseudo-correlation relationships between parameters; confidence assessment is performed on the causal arrows in the graph, and statistically significant causal connections are retained.

[0014] Furthermore, the logical process of comparing the causal arrows identified in the graph with the corresponding influence relationship edges in the knowledge graph to ensure directional consistency, and generating weight adjustment instructions based on the comparison results to drive the dynamic update of the time-varying weight attributes of the influence relationship edges is as follows: Causal arrows in some directed acyclic graphs are matched one-to-one with influence relationship edges stored in the knowledge graph. For causal relationships with consistent directions, a weight enhancement instruction is generated, increasing the dynamic weight of the corresponding edge according to the confidence level. For causal relationships with conflicting directions, a weight correction instruction is generated, reducing the weight of the corresponding edge or recalibrating it according to the degree of conflict. For newly discovered causal relationships driven by data, new influence relationship edges are created in the knowledge graph and marked as algorithm discoveries. A weight update log is established to record the basis and time of each adjustment, achieving traceable dynamic updates of the influence relationship edge weights.

[0015] The present invention has the following beneficial effects:

[0016] (1) The knowledge graph-based industrial data analysis and mining system effectively solves the problems of data silos and difficulty in identifying time delay effects in traditional analysis methods by constructing an industrial knowledge graph that integrates time-varying weight attributes. The system uses a time-delay mutual information algorithm to accurately capture the cross-process delay correlation between the oven temperature in the coating process and the DC internal resistance of the battery cell in the formation and capacity testing process. It verifies the correlation direction by combining predefined process rules to ensure the physical rationality of parameter correlation. On this basis, by traversing the process path in the knowledge graph in reverse, it integrates parameter deviation and dynamic weight to achieve multi-dimensional comprehensive analysis, which significantly improves the accuracy and comprehensiveness of the root cause localization of quality anomalies.

[0017] (2) The knowledge graph-based industrial data analysis and mining system realizes the transformation from passive analysis to proactive early warning through automatic mining of fault propagation patterns and continuous discovery of causal relationships. The system instantiates the identified abnormal parameters as event nodes, constructs an abnormal event graph, and mines recurring propagation patterns through subgraph isomorphism detection, thus establishing a reusable fault knowledge base. Furthermore, it adopts the FCI causal discovery algorithm to identify the true causal structure between parameters, and achieves dynamic weight adjustment by comparing with existing relationships in the knowledge graph, enabling the system to have the ability to self-evolve and continuously optimize, significantly improving the intelligence level of industrial data analysis.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] Figure 1 This is a flowchart of the knowledge graph-based industrial data analysis and mining system of the present invention. Detailed Implementation

[0020] This application's embodiments address the limitations of traditional industrial data analysis caused by data silos, the neglect of time lag effects in parameter correlation, and the lack of effective causal reasoning mechanisms through a knowledge graph-based industrial data analysis and mining system.

[0021] The overall approach of the solution in this application is as follows: by constructing an industrial knowledge graph with time-varying weights, a parameter association network across processes is established; based on this graph, accurate root cause localization of quality anomalies is achieved, and fault propagation patterns are mined; finally, through causal discovery and graph self-evolution mechanism, a complete closed loop of "data analysis - knowledge mining - system optimization" is formed, realizing intelligent industrial data analysis that combines data-driven and knowledge-guided approaches.

[0022] Please see Figure 1 This invention provides a technical solution: an industrial data analysis and mining system based on a knowledge graph, comprising the following modules: an industrial knowledge graph construction module, used to calculate the correlation strength between the oven temperature in the coating process and the DC internal resistance of the battery cell in the formation and capacity testing process using a time-delay mutual information algorithm, and to verify the correlation direction based on predefined temperature-internal resistance negative correlation process rules, generating influence relationship edges with time-varying weight attributes; and a cross-process quality root cause localization module, used to start from the batch with abnormal DC internal resistance, and to traverse the coating-rolling-slitting process path in the knowledge graph in reverse along the material flow direction, locate all upstream process parameters affecting DC internal resistance, and calculate the Z-score deviation of coating speed and areal density parameters and the dynamic correlation of the corresponding influence relationship in the knowledge graph. The system performs weighted fusion of state weights to generate a root cause parameter list sorted by comprehensive suspicion level. The fault propagation mode mining module instantiates abnormal parameters in the root cause parameter list into abnormal event nodes, constructs an abnormal event graph connecting the coating and rolling processes, and identifies recurring star and chain propagation topologies in the graph through a subgraph isomorphism detection algorithm to form a matching standardized fault mode. The causal discovery and graph evolution module executes the FCI causal discovery algorithm on the coating and rolling parameter sets involved in the fault mode, constructs a partially directed acyclic graph, compares the causal arrows identified in the graph with the corresponding influence relationship edges in the knowledge graph for directional consistency, and generates weight adjustment instructions based on the comparison results to drive the dynamic update of the time-varying weight attributes of the influence relationship edges.

[0023] In this implementation plan, an industrial knowledge graph construction module is used to perform semantic modeling and structured association of multi-process and multi-dimensional data in industrial production. The correlation strength between "oven temperature" in the coating process and "cell DC internal resistance" in the formation and capacity testing process is calculated using a time-delay mutual information algorithm. The time-delay mutual information algorithm is a statistical method used to analyze the time-delay correlation between different variables, revealing the time-dependent impact of upstream process variables on downstream quality indicators. Subsequently, the system verifies the rationality of the correlation direction based on a predefined "temperature-internal resistance negative correlation process rule" (i.e., the physical law that increased temperature usually leads to decreased internal resistance), ultimately generating influence relationship edges in the knowledge graph that include time-varying weight attributes (i.e., the correlation strength dynamically changes over time). This graph provides a structured semantic foundation for subsequent root cause localization and causal reasoning. A cross-process quality root cause localization module is used to automatically locate upstream process parameters causing quality anomalies in a multi-process production chain. Starting with "abnormal DC internal resistance batches," the system traverses the knowledge graph in reverse along the material flow direction, analyzing the dependencies between upstream process nodes such as coating, rolling, and slitting. By calculating the Z-score deviation of key parameters such as coating speed and areal density (i.e., the standard deviation multiple of the parameter from its historical mean, used to measure the degree of anomaly), and combining this with the dynamic weights of corresponding influence edges in the knowledge graph, a weighted fusion is performed to generate a root cause parameter list sorted by overall suspicion level. This module enables the quantitative localization of anomaly sources in complex process chains. The fault propagation pattern mining module identifies the propagation patterns of abnormal parameters between different processes. The system instantiates the abnormal parameters identified in the root cause parameter list as abnormal event nodes and constructs an abnormal event graph connecting the coating and rolling processes. In this graph, nodes represent specific abnormal events, and edges represent the temporal or logical relationships between anomalies. Subsequently, using a subgraph isomorphism detection algorithm (an algorithm for identifying recurring patterns in graph structures), the system analyzes recurring star-shaped structures (i.e., one core anomaly triggers multiple downstream anomalies) and chain-like structures (i.e., anomalies propagate sequentially along the process flow) in the anomaly event graph. These structures are refined into matchable standardized fault modes for subsequent pattern recognition and early warning. The causal discovery and graph evolution module reveals the true causal relationships between process parameters and dynamically updates the knowledge graph. For the coating and rolling parameter sets involved in the fault modes, the system applies the FCI (Fast Causal Inference) causal discovery algorithm. This algorithm can identify partial directed acyclic graphs (PAGs) even with potentially unobserved variables, thereby inferring possible causal directions between variables. The system then compares the identified causal arrows with the existing influence relationship edges in the knowledge graph for directional consistency. Based on the comparison results, it generates weight adjustment instructions to dynamically update the time-varying weights of the relationship edges in the knowledge graph. In this way, the knowledge graph can continuously evolve with new data and inference results, transforming from a static knowledge structure into a dynamic learning system.This embodiment constructs an evolvable industrial knowledge graph, realizing a closed-loop process from data association analysis to causal mechanism discovery, which can significantly improve the ability to diagnose and predict quality anomalies in complex process scenarios.

[0024] Specifically, the process of calculating the correlation strength between the oven temperature in the coating process and the DC internal resistance of the battery cell in the formation and capacity testing process using the time-delay mutual information algorithm is as follows: Based on the production cycle time between the coating and formation and capacity testing processes, a batch-aligned time-series data mapping relationship is established. Within a preset time delay range, the oven temperature sequence and DC internal resistance sequence under different time delays are scanned at a fixed sampling interval. For each time delay point, the normalized mutual information value of the oven temperature sequence and the DC internal resistance sequence is calculated to form a time-delay-mutual information strength relationship sequence. From the time-delay-mutual information strength relationship sequence, the time delay value and correlation strength corresponding to the mutual information peak are extracted as the optimal time-delay correlation feature between parameters.

[0025] In this implementation scheme, the calculation of the cross-process parameter correlation strength based on the time-delay mutual information algorithm is specifically as follows: The process of calculating the correlation strength between the oven temperature in the coating process and the DC internal resistance of the battery cell in the formation and capacity testing process using the time-delay mutual information algorithm is as follows: Based on the production time difference between the coating and formation and capacity testing processes, a batch-aligned time-series data mapping relationship is first established, so that the temperature data of each coating batch can be aligned with the internal resistance data of the corresponding formation and capacity testing batch on the time axis. Then, within a preset time delay range Δτ, the oven temperature sequence and DC internal resistance sequence under different time delays are scanned at a fixed sampling interval δ. At each time delay... Next, calculate the normalized mutual information value between the two. Its expression is: ;in: In time delay The joint probability distribution of the temperature range m of the lower oven and the DC internal resistance range n; The marginal probability distribution of the oven temperature sequence; The marginal probability distribution of the DC internal resistance sequence; Indicates time delay The degree of information sharing between the next two sequences is used to measure the strength of the statistical dependency between parameters. This is achieved by traversing the entire time delay. Forming a time-delay-mutual information strength relationship sequence The time delay τ and its peak intensity corresponding to the mutual information peak are extracted from the sequence and used as the optimal time delay correlation feature between "oven temperature" and "DC internal resistance". The purpose of this step is to identify the most significant time dependence between cross-process variables, thereby extending the traditional static parameter correlation into a dynamic correlation with time series significance.

[0026] Specifically, the generation logic of the influence relationship edge with time-varying weight attributes is as follows, based on the predefined temperature-internal resistance negative correlation process rule to verify the correlation direction: the correlation direction in the optimal time-delay correlation feature is matched and verified with the predefined temperature-internal resistance negative correlation rule. When the correlation direction is found to be consistent with the negative correlation rule, the mutual information peak intensity is used as the basic weight and a process consistency enhancement coefficient is added. When the correlation direction is found to conflict with the negative correlation rule, a process constraint penalty coefficient is applied to the basic weight to generate the weighted correlation weight. Based on the time window sliding mechanism, the above verification process is repeated to assign a time-varying weight attribute that evolves with time to the influence relationship edge, and a process credibility label is marked for each weight.

[0027] In this implementation scheme, the logic for verifying the correlation direction and generating weights based on the negative correlation process rules, and verifying the correlation direction based on the predefined "temperature-internal resistance negative correlation process rules," generates the influence relationship edge with time-varying weight attributes as follows: The correlation direction in the optimal time-delay correlation feature (i.e., the direction of internal resistance change when temperature rises) is matched and verified with the process rules. When the correlation direction is detected to be consistent with the negative correlation rule, the system will... As the basic weight value And introduce a process consistency enhancement coefficient. Weighted amplification: ;in: The corrected weights are those that have passed the process consistency verification. This is the process consistency enhancement coefficient, and its value is determined by the historical process stability index. The distribution is determined, and the calculation method is as follows: ; Characterizes the stability ratio of temperature to internal resistance under the same process conditions in the past; The system represents the average stability over the entire historical period. When a conflict is detected between the correlation direction and the negative correlation process rule, the system applies a process constraint penalty coefficient to the basic weights. Generate the reduced association weights: ;in: The penalty coefficient is determined by the deviation threshold. Deviation from process Determined, the formula for determining the value is: ; This represents the relative proportion by which the current parameters deviate from the expected process trend. To determine the reasonable deviation of the process from the upper limit threshold (taking the 95% confidence interval boundary) through statistical learning, the generation and confidence labeling of time-varying weight attributes are based on a sliding time window mechanism. The system repeatedly executes the above verification and weighting process within each window length Δt, forming a time-series weight sequence. Subsequently, the fluctuation coefficient of the weights over time is calculated. Based on its stability, a process reliability label is assigned to each weight. :when (in When the threshold is stable, it is marked as "high confidence"; when... When the value is low, it is marked as "low credibility". Finally, influence relationship edges with time-varying weights and credibility labels are generated in the knowledge graph, representing the dynamic influence strength and process consistency of "oven temperature" on "DC internal resistance". This part aims to transform static process knowledge into an adaptive and evolving dynamic knowledge structure, enabling the knowledge graph to reflect the temporal correlation changes of the real production process, providing dynamic support for subsequent root cause localization and causal reasoning.

[0028] Specifically, starting with the batch of abnormal DC internal resistance, the process of traversing the coating-rolling-slitting process path in the knowledge graph in reverse along the material flow direction to locate all upstream process parameters affecting DC internal resistance is as follows: Starting with the abnormal batch as the starting node, a reverse breadth-first traversal is performed along the material flow edge in the knowledge graph. During the traversal, all process nodes and corresponding equipment nodes are recorded. For each process node, all process parameters that are directly or indirectly related to DC internal resistance are retrieved through the parameter influence relationship edge, and a parameter-process-equipment mapping relationship table is established.

[0029] In this implementation, the abnormal batch node is used as the starting point, and a reverse breadth-first search (BFS) is performed along the material flow edges in the knowledge graph. During the traversal, the system sequentially visits each upstream process node and its corresponding equipment node, recording the node sequence that appears in the traversal path. ,in Indicates the first Each visited process node. The system further retrieves all sets of process parameters that are directly or indirectly related to DC internal resistance by examining the influence relationships of their associated parameters. Based on this, establish a parameter-process-equipment mapping table. ,in Indicate process The corresponding equipment nodes. This mapping table records the source process and equipment information for each parameter, providing data support for subsequent root cause suspicion calculation. The purpose of this step is to transform the topological structure in the knowledge graph into a computable parameter association path, enabling the cross-process tracing process to have clear process semantics and data source relationships, thus solving the problem of "root cause omission due to single-process limitations" in traditional methods.

[0030] Specifically, the process of generating a root cause parameter list by weighting and fusing the Z-score deviation of coating speed and areal density parameters with the dynamic weights of the corresponding influence relationships in the knowledge graph, and sorting them according to comprehensive suspicion is as follows: For each process parameter in the parameter-process-equipment mapping table, obtain its feature value sequence during the current abnormal batch production period, and calculate the Z-score statistical deviation of each parameter's feature value sequence relative to the recent golden batch benchmark distribution; read the current dynamic weight value of the influence relationship edge between each parameter and DC internal resistance from the knowledge graph, and use a multi-factor fusion algorithm to weight and calculate the Z-score deviation, dynamic weight value, and process distance attenuation factor to generate a comprehensive suspicion score for each parameter; sort the parameters in descending order according to the comprehensive suspicion score to generate a priority root cause parameter list, and label the associated process and equipment information for each parameter item.

[0031] In this implementation plan, the specific process of generating a root cause parameter list by weighting and fusing the Z-score deviation of coating speed and areal density parameters with the dynamic weights of the corresponding influence relationships in the knowledge graph, and sorting them according to the comprehensive suspicion degree, is as follows: For each process parameter in the parameter-process-equipment mapping relationship table... Obtain its feature value sequence during the current abnormal batch production period. And select the baseline distribution from the recent stable production period (i.e., gold batches). , As a baseline, calculate the Z-score deviation of each parameter in the abnormal batches. The expression is as follows: ;in: For parameters The mean during the period of abnormal batches; This represents the expected mean of this parameter across the gold batches. This represents the standard deviation of this parameter across the gold batch; This characterizes the standardized deviation of the parameter under relatively stable operating conditions. Subsequently, the current dynamic weight values ​​of the influence relationship edges between each parameter and the DC internal resistance are retrieved from the knowledge graph. And calculate the process distance attenuation factor based on the topological distance between the process in which the parameter is located and the target process. ,in: Indicates the shortest path length between processes; The distance attenuation coefficient is determined through empirical optimization to give higher weight to the influence of closer processes. Taking into account parameter deviation, dynamic weights, and the distance attenuation effect, a comprehensive suspicion score for each parameter is calculated. The expression is as follows: ;in: For parameters The overall level of suspicion; , This is a weighting balancing coefficient used to control the proportion of influence between statistical deviation and the strength of the correlation in the graph; where and The value of is optimized by minimizing the false positive rate of historical abnormal batches. Once the entire parameter set has been calculated, the system will... Arrange the parameters in order to generate a root cause parameter list, and attach the associated process number, equipment identifier, and confidence level label to each parameter. The confidence level is determined based on the parameter sample variance. Stability index of process status ,when (in When the threshold is empirical, it is marked as a high-confidence parameter; otherwise, it is marked as a low-confidence parameter.

[0032] Specifically, the construction logic of the abnormal event graph connecting the coating and rolling processes is as follows: Extract the attribute information of each abnormal parameter from the root cause parameter list, including parameter type, deviation degree, and the process equipment to which it belongs. Use the parameter deviation degree as the node weight, instantiate each abnormal parameter as an abnormal event node, and mark the node with a timestamp and process context. Based on the process flow sequence relationship between the coating and rolling processes, establish directed connection edges between consecutively occurring abnormal event nodes. Align the time of cross-process abnormal event nodes using a time window to construct a complete coating-rolling abnormal event propagation graph.

[0033] In this implementation, abnormal parameter items are extracted one by one from the root cause parameter list. The attribute information of each item is read, including parameter identifier, associated process, associated equipment, deviation index, and time interval of occurrence. Using the parameter deviation as the initial node weight, each abnormal parameter is instantiated as an abnormal event node, and a timestamp, process context (identifiers of preceding and following processes), and equipment information are recorded for that node. Based on the process flow sequence relationship between the coating and rolling processes, directed connections are established between consecutively occurring abnormal event nodes according to their time sequence. If the occurrence time of event AQ precedes event BQ and there is material flow reachability between the two processes, a directed edge is established between AQ and BQ. Time window alignment is applied: the timestamps of cross-process abnormal event nodes are aligned according to a preset sliding window, thereby identifying close temporal relationships across processes and forming a complete coating-rolling abnormal event propagation graph. The purpose of this step is to convert the static list of root cause parameters into an abnormal event network with temporal and topological semantics, enabling subsequent propagation pattern detection to be performed on a structured event graph, ensuring that causal pattern recognition has a clear temporal and process semantic basis.

[0034] Specifically, the process of identifying recurring star-shaped and chain-shaped propagation topologies in a graph using a subgraph isomorphism detection algorithm to form matching standardized fault modes is as follows: Candidate propagation subgraphs are extracted from historical anomaly event graphs, a candidate subgraph pattern library is established, and topological structure analysis is performed on each candidate subgraph to identify star-shaped radiation structures and chain-shaped propagation structures. The structural similarity between candidate subgraphs and known fault modes is calculated using the subgraph isomorphism detection algorithm. The detected recurring propagation topologies are then normalized and encoded to form standardized fault modes that include topological structure, propagation path, and typical parameter features. Finally, a fault mode matching index is established, assigning a unique pattern identifier and matching threshold to each standardized fault mode.

[0035] In this implementation, candidate propagation subgraphs are extracted from the historical anomaly event graph set via a sliding window. These candidate subgraphs are generated with a fixed upper limit on the number of nodes or a fixed time length, forming a candidate subgraph set. Topological structure analysis is performed on each candidate subgraph to determine whether it exhibits a star-shaped radial structure (a single core node pointing to multiple downstream nodes) or a chain-like propagation structure (nodes propagating in a chain-like manner). Candidate subgraphs that conform to a certain template enter the matching and scoring stage. Subgraph similarity is calculated, and each candidate subgraph U is scored against the existing standard pattern V. The overall similarity score is defined as: Where: C represents the overall similarity score between the candidate subgraph and the standard pattern; This represents the matching score between two graphs in terms of pure topological structure. This represents the maximum possible score achievable through topological matching. This represents the score of the two graphs on path similarity (i.e., consistency between typical propagation path length and node sequence); This represents the upper limit of the path similarity score; This represents the score of the two graphs on the similarity of node attributes (e.g., deviation distribution, device type distribution); The upper limit of the attribute similarity score; A, B, and H are the combination coefficients of the three scores, used to balance the importance of topology, path, and attribute; parameter explanation (each item is separated by a semicolon): Topological matching scores between candidate subgraphs and standard patterns in terms of structural equivalence or maximum isomorphic subset size; : The theoretical maximum value used for normalizing the topological score; The score reflects the similarity between the order of nodes and the length of the propagation path. : The upper limit used to normalize path similarity; Attribute similarity score calculated based on the numerical attributes of nodes (such as mean deviation and distribution of dominant parameter types); : Upper limit for normalizing attribute similarity; A: Topological similarity weight coefficient; B: Path similarity weight coefficient; H: Attribute similarity weight coefficient. The weight coefficients are determined by pairing historically labeled failure modes with candidate subgraphs and using cross-validation to search for the optimal combination of A, B, and H, maximizing the weighted sum of recall and precision for pattern recognition on the validation set. This search can be performed using grid search or Bayesian optimization. When the similarity C between a candidate subgraph U and a pattern V is greater than the matching threshold M, U is determined to be an instantiation event of pattern V. The matching threshold M is determined by selecting the optimal point based on the ROC curve of historical true positive / false positive balance. For candidate subgraphs determined to be instantiated, normalization encoding is performed: the encoding includes the pattern identifier Q, topological encoding (e.g., serialized adjacency signature), typical propagation path, and typical node attribute vector Φ under that pattern. The normalized pattern is added to the standard pattern library, and a matching index (e.g., an inverted index based on topological signature) is created for the pattern for subsequent real-time matching queries. Parameter explanation: M; matching threshold, determined by ROC curve analysis of historical labeled samples; Q; a unique pattern identifier assigned to the standardized fault mode; Φ; a typical node attribute vector representing the pattern, used for attribute similarity calculation. The purpose of this step: through structured similarity calculation and standardized coding, the recurring fault propagation topology in history is abstracted into identifiable and indexable standardized patterns, thereby achieving the accumulation and rapid matching of fault knowledge.

[0036] Specifically, the process of constructing a partial directed acyclic graph (DAG) by applying the FCI causal discovery algorithm to the coating and rolling parameter sets involved in the failure modes is as follows: Key parameter sets for the coating and rolling processes are extracted from the standardized failure modes, a multivariate time series dataset is constructed, and conditional independence tests are performed on the parameter sets to determine the dependencies between variables; the FCI causal discovery algorithm is executed to explore the causal structure between parameters in the presence of potential confounding factors; based on the separation set and collision node identification results, a partial DAG is constructed to clarify the direct causal, indirect causal, and pseudo-correlation relationships between parameters; confidence assessment is performed on the causal arrows in the graph, and statistically significant causal connections are retained.

[0037] In this implementation plan, key parameter sets for the coating and rolling processes are extracted from standardized failure modes, and a multivariate time series dataset is constructed. For each pair of parameter variables... , Perform a conditional independence test and calculate the conditional independence of the control variable set. Partial correlation coefficient .like Then it is believed and Conditional independence, among which The independence threshold can be determined using K-fold cross-validation on historical normal batches of samples to maximize the discrimination accuracy. Through multiple conditional independence tests, the dependency matrix between variables is obtained. The Fast Causal Inference (FCI) algorithm is executed, and in the presence of potential confounding factors, it progressively determines possible causal directions based on the dependency matrix and the separation set results. The partially directed acyclic graph (DAG) output by the FCI algorithm contains three types of edges: directly causal edges with defined directions, undirected potentially correlated edges, and pseudo-correlated edges caused by confounding factors. For each causal edge in the obtained causal graph... Calculate its confidence level The confidence level is obtained by repeatedly performing FCI in different sampling subsets and counting the frequency of edge occurrences. When a causal connection is determined to be statistically significant, then... The confidence threshold was set empirically. This resulted in a partially directed acyclic graph (DAG) clearly labeled with direct causal relationships, indirect causal relationships, and spurious correlations. By introducing the FCI causal discovery algorithm, accurate identification of the causal structure between coating and rolling parameters was achieved in complex process systems with potential confounding factors, providing a data-driven basis for subsequent dynamic updates of the knowledge graph weights.

[0038] Specifically, the logical process of comparing the causal arrows identified in the graph with the corresponding influence relationship edges in the knowledge graph to ensure directional consistency, and generating weight adjustment instructions based on the comparison results to drive the dynamic update of the time-varying weight attributes of the influence relationship edges is as follows: Causal arrows in some directed acyclic graphs are matched one-to-one with influence relationship edges stored in the knowledge graph. For causal relationships with consistent directions, a weight enhancement instruction is generated, increasing the dynamic weight of the corresponding edge according to the confidence level. For causal relationships with conflicting directions, a weight correction instruction is generated, reducing the weight of the corresponding edge or recalibrating it according to the degree of conflict. For newly discovered causal relationships driven by data, new influence relationship edges are created in the knowledge graph and marked as algorithm discoveries. A weight update log is established to record the basis and time of each adjustment, achieving traceable dynamic updates of the influence relationship edge weights.

[0039] In this implementation scheme, each causal edge in a partially directed acyclic graph will be... The influence relationship between the corresponding parameters in the knowledge graph Perform matching. If the directions are consistent, generate a weight enhancement instruction; if the directions are opposite, generate a weight correction instruction. For consistent relation edges, calculate the weight based on the confidence level of the causal edge. The dynamic weights are updated proportionally, using the following formula: ;in, This indicates the weight of the edge currently influencing the relationship. Indicates the updated weights; The weighting enhancement coefficients are determined on the validation set via grid search. For causal relationships with conflicting directions, a weight reduction correction is performed. ;in, The penalty coefficient was determined through experimental optimization. The directional conflict degree is quantified by the degree to which the causal arrow and the edge direction of the knowledge graph are inconsistent. If a new causal relationship is detected... If an edge does not already exist in the knowledge graph, a new edge will be automatically created in the knowledge graph. and with initial weights Initialization, where This represents the average edge weight in the current graph. The system establishes an update log for each weight adjustment, recording the update type, basis confidence level, execution time, and affected edge identifier, enabling traceable and dynamic evolution of causal edge weights in the knowledge graph. In this implementation scheme, through causal direction comparison and dynamic weight adjustment, an adaptive fusion of data-driven approaches and knowledge graph process priors is achieved. This ensures that the time-varying weights of influencing relationship edges continuously reflect changes in real process mechanisms, improving the system's causal interpretability and evolutionary autonomy.

[0040] In summary, this application has at least the following effects:

[0041] This knowledge graph-based industrial data analysis and mining system achieves closed-loop optimization from data-driven anomaly identification to mechanistic-level causal reasoning through a knowledge graph-based causal mining and dynamic weight update mechanism. On one hand, it utilizes the FCI causal discovery algorithm to accurately identify the true causal relationships between parameters in complex process environments with potential confounding factors, avoiding false associations and misjudgments caused by traditional correlation analysis. On the other hand, through bidirectional comparison of causal directions and influence edges in the knowledge graph, it achieves adaptive dynamic updates of edge weights in the knowledge graph, continuously optimizing the graph structure as the production process evolves. Furthermore, the system features causal update logs and weight evolution tracking functions, enabling traceable verification of each relationship correction and ensuring the consistency of the knowledge graph in terms of interpretability, timeliness, and reliability. This significantly improves the accuracy and intelligence level of industrial process anomaly diagnosis and source tracing analysis.

[0042] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0043] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0044] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0045] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0046] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0047] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A knowledge graph-based industrial data analysis and mining system, characterized in that, Includes the following modules: The industrial knowledge graph construction module is used to calculate the correlation strength between the oven temperature of the coating process and the DC internal resistance of the battery cell in the formation and capacity test process using a time-delay mutual information algorithm. Based on the predefined temperature-internal resistance negative correlation process rules, the module verifies the correlation direction between the oven temperature of the coating process and the DC internal resistance of the battery cell in the formation and capacity test process calculated by the time-delay mutual information algorithm, and generates influence relationship edges with time-varying weight attributes. The cross-process quality root cause localization module is used to start with the batch with abnormal DC internal resistance, and traverse the coating-rolling-slitting process path in the knowledge graph in reverse along the material flow direction to locate all upstream process parameters that affect DC internal resistance. By calculating the Z-score deviation of coating speed and areal density parameters and the dynamic weight of the corresponding influence relationship in the knowledge graph, a weighted fusion is performed to generate a root cause parameter list sorted by comprehensive suspicion degree. The fault propagation mode mining module is used to instantiate abnormal parameters in the root cause parameter list into abnormal event nodes, construct an abnormal event graph connecting the coating and rolling processes, and identify the recurring star and chain propagation topologies in the graph through the subgraph isomorphism detection algorithm to form a matching standardized fault mode. The causal discovery and graph evolution module is used to execute the FCI causal discovery algorithm on the coating and rolling parameter set involved in the failure mode, construct a partially directed acyclic graph, compare the causal arrows identified in the graph with the corresponding influence relationship edges in the knowledge graph for directional consistency, generate weight adjustment instructions based on the comparison results, and drive the time-varying weight attributes of the influence relationship edges to be dynamically updated.

2. The knowledge graph-based industrial data analysis and mining system according to claim 1, characterized in that: The specific process of calculating the correlation strength between the oven temperature in the coating process and the DC internal resistance of the battery cell in the formation and capacity testing process using the time-delay mutual information algorithm is as follows: Based on the production cycle between coating and formulation / compression processes, a batch-aligned time-series data mapping relationship is established. Within a preset time lag range, the oven temperature sequence and DC internal resistance sequence under different time lags are scanned at a fixed sampling interval. For each time delay point, the normalized mutual information value of the oven temperature sequence and the DC internal resistance sequence is calculated to form a time delay-mutual information intensity relationship sequence. From the time delay-mutual information intensity relationship sequence, the time delay value and correlation intensity corresponding to the mutual information peak are extracted as the optimal time delay correlation feature between parameters.

3. The knowledge graph-based industrial data analysis and mining system according to claim 2, characterized in that: The correlation direction is verified based on predefined temperature-internal resistance negative correlation process rules. The generation logic process for generating influence relationship edges with time-varying weight attributes is as follows: The correlation direction in the optimal time delay correlation feature is matched and verified with the predefined temperature-internal resistance negative correlation rule. When the correlation direction is found to be consistent with the negative correlation rule, the mutual information peak intensity is used as the basic weight and a process consistency enhancement coefficient is added. When the correlation direction is found to conflict with the negative correlation rule, a process constraint penalty coefficient is applied to the basic weight to generate the weighted correlation weight. Based on the time window sliding mechanism, the above verification process is repeated, and time-varying weight attributes that evolve with time are assigned to the influencing relationship edges, and a process credibility label is marked for each weight.

4. The knowledge graph-based industrial data analysis and mining system according to claim 1, characterized in that: Starting with the batch with abnormal DC internal resistance, the specific process of traversing the coating-rolling-slitting process path in the knowledge graph in reverse along the material flow direction to locate all upstream process parameters affecting DC internal resistance is as follows: Starting with the abnormal batch, perform a reverse breadth-first traversal along the material flow edges in the knowledge graph. During the traversal, record all process nodes and corresponding equipment nodes. For each process node, all process parameters that are directly or indirectly related to DC internal resistance are retrieved through the parameter influence relationship edge, and a parameter-process-equipment mapping relationship table is established.

5. The knowledge graph-based industrial data analysis and mining system according to claim 4, characterized in that: The specific process of generating a root cause parameter list by weighting and fusing the Z-score deviation of coating speed and areal density parameters with the dynamic weights of the corresponding influence relationships in the knowledge graph, and sorting them according to the overall suspicion level, is as follows: For each process parameter in the parameter-process-equipment mapping table, obtain its characteristic value sequence during the current abnormal batch production period, and calculate the Z-score statistical deviation of each parameter characteristic value sequence relative to the recent golden batch benchmark distribution; The current dynamic weight value of the influence relationship edge between each parameter and DC internal resistance is read from the knowledge graph. The Z-score deviation, dynamic weight value and process distance attenuation factor are weighted and calculated by a multi-factor fusion algorithm to generate a comprehensive suspicion score for each parameter. The parameters are sorted in descending order based on the overall suspicion score, and a priority list of root cause parameters is generated. The process and equipment information associated with each parameter item are labeled.

6. The knowledge graph-based industrial data analysis and mining system according to claim 5, characterized in that: The logic for instantiating the abnormal parameters in the root cause parameter list into abnormal event nodes and constructing the abnormal event graph connecting the coating and rolling processes is as follows: Extract the attribute information of each abnormal parameter from the root cause parameter list, including parameter type, degree of deviation and the process equipment to which it belongs. Use the degree of parameter deviation as the node weight, instantiate each abnormal parameter as an abnormal event node, and mark the node with a timestamp and process context. Based on the process flow sequence relationship between coating and rolling processes, directed connection edges are established between consecutive abnormal event nodes. The time of cross-process abnormal event nodes is aligned through time windows to construct a coating-rolling abnormal event propagation graph.

7. The knowledge graph-based industrial data analysis and mining system according to claim 6, characterized in that: The specific process of identifying recurring star and chain propagation topologies in a graph using a subgraph isomorphism detection algorithm to form a matching standardized fault mode is as follows: Candidate propagation subgraphs are extracted from the propagation graph of historical anomalies, a candidate subgraph pattern library is established, and topological structure analysis is performed on each candidate subgraph to identify star-shaped radiation structures and chain-like transmission structures. The structural similarity between candidate subgraphs and known fault modes is calculated using a subgraph isomorphism detection algorithm. The detected repeated propagation topologies are then normalized and encoded to form standardized fault modes that include topological structure, propagation path, and typical parameter features. Establish a fault mode matching index and assign a unique mode identifier and matching threshold to each standardized fault mode.

8. The industrial data analysis and mining system based on knowledge graphs according to claim 1, characterized in that: The specific process of performing the FCI causal discovery algorithm on the coating and rolling parameter set involved in the failure modes to construct a partially directed acyclic graph is as follows: Key parameter sets for coating and rolling processes are extracted from standardized failure modes, a multivariate time series dataset is constructed, and conditional independence tests are performed on the parameter sets to determine the dependencies between variables. The FCI causal discovery algorithm is executed to explore the causal structure between parameters in the presence of potential confounding factors; Based on the separation set and collision node identification results, a partial directed acyclic graph is constructed to clarify the direct causal, indirect causal, and pseudo-correlation relationships between parameters; Confidence assessments were performed on the causal arrows in the diagram, and statistically significant causal connections were retained.

9. The industrial data analysis and mining system based on knowledge graphs according to claim 8, characterized in that: The logical process of comparing the directional consistency of the causal arrows identified in the graph with the corresponding influence relationship edges in the knowledge graph, and generating weight adjustment instructions based on the comparison results to drive the dynamic update of the time-varying weight attributes of the influence relationship edges is as follows: The causal arrows in some directed acyclic graphs are matched one by one with the influence relationship edges stored in the knowledge graph. For causal relationships with consistent directions, a weight enhancement instruction is generated to increase the dynamic weight of the corresponding edge according to the confidence ratio. For causal relationships with conflicting directions, a weight correction instruction is generated to reduce the weight of the corresponding edge or recalibrate it according to the degree of conflict. For newly discovered causal relationships driven by data, new influence relationship edges are created in the knowledge graph and marked as algorithmic discoveries. A weight update log is established to record the basis and time of each adjustment, thereby enabling traceable and dynamic updates of the influence relationship edge weights.

Images