Complex equipment fault predictive maintenance method and system based on knowledge graph

Abstract

The invention provides a complex equipment fault predictive maintenance method and system based on a knowledge graph, and relates to the technical field of industrial equipment fault prediction.The complex equipment fault predictive maintenance method comprises the steps that equipment standard data are obtained, a standard data triple is generated, the confidence coefficient is calculated, and a mining data triple is constructed in combination with structured time series data collected in real time, and the confidence coefficient is calculated; after mapping the two, calculating a comprehensive confidence coefficient, and screening high-quality triples to construct a knowledge graph; on the basis, a graph structure is constructed with entities as nodes and relations as edges, attribute features are embedded, an attention mechanism is adopted to aggregate neighbor node information, and node comprehensive feature vectors are generated; the node fault probability is predicted based on a logistic regression model, high-risk nodes are identified, and a root cause and potential influence are determined through forward and reverse tracking, so that an accurate maintenance strategy is formulated. According to the system, deep fusion and dynamic reasoning of multi-source data are realized, and the accuracy of fault prediction and the pertinence of maintenance are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault prediction technology, specifically to a method and system for predictive maintenance of complex equipment faults based on knowledge graphs. Background Technology

[0002] In the operation of complex equipment, timely fault prediction and maintenance are crucial to ensuring the safe and stable operation of the equipment. With the increasing complexity of industrial equipment, traditional fault diagnosis methods based on rules or single data sources are no longer sufficient to meet the high reliability and real-time maintenance requirements. Existing methods often rely on static knowledge or isolated data analysis, lacking the ability to deeply integrate and dynamically reason with multi-source heterogeneous data. This results in delayed fault warnings, high false alarm rates, and a lack of targeted maintenance strategies. Especially in dynamic operating environments, existing methods lack in-depth mining of the semantic relationships between data, making it difficult to accurately trace the root causes of faults and proactively warn of potential risks, thus limiting the real-time performance and reliability of predictive maintenance.

[0003] In the prior art, CN113887751A discloses a method and system for predictive maintenance of mechanical faults based on knowledge graphs, including: obtaining the names of the mechanical equipment components to be maintained and the information of the mechanical equipment to be maintained; initially constructing a mechanical fault knowledge graph for the mechanical equipment to be maintained from top to bottom; supplementing and improving the initially constructed mechanical fault knowledge graph in a bottom-up manner; selecting the corresponding dataset for the mechanical equipment to be maintained; assigning multiple features to each state of each component in the dataset according to the supplemented and improved mechanical fault knowledge graph; defining fuzzy variables, membership functions, and fuzzy rules according to the features and feature values; using the values ​​of the fuzzy variables as input parameters, performing fuzzification processing on the input parameters using the membership function, performing fuzzy inference on the fuzzified data using fuzzy rules, and defuzzifying the data obtained from the fuzzy inference to obtain the fault severity priority, thereby realizing predictive maintenance of mechanical faults.

[0004] The main problems with the above solutions are: the knowledge graph is not precisely filtered, which may introduce a lot of noise or low-quality data, affecting the accuracy of reasoning; and the fuzzy system only performs independent reasoning on multiple features of a single component, and cannot comprehensively judge the risk by analyzing the relationship between a node and its neighboring nodes, and has a weak ability to mine complex relationships; it only knows that a certain component has a high priority, but cannot quickly locate the root cause of the abnormality of the component, nor can it predict which other parts of the system will be affected by the failure of the component, and the prediction results are independent, lacking systematic root cause analysis and impact chain tracing.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predictive maintenance of complex equipment based on knowledge graphs, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A knowledge graph-based predictive maintenance method for complex equipment faults includes the following steps: Step 1: Obtain the standard data of complex equipment, define entity types and relation types respectively, generate standard data triples, and calculate the standard data confidence of each standard data triple; Step 2: Collect structured time-series data of complex equipment in real time, generate mining data triples, and calculate the mining data confidence of each mining data triple; Step 3: Map the mined data triples to the normalized data triples, construct mapping groups based on the mapped triples, calculate the overall confidence of the mapping groups, and pre-set a confidence threshold. Incorporate all mapping groups with an overall confidence higher than the confidence threshold into the knowledge graph, thereby completing the construction of the knowledge graph. Step 4: Using entities in the knowledge graph as nodes and relationships as edges, determine the nodes corresponding to device, component and sensor entities and set them as first-level nodes. Embed attribute features for each first-level node, embed the relationship type and confidence level for each edge, construct the graph structure, and generate the initial feature vector for each first-level node. Step 5: Calculate the attention coefficient between each primary node and its neighboring nodes, normalize the attention coefficient to generate attention weights, update the initial feature vector based on the attention weights between each primary node and its neighboring nodes to generate the comprehensive feature vector of the primary node, predict the failure probability of the primary node based on the comprehensive feature vector, identify high-risk nodes based on the failure probability, and perform forward and backward tracing of the graph structure to determine the maintenance strategy.

[0008] Furthermore, the specification data for complex equipment includes equipment drawings, fault manuals, and knowledge of the complex equipment from experts in the field; entity types include equipment, components, sensors, fault sources, fault symptoms, and maintenance measures; relationship types include composition, deployment, generation, cause, and mitigation; specification data triples include: components constitute equipment, components deploy sensors, components generate fault sources, fault sources cause fault symptoms, and maintenance measures mitigate fault symptoms. The specification data confidence score for each specification data triple is calculated using the following formula: ; in, Indicates the first The normalized confidence level of a normalized triplet. Indexes representing canonical triples Indicates the first Document authority score for a set of canonical data triples. Indicates the first Expert-authoritative scoring of a set of standardized data triples. Represents the consensus factor. The range of values ​​is , represent the weighting coefficients for the document authority score and the expert authority score, respectively. .

[0009] Furthermore, the entity types of the structured time-series data acquired in real time by complex equipment include: sensors deployed on the equipment, data acquired by the sensors, and fault symptoms identified from the data acquired by the sensors; the relationship types include cause and reflection; the mined data triples include: sensor data reflecting fault symptoms, and changes in sensor data causing changes in other sensor data; the formula for calculating the mined data confidence of each mined data triple is: ; in, Indicates the first Mining confidence of a triplet of data This represents the index for mining triples. Indicates the first The causal strength score of each data triplet is mined. Indicates the first The association strength score of each data triplet is mined. represent the weight coefficients of the causal strength score and the association strength score, respectively, and .

[0010] Furthermore, the principle underlying the mapping of mined data triples to canonical data triples is as follows: For each mined data triple whose data type is sensor data reflecting fault symptoms, query the deployment relationship in the normalized data triple. If a deployment relationship is found, and the fault symptom entity in the mined data and the fault source entity in the normalized data form a causal relationship, then it is considered a successful mapping. For a triple where changes in sensor data cause changes in other sensor data, query the deployment relationship of the two sensor entities in the normalized data triple respectively. If both are found to form a deployment relationship with the same component, then it is considered a successful mapping.

[0011] Furthermore, the principle for calculating the overall confidence level is as follows: A mapping group is generated by combining a canonical data triple with its successfully mapped mined data triples. The overall confidence of the mapping group is then calculated using the following formula: ; in, Indicates the first The overall confidence of each mapping group Indicates the index of the mapping group. Indicates the first In the mapping group, the first The normalized confidence level of a normalized triplet. Represents the index of the canonical data triple in the mapping group, and , This represents the number of canonical triples in a mapping group. Indicates the first Mining data confidence in each mapping group These represent the weights of the confidence scores for the normalized data and the mining data, respectively. , Indicates the first The conflict weights for each mapping group are set with a high confidence threshold of 0.8 and a low confidence threshold of 0.2. and hour, ,otherwise , Indicates the conflict penalty coefficient; A confidence threshold is set based on expert scoring, when... If the confidence level is not lower than the threshold, the corresponding mapping group is added to the knowledge graph until all mapping groups are traversed.

[0012] Furthermore, the principle underlying the generation of the initial feature vector for each first-level node is as follows: The attribute characteristics of first-level nodes include: historical fault correlation, calculated by dividing the number of times the first-level node is recorded as an anomaly in historical fault events by the total number of historical fault events; degree centrality, calculated by dividing the degree of the first-level node by the maximum degree in the entire graph structure; and lifetime depreciation, calculated as the ratio of the entity's used time to its design life. Therefore, the initial feature vector of the first-level node is represented as follows: ,in, Indicates the first The initial feature vectors of each first-level node. Indicates the index of the first-level node. Indicates the first Historical fault correlation of each primary node Indicates the first Degree centrality of each first-level node, Indicates the first The lifetime loss rate of each primary node.

[0013] Furthermore, the principle underlying the generation of the comprehensive feature vector of nodes is as follows: The formula used to calculate the attention coefficient between each first-level node and its neighboring nodes is as follows; ; in, Represents a first-level node Its neighboring nodes; Attention coefficient between them Indicates the index of the neighboring node. , Indicates the number of neighboring nodes. express and The confidence level of the edges between them. Representing neighboring nodes Degree centrality; The formula for normalizing the attention coefficients to generate attention weights is: ; ; in, Represents a first-level node Its neighboring nodes Attention weights between them Represents a first-level node The weight of one's own attention; The formula for generating the comprehensive feature vector is: ; in, Represents a first-level node The comprehensive feature vector, Representing neighboring nodes The initial eigenvectors.

[0014] Furthermore, the principle underlying the prediction of node failure probability based on comprehensive feature vectors is as follows: Using a logistic regression model, the comprehensive feature vector of the first-level nodes is used as the model input to calculate the failure probability. The formula used is as follows: ; in, Represents a first-level node Predicted failure probability, This represents the weight vector of the model. This represents the bias term of the model. Represents the sigmoid function; The principle of training weight vectors and bias terms is as follows: take the comprehensive feature vector of nodes in historical fault data as input and the binary fault label of nodes as output to train weight vectors and bias terms. The binary fault label indicates whether the entity represented by the node has failed when prediction is made after generating the comprehensive feature vector of the node in historical fault data. A value of 1 indicates that the entity has failed and a value of 0 indicates that the entity has not failed. Set a failure probability threshold. When the predicted failure probability is greater than or equal to the failure probability threshold, mark the node as a high-risk node.

[0015] Furthermore, the principles upon which the maintenance strategy is based are as follows: Reverse tracing means starting from a high-risk node and tracing upstream along the causal relationships in the graph structure; forward tracing means starting from a high-risk node and tracing downstream along the causal relationships in the graph structure; based on the tracing results, the root causes and potential consequences of the high-risk node are determined, and maintenance strategies are adopted based on the root causes and potential consequences.

[0016] This invention also provides a knowledge graph-based predictive maintenance system for complex equipment failures. The system is used to implement the aforementioned knowledge graph-based predictive maintenance method for complex equipment failures, specifically including: The standardized data acquisition module is used to acquire standardized data of complex equipment, define entity types and relation types respectively, generate standardized data triples, and calculate the standardized data confidence of each standardized data triple; The data acquisition module is used to collect structured time-series data of complex equipment in real time, generate data triplet sets, and calculate the data confidence of each data triplet set. The knowledge graph construction module is used to map the mined data triples to the normalized data triples, construct mapping groups based on the mapped triples, calculate the overall confidence of the mapping groups, and pre-set a confidence threshold. All mapping groups with an overall confidence higher than the confidence threshold are included in the knowledge graph, thereby completing the construction of the knowledge graph. The feature extraction module is used to identify the nodes corresponding to the device, component and sensor entities in the knowledge graph as nodes and the relations as edges, and set them as first-level nodes. It embeds attribute features for each first-level node, embeds the relation type and confidence of each edge, constructs the graph structure, and generates the initial feature vector of each first-level node. The fault prediction module is used to calculate the attention coefficient between each primary node and its neighboring nodes, normalize the attention coefficient to generate attention weights, update the initial feature vector based on the attention weights between each primary node and its neighboring nodes to generate the comprehensive feature vector of the primary node, predict the fault probability of the primary node based on the comprehensive feature vector, identify high-risk nodes based on the fault probability, and perform forward and backward tracing of the graph structure to determine the maintenance strategy.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention transforms fragmented, text-based domain knowledge into a unified triplet format, generating both standardized data triples and mined data triples. This lays the foundation for subsequent screening of high-quality knowledge to construct a knowledge graph. By combining the knowledge base and mined data, and comprehensively considering the confidence levels of both the standardized design and the mined data, a preset confidence threshold is used for screening. This ensures that only high-quality, highly credible knowledge is included in the final knowledge graph, guaranteeing that the knowledge graph can cover multi-source data and maintain accuracy.

[0018] This invention also constructs a graph structure with entities as nodes and relationships as edges, embedding data. This allows node representation to not only include topological information but also integrate key dimensions such as operational status and historical performance. It distinguishes not only node types but also the strength of different nodes, providing a rich information foundation for accurately assessing node risks. This enables fault prediction to move beyond single components and make comprehensive judgments based on the relationships within the entire equipment system. Based on the graph structure, it calculates the attention coefficient between nodes and their neighbors. The calculation of the attention coefficient integrates the topological structure and knowledge reliability of the graph, avoiding biases caused by relying solely on topological structure or independently analyzing node characteristics. Furthermore, it weights the neighbor information based on the comprehensive confidence of edges and the degree centrality of neighbor nodes, achieving differentiated aggregation of neighbor information. After identifying high-risk nodes, it performs bidirectional tracking, realizing root cause analysis and impact chain analysis, improving the accuracy and foresight of maintenance. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of an embodiment of the present invention; Figure 2 This is a schematic diagram of the system modules in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0021] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] Example: Please see Figure 1 The present invention provides a technical solution: A knowledge graph-based predictive maintenance method for complex equipment faults includes the following steps: Step 1: Obtain the standard data of complex equipment, define entity types and relation types respectively, generate standard data triples, and calculate the standard data confidence of each standard data triple; In this embodiment, the standard data for complex equipment includes equipment drawings, fault manuals, and knowledge of complex equipment from experts in the field; entity types include equipment, components, sensors, fault sources, fault symptoms, and maintenance measures; relationship types include composition, deployment, generation, cause, and mitigation; the standard data triple includes: components constitute equipment, components deploy sensors, components generate fault sources, fault sources cause fault symptoms, and maintenance measures mitigate fault symptoms. The standard data confidence score for each standard data triple is calculated using the following formula: ; in, Indicates the first The normalized confidence level of a normalized triplet. Indexes representing canonical triples Indicates the first Document authority score for a set of canonical data triples. Indicates the first Expert-authoritative scoring of a set of standardized data triples. Represents the consensus factor. The range of values ​​is , represent the weighting coefficients for the document authority score and the expert authority score, respectively. .

[0023] Standardized data refers to data obtained from authoritative, structured, and static domain knowledge sources that describes information such as design specifications, structural components, failure mechanisms, and maintenance standards for complex equipment. This data typically originates from official documents, industry standards, and expert experience, and is characterized by high reliability, standardization, and stability. Specifically, it includes the following three data sources: equipment drawings and structural documents, fault manuals and maintenance guides, and domain expert knowledge. Equipment drawings and structural documents describe the physical structure, component composition, and connection relationships of the equipment, such as assembly drawings, BOMs, and wiring diagrams. Fault manuals and maintenance guides describe known fault types, symptoms, causes, and maintenance measures, such as fault code manuals, maintenance operation manuals, and fault tree analysis documents. Domain expert knowledge represents fault associations summarized by experts based on their experience. Rules and maintenance recommendations exist in the form of rule bases, knowledge bases, and academic papers. To transform unstructured normative data into structured knowledge, entity types and relation types are first defined. Entity types include equipment, components, sensors, fault sources, fault symptoms, and maintenance measures. Relationship types include composition, deployment, generation, cause, and mitigation. Normative data triples are constructed based on entity types and relation types. Each normative data triple takes the form of (head entity, relation, tail entity), such as (bearing, composition, motor), (motor, deployment, temperature sensor), (bearing wear, cause, abnormal vibration), (replace bearing, mitigate, abnormal vibration). Knowledge from different sources may be inconsistent, ambiguous, or differ in authority; therefore, a confidence score is calculated for each triple. This is used to assess the feasibility of each triplet for subsequent screening of high-quality knowledge. Confidence score reflects the reliability of the canonical data triplet, calculated based on the authority of the data source. Data sources primarily include document acquisition and expert evaluation, reflected by document authority score and expert authority score respectively. The document authority score ranges from [value missing]. The higher the value, the more authoritative the document. Document types ranked from highest to lowest authority score are: national standards and industry specifications, original equipment manufacturer (OEM) manuals, company standardization documents, publicly published papers or reports, and unofficial online resources. Expert authority scores are determined based on the qualifications and experience of the experts during the evaluation; higher qualifications and more experience result in higher expert authority scores. A consensus factor G is also introduced to evaluate the consensus level of the data. Even if both the document and expert scores are high, if the knowledge contradicts itself across different sources, its consensus level is low. The value range of G is... A higher value indicates a better degree of consensus. For a given triple, if all documents and expert evaluations support the triple, its consensus factor is 1. When disagreements or contradictions arise, the consensus factor is gradually reduced. The confidence level of the data is directly proportional to the document authority score, the expert authority score, and the consensus factor. These represent the weighting coefficients for document authority score and expert authority score, respectively, used to reflect the importance of document authority and expert authority in the calculation of data confidence. Existing documents are typically summarized by experts in the field and largely reflect expert opinions. ,Pick , .

[0024] Step 2: Collect structured time-series data of complex equipment in real time, generate mining data triples, and calculate the mining data confidence of each mining data triple; In this embodiment, the entity types of the structured time-series data collected in real time by complex equipment include: sensors deployed on the equipment, data collected by the sensors, and fault symptoms identified from the data collected by the sensors; the relationship types include cause and reflection; the mined data triples include: sensor data reflecting fault symptoms, and changes in sensor data causing changes in other sensor data; the formula for calculating the mined data confidence of each mined data triple is: ; in, Indicates the first Mining confidence of a triplet of data This represents the index for mining triples. Indicates the first The causal strength score of each data triplet is mined. Indicates the first The association strength score of each data triplet is mined. represent the weight coefficients of the causal strength score and the association strength score, respectively, and .

[0025] Structured time-series data represents a series of data collected chronologically from sensors or monitoring systems installed on complex equipment. These data are presented in a fixed structured format, including numerical values ​​and category labels, reflecting the equipment's operating status and performance parameters. Each data point is collected and recorded at set time intervals and has a clear meaning, such as temperature, pressure, or vibration frequency. Data is continuously collected during equipment operation, reflecting the equipment's current or recent state. Specific data types include: physical state parameters (temperature, pressure, speed, current, voltage, etc.); performance index parameters (efficiency, energy consumption, output power, control accuracy, etc.); and environmental parameters (ambient temperature and humidity). The knowledge mined from structured time-series data is called mined data. Mined data triples define two core relationships: sensor data reflects fault symptoms, and changes in sensor data lead to changes in other sensor data. The first relationship indicates that data collected by a certain sensor reflects the existence of a certain fault symptom; the second relationship indicates that a change in one sensor's data will cause a change or impending change in another sensor's data.

[0026] Data confidence represents the degree of credibility of knowledge triples mined from structured time-series data acquired in real time; among which, the causal strength score... It is used to evaluate the strength of the causal relationship between two entities in a triplet of mined data, not only reflecting that the two entities occur simultaneously, but also reflecting that the occurrence of one entity leads to the occurrence of the other entity; The range of values ​​is The Granger causality test is used to calculate the strength of the causal effect between two entities. The stronger the causal effect, the stronger the causal effect. The larger the value, the stronger the association score. This is used to assess the strength of the statistical association between two entities in a mined data triplet, i.e., whether the two entities exhibit a consistent pattern of change in the data, regardless of the time sequence. The association strength score of the mined data triplet is obtained by calculating the Pearson correlation coefficient between the two entities in the same mined data triplet and normalizing the Pearson correlation coefficient. The confidence level of the mined data is proportional to both the causal strength score and the association strength score. These respectively reflect the causal strength scores and correlation strength score Regarding the impact on the confidence level of data mining, causal relationships provide a mechanistic explanation of "one entity causing a change in another entity," while statistical associations only indicate whether two entities occur simultaneously or sequentially. Simple associations can only issue early warnings, reflecting data changes, while causal relationships can guide the tracing of the source of data changes. ,Pick , .

[0027] Step 3: Map the mined data triples to the normalized data triples, construct mapping groups based on the mapped triples, calculate the overall confidence of the mapping groups, and pre-set a confidence threshold. Incorporate all mapping groups with an overall confidence higher than the confidence threshold into the knowledge graph, thereby completing the construction of the knowledge graph. In this embodiment, the principle underlying the mapping of mined data triples to canonical data triples is as follows: For each mined data triple whose data type is sensor data reflecting fault symptoms, query the deployment relationship in the normalized data triple. If a deployment relationship is found, and the fault symptom entity in the mined data and the fault source entity in the normalized data form a causal relationship, then it is considered a successful mapping. For a triple where changes in sensor data cause changes in other sensor data, query the deployment relationship of the two sensor entities in the normalized data triple respectively. If both are found to form a deployment relationship with the same component, then it is considered a successful mapping.

[0028] The purpose of mapping is to semantically align dynamically observed mining data with static, standardized data, thereby constructing a fused knowledge graph that includes both prior knowledge and reflects the actual operating state. For the type of mining triplet where "sensor data reflects fault symptoms," the deployment relationship is first searched in the standard data to determine whether the sensor is deployed on a certain component. Then, the causal relationship is searched in the standard data. Taking the mining data triplet (temperature sensor reading too high, reflects, motor overheating), the standard data triplet (motor, deployment, temperature sensor), and (bearing wear, causes, motor overheating) as an example, for the temperature sensor in the mining data triplet, there is a deployment relationship triplet in the standard data (motor, deployment, temperature sensor), thus determining that the temperature sensor is deployed on the motor. And for the fault symptom of motor overheating in the mining data triplet, there is a triplet in the standard data (bearing wear, causes, motor overheating), thus determining that the fault symptom in the mining data is caused by an entity in the standard data. Therefore, the mining data triplet (temperature sensor reading too high, reflects, motor overheating) can be considered a successful mapping with the standard data triplet. For the triple "Changes in sensor data lead to changes in other sensor data", query whether the two sensor entities have deployment relationships with other components in the specification data to determine whether the two sensor entities are deployed on the same system or related components. If deployment relationships are found in both cases, the mapping is considered successful.

[0029] The principle for calculating the overall confidence level is as follows: A mapping group is generated by combining a canonical data triple with its successfully mapped mined data triples. The overall confidence of the mapping group is then calculated using the following formula: ; in, Indicates the first The overall confidence of each mapping group Indicates the index of the mapping group. Indicates the first In the mapping group, the first The normalized confidence level of a normalized triplet. Represents the index of the canonical data triple in the mapping group, and , This represents the number of canonical triples in a mapping group. Indicates the first Mining data confidence in each mapping group These represent the weights of the confidence scores for the normalized data and the mining data, respectively. , Indicates the first The conflict weights for each mapping group are set with a high confidence threshold of 0.8 and a low confidence threshold of 0.2. and hour, ,otherwise , Indicates the conflict penalty coefficient; A confidence threshold is set based on expert scoring, when... If the confidence level is not lower than the threshold, the corresponding mapping group is added to the knowledge graph until all mapping groups are traversed.

[0030] Overall confidence level The goal is to integrate the credibility of normalized data and mined data, and to penalize data conflicts, in order to achieve a comprehensive assessment of the overall credibility of knowledge triples. The overall confidence reflects the overall credibility and consistency of knowledge. The higher the overall confidence, the better the entity relationship represented by the mapping group is supported by both prior knowledge and real-time data, and the higher the credibility. Conflict penalties are introduced. When the confidence of data in normalized data triples and mined data triples differs significantly, it indicates that there may be conflicts or contradictions in the data, and the overall confidence is reduced. For each mined data triple, it may correspond to more than one triple in normalized data. Therefore, a mapping group consists of multiple normalized data triples and one mined data triple. Used to reflect the data confidence of the canonical data triples in the mapping group, the average confidence of all canonical data triples in the mapping group is taken; This represents a conflict penalty term, used to reduce the overall confidence of a mapping group when the confidence differences between triples in the mapping group are large. The quality of a knowledge graph depends on the consistency of knowledge. If the confidence of the normative data is very high, exceeding the high-order threshold, while the confidence of the mined data is very low, not yet reaching the low-order threshold, it indicates a significant conflict between prior knowledge and actual observations. This suggests that the knowledge may be outdated, inapplicable, or contain false alarms. The conflict penalty term reduces the overall confidence of such mapping groups. Normative data has high prior authority and stability, is not easily affected by noise, and serves as a benchmark reference for data; therefore, it has a higher weight in the mapping group. Although mined data reflects the current state, it is easily affected by environmental interference, sensor errors, data gaps, etc., and its weight is relatively lower. , The conflict penalty coefficient is set to a value of ; To ensure the conflict penalty measures are highly effective, while avoiding a single conflict causing the overall confidence level to drop to zero, if... If the value is too large, it may lead to excessive conflict penalties, where even a small number of conflicts could result in the complete exclusion of triples, affecting the integrity of the knowledge graph. If the value is too small, the conflict penalty will be insufficient, potentially including contradictory and invalid data in the knowledge graph, affecting its accuracy. .

[0031] Table 1 reflects the impact of different average normalized data confidence levels and mined data confidence levels on the overall confidence level. When both the average normalized data confidence level and the mined data confidence level are high, the overall confidence level is high; when both are low, the overall confidence level is low; when one of them is high, the overall confidence level is low due to the impact of the conflict penalty term. Table 1. Changes in Overall Confidence Level

[0032] The overall confidence threshold is determined based on expert scoring. The overall confidence threshold is then used to determine whether a mapping group can be included in the knowledge graph. When the overall confidence of a mapping group is greater than the overall confidence threshold, it means that the data quality and accuracy in the mapping group meet the requirements, and each triple in it is added to the knowledge graph. Otherwise, it means that the data accuracy of the mapping group is insufficient, and the triples in these mapping groups are removed.

[0033] Step 4: Using entities in the knowledge graph as nodes and relationships as edges, determine the nodes corresponding to device, component and sensor entities and set them as first-level nodes. Embed attribute features for each first-level node, embed the relationship type and confidence level for each edge, construct the graph structure, and generate the initial feature vector for each first-level node. In this embodiment, the graph structure built based on the knowledge graph is composed of entities and relationships from the normative data and the mined data. The nodes corresponding to the entities of equipment, components and sensors are selected as first-level nodes. The focus is on analyzing these nodes because their changes directly reflect the equipment failure situation. The confidence level of each edge is the confidence level of the triple formed by the edge and the two entities it connects to. In this embodiment, the principle underlying the generation of the initial feature vector for each first-level node is as follows: The attribute characteristics of first-level nodes include: historical fault correlation, calculated by dividing the number of times the first-level node is recorded as an anomaly in historical fault events by the total number of historical fault events; degree centrality, calculated by dividing the degree of the first-level node by the maximum degree in the entire graph structure; and lifetime depreciation, calculated as the ratio of the entity's used time to its design life. Therefore, the initial feature vector of the first-level node is represented as follows: ,in, Indicates the first The initial feature vectors of each first-level node. Indicates the index of the first-level node. Indicates the first Historical fault correlation of each primary node Indicates the first Degree centrality of each first-level node, Indicates the first The lifetime loss rate of each primary node.

[0034] Nodes correspond to entities in the knowledge graph, and edges correspond to relationships between entities. Each first-level node is embedded with its attribute features, specifically including historical fault correlation, degree centrality, and lifetime loss. The historical fault correlation is calculated as the ratio of the number of times a first-level node is recorded as an anomaly in historical fault events to the total number of all historical fault events. This reflects the frequency of an entity's failure; a higher value indicates a greater likelihood of the first-level node failing and a higher prior probability of future failure. Degree centrality is calculated as the ratio of the degree of a first-level node to the maximum degree of all nodes in the entire graph structure. Degree represents the degree of a node. The number of edges a node connects to in a graph structure represents the number of direct relationships between that primary node and other nodes. Degree centrality reflects the pivotal nature of a node in the network. A primary node with high degree centrality is more likely to experience a rapid spread of anomalies through multiple relationship chains, affecting more other nodes. Highly central nodes are often critical components of the system, and their failures can lead to systemic risks. Lifetime depreciation is calculated as the ratio of the usage time of the entity corresponding to the primary node to its designed lifespan. This reflects the aging degree of the entity over time; a higher lifetime depreciation indicates that the entity is closer to its theoretical lifespan end and has a relatively higher risk of failure. These three dimensions are integrated into a single feature vector. Entities with higher historical failure frequencies, more critical locations, and longer operating times are given higher maintenance priority.

[0035] Step 5: Calculate the attention coefficient between each primary node and its neighboring nodes, normalize the attention coefficient to generate attention weights, update the initial feature vector based on the attention weights between each primary node and its neighboring nodes to generate the comprehensive feature vector of the primary node, predict the failure probability of the primary node based on the comprehensive feature vector, identify high-risk nodes based on the failure probability, and perform forward and backward tracing of the graph structure to determine the maintenance strategy.

[0036] In this embodiment, the principle underlying the generation of the comprehensive feature vector of the first-level node is as follows: The formula used to calculate the attention coefficient between each first-level node and its neighboring nodes is as follows; ; in, Represents a first-level node Its neighboring nodes Attention coefficient between them Indicates the index of the neighboring node. , Indicates the number of neighboring nodes. express and The confidence level of the edges between them. Representing neighboring nodes Degree centrality; Reflects the nodes For first-level nodes The importance of a first-level node in the information aggregation process is calculated based on confidence level and the degree centrality of its neighboring nodes. A higher confidence level indicates a stronger first-level node. with neighboring nodes The higher the feasibility of the constructed triple, the more reliable the triple is in the knowledge graph, and the better its corresponding neighbor nodes. In the The more important it is when performing information aggregation, the more important it is. Degree centrality reflects the hub status of neighboring nodes in the graph structure. The higher the degree centrality, the stronger the neighboring nodes are. The larger the scope of influence in a graph structure, the greater the impact. The more important it is when aggregating information; the attention coefficient takes into account both the credibility of the relationship and the importance of the node, and realizes the integration of knowledge reliability and structural importance. If the confidence is low, it means that the credibility of this triple is low. Even if the degree centrality of the neighboring node is high, its influence will be suppressed. If the degree centrality of the neighboring node is low, even if the credibility is high, the influence of the neighboring node in the whole graph structure is low and the influence on the first-level node is also low. The formula for normalizing the attention coefficients to generate attention weights is: ; ; in, Represents a first-level node Its neighboring nodes Attention weights between them Represents a first-level node The weight of one's own attention; For each first-level node, the attention weights of its neighboring nodes and its own attention weight are considered separately. While considering the aggregation of information from neighboring nodes, the node's own characteristics are preserved. Represents a first-level node For nodes with degree centrality and height centrality, even if the attention weights of their neighbors are relatively weak, their own characteristics should still be given priority consideration; calculate At the same time, it measures both relationship reliability and neighbor importance, and reliable and important neighbor pairs are considered for nodes. The impact is greater; calculation At that time, the node's confidence level about itself is 1, therefore In reality , The greater the degree centrality, the greater the attention weight it has on itself; the denominator This indicates that the first-level node The attention coefficient of the first-level node is added to the attention coefficients of all neighboring nodes; this avoids the first-level node's own information being ignored in the information aggregation of neighboring nodes, and balances the first-level node's own characteristics with the influence of its neighbors.

[0037] The formula for generating the comprehensive feature vector is: ; in, Represents a first-level node The comprehensive feature vector, Representing neighboring nodes The initial eigenvectors.

[0038] The comprehensive feature vector integrates the node's own features with the features of its neighboring nodes. This indicates that the first-level node The feature vectors of all neighboring nodes are summed using weighted methods. This indicates the first-level nodes retained during the fusion process. The node's own information, combined with its feature vector, reflects both its state attributes (preserving historical failures, structural importance, and aging characteristics) and the impact of neighbor node distances (specifically, failure propagation paths). For example, a high number of historical failures among neighboring nodes indicates... It also faces risks; structural correlation, anomalies in highly central neighboring nodes may trigger a chain reaction; semantic correlation strength, the state of neighboring nodes with high confidence is also a factor. The impact is more credible; the integrated feature vector merges the feature vector of the first-level node with the global semantic association into a higher-dimensional vector representation, reflecting the contextual semantics of the first-level node in the entire knowledge graph. It not only includes the node's own state, but also implies information such as fault propagation path and system-level associated risks, enabling subsequent predictions based on richer features in the regression model and reducing false positives and false negatives.

[0039] The principle underlying the prediction of first-level node failure probability based on comprehensive feature vectors is as follows: Using a logistic regression model, the comprehensive feature vector of the first-level nodes is used as the model input to calculate the failure probability. The formula used is as follows: ; in, Represents a first-level node Predicted failure probability, This represents the weight vector of the model. This represents the bias term of the model. Represents the sigmoid function; The principle of training weight vectors and bias terms is as follows: take the comprehensive feature vector of nodes in historical fault data as input and the binary fault label of nodes as output to train weight vectors and bias terms. The binary fault label indicates whether the entity represented by the node has failed when prediction is made after generating the comprehensive feature vector of the node in historical fault data. A value of 1 indicates that the entity has failed and a value of 0 indicates that the entity has not failed. Logistic regression is used as the basic model for fault probability prediction. The output of the logistic regression model is a probability value, which is convenient for combining with a threshold to determine risk. Its expression is: , This is a weight vector with the same dimension as the comprehensive feature vector, used to weight the comprehensive feature vector. This represents the bias term, used to adjust the output offset of the model. Used to map linear output to interval, ; The target learning parameters for training a logistic regression model and To ensure the predicted probability closely approximates the true fault label, the cross-entropy loss function is chosen as the loss function during training, and parameters are adjusted by minimizing this loss function. Training data is derived from historical fault event records, and the output features are the comprehensive feature vectors of nodes, labeled with binary fault labels. A binary fault label of 1 indicates that the entity corresponding to that node has experienced a fault in the past, while a binary fault label of 0 indicates that no fault has occurred. The comprehensive feature vectors of all first-level nodes are extracted from the historical running data and labeled with their corresponding fault labels. The weight vectors and bias terms are randomly initialized, and the cross-entropy loss function is used to measure the difference between the model's prediction and the true label. Gradient descent is used to update the model's predictions. and The training stops when the maximum number of iterations is reached, minimizing the loss function, and outputs the result at that point. and According to the current situation and Determine the predicted failure probability; Set a failure probability threshold. When the predicted failure probability is greater than or equal to the failure probability threshold, mark the node as a high-risk node.

[0040] The failure probability threshold is determined based on expert scoring. The higher the failure probability threshold, the more stringent the screening of high-risk nodes.

[0041] The principles upon which the maintenance strategy is determined are: Reverse tracing means starting from a high-risk node and tracing upstream along the causal relationships in the graph structure; forward tracing means starting from a high-risk node and tracing downstream along the causal relationships in the graph structure; based on the tracing results, the root causes and potential consequences of the high-risk node are determined, and maintenance strategies are adopted based on the root causes and potential consequences.

[0042] The reverse tracing is based on the causal relationship chain already constructed in the knowledge graph. Starting from the high-risk node, it queries the knowledge graph for all upstream nodes connected to the high-risk node with an edge relationship of "caused". Starting from the upstream node at this level, it searches for upstream nodes at higher levels using the same method until no "caused" relationship can be found in the upstream nodes. Taking the high-risk node as the top node, it obtains all upstream end nodes and constructs a root cause candidate set based on the path between each upstream end node and the high-risk node. Forward tracing starts from the high-risk node and queries the knowledge graph for all downstream nodes that are connected to the high-risk node and whose edge relationship is "cause". Starting from each downstream node found, the same method is used to find the next level node until no "cause relationship" can be found in the downstream nodes. All downstream end nodes are obtained, and a potential consequence candidate set is constructed based on the path between the high-risk node and each downstream end node. The confidence level of each path in the root cause candidate set and the potential consequence candidate set is calculated. The calculation logic is the sum of the confidence levels among all nodes on each path. In each candidate set, the path with the highest confidence level is selected, and the corresponding upstream and downstream end nodes are regarded as the root cause node and potential consequence node of the high-risk node, respectively. Based on the degree of damage of the root cause node and the severity of the potential consequence node, the maintenance strategy is determined based on expert evaluation.

[0043] Please see Figure 2 The present invention also provides a knowledge graph-based predictive maintenance system for complex equipment failures. This system is used to implement the aforementioned knowledge graph-based predictive maintenance method for complex equipment failures, specifically including: The standardized data acquisition module is used to acquire standardized data of complex equipment, define entity types and relation types respectively, generate standardized data triples, and calculate the standardized data confidence of each standardized data triple; The data acquisition module is used to collect structured time-series data of complex equipment in real time, generate data triplet sets, and calculate the data confidence of each data triplet set. The knowledge graph construction module is used to map the mined data triples to the normalized data triples, construct mapping groups based on the mapped triples, calculate the overall confidence of the mapping groups, and pre-set a confidence threshold. All mapping groups with an overall confidence higher than the confidence threshold are included in the knowledge graph, thereby completing the construction of the knowledge graph. The feature extraction module is used to identify the nodes corresponding to the device, component and sensor entities in the knowledge graph as nodes and the relations as edges, and set them as first-level nodes. It embeds attribute features for each first-level node, embeds the relation type and confidence of each edge, constructs the graph structure, and generates the initial feature vector of each first-level node. The fault prediction module is used to calculate the attention coefficient between each primary node and its neighboring nodes, normalize the attention coefficient to generate attention weights, update the initial feature vector based on the attention weights between each primary node and its neighboring nodes to generate the comprehensive feature vector of the primary node, predict the fault probability of the primary node based on the comprehensive feature vector, identify high-risk nodes based on the fault probability, and perform forward and backward tracing of the graph structure to determine the maintenance strategy.

[0044] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0045] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0046] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0047] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A knowledge graph-based complex equipment failure predictive maintenance method, characterized in that, The specific steps include: Step 1: Obtain the specification data of complex equipment, define entity types and relationship types respectively, and generate specification data triples, and calculate the specification data confidence of each specification data triple; Step 2: Real-time acquisition of structured time series data of complex equipment, generation of mining data triples, and calculation of the mining data confidence of each mining data triple; Step 3: Map the mining data triples to the specification data triples, construct a mapping group based on the triples that have completed mapping, calculate the comprehensive confidence of the mapping group, and pre-set the confidence threshold, and all mapping groups with a comprehensive confidence higher than the confidence threshold are included in the knowledge graph, and the construction of the knowledge graph is completed; Step 4: Take the entities in the knowledge graph as nodes, the relationships as edges, and determine the nodes corresponding to the equipment, components and sensor entities as first-level nodes, set the attribute features for each first-level node, embed the relationship type and confidence for each edge, construct the graph structure, and generate the initial feature vector of each first-level node; Step 5: Calculate the attention coefficient between each first-level node and its neighbor nodes, normalize the attention coefficient to generate the attention weight, update the initial feature vector based on the attention weight between each first-level node and its neighbor nodes, generate the comprehensive feature vector of the first-level node, predict the failure probability of the first-level node based on the comprehensive feature vector, determine the high-risk nodes according to the failure probability, and perform forward and reverse tracking on the graph structure to determine the maintenance strategy. 2.The complex equipment failure predictive maintenance method based on a knowledge graph according to claim 1, characterized in that: The specification data of the complex equipment in step 1 includes equipment drawings, fault manuals and the knowledge of experts in the field about the complex equipment; the entity types include equipment, components, sensors, fault sources, fault symptoms and maintenance measures; the relationship types include composition, deployment, generation, cause and mitigation; the specification data triples include: component composition equipment, component deployment sensor, component generation fault source, fault source cause fault symptom, and maintenance measure mitigation fault symptom; the specification data confidence of each specification data triple is calculated, and the specific formula is: ； wherein, denotes the normative data confidence of the th normative data triple, denotes the index of the normative data triple, denotes the document authority score of the th normative data triple, denotes the expert authority score of the th normative data triple, denotes the consensus degree factor, the value ranges of , respectively denote the weight coefficients of the document authority score and the expert authority score, and . 3.The complex equipment failure predictive maintenance method based on a knowledge graph according to claim 1, characterized in that: The entity types of the structured time series data of the complex equipment collected in real time in step 2 include: sensors deployed on the equipment, data collected by the sensors, and fault symptoms identified from the data collected by the sensors; the relationship types include cause and reflection; the mining data triples include: sensor data reflecting fault symptoms, and sensor data changes causing changes in other sensor data; the formula for calculating the mining data confidence of each mining data triple is: ； wherein, represents a mining data confidence of the th mining data triple, represents an index of the mining data triple, represents a causal strength score of the th mining data triple, represents an association strength score of the th mining data triple, respectively represent weight coefficients of the causal strength score and the association strength score, and .

4. The knowledge graph-based complex equipment failure predictive maintenance method of claim 1, wherein: The principle for mapping the mining data triples to the specification data triples in step 3 is: For each mined data triple reflecting a fault symptom of the data type of sensor data, the deployment relationship is queried in the canonical data triple, if the deployment relationship is queried and the fault symptom entity in the mined data and the fault source entity in the canonical data constitute a causing relationship, it is considered as successful mapping; for the triple that the change of sensor data leads to the change of other sensor data, the deployment relationships of the two sensor entities are respectively queried in the canonical data triple, if they are both queried to constitute a deployment relationship with the same component, it is considered as successful mapping.

5. The knowledge graph-based complex equipment failure predictive maintenance method of claim 1, wherein: The principle of calculating the comprehensive confidence is: The principle of generating the initial feature vector of each primary node in step 4 is: ； wherein, represents the comprehensive confidence of the th mapping group, represents the index of the mapping group, represents the comprehensive confidence of the th mapping group, represents the canonical data confidence of the th canonical data triple in the mapping group, and , represents the number of canonical data triples in a mapping group, represents the mining data confidence of the th mapping group, respectively represent the weights of the canonical data confidence and the mining data confidence, , represents the conflict weight of the th mapping group, the high confidence threshold is set as 0.8, the low confidence threshold is set as 0.2, when and , , otherwise , represents the conflict penalty coefficient; Based on the expert scoring method, a confidence threshold is set, and when When the confidence threshold is not less than the confidence threshold, the corresponding mapping group is added to the knowledge graph, and all mapping groups are traversed.

6. The knowledge graph-based complex equipment failure predictive maintenance method of claim 1, wherein: The principle of generating the comprehensive feature vector of the node in step 5 is: The attribute characteristics of first-level nodes include: historical fault correlation, calculated by dividing the number of times the first-level node is recorded as an anomaly in historical fault events by the total number of historical fault events; degree centrality, calculated by dividing the degree of the first-level node by the maximum degree in the entire graph structure; and lifetime depreciation, calculated as the ratio of the entity's used time to its design life. Therefore, the initial feature vector of the first-level node is represented as follows: ,in, Indicates the first The initial feature vectors of each first-level node. Indicates the index of the first-level node. Indicates the first Historical fault correlation of each primary node Indicates the first Degree centrality of each first-level node, Indicates the first The lifetime loss rate of each primary node.

7. The knowledge graph-based complex equipment failure predictive maintenance method of claim 6, wherein: The formula for calculating the attention coefficient between each primary node and its neighbor node is: The formula for normalizing the attention coefficient to generate the attention weight is: ； wherein, denotes a primary node and its neighbor nodes between which the attention coefficient, denotes the index of the neighbor node, , denotes the number of neighbor nodes, denotes and the confidence of the edge between denotes the degree centrality of the neighbor node ; The formula for generating the comprehensive feature vector is: ； ； wherein, represents the attention weight between the primary node and its neighbor nodes , represents the attention weight of the primary node to itself; The principle of predicting the fault probability of the node based on the comprehensive feature vector in step 5 is: ； wherein, represents a primary node comprehensive feature vector, represents an initial feature vector of a neighbor node .

8. The knowledge graph-based complex equipment failure predictive maintenance method of claim 7, wherein: Using a logistic regression model, the comprehensive feature vector of the primary node is used as the input of the model to calculate the fault probability, and the formula is: The principle of training the weight vector and the bias term is: the comprehensive feature vector of the node in the historical fault data is used as the input, and the binary fault label of the node is used as the output, the weight vector and the bias term are trained, and the binary fault label represents whether the entity represented by the node has failed when the comprehensive feature vector of the node is generated and predicted in the historical fault data, and the value is 1 when the entity has failed, and the value is 0 when the entity has not failed; ； in, Represents a first-level node Predicted failure probability, This represents the weight vector of the model. This represents the bias term of the model. Represents the sigmoid function; Set the fault probability threshold, when the predicted fault probability is greater than or equal to the fault probability threshold, mark the node as a high-risk node. The principle of determining the maintenance strategy in step 5 is:

9. The knowledge graph-based complex equipment failure predictive maintenance method of claim 8, wherein: Backward tracking means tracing upstream along the causing relationship in the graph structure from the high-risk node; forward tracking means tracing downstream along the causing relationship in the graph structure from the high-risk node; According to the results of the tracing, determine the root cause and potential consequences of the high-risk node, and take maintenance strategies based on the root cause and potential consequences. The system is used to implement the knowledge graph-based complex equipment fault predictive maintenance method of any one of claims 1-9, and specifically comprises:

10. A knowledge graph based complex equipment failure predictive maintenance system, characterized in that: A canonical data acquisition module is configured to obtain canonical data of complex equipment, define entity types and relationship types respectively, and generate canonical data triples, and calculate the canonical data confidence of each canonical data triple; A mined data acquisition module is configured to acquire real-time structured time series data of complex equipment, generate mined data triples, and calculate the mined data confidence of each mined data triple; ​ The atlas construction module is configured to map the mined data triples to standard data triples, construct a mapping group based on the mapped triples, calculate a comprehensive confidence of the mapping group, set a confidence threshold, and include all mapping groups with a comprehensive confidence higher than the confidence threshold into the knowledge graph, thereby completing construction of the knowledge graph; The feature extraction module is configured to take entities in the knowledge graph as nodes and relationships as edges, determine nodes corresponding to device, component, and sensor entities, set the nodes as first-level nodes, embed attribute features for each first-level node, embed a relationship type and a confidence for each edge, construct a graph structure, and generate an initial feature vector for each first-level node; The fault prediction module is configured to calculate attention coefficients between each first-level node and its neighbor nodes, normalize the attention coefficients to generate attention weights, update the initial feature vector based on the attention weights between each first-level node and its neighbor nodes to generate a comprehensive feature vector of the first-level node, predict a fault probability of the first-level node based on the comprehensive feature vector, determine a high-risk node according to the fault probability, and perform forward and reverse tracking on the graph structure to determine a maintenance strategy.

Images