A Knowledge Graph-Based Causal Analysis Method for Mining Equipment Failures

Abstract

This invention discloses a causal analysis method for mining equipment failures based on a knowledge graph, relating to the field of industrial diagnostic technology. The method includes: collecting core data packets, establishing a unified correspondence and aligning them temporally to obtain time-series data, data quality identifiers, and operating condition labels; constructing an equipment failure knowledge graph based on the above data and generating a causal model template by combining structural data; weighting candidate causal relationships according to operating condition and quality weights to obtain operating condition parameters; filtering candidate edges based on operating condition activation consistency and cross-operating condition stability to obtain a structural causal model and causal relationship reliability; forming a confidence ranking through online inference; selecting executable diagnostic actions for verification when uncertain and outputting the root cause. By constraining the causal structure search space with allowed edge sets, prohibited edge sets, and mandatory edge sets, the method suppresses the spread of false causality and improves the robustness of multi-operating-condition inference. Simultaneously, it introduces causal relationship reliability weighting and a falsification loop to improve the deterministic location of the cause.

Description

Technical Field

[0001] This invention relates to the field of industrial diagnostic technology, and in particular to a method for causal analysis of mining equipment failures based on knowledge graphs. Background Technology

[0002] Mining equipment operates in environments characterized by high dust, strong vibration, and heavy load fluctuations. The condition assessment of its critical components typically combines structural data, maintenance work orders, historical fault records, and sensor and operational parameter data. In engineering practice, feature extraction and threshold discrimination, fault tree / empirical rule-based systems based on expert knowledge, and probabilistic graphical or regression models incorporating historical samples are commonly employed to identify abnormal symptoms, locate fault locations, and generate handling suggestions. This allows for closed-loop management that can be integrated with maintenance processes.

[0003] However, existing methods still have two limitations: First, they mostly generate candidate relations based on correlation or local rules, lacking structural constraints such as allowed edge sets, prohibited edge sets, and mandatory edge sets, as well as the limitation of the effective range of edges by the working condition activation relation, which easily introduces spurious causality that is inconsistent with the direction of structural transmission; Second, when the root cause order is close, there is a lack of consistency falsification of diagnostic actions and response data, making it difficult to complete disambiguation within the safety boundary. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a knowledge graph-based method for causal analysis of mining equipment failures, which solves the problems in existing technologies, such as the lack of structural constraints on candidate causal relationships leading to the inclusion of false causal factors, and the lack of a consistent falsification mechanism when the root cause is uncertain, making it difficult to disambiguate.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a knowledge graph-based method for causal analysis of mining equipment failures. The method includes: collecting core data packets, establishing a unified correspondence and performing time alignment, obtaining aligned time-series data and data quality identifiers, and generating operating condition labels; constructing an equipment failure knowledge graph based on the time-series data, data quality identifiers, and operating condition labels, and generating a causal model template according to the relationship constraints between structural data and the equipment failure knowledge graph; calibrating candidate causal relationships using the causal model template, time-series data, and operating condition labels, obtaining operating condition parameters, and filtering out causal relationships that do not meet the constraints, obtaining the structural causal model and the reliability of the causal relationships; performing online root cause inference based on the structural causal model and the reliability of the causal relationships, obtaining the confidence ranking of candidate failure mechanisms, performing executable filtering, executing the filtered diagnostic actions, and outputting the final root cause.

[0007] As a preferred embodiment of the knowledge graph-based causal analysis method for mining equipment failures described in this invention, the specific steps for obtaining aligned time-series data and data quality identifiers are as follows: Collect core data packets; match the equipment number, measurement point type, and installation location fields; when multiple candidate components appear, select a unique component according to a fixed priority rule to establish a unified correspondence; pair the entire data source event sequence with the candidate shift amount and the anchor point sequence, and map it to a common time axis according to the nearest neighbor method to complete time alignment, obtaining aligned time-series data; convert missing, jump, and drift identifiers into quality weights according to a fixed mapping rule to obtain data quality identifiers; the fixed priority rule prioritizes consistent installation locations, followed by the shortest structural connection link, and then those with consistent historical fault directions.

[0008] As a preferred embodiment of the knowledge graph-based causal analysis method for mining equipment failures described in this invention, the generation of working condition labels includes: extracting fixed fields from operating parameters and normalizing them according to a unified scale to form a sequence of working condition feature vectors; performing coarse and fine working condition determinations to generate working condition labels; the fine working condition determination includes: within the determined coarse working condition category, setting a fine working condition prototype vector, using quality weights to weight and accumulate the distance within the time window, and selecting the fine working condition number with the smallest distance as the working condition label.

[0009] As a preferred embodiment of the knowledge graph-based causal analysis method for mining equipment failures described in this invention, the construction of the equipment failure knowledge graph includes: uniquely attaching each sensor tag to the corresponding component node, and uniquely attaching each maintenance record and historical failure event to the corresponding component node; calculating observation features and generating observation feature nodes within a fixed sliding window for the aligned time-series data of each sensor; extracting failure mechanism terms and handling action terms, generating failure mechanism nodes from the mechanism terms, and generating action nodes from the handling action terms; using component nodes, observation feature nodes, failure mechanism nodes, operating condition nodes, and action nodes as nodes in the equipment failure knowledge graph; and solidifying the relationship types into graph edges and writing them into the equipment failure knowledge graph to construct the equipment failure knowledge graph.

[0010] As a preferred embodiment of the knowledge graph-based causal analysis method for mine equipment failures described in this invention, the following steps are included: generating a causal model template by generating candidate causal edges according to causal rules; including candidate causal edges that violate structural direction, energy transfer direction, and control link direction, and candidate causal edges that lead to mechanisms in the opposite direction of observation features in a prohibited edge set; including links that are explicitly stated to be transmitted by structural data and action-to-mechanism edges that are confirmed to be effective by operation and maintenance records in a mandatory edge set; for candidate causal edges that are not directly prohibited and are not mandatory, calculating the candidate causal edge admission score under the constraints of current operating condition label and data quality identifier; when the candidate causal edge admission score is greater than or equal to the admission threshold, determining that the current candidate causal edge meets the admission requirements and is written into the allowed edge set; otherwise, it is not written into the allowed edge set; the allowed edge set, prohibited edge set, mandatory edge set, and the structural equation skeleton corresponding to each type of variable are combined to form a structurally constrained causal model template; the causal rules include candidate causal edges from mechanism nodes to observation feature nodes, from upstream component states to downstream component states, and from actions to mechanisms or from actions to component states, with operating condition activation relationships serving only as the effective condition for candidate causal edges.

[0011] As a preferred embodiment of the knowledge graph-based causal analysis method for mining equipment failures described in this invention, the method for obtaining operating condition parameters includes: weighting and recording operating condition labels according to their quality weights to form an effective sample set under each operating condition; based on the structural equation skeleton, reading the set of parent variables and the lag time for each target variable, and extracting the values ​​of the target variable and the parent variables on the common time axis to form a sample record; and using weighted least squares to solve for the coefficients in the effective sample set of each operating condition to obtain the operating condition parameters.

[0012] As a preferred embodiment of the knowledge graph-based causal analysis method for mining equipment failures described in this invention, the acquisition of the structural causal model and causal relationship reliability includes: filtering out candidate edges that do not meet the consistency of operating condition activation based on the operating condition activation relationship; comparing the operating condition parameters under multiple active operating conditions for the remaining candidate edges; filtering out candidate edges that do not meet the stability across operating conditions; calculating the causal relationship reliability for the finally retained causal edges; and combining the causal structure, variable definitions, variable mappings, operating condition parameters, and causal relationship reliability to form a structural causal model.

[0013] As a preferred embodiment of the knowledge graph-based causal analysis method for mining equipment failures described in this invention, the online root cause inference includes: calculating the current observation value and deviation for each sensor feature; identifying features with deviations exceeding the upper quantile of the healthy segment distribution as anomalous evidence; weighting the contribution of each causal edge according to the reliability of the causal relationship; using the set of anomalous evidence as the observation condition, calculating the explanatory power and forming a confidence level for each mechanism node related to the anomaly; sorting all candidate failure mechanisms from high to low explanatory power to obtain a confidence ranking; when the difference between the explanatory power of the top two candidate failure mechanisms is less than the difference threshold or the top two alternate in leading positions, determining that the current root cause inference is uncertain and triggering an active verification process; otherwise, directly outputting the failure mechanism with the highest confidence level as the final root cause.

[0014] As a preferred embodiment of the knowledge graph-based causal analysis method for mining equipment failures described in this invention, the computational explanatory power includes: selecting effective causal edges in the structural causal model, extracting directed paths pointing to abnormal observation features, and performing forward inference according to path parameters to obtain the predicted deviation; comparing the predicted deviation with the real-time observed deviation to obtain the path explanatory strength; multiplying the reliability of causal relationships on the path as the path weight and reducing the weight of evidence contributions according to quality weight; taking the path with the largest product of path weight and path explanatory strength as the explanatory contribution for each abnormal observation feature, and summing all explanatory contributions as the explanatory power.

[0015] As a preferred embodiment of the knowledge graph-based causal analysis method for mining equipment failures described in this invention, the active verification process includes: retrieving action nodes in the equipment failure knowledge graph that have an action intervention relationship with candidate failure mechanisms and filtering to obtain a set of executable diagnostic actions; when the set of executable diagnostic actions contains multiple actions, determining the diagnostic priority according to the number of candidate failure mechanisms and the mean absolute difference of predicted responses, and selecting a single action to execute; executing the diagnostic action to obtain the measured response sequence and the predicted response sequence inferred from the structural causal model; calculating the normalized difference index based on the measured response sequence, the predicted response sequence, and the quality weight, and using the exponential decay mapping index consistency score, eliminating candidate failure mechanisms whose index consistency scores are lower than the consistency threshold, and outputting the root causes of the remaining candidate failure mechanisms in order of index consistency scores.

[0016] The beneficial effects of this invention are as follows: By constructing a set of allowed edges, a set of prohibited edges, and a set of required edges, and using the activation relationship of operating conditions as the effective condition for the edges to form a structurally constrained causal model template, the search space for constrained causal structures is realized and the spread of pseudo-causality is suppressed; by constructing a set of allowed edges, a set of prohibited edges, and a set of required edges, and using the activation relationship of operating conditions as the effective condition for the edges to form a structurally constrained causal model template, the search space for constrained causal structures is realized and the spread of pseudo-causality is suppressed; by obtaining operating condition parameters through weighted parameter estimation based on operating condition labels and quality weights, and by filtering out candidate edges based on the consistency of operating condition activation and cross-operating condition stability, a stable and reliable structural causal model under multiple operating conditions is obtained and the robustness of root cause inference is improved. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a knowledge graph-based method for causal analysis of mining equipment failures.

[0019] Figure 2 A flowchart for data access and time alignment.

[0020] Figure 3 A flowchart for constructing a fault knowledge graph and generating a structurally constrained causal model template.

[0021] Figure 4 This is a flowchart for online root cause inference. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for causal analysis of mining equipment failures based on knowledge graphs, including the following steps: S1. Collect core data packets, establish a unified correspondence and perform time alignment, obtain aligned time-series data and data quality identifiers, and generate operating condition labels.

[0026] Access structural data, operation and maintenance records, historical faults, and real-time data; combine the equipment number and component number in the structural data as the primary key to form a component directory; standardize the part descriptions in the operation and maintenance records and historical faults, and uniquely match them with the component directory; match the sensor label using the equipment number, measurement point type, and installation location fields; when multiple candidate components appear, select a unique component according to a fixed priority rule; map each work order to a unique component number based on the equipment number and standardized part description, forming a correspondence between work order number and component number, and establishing a unified correspondence.

[0027] The fixed priority rule is as follows: first, the installation location is consistent; second, the structural connection link is shortest; and third, the historical fault points are consistent.

[0028] A common time axis with a fixed sampling period is generated within the analysis interval. Events for alignment (e.g., alarm edges, shutdown moments, current surges, and valve position step points) are extracted from real-time data. Corresponding event timestamps are extracted from historical fault and maintenance records as alignment anchors. Candidate shift amounts are enumerated one by one at fixed time steps within the allowed shift range. The entire data source event sequence is paired with the candidate shift amounts and then paired with the anchor sequence. Specifically, for each shifted event moment, the anchor with the smallest time difference is selected from the anchor sequence, and the minimum time difference is calculated. When the minimum time difference does not exceed the tolerance threshold, it is recorded as a valid match. At the same time, the number of all valid matches is used as a consistency score. The candidate shift amount with the largest consistency score is taken as the offset of the data source. The timestamps of all sampled records in the data source are uniformly corrected. The timestamp of each sampled record is replaced with the sum of the original timestamp and the offset. Based on the corrected timestamps, the sampled values ​​are mapped to the common time axis in a nearest neighbor manner to complete the time alignment, resulting in aligned time series data.

[0029] It should be noted that the tolerance threshold is determined by statistically analyzing the absolute value distribution of the timestamp difference between sample pairs that are historically known to be of the same type of event, and taking the high quantile of the distribution as the tolerance threshold, such as the time difference corresponding to the first 95% of positions as the tolerance threshold, with a value range of [0.5s, 60s].

[0030] If there is no valid sample value at a certain moment on the common time axis, it is marked as missing; if the difference between adjacent samples exceeds the jump threshold, it is marked as jump; if the sensor is in an uncalibrated state or deviates from the healthy baseline for a long time, it is marked as drift; the missing, jump, and drift labels are converted into a quality weight sequence between 0 and 1 according to a fixed mapping rule to obtain the data quality label.

[0031] It should be noted that the jump threshold is determined separately for each sensor feature. The distribution of adjacent sampling differences is statistically analyzed within the historical period when the sensor is confirmed to be in a healthy state, and the high quantile of the absolute value distribution is taken as the jump judgment threshold, such as the 90th percentile. The fixed mapping rule uses the penalty coefficients corresponding to the three types of identifiers—missing, jumping, and drift—as input. A value of 1 is assigned for no identifier, 0 for missing identifiers, and the low quantile value between 0 and 1 (e.g., 0.2) is assigned for no missing identifiers and no jumping identifiers but drifting identifiers. A constant value between the quality weights corresponding to jumping and missing identifiers (e.g., 0.6) is assigned. When multiple identifiers are present, the minimum value in the candidate weight value set corresponding to each identifier triggered at the current moment is used to generate the quality weight. The health baseline is obtained by selecting historical data segments where the device is confirmed to be in normal operation and statistically averaging the values ​​of the same sensor feature under operating conditions.

[0032] Fixed fields (e.g., speed, load, current, valve position, pump speed) are extracted from operating parameters and normalized according to a uniform scale to form a sequence of operating condition feature vectors. Coarse operating condition determination is performed, classifying time points into one of the following categories based on fixed threshold rules: start-stop, no-load, steady-state, or heavy-load. Fine operating condition determination is then performed. Within the determined coarse operating condition categories, fine operating condition prototype vectors are defined. Mass weights are used to weight and accumulate distances within a time window, and the fine operating condition number with the smallest distance is selected as the label. The expression is: ; in, Indicates time Detailed working condition labels, This indicates the total number of detailed operating conditions. This indicates the length of the integration window for determining operating conditions. Indicates time Quality weight, Indicates time The working condition feature vector, Indicates the first A detailed prototype vector for each working condition.

[0033] It should be noted that the fixed threshold rule is to first select a labeled sample segment in the historical operating data of the target equipment where the operating conditions can be determined (for example, start-stop is determined by start-stop command and zero speed segment, no-load is determined by load close to zero and speed stable segment, heavy load is determined by load close to rated and current close to rated segment, steady state is determined by continuous segment where speed and load fluctuations are small). The operating parameters (speed, load, current, valve position and pump speed) in the sample segment are statistically distributed and the distribution boundary point between adjacent operating conditions is taken as the segmentation threshold (take the line connecting the medians of the two types of samples that can be used) to obtain the operating condition judgment threshold. This is achieved by selecting multiple stable operating conditions with clear start and end points from historical data (e.g., entering steady state after start-stop, entering heavy load from steady state), statistically analyzing the convergence time distribution required for the operating parameters to enter the stable range of the new operating condition from the start of switching, and taking the corresponding percentile of the convergence time as the length of the operating condition determination integration window, such as the 95th percentile.

[0034] S2. Construct an equipment fault knowledge graph based on time-series data, data quality identifiers, and operating condition labels, and generate a causal model template based on the relationship constraints between the structural data and the equipment fault knowledge graph.

[0035] The nodes in the graph are fixed into five categories: component nodes, observation feature nodes, fault mechanism nodes, operating condition nodes, and action nodes. Component nodes include equipment number, component number, component name, installation location, and parent component number. Observation feature nodes include sensor label, feature name, calculation window length, and unit. Fault mechanism nodes include mechanism name, mechanism category, and description of the direction of influence. Operating condition nodes include operating condition number, coarse operating condition category, and fine operating condition number. Action nodes include action number, action type (diagnostic action and maintenance action), description of the object of action, and description of execution constraints.

[0036] Each sensor tag is uniquely attached to the corresponding component node, and each maintenance record and historical fault event is uniquely attached to the corresponding component node. For records that cannot be uniquely matched, a unique attachment object is selected according to a fixed priority rule: priority is given to consistent installation location, shortest structural connection link, and consistent historical confirmation location.

[0037] For the aligned time-series data of each sensor, the observation features are calculated and observation feature nodes are generated within a fixed sliding window. The observation features include one or more of the following: mean, variance, rate of change, kurtosis, and bandpass energy percentage. Each observation feature node records the calculation window length and the corresponding sensor label, and the observation feature node is connected to the corresponding component node through component and monitoring relationship.

[0038] Fault mechanism terms and handling action terms are extracted from operation and maintenance records and historical fault confirmation conclusions. Different expressions are merged into unified mechanism names and unified action names using a fixed dictionary and fixed synonym merging rules. Fault mechanism nodes are generated from mechanism terms and action nodes are generated from handling action terms. Three types of relationships, namely mechanism and action, mechanism and phenomenon (characteristic) and operating condition and mechanism activation, are selected as candidate relationships.

[0039] The relationship types are solidified into graph edges and written into the graph. Specifically, the monitoring relationship is the monitoring relationship between component nodes and observed feature nodes; the mechanism representation relationship is the representation relationship between fault mechanism nodes and observed feature nodes, including directional fields (e.g., increase, decrease, increased fluctuation); the operating condition activation relationship is the activation relationship between operating condition nodes and fault mechanism nodes, used to define under what operating conditions a fault mechanism or causal relationship between mechanisms is considered valid. The operating conditions required for the mechanism to be established are first determined from the fault mechanism data of the mining equipment, and then samples that meet the operating conditions are screened using historical operating data for verification. After verification, the operating condition node and fault mechanism node corresponding to the condition are solidified into an activation relationship; the action intervention relationship is the intervention relationship between action nodes and fault mechanism nodes or component nodes, used to represent objects that can be directly changed by actions; the structural connection relationship is the connection relationship established between component nodes based on structural data (e.g., transmission connection, hydraulic connection, electrical control connection), and the connection type and direction field are recorded to obtain an equipment fault knowledge graph containing five types of nodes and fixed relationship types.

[0040] Based on the equipment fault knowledge graph, candidate causal edges are generated according to causal rules. Specifically, for any mechanism node and the observed feature node, a candidate causal edge from mechanism to feature is generated; for adjacent components in the structural connection relationship with energy, material or control transmission direction, a candidate causal edge from upstream component state to downstream component state is generated; for action intervention relationship, a candidate causal edge from action to mechanism or action to component state is generated; for operating condition activation relationship, it is only used as the effective condition of the edge, and the operating condition is not directly used as the causal edge object to be learned (to avoid treating the operating condition as a common cause and causing false causal diffusion).

[0041] Generate a set of prohibited edges and a set of required edges. Specifically, candidate edges that violate the structural direction, energy transfer direction, or control link direction are directly included in the set of prohibited edges; edges whose observed characteristics lead to mechanisms are included in the set of prohibited edges; links that are clearly stated to be transmitted by structural data (e.g., vibration transmission caused by coaxial rigid connections) are included in the set of required edges; and action-to-mechanism edges whose action intervention objects are clear and whose operation and maintenance records confirm their effectiveness are included in the set of required edges.

[0042] For candidate edges that are not directly prohibited and are not mandatory, under the constraints of the current operating condition label and data quality identifier, we examine whether the edge is consistent with the influence direction described in the equipment fault knowledge graph, and map the consistency to an admission degree of 0 to 1. We calculate the admission score of the candidate causal edge to determine whether to include it in the set of allowed edges. The expression is as follows: ; in, This represents the admission score for candidate causal edges. Indicates the start time of the statistics. Indicates the end time of the statistics. The candidate edge endpoint variable represents the time... The normalized change The candidate edge origin variable at time t is represented. The normalized lag change Indicates the sign constant of direction. Indicates the time lag.

[0043] It should be noted that, This is achieved by taking the time from the aligned endpoint variable time series. The current value is the difference between the current value and the baseline value obtained from the historical healthy segment under the same working condition, and the difference is normalized to be the normalized change at the current moment. The method involves selecting sample segments from historical data that have been confirmed to contain candidate causal edges under the same working conditions, calculating the time alignment correlation between the starting variable change sequence and the ending variable change sequence, and using the time shift that maximizes the similarity between the two sequences as the delay estimate of the sample segment. The median of the delay estimates for multiple sample segments is then used as the lag time. The direction sign constant takes the value of 1 or -1, determined by the mechanism representation relationship in the equipment fault knowledge graph. If the equipment fault knowledge graph record shows that an increase in the starting point leads to an increase in the ending point, then... If the record shows that an increase in the starting point leads to a decrease in the ending point, then the value is 1. It is -1.

[0044] When the admission score of a candidate causal edge is greater than or equal to the admission threshold, the current candidate causal edge is determined to meet the admission requirements and is written into the allowed edge set; otherwise, it is not written into the allowed edge set.

[0045] It should be noted that the admission threshold is determined by collecting candidate causal edge samples that have been manually confirmed as eligible and ineligible for admission in the mining area, calculating the admission score for each sample and arranging them in ascending order, and taking the arithmetic mean of the minimum score sequence of eligible samples and the maximum score sequence of ineligible samples as the admission threshold, with a value ranging from 0 to 1.

[0046] The allowed edge set, the prohibited edge set, the required edge set, and the structural equation skeleton corresponding to each type of variable are combined to form a structurally constrained causal model template.

[0047] It should be noted that the structural equation skeleton refers to fixing only the parent-child dependency structure and variable correspondence of each variable in the causal model template.

[0048] S3. Using causal model templates, time series data, and operating condition labels, perform parameter calibration on candidate causal relationships, obtain operating condition parameters, and screen out causal relationships that do not meet the constraints, thereby obtaining the reliability of the structural causal model and causal relationships.

[0049] For each working condition label, filter the continuous sample segments in the aligned time series data whose labels are equal to the working condition, and delete the sampling points with a quality weight of 0; then, weight the remaining sampling points according to their quality weights to form an effective sample set under the working condition.

[0050] For each target variable in the causal model template, the set of parent variables corresponding to the target variable is read according to the structural equation skeleton, and it is determined whether each parent variable contains lagged terms and their corresponding lagged times. The target variable at time t is then extracted on the common time axis. The values ​​of and the parent variables at time or The values ​​of are used to form sample records that correspond one-to-one with the timestamp. In each working condition sample set, the quality weight of each sample record is used as the error weight of the sample. The coefficients corresponding to the parent variable are solved by weighted least squares so that the sum of the squares of the differences between the predicted and observed values ​​of all samples is minimized in the sense of weight. The parameter estimates of each candidate causal relationship under the working condition are obtained, and the parameter estimates obtained by each candidate edge within the working condition are recorded as the working condition parameters of the current edge.

[0051] For each candidate edge, based on the working condition activation relationship, the working condition labels that have a working condition activation relationship with the candidate edge are recorded as the active working condition set, and the remaining working condition labels are recorded as the inactive working condition set. For each working condition in the active working condition set, the working condition parameters corresponding to the candidate edge are read and an active working condition parameter sequence is formed. For each working condition in the inactive working condition set, the working condition parameters corresponding to the candidate edge are read and an inactive working condition parameter sequence is formed. When the absolute value of the working condition parameters of a candidate edge in the active working condition set is not greater than the zero decision threshold, or when there is a working condition in the inactive working condition set with a working condition parameter absolute value greater than the zero decision threshold, the candidate edge is determined not to satisfy the working condition activation consistency and is removed from the causal structure.

[0052] It should be noted that the zero decision threshold is determined by using the corresponding percentile of the absolute value set of the inactive operating condition parameter sequence as the zero decision threshold, such as the 95th percentile.

[0053] For candidate edges that have been screened out by working condition activation consistency, the working condition parameter sequences of the candidate edges are compared only within the set of activated working conditions. The dispersion of the working condition parameter sequences is calculated using the interquartile range and used as the cross-working condition volatility of the candidate edges. When the cross-working condition volatility of a candidate edge is greater than the stability threshold, the candidate edge is determined to not meet the cross-working condition stability and is removed from the causal structure; otherwise, it is retained.

[0054] It should be noted that the stability threshold is determined by using the corresponding percentile of the cross-condition volatility set of the retained edges in the dataset of historically confirmed root causes as the stability threshold, such as the 95th percentile.

[0055] For each causal edge that is ultimately retained, the reliability of the causal relationship is generated by combining the activation consistency bias and the cross-condition parameter dispersion, and the expression is: ; in, Indicates the reliability of causal relationships. This indicates the activation consistency penalty amount. This represents the penalty for dispersion across operating conditions.

[0056] It should be noted that, For each candidate causal edge, the set of active and inactive operating conditions is determined based on the operating condition activation relationship in the equipment fault knowledge graph. The parameter estimates of the candidate causal edge under each operating condition are obtained from the sample sets of the two types of operating conditions. Then, the absolute values ​​of the parameter estimates in the inactive operating conditions are averaged, and the ratio of the average value to the average absolute value of the parameter estimates in the active operating conditions is used as the activation consistency penalty. The process involves determining the set of active operating conditions for each candidate causal edge based on the operating condition activation relationships in the equipment fault knowledge graph. Then, for each active operating condition, the operating condition parameters of the candidate causal edges are estimated from the sample set, resulting in a sequence of operating condition parameters. The dispersion of this sequence is then calculated and used as the basis for further analysis. The dispersion is fixed and calculated using the interquartile range.

[0057] The structural causal model is composed of causal structure (set of edges), variable definition and mapping, operating parameters, and the reliability of causal relationships corresponding to each edge.

[0058] It should be noted that the training process of the structural causal model is as follows: using the edge set and structural equation skeleton given by the causal model template as constraints, the aligned time series data is divided into several working condition sample sets according to the working condition labels, and a quality weight is generated for each sampling point. In each working condition sample set, for each objective variable equation allowed by the causal model template, the objective quantity is taken as the value of that variable at time step [time value missing]. The observed value, taking the input quantity as its parent variable at time t. or The values ​​of are obtained by using the least square error fitting with quality weights to obtain the coefficients corresponding to each parent variable, thereby obtaining the parameters of each candidate causal edge under the current operating condition. The significance consistency of the operating condition parameters in the active and inactive operating conditions is checked according to the operating condition activation relationship of the equipment fault knowledge graph. The cross-operating condition stability is evaluated according to the magnitude of parameter fluctuations between different active operating conditions. Edges that do not meet the activation consistency or stability are eliminated. The operating condition parameters and causal relationship reliability of the retained edges are solidified together into the calibrated structural causal model.

[0059] S4. Based on the structural causal model and the reliability of causal relationships, perform online root cause inference, obtain the confidence ranking of candidate failure mechanisms, perform executability filtering, execute the filtered diagnostic actions, and output the final root cause.

[0060] The aligned real-time sensors and operating parameters are continuously updated using a fixed-length sliding time window (e.g., a sliding window length of 60s and an update step size of 1s). For each sensor feature, the current observation value (window mean) and deviation (the difference between the observation value and the healthy baseline) are calculated within the window. Features with deviations exceeding the upper quantile of the healthy segment distribution (e.g., the 99th percentile) are identified as abnormal evidence. The abnormal evidence and the corresponding component number are mapped to the observation feature nodes and component nodes in the equipment fault knowledge graph through a unified correspondence, forming an evidence set for the current abnormal event.

[0061] Under the activation conditions corresponding to the current working condition label, causal edges and corresponding working condition parameters that are effective in the working condition are selected from the structural causal model. The contribution of each causal edge is weighted according to the reliability of the causal relationship. Using the abnormal evidence set as the observation condition, the explanatory power of each mechanism node related to the abnormality is calculated and a confidence level is formed. All candidate fault mechanisms are sorted from high to low explanatory power to obtain the confidence ranking of candidate fault mechanisms.

[0062] The computational explanatory power is achieved by selecting causal edges that are effective under the current operating conditions in the calibrated structural causal model, extracting local directed paths (with a path length not exceeding 3) from the current mechanism to each anomalous observation feature, reading the parameters of each causal edge on the path under the current operating conditions, using the current deviation of the mechanism variable as input, sequentially tracing forward along the path to obtain the predicted deviation for each anomalous feature, comparing the predicted deviation with the real-time observed deviation, forming the explanatory strength of the current path for anomalous evidence based on directional consistency and the magnitude of the difference, multiplying the causal relationship reliability of each causal edge on the path as the path weight, and using quality weights to reduce the weight of low-confidence evidence, taking only the path with the largest contribution for each anomalous feature as the explanatory contribution of the current mechanism, and summing the explanatory contributions of all anomalous features to obtain the explanatory power of the current mechanism.

[0063] When the confidence ranking of candidate fault mechanisms shows that the difference in explanatory power between the top two is less than the difference threshold, or when the top two alternate in leading the ranking within multiple consecutive windows (e.g., 3), the current root cause inference is determined to be uncertain and an active verification process is triggered. Otherwise, the fault mechanism with the highest confidence is directly output as the final root cause.

[0064] It should be noted that the gap threshold is calculated by taking the percentile of the difference between the explanatory power of the true root cause and the second-highest explanatory power for each event in the historically confirmed root cause dataset, and taking the 95th percentile as the gap threshold. This ensures that the vast majority of true root cause events can still maintain the correct lead when no active verification is required.

[0065] In the equipment fault knowledge graph, action nodes that have an action intervention relationship with candidate fault mechanisms are retrieved to form a diagnostic action candidate set. For each candidate action, the target object, execution constraints, and interlocking conditions are read. The current operating parameters are compared with the safety boundary table, and actions that do not meet the safety boundary are eliminated to obtain an executable diagnostic action set. When the executable diagnostic action set contains multiple actions, a unique action is selected and executed according to the diagnostic priority.

[0066] The diagnostic priority includes, for each action, counting the number of candidate fault mechanisms that can be directly affected through action intervention relationships based on the equipment fault knowledge graph, and prioritizing the action with the smallest number of candidate fault mechanisms; if multiple actions have the same number of candidate fault mechanisms, then for each action, assume that each candidate mechanism is valid and use a structural causal model to deduce the predicted response of key observation features under the current action, calculate the average absolute difference between the predicted responses of each pair of candidate mechanisms, and prioritize the action with the largest average absolute difference.

[0067] It should be noted that the safety boundary table is formed by the permissible operating conditions range in the manufacturer's technical manual and commissioning and acceptance data of the mining equipment, the interlocking and limiting parameters configured in the field control system (PLC), and the limiting conditions for diagnostic operations in the operation and maintenance procedures; key observation features refer to the set of observation features that are directly related to the object of the diagnostic action and are marked as observable changes by the candidate mechanism in the equipment fault knowledge graph; the predicted response refers to the sequence of changes that the key observation features will appear in the action window under the assumption that a certain candidate mechanism is valid and a diagnostic action is executed.

[0068] Diagnostic actions are performed, during which aligned real-time data is continuously collected and a response data segment is formed. Key observation features in the response data segment are calculated into a measured response sequence. For each candidate fault mechanism, forward inference is performed using a structural causal model to obtain a predicted response sequence. The mass weight corresponding to each sampling point within the action window is read. For each sampling point within the action window, the absolute difference between the measured response value and the predicted response value is calculated. The absolute difference is weighted by the mass weight and then accumulated to obtain the weighted difference accumulation. The mass weights within the action window are accumulated to obtain the effective weight accumulation. The weighted difference is then calculated. The ratio of heterogeneous cumulant to effective weighted cumulant is used as the normalized difference index of candidate mechanisms under diagnostic actions. The normalized difference index is mapped using exponential decay according to a monotonically decreasing relationship. Specifically, in the historical confirmed root cause dataset, the median of the normalized difference index set of non-true root cause mechanisms is statistically obtained as the scaling parameter. The normalized difference index of the current candidate mechanism is scaled according to the scaling parameter, and the negative power operation of the natural exponent is performed to obtain the consistency score. This ensures that the consistency score monotonically decreases as the normalized difference index increases and the consistency score is limited to between 0 and 1, thus obtaining the indicator consistency score.

[0069] The consistency score of the indicator is compared with the indicator threshold. Candidate fault mechanisms with consistency scores lower than the indicator threshold are judged as inconsistent and removed from the candidate set. The candidate fault mechanisms that are not removed are sorted from high to low according to their consistency scores, and the one ranked first is output as the root cause.

[0070] It should be noted that the indicator threshold is calculated by taking each sample from the historically confirmed root cause dataset, using the true root cause mechanism as the candidate mechanism and the non-true root cause mechanism as the candidate mechanism, and calculating their respective indicator consistency scores. This results in a set of true root cause indicator consistency scores and a set of non-true root cause indicator consistency scores. The percentile of the latter set is used as the indicator threshold, such as the 99th percentile.

[0071] In summary, this invention achieves a constrained causal structure search space and suppresses pseudo-causal diffusion by constructing allowed edge sets, prohibited edge sets, and required edge sets, and using operating condition activation relationships as edge activation conditions to form a structurally constrained causal model template; it also achieves a constrained causal structure search space and suppresses pseudo-causal diffusion by constructing allowed edge sets, prohibited edge sets, and required edge sets, and using operating condition activation relationships as edge activation conditions to form a structurally constrained causal model template; and it obtains operating condition parameters by weighted parameter estimation based on operating condition labels and quality weights, and filters candidate edges based on operating condition activation consistency and cross-operating condition stability, thereby obtaining a stable and reliable structural causal model under multiple operating conditions and improving the robustness of root cause inference.

[0072] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A causal analysis method for mining equipment failures based on knowledge graphs, characterized in that: include, Collect core data packets, establish a unified correspondence and perform time alignment, obtain aligned time-series data and data quality identifiers, and generate operating condition labels; A knowledge graph of equipment failures is constructed based on time-series data, data quality identifiers, and operating condition labels, and a causal model template is generated based on the relationship constraints between the structural data and the knowledge graph of equipment failures. By using causal model templates, time series data, and operating condition labels, the parameters of candidate causal relationships are calibrated, operating condition parameters are obtained, and causal relationships that do not meet the constraints are screened out, thereby obtaining the reliability of the structural causal model and causal relationships. Online root cause inference is performed based on structural causal model and causal relationship reliability. The confidence ranking of candidate failure mechanisms is obtained, and executability filtering is performed. The filtered diagnostic actions are executed, and the final root cause is output.

2. The knowledge graph-based causal analysis method for mining equipment failures as described in claim 1, characterized in that: The specific steps for obtaining the aligned time-series data and data quality identifier are as follows: Collect core data packets and match the fields of device number, measurement point type and installation location. When multiple candidate components appear, select a unique component according to a fixed priority rule and establish a unified correspondence. The entire event sequence from the data source is paired with the candidate shift amount and the anchor sequence, and then mapped to the common time axis according to the nearest neighbor method to complete the time alignment, resulting in aligned time series data. The missing, jump, and drift identifiers are converted into quality weights according to a fixed mapping rule to obtain data quality identifiers. The fixed priority rule is as follows: first, the installation location is consistent; second, the structural connection link is shortest; and third, the historical faults are consistent.

3. The knowledge graph-based causal analysis method for mining equipment failures as described in claim 2, characterized in that: The generation of operating condition labels includes extracting fixed fields from operating parameters and normalizing them according to a uniform scale to form a sequence of operating condition feature vectors. Perform coarse and fine working condition assessments and generate working condition labels; The detailed working condition determination includes setting a detailed working condition prototype vector within the determined coarse working condition category, using quality weights to weight and accumulate the distance within the time window, and selecting the detailed working condition number with the smallest distance as the working condition label.

4. The knowledge graph-based causal analysis method for mining equipment failures as described in claim 3, characterized in that: The construction of the equipment fault knowledge graph includes uniquely attaching each sensor tag to the corresponding component node, and uniquely attaching each maintenance record and historical fault event to the corresponding component node. For the aligned time-series data of each sensor, the observation features are calculated and observation feature nodes are generated within a fixed sliding window; Extract fault mechanism terms and handling action terms, generate fault mechanism nodes from the mechanism terms, and generate action nodes from the handling action terms; Component nodes, observation feature nodes, fault mechanism nodes, operating condition nodes, and action nodes are used as nodes in the equipment fault knowledge graph. The relationship types are solidified as graph edges and written into the device fault knowledge graph to construct the device fault knowledge graph.

5. The knowledge graph-based causal analysis method for mining equipment failures as described in claim 4, characterized in that: The generated causal model template includes generating candidate causal edges according to causal rules; Candidate causal edges that violate the structural direction, energy transfer direction, and control link direction, and candidate causal edges that lead to the mechanism in the opposite direction of observation features, are included in the prohibited edge set. Links that are clearly expressed by structural data and actions confirmed by operation and maintenance records to the mechanism are included in the mandatory edge set. For candidate causal edges that are not directly prohibited and are not mandatory, calculate the admission score of the candidate causal edge under the constraints of the current working condition label and data quality identifier. When the admission score of the candidate causal edge is greater than or equal to the admission threshold, determine that the current candidate causal edge meets the admission requirements and write it into the allowed edge set; otherwise, do not write it into the allowed edge set. The allowed edge set, the prohibited edge set, the required edge set, and the structural equation skeleton corresponding to each type of variable are combined to form a structurally constrained causal model template. The causal rules include candidate causal edges from mechanism nodes to observed feature nodes, from upstream component states to downstream component states, and from actions to mechanisms or from actions to component states. The operating condition activation relationship is only used as the effective condition for candidate causal edges.

6. The knowledge graph-based causal analysis method for mining equipment failures as described in claim 5, characterized in that: The acquisition of working condition parameters includes weighting and recording the working condition labels according to their quality weights to form an effective sample set under the working conditions. Based on the structural equation skeleton, the set of parent variables and lag time are read for each target variable, and the values ​​of the target variable and the parent variables are extracted on the common time axis to form a sample record; The coefficients are solved by weighted least squares in the effective sample set of each working condition to obtain the working condition parameters.

7. The knowledge graph-based causal analysis method for mining equipment failures as described in claim 6, characterized in that: The acquisition of the structural causal model and causal relationship reliability includes: filtering out candidate edges that do not meet the working condition activation consistency based on the working condition activation relationship, and comparing the working condition parameters under multiple active working conditions for the remaining candidate edges to filter out candidate edges that do not meet the cross-working condition stability. The reliability of causal relationships is calculated for the final retained causal edges, and the causal structure, variable definitions, variable mappings, operating parameters, and causal relationship reliability are combined to form a structural causal model.

8. The knowledge graph-based causal analysis method for mining equipment failures as described in claim 7, characterized in that: The online root cause inference includes calculating the current observation value and deviation for each sensor feature, and determining features whose deviation exceeds the upper quantile of the healthy segment distribution as abnormal evidence. The contribution of each causal edge is weighted according to the reliability of the causal relationship. Using the set of anomalous evidence as the observation condition, the explanatory power of each mechanism node related to the anomalous is calculated and a confidence level is formed. All candidate failure mechanisms are sorted from highest to lowest explanatory power to obtain a confidence ranking. When the confidence ranking of candidate fault mechanisms shows that the difference in explanatory power between the top two is less than the difference threshold or the top two alternate in leading, the current root cause inference is determined to be uncertain and an active verification process is triggered; otherwise, the fault mechanism with the highest confidence is directly output as the final root cause.

9. The knowledge graph-based causal analysis method for mining equipment failures as described in claim 8, characterized in that: The computational interpretation capability includes selecting effective causal edges in the structural causal model, extracting directed paths that point to abnormal observation features, and performing forward inference based on path parameters to obtain the predicted deviation. The path interpretation strength is obtained by comparing the predicted deviation and the real-time observed deviation. The reliability of causal relationships on the path is multiplied together as the path weight, and the contribution of evidence is reduced according to the quality weight. For each anomalous observation feature, the path with the largest product of path weight and path explanatory strength is taken as the explanatory contribution, and the sum of all explanatory contributions is taken as the explanatory power.

10. The knowledge graph-based causal analysis method for mining equipment failures as described in claim 9, characterized in that: The active verification process includes retrieving action nodes in the equipment fault knowledge graph that have an action intervention relationship with the candidate fault mechanism and filtering to obtain a set of executable diagnostic actions; When the set of executable diagnostic actions contains multiple actions, the diagnostic priority is determined by the number of candidate fault mechanisms and the mean absolute difference of the predicted response, and a single action is selected for execution. Perform diagnostic actions to obtain measured response sequences and predicted response sequences inferred from structural causal models; Based on the measured response sequence, the predicted response sequence, and the quality weight, a normalized difference index is calculated and an exponential decay mapping index consistency score is used. Candidate fault mechanisms with index consistency scores below the consistency threshold are eliminated, and the remaining candidate fault mechanisms are sorted by index consistency score to output the root cause.

Images