An industrial equipment operation and maintenance fault tracing multi-granularity causal diagram learning method

Abstract

This application relates to a multi-granularity causal graph learning method for fault tracing in industrial equipment operation and maintenance. The method includes: constructing a three-layer system of macro-mechanism, meso-process, and micro-indicators, obtaining the mapping relationship between the node sets and hierarchical variables at each layer; acquiring domain text corpus, extracting and granularly labeling causal statements using a large language model, and mapping and weighting them to form a set of weighted language candidate edges at each layer; using a scoring-search algorithm based on the Bayesian information criterion to learn the macro-layer causal structure, obtaining macro-level reliable causal edges consistent with the language and data; transforming these edges into micro-layer cross-cluster candidate edges and directional constraints through a projection mechanism, forming a micro-level search space constraint; learning the micro-layer causal structure under this constraint, obtaining a directed acyclic causal graph to realize fault tracing and propagation path analysis. This method can meet the practical needs of complex industrial system fault tracing for high-precision and high-efficiency causal structure learning.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance technology, and in particular to a multi-granularity cause-effect graph learning method for tracing the source of faults in industrial equipment operation and maintenance. Background Technology

[0002] In the field of intelligent manufacturing and industrial equipment operation and maintenance, fault tracing and fault propagation path analysis are core links to ensure stable operation of production lines and reduce equipment downtime losses. With the development of industrial Internet of Things (IoT) technology, massive sensors have been deployed in complex industrial systems such as production lines, rotating machinery, semiconductors, and chemical processes, which can collect multi-dimensional micro-indicator data. At the same time, industrial scenarios have accumulated a wealth of textual materials such as FMEA reports, maintenance work orders, equipment operating procedures, and fault mechanism analysis documents, which contain a large amount of macro-causal knowledge summarized by experts. How to integrate multi-granular data and textual knowledge to achieve accurate causal structure learning has become a key technical challenge for industrial fault tracing. Traditional causal structure learning methods only model at a single granular level, constructing a directed acyclic graph with micro-variables such as sensor indicators as nodes, and searching for causal structures in the graph space through conditional independence tests or scoring functions. However, industrial fault tracing scenarios generally have problems such as high variable dimensionality, scarce fault samples, and high data noise. Such methods are prone to problems such as unstable determination of the existence and direction of causal edges, huge search space, and high computational cost. Moreover, they cannot effectively absorb and integrate macro-mechanism knowledge from the text, making it difficult to meet the demand for strong explanatory power of causal structures in industrial maintenance decisions.

[0003] To alleviate computational and sample pressures in high-dimensional scenarios, multi-granularity causal structure learning methods have been applied to industrial data analysis. These methods compress a large number of microscopic sensor indicators into a small number of macroscopic representations through clustering and other means, and then jointly learn causal structures on multiple granular variables, reflecting the structural relationships between the macroscopic and microscopic levels to some extent. However, the macroscopic variables in these methods are mostly automatically abstracted by the algorithm, lacking semantic correspondence with the natural hierarchical system of "mechanism, process, and indicator" in industrial scenarios, resulting in insufficient interpretability. Furthermore, the structures learned at the macroscopic level are difficult to directly translate into explicit search constraints at the microscopic level, making it impossible to systematically connect with causal statements in industrial texts. At the same time, these methods often rely on large-scale deep model training, requiring high computing power, complex parameter tuning, and incurring high deployment and auditing costs, which contradicts the actual needs of industrial scenarios for lightweight and interpretable algorithms.

[0004] In recent years, utilizing large language models to assist causal discovery has become a research hotspot in industrial fault diagnosis. Related methods extract causal information from industrial texts using large language models, combining this as a priori constraint with traditional statistical causal discovery algorithms to improve the accuracy of fault tracing. However, these methods generally implicitly assume that the causal statements in the text and the observed data variables are at the same granularity level. Industrial textual knowledge is mostly macroscopic descriptions of mechanisms, components, and processes, while the data side consists of massive microscopic sensor indicators. The same macroscopic fault mechanism often corresponds to multiple combinations of microscopic indicators, and direct mapping easily leads to the problem of "semantic correctness but granularity mismatch," causing a superficial conflict between linguistic priors and statistical evidence. Existing methods lack effective means to reconcile such conflicts, often simply discarding linguistic knowledge or forcibly covering data with language, resulting in a waste of valuable causal knowledge in the industrial field. Furthermore, they only treat domain constraints as soft weights or rule adjustments, failing to structurally shrink the search space of the microscopic causal graph, making it difficult to meet the actual needs of complex industrial system fault tracing for high-precision and high-efficiency causal structure learning. Summary of the Invention

[0005] Based on this, it is necessary to provide a multi-granularity causal graph learning method for industrial equipment operation and maintenance fault tracing that can meet the practical needs of high-precision and high-efficiency causal structure learning for fault tracing in complex industrial systems.

[0006] A multi-granularity cause-effect graph learning method for tracing the causes of operational and maintenance faults in industrial equipment, the method comprising:

[0007] Step 1: Obtain micro-observation data of industrial equipment operation and maintenance, and construct a multi-granularity hierarchical system based on the micro-observation data, consisting of a macro-mechanism layer, a meso-process layer, and a micro-indicator layer, to obtain the node set of each granularity layer and the variable mapping relationship between the layers; Step 2: Obtain domain text corpus of industrial equipment operation and maintenance, use large language model to extract causal statements from domain text corpus, perform granular labeling on each causal statement and map and weight the causal concepts in the causal statement to the corresponding granularity layer variable clusters to form a weighted language candidate edge set for each granularity layer. Step 3: Use a scoring-search causal discovery algorithm based on Bayesian information criterion to learn the causal structure of the node data and weighted language candidate edge set at the macroscopic mechanism layer, and obtain a macroscopic credible causal edge set that is consistent with the language and data. Step 4: Based on the variable mapping relationship between levels, the macroscopic credible causal edge set is transformed into the cross-cluster candidate edge set and directional constraints of the microscopic index layer through the projection mechanism from macroscopic edges to microscopic candidate edge set, thus forming the search space constraints of the microscopic index layer. Step 5: Perform causal structure learning on the observation data of the micro-indicator layer, the set of weighted language candidate edges, and the search space constraints to obtain the directed acyclic causal graph of the micro-indicator layer. Based on the directed acyclic causal graph of the micro-indicator layer, realize fault tracing and fault propagation path analysis for industrial equipment operation and maintenance.

[0008] The aforementioned multi-granularity causal graph learning method for tracing faults in industrial equipment operation and maintenance, based on micro-observation data, constructs a multi-granularity hierarchical system comprising a macro-mechanism layer, a meso-process layer, and a micro-indicator layer, and establishes hierarchical variable mapping. This breaks the limitations of traditional single-granularity modeling, allowing causal learning to naturally align with the "mechanism, process, and indicator" hierarchy of the industrial system, and laying a data association foundation for subsequent cross-layer constraints. It utilizes a large language model to extract causal statements from domain texts, perform granular labeling, and assign weights to the mapping from concepts to variable clusters, forming a weighted language candidate edge set for each granularity layer. This effectively solves the problem of granular alignment between macro-language knowledge and micro-data, realizing the structured transformation and effective utilization of causal knowledge in industrial texts. A rating-search algorithm based on the Bayesian information criterion is used to learn the macro-layer causal structure, obtaining macro-credible causal edges consistent with language and data. Leveraging the characteristics of fewer nodes and ease of learning stability with fewer fault samples in the macro-layer, it integrates data statistical regularities and domain language priors, outputting... The causal edges possess both scientific validity and domain relevance. Then, through a projection mechanism, macroscopically reliable causal edges are transformed into a set of cross-cluster candidate edges and directional constraints at the microscopic level. This structurally prunes the microscopic search space, completely resolving the problems of large search space and high computational cost caused by traditional methods relying solely on soft weight constraints and blind microscopic search, significantly improving search efficiency. Under these constraints, causal structure learning at the microscopic level yields an accurate directed acyclic causal graph. This integrates microscopic data, macroscopic projection constraints, and microscopic linguistic priors, allowing microscopic causal edges to be traced back to macroscopic mechanism edges. This enables precise and interpretable localization of fault root causes and propagation paths. Furthermore, the complete system formed by the interconnected steps provides a foundation for subsequent granular adaptive transfer mechanisms, interpreting surface conflicts between language and data as granularity mismatches. Ultimately, without increasing the sample size, it achieves a dual improvement in microscopic causal structure learning accuracy and search efficiency, adapting to the core needs of high-dimensionality, low-sample, and highly interpretable industrial scenarios. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a multi-granularity cause-effect graph learning method for tracing the source of operational and maintenance faults in industrial equipment, as shown in one embodiment. Figure 2 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] In one embodiment, such as Figure 1 As shown, a multi-granularity cause-effect graph learning method for tracing the causes of operational and maintenance faults in industrial equipment is provided, including the following steps: Step 1: Obtain micro-observation data of industrial equipment operation and maintenance, and construct a multi-granularity hierarchical system based on the micro-observation data, consisting of a macro-mechanism layer, a meso-process layer, and a micro-indicator layer, to obtain the node set of each granularity layer and the variable mapping relationship between the layers.

[0012] Industrial equipment includes one or more of the following: production lines, rotating machinery, semiconductor manufacturing equipment, and chemical process equipment. Microscopic observation data refers to time-series or batch data collected by sensors during the operation and maintenance of industrial equipment, specifically including one or more of the following: vibration RMS, bearing temperature, lubricating oil pressure, coolant flow rate, spindle current, cavity pressure, yield KPI, and alarm code count. The macroscopic mechanism layer consists of nodes related to the fault mechanisms and failure modes of industrial equipment. The mesoscopic process layer consists of nodes related to the subsystems, process sections, operation levels, or machine / station levels of industrial equipment. The microscopic indicator layer consists of nodes related to sensors and KPIs. The variable mapping relationship between the layers includes aggregation mapping and inverse mapping, realizing the corresponding association between nodes at each granular level and microscopic indicators, which is the basis for subsequent cross-layer projection.

[0013] Step 2: Obtain domain text corpus of industrial equipment operation and maintenance, extract causal statements from the domain text corpus using a large language model, perform granular labeling on each causal statement and map and weight the causal concepts in the causal statement to the corresponding granularity layer variable clusters, forming a weighted language candidate edge set for each granularity layer.

[0014] The domain text corpus is a high-value causal statement library accumulated in industrial scenarios, which is the core source for extracting causal knowledge. Causal statements are statements in the text that contain cause-effect logical relationships. Granularity labeling determines the macro-mechanism layer, meso-process layer, or micro-indicator layer to which each causal statement belongs, avoiding forcibly constraining causal statements at the mechanism layer to specific sensor indicator pairs. The concept-to-variable cluster mapping and weighting associates the abstract concepts in the causal statements with semantic variable clusters at each granularity layer and assigns corresponding weights. The weighted language candidate edge set transforms causal knowledge into prior information that can be used for causal structure learning at each granularity layer, laying the foundation for subsequent fusion of data and language knowledge.

[0015] Step 3: Use a scoring-search causal discovery algorithm based on Bayesian information criterion to learn the causal structure of the node data and weighted language candidate edge set at the macroscopic mechanism layer, and obtain a macroscopic credible causal edge set that is consistent with the language and data.

[0016] The scoring-search causal discovery algorithm based on Bayesian information criteria is a causal structure learning method that combines statistical features of data with prior linguistic knowledge. It can take into account both the statistical regularity of the data and the expert knowledge of the domain. The number of nodes in the macro-mechanism layer is on the order of dozens, far less than that in the micro-indicator layer, making it easier to learn the causal structure stably when there are limited industrial fault samples. The set of macro-credible causal edges that are consistent with language and data are causal edges that are supported by prior linguistic knowledge and verified by statistical data. They are consistent with the domain mechanism and match the actual data features, and are the core basis for subsequent constraints on the micro-level search.

[0017] Step 4: Based on the variable mapping relationship between levels, the macroscopic credible causal edge set is transformed into the cross-cluster candidate edge set and directional constraints of the microscopic index layer through the projection mechanism from macroscopic edges to microscopic candidate edge set, thus forming the search space constraints of the microscopic index layer.

[0018] The projection mechanism is the core means of transforming macroscopic causal relationships into microscopic explicit search constraints, realizing the structured constraint of macroscopic mechanisms on microscopic indicators; the cross-cluster candidate edge set is the combination of microscopic indicators corresponding to macroscopic credible causal edges, and the directional constraint locks the causal direction of microscopic cross-cluster edges; the search space constraints of the microscopic indicator layer structurally trim the search range of the microscopic causal graph, avoiding blind search among all sensor indicators, and is the key to solving the problem of high-dimensional variable search in industrial scenarios.

[0019] Step 5: Perform causal structure learning on the observation data of the micro-indicator layer, the set of weighted language candidate edges, and the search space constraints to obtain the directed acyclic causal graph of the micro-indicator layer. Based on the directed acyclic causal graph of the micro-indicator layer, realize fault tracing and fault propagation path analysis for industrial equipment operation and maintenance.

[0020] Causal structure learning under search space constraints can significantly reduce the number of candidate edges and the number of search operation evaluations at the micro level, improving computational efficiency and statistical stability. The directed acyclic causal graph at the micro index layer clearly presents the causal relationships and directions between micro indicators such as sensors and KPIs. Based on this causal graph, it is possible to trace back from abnormal indicators to the upstream causal chain to locate the root cause of the fault and identify the fault propagation path. At the same time, all micro cross-cluster edges can be traced back to the corresponding macro mechanism edges, which can provide interpretable maintenance and intervention guidance for industrial equipment operation and maintenance.

[0021] The aforementioned multi-granularity causal graph learning method for tracing faults in industrial equipment operation and maintenance, based on micro-observation data, constructs a multi-granularity hierarchical system comprising a macro-mechanism layer, a meso-process layer, and a micro-indicator layer, and establishes hierarchical variable mapping. This breaks the limitations of traditional single-granularity modeling, allowing causal learning to naturally align with the "mechanism, process, and indicator" hierarchy of the industrial system, and laying a data association foundation for subsequent cross-layer constraints. It utilizes a large language model to extract causal statements from domain texts, perform granular labeling, and assign weights to the mapping from concepts to variable clusters, forming a weighted language candidate edge set for each granularity layer. This effectively solves the problem of granular alignment between macro-language knowledge and micro-data, realizing the structured transformation and effective utilization of causal knowledge in industrial texts. A rating-search algorithm based on the Bayesian information criterion is used to learn the macro-layer causal structure, obtaining macro-credible causal edges consistent with language and data. Leveraging the characteristics of fewer nodes and ease of learning stability with fewer fault samples in the macro-layer, it integrates data statistical regularities and domain language priors, outputting... The causal edges possess both scientific validity and domain relevance. Then, through a projection mechanism, macroscopically reliable causal edges are transformed into a set of cross-cluster candidate edges and directional constraints at the microscopic level. This structurally prunes the microscopic search space, completely resolving the problems of large search space and high computational cost caused by traditional methods relying solely on soft weight constraints and blind microscopic search, significantly improving search efficiency. Under these constraints, causal structure learning at the microscopic level yields an accurate directed acyclic causal graph. This integrates microscopic data, macroscopic projection constraints, and microscopic linguistic priors, allowing microscopic causal edges to be traced back to macroscopic mechanism edges. This enables precise and interpretable localization of fault root causes and propagation paths. Furthermore, the complete system formed by the interconnected steps provides a foundation for subsequent granular adaptive transfer mechanisms, interpreting surface conflicts between language and data as granularity mismatches. Ultimately, without increasing the sample size, it achieves a dual improvement in microscopic causal structure learning accuracy and search efficiency, adapting to the core needs of high-dimensionality, low-sample, and highly interpretable industrial scenarios.

[0022] In one embodiment, a multi-granularity hierarchical system is constructed based on microscopic observation data, comprising a macroscopic mechanism layer, a mesoscopic process layer, and a microscopic indicator layer, including: Correlation coefficients or mutual information matrices are calculated for micro-observation data. Hierarchical clustering is then performed using the process topology of industrial equipment and expert rules to aggregate micro-indicators into semantically clear variable clusters. Linear combination or principal component analysis methods are then used to construct node data for the meso-process and macro-mechanism layers from these variable clusters. Simultaneously, an aggregation mapping relationship is established from macro-mechanism and meso-process layer nodes to micro-indicator layer variables. and set of micro variables ,in, Indicates the index of micro variables to the first... k Mapping of granular layer nodes, The text indicates the first k Granularity layer nodes The corresponding set of micro variables, For the number of micro-indicators, For the first k The set of nodes in the granularity layer.

[0023] In a specific embodiment, the observation data matrix is ​​denoted as... The rows correspond to samples, and the columns correspond to sets of micro-indicator variables. The set of granular layers is denoted as Smaller indexes indicate higher-granularity layers, with the lowest layer being the micro-level. Each layer is given an aggregation mapping from micro-variable indices to nodes within that layer. For any high-level node ( ), and define its corresponding set of micro variables as The set of textual causal statements is denoted as in These are phrases representing the concepts of cause and effect, respectively. This is the original sentence. Through a large language model and rules, we obtain the alignment mapping from concepts to sets of micro-variables. Give each statement a granular label In each granularity layer (k Define a cause-effect graph on the ) in This is the set of nodes at this layer. It is a set of directed acyclic edges. For any node... The set of its parent nodes is denoted as .

[0024] Specifically, the correlation between indicators in the microscopic observation data is first analyzed using correlation coefficients or mutual information matrices. Then, hierarchical clustering is performed by combining existing process topologies and expert rules such as P&ID, equipment BOM, process flow diagrams, and FMEA classifications for industrial equipment. Compared to pure algorithm clustering, this method makes the clustering results more closely resemble actual industrial scenarios, ensuring that the aggregated variable clusters have clear industrial semantics. For example, lubricating oil pressure, oil temperature, and flow rate are aggregated into a "lubrication subsystem cluster," and bearing temperature and vibration are aggregated into a "bearing subsystem cluster." Subsequently, linear methods such as linear combination or principal component analysis are used to construct mesoscopic and macroscopic node data, avoiding reliance on large-scale deep models and reducing computing power and parameter tuning costs. Simultaneously, an aggregation mapping relationship is established. and set of micro variables This approach achieves precise mapping from high-level nodes to micro-level indicators, providing a data association foundation for the subsequent implementation of cross-layer projection mechanisms. By constructing a multi-granularity hierarchical system that naturally matches the "mechanism, process, and indicator" hierarchy of industrial scenarios, each granularity layer node possesses clear industrial semantics, and the mapping relationships between levels are clear. This not only solves the problem of insufficient interpretability of macro-variable semantics in traditional multi-granularity methods but also adapts to the industrial scenario's need for lightweight and interpretable algorithms.

[0025] In one embodiment, the domain text corpus includes one or more of the following: FMEA reports, maintenance work orders, equipment operating procedures, fault mechanism analysis reports, and inspection records. Each causal statement is granularly labeled, categorizing it as a macroscopic mechanism layer, a mesoscopic process layer, or a microscopic indicator layer. A weighted language candidate edge set is formed for each granularity layer, assigning weights to the causal edges within each granularity layer. ,in k Number the grain size layer. u , v These are the nodes corresponding to the granularity layer.

[0026] Specifically, the FMEA report includes a failure mode-cause-effect chain; maintenance work orders and inspection records document the actual fault repair and equipment inspection processes; equipment operating procedures clarify the operational logic of the process and equipment; and the fault mechanism analysis report is a professional summary of the fault occurrence mechanism by experts. All of the above textual corpora are readily available, high-quality sources of causal knowledge in industrial scenarios. Using a large language model and combining pattern rules, the domain textual corpora are processed sentence by sentence. First, "cause and effect" concept phrases are extracted to form causal statements. Then, with example-based prompts, the large language model completes granular annotation. Simultaneously, based on existing industrial vocabularies / synonyms / encoding systems (including part numbers, alarm codes, and measurement point names), the large language model completes semantic alignment and weighting of causal concepts to corresponding granularity-level variable clusters. Causal concepts and variable clusters with high matching degrees are assigned higher weights, while those with low matching degrees are assigned lower weights. Finally, based on the granular annotation results, the weighted causal edges are allocated to the corresponding granularity layers, forming a weighted language candidate edge set for each granularity layer. .

[0027] This method transforms unstructured industrial text corpora into a structured, granularly labeled set of weighted language candidate edges that can be directly used for causal structure learning. This enables the effective extraction and structured representation of domain causal knowledge. At the same time, the granular labeling solves the problem of semantically correct but granular mismatch, laying the foundation for subsequent integration of language and data knowledge.

[0028] In one embodiment, the scoring-search causal discovery algorithm based on the Bayesian information criterion includes: Construct an allowable edge mask for the macro-mechanism layer, encode the process sequence and energy and material flow of industrial equipment as hard constraints, calculate the total score of the macro-mechanism layer, and solve the directed acyclic graph that maximizes the total score by greedy-multiple-startup search or integer programming. The overall score is:

[0029] in, For the first Granularity layer cause-effect graph Overall score Rate the data. Scoring for language priors, For language prior weight coefficients, Take the number corresponding to the macroscopic mechanism layer.

[0030] Specifically, when constructing the allowable edge mask for the macro-mechanism layer, based on industrial common sense such as the actual process sequence, energy / material flow, and assembly sequence of industrial equipment, causal edge directions that are clearly inconsistent with reality are prohibited. For example, "consumption process" is prohibited from pointing to "raw material processing process," and "cooling subsystem" is prohibited from being driven by "bearing subsystem." These industrial common sense principles are encoded as hard constraints to reduce unreasonable causal structure searches. Subsequently, the numerical scores of the macro-mechanism layer are calculated respectively. and language prior score Through language prior weight coefficients Adjust the weighting of the two factors in the overall score. Setting ≤0 allows the total score to take into account both statistical data features and prior linguistic knowledge. Finally, a combination optimization method of greedy-multiple-start search or integer programming is adopted to solve the directed acyclic graph that maximizes the total score under the hard constraints of acyclic constraints and allowed edge masks. This directed acyclic graph is the causal structure of the macroscopic mechanism layer.

[0031] At the macro-mechanism level, the system integrates hard constraints of industrial common sense, statistical data patterns, and prior linguistic knowledge. The learned macro-causal structure is more in line with industrial realities. Furthermore, with fewer nodes at the macro level, the system exhibits higher learning stability in industrial scenarios with limited fault samples. At the same time, the output fault propagation mechanism chain / subsystem chain aligns with the understanding of industrial maintenance personnel.

[0032] In one embodiment, the data score is:

[0033] in, For the first k Grain size layer j Nodes Local data scoring based on Bayesian information criteria For the first k Grain size layer j The set of parent nodes of each node.

[0034] Specifically, local data scoring based on Bayesian information criteria. It was calculated using a linear regression model, with the first... k Grain size layer j Nodes The set of parent nodes is the dependent variable. A linear regression model is constructed for the independent variables. The model fit and complexity are comprehensively evaluated using the Bayesian information criterion to obtain the local data score for each node. A higher local data score indicates a stronger statistical significance of the causal relationship between that node and its parent node set. Subsequently, the local data scores of all nodes in the macro-mechanism layer are summed to obtain the overall data score for the macro-mechanism layer. A higher overall data score indicates a higher degree of matching between the causal graph structure and the statistical characteristics of the macroscopic data. By using linear regression scoring based on the Bayesian information criterion, a quantitative assessment of the matching degree between the causal structure and the data is achieved. The use of a linear model ensures the interpretability of the score, adapting to the interpretability requirements of industrial scenarios. Simultaneously, the Bayesian information criterion effectively balances model fit and complexity, avoiding overfitting.

[0035] In one embodiment, the language prior score is:

[0036] in, For the first k The set of weighted language candidate edges at the granular level [] indicates an indicator function. For the first k The set of directed acyclic edges in a granular layer causal graph. Represents the edges of the weighted language candidate edge set.

[0037] Specifically, in actual calculations, edges consistent with the causal statements in the text are assigned positive weights, while those opposite are assigned negative weights. This allows the language prior score to effectively distinguish between causal edges that conform to and violate domain knowledge. This embodiment achieves a quantitative evaluation of the matching degree between causal structures and domain language prior knowledge, transforming unstructured domain knowledge into computable numerical scores, enabling causal structure learning to effectively absorb expert knowledge and practical operational experience from the industrial field.

[0038] In one embodiment, based on the variable mapping relationship between levels, a projection mechanism from macroscopic edges to microscopic candidate edge sets is used to transform the macroscopic credible causal edge set into a cross-cluster candidate edge set and directional constraints at the microscopic index layer, including: For a causal edge at the macroscopic level ( The set of candidate edges projected onto the microscopic level by causal edges is defined as follows:

[0039] in, For macroscopic credible causal edge The corresponding set of microscopic cross-cluster candidate edges, , As a node at the macro-mechanism level, i For nodes Corresponding micro-variable index, j For nodes Corresponding micro-variable index, , They are nodes , The corresponding set of micro variables; The directional constraint allows only the micro-indicator layer to be oriented from... arrive Cross-cluster edges are allowed, but cross-cluster edges in the opposite direction are prohibited.

[0040] Specifically, based on the set of micro variables , Find the macroscopic credible causal edge Middle node , For each of the corresponding micro-indices, these micro-indices are paired to form a set of micro-cross-cluster candidate edges corresponding to the macro-credible causal edge. For example, the macroscopically reliable causal edge "lubrication subsystem → bearing subsystem" corresponds to a set of microscopic cross-cluster candidate edges that are all combinations of {lubricating oil pressure, oil temperature, flow rate} × {bearing temperature, vibration}; and based on the macroscopically reliable causal edge... The causal direction, locking the direction of the microscopic cross-cluster edge as from The corresponding micro indicators point to For the corresponding micro-indicators, reverse cross-cluster edges are prohibited; for example, bearing temperature is prohibited from pointing to lubricating oil pressure. The above operation is performed on all causal edges in the macro-credible causal edge set to obtain the cross-cluster candidate edge set and directional constraints of the micro-indicator layer.

[0041] Transforming credible causal relationships at the macro level into structural hard constraints at the micro level, rather than the soft weights / rules adjustments of traditional methods, fundamentally shrinks the search space at the micro level while retaining the flexibility of different micro-level implementation paths under the same macro mechanism. All micro-level cross-cluster edges can be traced back to the corresponding macro-level mechanism edges, thus improving the interpretability of the micro-level causal structure.

[0042] In one embodiment, the search space constraint for the micro-index layer is as follows: intra-cluster edges of the micro-index layer are allowed to be searched by default, while cross-cluster edges are only allowed in the cross-cluster candidate edge set. For internal search, all cross-cluster edges that are opposite in direction to the macroscopically reliable causal edges are directly marked as prohibited from search.

[0043] Specifically, the variable clusters at the micro-indicator layer are semantically defined clusters formed by clustering in step 1. Intra-cluster edges are causal edges between micro-indicators within the same variable cluster. These edges are allowed to be searched by default to preserve the internal correlations between indicators within the cluster. Cross-cluster edges are causal edges between micro-indicators from different variable clusters, and are only allowed in the cross-cluster candidate edge set formed by the projection of macroscopically reliable causal edges. The search is performed within the set, and cross-cluster edges outside the set are not included in the search scope. Simultaneously, all cross-cluster edges whose direction is opposite to that of the macroscopic reliable causal edge are directly marked as prohibited from search, and no edge addition, deletion, or reversal operations are performed in subsequent causal structure learning. By rigidly pruning the search space of the microscopic index layer through the above constraints, the number of candidate edges in the microscopic layer can be significantly reduced, greatly reducing computational overhead. At the same time, locking the direction of cross-cluster edges reduces the interference of noise and collinearity on the determination of microscopic causal direction, improving the statistical stability of causal structure learning.

[0044] In one embodiment, the causal structure learning process includes: The total prior weight matrix is ​​obtained by adding the projected prior weights of the micro-indicator layer to the prior weights of the micro-linguistic statements. The total score is then obtained by combining the Bayesian information criterion data scoring of the micro-observation data. Under search space constraints, a greedy-multiple-start search or integer programming approach is used to solve the directed acyclic causal graph that maximizes the total score. The projected prior weights are calculated for each cross-cluster candidate edge set. The positive preference weights assigned to the micro-edges within the micro-level are the weights assigned to the weighted language candidate edges of the micro-index layer.

[0045] Specifically, the projection prior weights assign positive preference weights to all micro-edges within the cross-cluster candidate edge set Γ(A,B), reflecting the prior constraints of macro-credible causal edges on the micro-level. The micro-linguistic statement prior weights are the weights of the weighted linguistic candidate edges of the micro-index layer formed in step 2, reflecting the constraints of prior linguistic knowledge at the micro-level. The two are added together according to their corresponding edges to obtain the total prior weight matrix of the micro-index layer, which integrates macro-projection constraints and micro-linguistic knowledge. Subsequently, a linear regression scoring method based on the Bayesian information criterion, consistent with the macro-level method, is used to calculate the data score of the micro-observation data. Combined with the total prior weight matrix, the total score of the micro-level is obtained, and the total score formula is consistent with that of the macro-level method. Finally, under the constraints of the search space, a combination optimization method of greedy-multiple-start search or integer programming is used to iteratively attempt to perform three types of operations on the allowed edges: "adding edges, deleting edges, and reversing edges." Under the premise of ensuring acyclicity and in-degree constraints, the operation that maximizes the increase in the total score is selected and applied until no further improvement is possible, thus solving for the directed acyclic causal graph that maximizes the total score.

[0046] Within the shrunk search space, the algorithm integrates micro-level data statistical features, macro-level projection priors, and micro-level language priors. This not only improves the efficiency and accuracy of learning micro-level causal structures but also makes the learned micro-level causal graphs highly interpretable. Furthermore, it employs scoring and search methods consistent with those used in the macro-level layer, ensuring the algorithm's uniformity and lightweight nature.

[0047] In one embodiment, the consistency of each causal statement at different granularity levels is calculated. When a causal statement lacks data support at the current labeling granularity level but exhibits high consistency at another granularity level, its granularity label is automatically adjusted and the weighted language candidate edge set at the corresponding granularity level is updated. Steps 3 to 5 are then re-executed to form a closed-loop optimization. The consistency calculation process is as follows:

[0048] In the formula, For the first m One causal statement, For the first m The causal statement in the first k The corresponding set of candidate edges at the granularity level. For the first k The set of directed acyclic edges in a granular layer causal graph. This represents the edges in the candidate edge set.

[0049] Specifically, firstly, for each causal statement... Based on its concept-variable mapping results at each granularity level, its value at each granularity level is determined. k The corresponding candidate edge set Subsequently, the consistency at each granularity level was calculated using a consistency formula. If the candidate edge set There exists a set of directed acyclic edges at least one of which appears in the set of directed acyclic edges at this granularity level. In this context, a consistency value of 1 indicates that the causal statement has data support at that granularity level, signifying high consistency. If no edge in the candidate edge set appears in the directed acyclic edge set, the consistency value is 0, indicating that the causal statement has no data support at that granularity level. When the consistency of a causal statement at its current granularity level is 0, while the consistency at another granularity level is 1, it is determined to be a granularity mismatch, and its granularity label is automatically adjusted to the granularity level with a consistency of 1. Subsequently, based on the adjusted granularity label, the weighted causal edges of the causal statement are removed from the original granularity level and added to the new granularity level, updating the weighted language candidate edge set at each granularity level. Finally, based on the updated weighted language candidate edge set, steps 3 to 5 are re-executed to complete a new round of causal structure learning, forming a closed-loop optimization of structure, granularity, and language.

[0050] Interpreting the apparent conflict between linguistic priors and data evidence as a granularity mismatch avoids the drawbacks of simply discarding linguistic knowledge or forcibly covering data with language, and maximizes the use of causal knowledge in the industrial field; at the same time, closed-loop optimization allows granular labels and causal structures to correct each other, continuously improving the accuracy of causal structure learning, and adapting to the actual characteristics of mixed hierarchical writing and inconsistent formats in industrial texts.

[0051] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0052] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 2 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a multi-granularity cause-effect graph learning method for tracing faults in industrial equipment operation and maintenance. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0053] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0054] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0055] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A multi-granularity cause-effect graph learning method for tracing the causes of operational and maintenance faults in industrial equipment, characterized in that, The method includes: Step 1: Acquire microscopic observation data for industrial equipment operation and maintenance. This microscopic observation data consists of time-series or batch data collected by sensors during the operation and maintenance process, including vibration RMS, bearing temperature, and lubricating oil pressure. Based on this microscopic observation data, construct a multi-granularity hierarchical system comprising a macroscopic mechanism layer, a mesoscopic process layer, and a microscopic indicator layer. The macroscopic mechanism layer comprises nodes related to the failure mechanisms and failure modes of the industrial equipment. The mesoscopic process layer comprises nodes related to subsystems, process sections, procedures, or machine / site-level processes of the industrial equipment. The microscopic indicator layer comprises nodes related to sensors and KPIs. Through hierarchical clustering, aggregate the microscopic indicators into semantically clear variable clusters to obtain the node sets of each granularity layer and the variable mapping relationships between layers. Step 2: Obtain domain text corpus of industrial equipment operation and maintenance, extract causal statements from the domain text corpus using a large language model, perform granular labeling on each causal statement and map and weight the causal concepts in the causal statement to the corresponding granular layer variable clusters to form a weighted language candidate edge set for each granular layer. Step 3: Use a scoring-search causal discovery algorithm based on Bayesian information criterion to learn the causal structure of the node data and weighted language candidate edge set of the macroscopic mechanism layer, and obtain a macroscopic credible causal edge set that is consistent with the language and data. Step 4: Based on the variable mapping relationship between the levels, the macroscopic reliable causal edge set is transformed into the cross-cluster candidate edge set and directional constraints of the microscopic index layer through the projection mechanism from macroscopic edges to microscopic candidate edge set, forming the search space constraints of the microscopic index layer; the search space constraints of the microscopic index layer are as follows: intra-cluster edges of the microscopic index layer are allowed to be searched by default, cross-cluster edges are only allowed to be searched within the cross-cluster candidate edge set, and all cross-cluster edges with directions opposite to those of the macroscopic reliable causal edges are directly marked as prohibited from being searched; Step 5: Perform causal structure learning on the observation data, weighted language candidate edge set, and search space constraints of the micro-indicator layer to obtain the directed acyclic causal graph of the micro-indicator layer. Based on the directed acyclic causal graph of the micro-indicator layer, realize fault tracing and fault propagation path analysis of industrial equipment operation and maintenance by tracing back the upstream causal chain of the abnormal indicators. Based on the aforementioned microscopic observation data, a multi-granularity hierarchical system is constructed, comprising a macroscopic mechanism layer, a mesoscopic process layer, and a microscopic indicator layer, including: Correlation coefficients or mutual information matrices are calculated for microscopic observation data. Hierarchical clustering is then performed using the process topology of industrial equipment and expert rules to aggregate microscopic indicators into semantically clear variable clusters. Linear combination or principal component analysis methods are then used to construct node data for the mesoscopic process layer and macroscopic mechanism layer from these variable clusters. Simultaneously, an aggregation mapping relationship is established from macroscopic mechanism layer and mesoscopic process layer nodes to microscopic indicator layer variables. and set of micro variables ,in, Indicates the index of micro variables to the first... k Mapping of granular layer nodes, Indicates the first k Granularity layer nodes The corresponding set of micro variables, For the number of micro-indicators, For the first k The set of nodes in the granularity layer; The scoring-search causal discovery algorithm based on the Bayesian information criterion includes: Construct an allowable edge mask for the macro-mechanism layer, encode the process sequence and energy and material flow of industrial equipment as hard constraints, calculate the total score of the macro-mechanism layer, and solve the directed acyclic graph that maximizes the total score by greedy-multiple-startup search or integer programming. The overall score is: in, For the first Granularity layer cause-effect graph Overall score Rate the data. Scoring for language priors, For language prior weight coefficients, Take the number corresponding to the macro-mechanism layer; The data score is: in, For the first k Grain size layer j Nodes Local data scoring based on Bayesian information criteria For the first k Grain size layer j The set of parent nodes of each node; The language prior score is: in, For the first k The set of weighted language candidate edges at the granular level [ ] indicates an indicator function. For the first k The set of directed acyclic edges in a granular layer causal graph. Denotes the edges of the weighted language candidate edge set. This indicates that causal edges are assigned weights.

2. The method according to claim 1, characterized in that, The domain text corpus includes one or more of the following: FMEA reports, maintenance work orders, equipment operating procedures, fault mechanism analysis reports, and inspection records; the granular labeling of each causal statement represents labeling each causal statement as a macroscopic mechanism layer, a mesoscopic process layer, or a microscopic indicator layer; the formation of weighted language candidate edge sets for each granularity layer involves assigning weights to the causal edges within each granularity layer. ,in k Number the grain size layer. u , v These are the nodes corresponding to the granularity layer.

3. The method according to claim 1, characterized in that, Based on the variable mapping relationship between the levels, the macroscopic reliable causal edge set is transformed into a cross-cluster candidate edge set and directional constraints of the microscopic index layer through a projection mechanism from macroscopic edges to microscopic candidate edge sets, including: For a causal edge at the macroscopic level The set of candidate edges projected onto the microscopic level by causal edges is defined as follows: in, For macroscopic credible causal edge The corresponding set of microscopic cross-cluster candidate edges, , As a node at the macro-mechanism level, i For nodes Corresponding micro-variable index, j For nodes Corresponding micro-variable index, , They are nodes , The corresponding set of micro variables; The directional constraint allows only elements from the micro-index layer to... arrive Cross-cluster edges are allowed, but cross-cluster edges in the opposite direction are prohibited.

4. The method according to claim 1, characterized in that, The process of learning the causal structure includes: The total prior weight matrix is ​​obtained by adding the projected prior weights of the micro-index layer to the prior weights of the micro-linguistic statements. The total score is then obtained by combining the Bayesian information criterion data scoring of the micro-observation data. Under search space constraints, a greedy-multiple-start search or integer programming approach is used to solve the directed acyclic causal graph that maximizes the total score. The projected prior weights are calculated for the cross-cluster candidate edge set. The positive preference weights assigned to the micro-edges within the micro-level, wherein the prior weights of the micro-language statements are the weights assigned to the weighted language candidate edges of the micro-index layer.

5. The method according to claim 1, characterized in that, The method further includes: Based on the alignment mapping and hierarchical aggregation mapping from causal concepts to micro-variables, the candidate edge set corresponding to each causal statement at each granularity level is determined; the consistency of each causal statement at different granularity levels is calculated. When a causal statement lacks data support at the current labeling granularity level but exhibits high consistency at another granularity level, its granularity label is automatically adjusted and the weighted language candidate edge set at the corresponding granularity level is updated. Steps 3 to 5 are then re-executed to form a closed-loop optimization. The consistency calculation process is as follows: In the formula, For the first m One causal statement, For the first m The causal statement in the first k The corresponding set of candidate edges at the granularity level. For the first k The set of directed acyclic edges in a granular layer causal graph. This represents the edges in the candidate edge set.

Images