Motor-oriented fault mode knowledge graph construction and updating method and system

Abstract

The invention discloses a motor-oriented fault mode knowledge graph construction and updating method and a motor-oriented fault mode knowledge graph construction and updating system. The motor-oriented fault mode knowledge graph construction and updating method comprises the following steps: constructing a motor fault domain ontology model: defining a core concept class hierarchy, an object attribute and a semantic relationship, obtaining structured data and unstructured data, defining a data attribute, and setting a logic constraint and an axiom; performing double-channel knowledge extraction: performing knowledge extraction on the unstructured data by adopting a deep learning model, performing knowledge extraction on the structured data by adopting a mapping method, and generating original triple knowledge; knowledge fusion: integrating knowledge from different data sources through entity alignment, attribute fusion and conflict resolution to construct a high-quality and unambiguous knowledge graph, and storing the high-quality and unambiguous knowledge graph in a graph database; and performing intelligent fault reasoning and dynamic updating based on the knowledge graph. According to the method, multi-source data are fused, long and difficult entities and complex semantic relationships are identified more accurately, and the accuracy and efficiency of troubleshooting are improved through a double weighting algorithm.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge graph construction, and in particular to a motor-oriented fault mode knowledge graph construction and updating method and system. BACKGROUND

[0002] The motor (especially permanent magnet synchronous motor) is the core power source of modern industrial production lines, new energy vehicles and aerospace equipment, and its operation reliability is directly related to the safety and efficiency of the entire system. With the advancement of intelligent manufacturing, the fault diagnosis technology for motors has shifted from traditional single signal threshold monitoring to knowledge-based intelligent diagnosis.

[0003] In the whole life cycle management of the motor, a large amount of fault-related data is accumulated, mainly divided into two categories: one is the structured data generated in the design and manufacturing stage, typical such as the Failure Mode and Effects Analysis (FMEA) table, which contains standard fault modes, failure mechanisms and theoretical risk levels (such as RPN values) defined by domain experts; the other is unstructured data generated in the operation and maintenance stage, such as maintenance work orders, fault logs, expert diagnosis reports, etc., which record the actual fault phenomena, processing procedures and replacement of spare parts information.

[0004] However, the existing technology in utilizing these data to construct fault diagnosis systems faces the following serious technical challenges:

[0005] First, the "island effect" of multi-source heterogeneous data leads to low knowledge utilization. Existing diagnostic systems are often single-based, FMEA-based systems lack support from field measurement data, leading to a disconnection between theoretical risk and actual fault distribution (for example, theoretically high-risk faults rarely occur in practice); while log mining-based systems lack expert theoretical guidance, making it difficult to discover deep fault modes, and are easily limited by non-standard text descriptions. The existing technology lacks an effective mechanism that can deeply semantically integrate "theoretical norms on the design side" and "practical experience on the operation side".

[0006] Second, there is a lack of deep semantic understanding and reasoning ability. Traditional fault knowledge bases are mostly based on keyword matching or shallow association. For example, when searching for "motor vibration is large", the existing system is difficult to discover through multi-hop reasoning that "bearing retainer fracture" will cause "rotor imbalance", and then cause "vibration is large" deep causal chain. In addition, there are many professional terms in the motor field (such as "tooth slot torque", "turn-to-turn short circuit"), and traditional word segmentation and entity recognition methods have low accuracy when facing complex maintenance texts.

[0007] Third, the knowledge base is static and rigid, lacking a dynamic adaptive updating mechanism. Existing expert system knowledge bases are usually built once. However, as the motor operating environment changes and ages, new fault modes may emerge (such as new bearing electric corrosion caused by improper setting of the carrier frequency of the frequency converter), or the frequency of occurrence of some faults may shift over time. The existing technology lacks a self-growth mechanism that can automatically "learn" new knowledge from a constant stream of new maintenance records and real-time correct fault correlation weights, resulting in a decrease in the timeliness of the diagnostic model over time.

[0008] In summary, there is an urgent need for a motor fault mode knowledge graph construction method that can deeply integrate design and operation data, has deep semantic reasoning capabilities, and can be dynamically updated with business data. SUMMARY

[0009] The present application overcomes the shortcomings of the above-mentioned technologies and provides a motor-oriented fault mode knowledge graph construction and updating method and system.

[0010] The technical solution adopted by the present application to overcome the technical problems is:

[0011] A motor-oriented fault mode knowledge graph construction and updating method, comprising the steps of:

[0012] Constructing a motor fault domain ontology model: defining a core concept class hierarchy for representing motor fault knowledge, defining object attributes and semantic relationships, obtaining structured and unstructured data and defining data attributes, and then setting logical constraints and axioms;

[0013] Dual-channel knowledge extraction: based on the motor fault domain ontology model, unstructured data is extracted using a deep learning model, and structured data is extracted using a mapping method to generate original triple knowledge;

[0014] Knowledge fusion: entity alignment, attribute fusion, and conflict resolution are performed on the original triple knowledge, and the knowledge from different data sources is integrated to construct a high-quality, unambiguous knowledge graph and store it in a graph database;

[0015] Intelligent fault reasoning and dynamic updating based on the knowledge graph: using the constructed knowledge graph, responding to user diagnostic requests, and based on a pre-set dual-weighting algorithm, intelligent reasoning and sorting from fault phenomena to root causes are achieved. At the same time, if the work order contains new components or new fault relationships that do not exist in the knowledge graph and the extraction confidence is higher than the pre-set new knowledge threshold, the main knowledge graph is incorporated, and the dynamic updating of the knowledge graph is achieved.

[0016] Further, the core concept class hierarchy is a top-down five-dimensional concept tree structure, and the five dimensions are equipment and component class, fault phenomenon class, fault mode class, diagnosis and detection class, and maintenance decision class; the object attributes and semantic relationships include at least a composition relationship for connecting equipment and components, a causal relationship for connecting fault modes and fault phenomena, a diagnosis relationship for connecting fault modes and detection methods, a location relationship for connecting fault modes and equipment components, and a governance relationship for connecting fault modes and maintenance measures; and the data attributes include at least attributes for the fault mode class and attributes for the maintenance decision class.

[0017] Further, the unstructured data is subjected to knowledge extraction by using a deep learning model, and specifically includes:

[0018] A BERT-BiLSTM-CRF model is constructed, and the BERT-BiLSTM-CRF model is used for entity recognition on the unstructured data, wherein the BERT-BiLSTM-CRF model includes a BERT layer for converting the unstructured data into context-related vectors, a BiLSTM layer for bidirectional extraction of sequence context features, and a CRF layer for label decoding of an entity tag sequence.

[0019] After the entities are recognized, a classification model based on an attention mechanism is used for semantic relationship extraction between entities, so as to generate a triple knowledge.

[0020] Further, the structured data is subjected to knowledge extraction by using a mapping method, and specifically includes:

[0021] Based on a preset field-entity mapping rule, field contents in the structured data are mapped to corresponding entities in the motor fault domain ontology model;

[0022] Based on an attribute extraction rule, a risk priority number RPN in the structured data is extracted and calculated as an attribute of a fault mode node;

[0023] Based on a relationship construction rule, (fault mode, cause, fault phenomenon) triples and (fault mode, located in, equipment component) triples are automatically instantiated according to the logical structure of the structured data.

[0024] Further, the entity alignment specifically includes:

[0025] The BERT-BiLSTM-CRF model is used to extract semantic vector representations of all entity names in the original triples;

[0026] The semantic similarity of the entities is calculated and judged, and if the similarity exceeds a preset alignment threshold, the entities are merged, the standard name of the entity is retained, and the alias is recorded.

[0027] Further, the attribute fusion adopts a master-slave strategy, specifically including:

[0028] The node generated by structured data is the main data source, and the prior theoretical attribute is retained, which at least includes risk priority number, severity, occurrence and detection;

[0029] The node generated by unstructured data is the slave data source, and the actual occurrence frequency and average repair time of the fault are counted;

[0030] The statistical attribute of the slave data source is mounted on the entity node determined by the main data source to realize attribute fusion.

[0031] Further, the conflict resolution specifically includes:

[0032] For theoretical parameter conflict, structured data is used as the standard;

[0033] For measured parameter conflict, unstructured data is used as the standard;

[0034] For fusion data with a confidence level below a preset threshold, data confirmed by manual review is used as the standard.

[0035] Further, the constructed knowledge graph is used to respond to the user's diagnosis request, and a preset double weighting algorithm is used to realize intelligent reasoning and ordering from fault phenomenon to root cause, specifically including:

[0036] A trained deep learning model is used to analyze the user's input query statement and locate the corresponding fault phenomenon node in the knowledge graph;

[0037] The fault phenomenon node is used as the starting point, and reverse breadth-first search is performed along the causal relationship to obtain a candidate fault set;

[0038] A preset "theoretical risk priority number and actual occurrence frequency" double weighting algorithm is used to calculate the comprehensive recommendation score of each candidate fault;

[0039] The candidate fault set is sorted from high to low according to the comprehensive recommendation score, and the fault cause and its corresponding repair measures and detection methods are output.

[0040] Further, if the work order contains a new component or a new fault relationship that does not exist in the knowledge graph and the extraction confidence is higher than a preset new knowledge threshold, the new component or new fault relationship is incorporated into the main knowledge graph to realize dynamic updating of the knowledge graph, specifically including:

[0041] The work order state of the maintenance management system is monitored in real time, and the updating process is triggered when the work order is completed and contains a confirmed cause;

[0042] If the confirmed cause in the work order exists in the existing knowledge graph, the corresponding fault mode node is retrieved in the knowledge graph, and the actual occurrence frequency attribute of the fault mode node is accumulated and corrected;

[0043] If the entity or relationship extracted from the new work order does not exist in the existing knowledge graph and the comprehensive confidence is higher than the preset new knowledge threshold, the corresponding new node and new edge are automatically created in the candidate area, and the expert review process is triggered, and after the review is passed, it is incorporated into the main knowledge graph, realizing the dynamic updating of the knowledge graph.

[0044] The application also discloses a motor-oriented fault mode knowledge graph construction and updating system, comprising:

[0045] The motor fault domain ontology model construction module is configured to define a core concept class hierarchy for representing motor fault knowledge, define object attributes and semantic relationships, obtain structured data and unstructured data and define data attributes, and then set logical constraints and axioms;

[0046] The double-channel knowledge extraction module is configured to extract knowledge from unstructured data using a deep learning model and extract knowledge from structured data using a mapping method based on the motor fault domain ontology model, and generate original triple knowledge;

[0047] The knowledge fusion module is configured to perform entity alignment, attribute fusion and conflict resolution on the original triple knowledge, integrate knowledge from different data sources to construct a high-quality and unambiguous knowledge graph, and store the knowledge graph in a graph database;

[0048] The intelligent fault reasoning and dynamic updating module is configured to utilize the constructed knowledge graph to respond to user's diagnosis request, and realize intelligent reasoning and sorting from fault phenomenon to root cause based on a preset double-weighting algorithm; at the same time, if the work order contains new components or new fault relationships that do not exist in the knowledge graph and the extraction confidence is higher than the preset new knowledge threshold, the main knowledge graph is incorporated to realize the dynamic updating of the knowledge graph.

[0049] The application has the following advantages:

[0050] 1. Full-dimensional knowledge fusion: the application innovatively proposes a fusion method combining structured data in the design and manufacturing stage and unstructured data in the operation and maintenance stage, adopts multi-source heterogeneous data, retains the theoretical rigor of the expert design stage (through FMEA mapping), and absorbs the field experience in the operation and maintenance stage (through text mining), solving the problem of one-sidedness of single data source information.

[0051] 2. High-precision knowledge extraction: To address the pain points of complex professional terms and large text description differences in the motor field, the application uses a BERT-BiLSTM-CRF model, which can more accurately identify long and difficult entities and complex semantic relationships compared to traditional dictionary matching or shallow machine learning methods, significantly improving the utilization of unstructured data.

[0052] 3. Intelligent weighted reasoning: The application discards simple Boolean queries and introduces a dual-weighted sorting algorithm based on "theoretical risk priority number" and "actual occurrence frequency", so that the diagnosis result not only indicates "what fault", but also informs "which fault to prioritize troubleshooting" according to risk and probability, greatly improving the accuracy and efficiency of on-site fault troubleshooting.

[0053] 4. Dynamic evolution mechanism: By introducing incremental update logic, the application makes the knowledge graph not a static database, but a system that can automatically correct weights and expand new knowledge with each maintenance record, effectively solving the problem of knowledge base aging and lag. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The principle diagram of the motor-oriented fault mode knowledge graph construction and updating method according to the embodiments of the application.

[0055] Figure 2 The principle structure diagram of the BERT-BiLSTM-CRF model according to the embodiments of the application.

[0056] Figure 3 The performance comparison diagram of different knowledge extraction models according to the embodiments of the application.

[0057] Figure 4 The diagnosis accuracy comparison diagram of different recommendation algorithms according to the embodiments of the application. DETAILED DESCRIPTION

[0058] In order to facilitate those skilled in the art to better understand the application, the application will be further described in detail below in conjunction with the drawings and specific embodiments, which are only exemplary and do not limit the scope of protection of the application.

[0059] I. Example 1:

[0060] The application discloses a motor-oriented fault mode knowledge graph construction and updating method, comprising the following steps:

[0061] Constructing a motor fault domain ontology model: defining a core concept class hierarchy for representing motor fault knowledge, defining object attributes and semantic relationships, obtaining structured data and unstructured data and defining data attributes, and then setting logical constraints and axioms;

[0062] Dual-channel knowledge extraction: based on the motor fault domain ontology model, a deep learning model is used for knowledge extraction on unstructured data, and a mapping method is used for knowledge extraction on structured data to generate original triple knowledge;

[0063] Knowledge fusion: entity alignment, attribute fusion and conflict resolution are performed on the original triple knowledge, and high-quality and unambiguous knowledge graphs are constructed from knowledge from different data sources and stored in a graph database;

[0064] Intelligent fault reasoning and dynamic updating based on knowledge graph: using the constructed knowledge graph, responding to the user's diagnosis request, based on the preset double weighting algorithm, intelligent reasoning and sorting from fault phenomenon to root cause are realized; at the same time, if the work order contains new components or new fault relationships that do not exist in the knowledge graph and the extraction confidence is higher than the preset new knowledge threshold, the main knowledge graph is integrated, and the dynamic updating of the knowledge graph is realized.

[0065] In order to better understand the above technical solutions, the following will be combined with the permanent magnet synchronous motor (PMSM) full life cycle operation and maintenance management system of a certain motor production and manufacturing enterprise as an example to elaborate the example embodiment of the motor-oriented fault mode knowledge graph construction and updating method described in the present application, as shown in Figure 1 The embodiment is only an example embodiment of the present application, but it should be understood that the present application can also be realized in various forms and is not limited to the embodiments described herein. The embodiment is to make those skilled in the art more clearly and more thoroughly understand the present application. The method aims to solve the problems of the existing system, such as the separation of structured data (such as FMEA data) generated in the design and manufacturing stage and unstructured data (such as maintenance logs) generated in the operation and maintenance stage, the lack of semantic reasoning ability in fault diagnosis, and the lag of knowledge base updating.

[0066] S1. Construct a motor fault domain ontology model.

[0067] In this step, Protégé 5.5 knowledge modeling tool and Web Ontology Language (OWL) are used to construct the ontology according to the "seven-step method" (determine the field, check reuse, list terms, define classes and hierarchy, define attributes, define constraints, and create instances).

[0068] The specific implementation process includes the following sub-steps:

[0069] S11. Define the core concept class hierarchy (classes and subclasses).

[0070] In order to comprehensively cover the motor fault knowledge, the embodiment constructs a five-dimensional concept tree structure from top to bottom, and all concepts are mapped to "class" in the ontology.

[0071] (1) Equipment and component class:

[0072] The root node is defined as the motor.

[0073] The first-level child node is defined as the permanent magnet synchronous motor, and is further subdivided into second-level child nodes according to the physical structure:

[0074] 1. Stator system: including stator core, stator winding, and insulation layer.

[0075] 2. Rotor system: including rotor core, permanent magnet, and rotating shaft.

[0076] 3. Support system: including bearing (further subdivided into inner ring, outer ring, rolling body, and retainer) and end cover.

[0077] 4. Cooling system and sensing system (including encoder and temperature sensor).

[0078] (2) Fault phenomenon class:

[0079] The observable indicators representing abnormal states of the motor are defined, including:

[0080] 1. Vibration anomaly: subdivided into high-frequency vibration and axial movement.

[0081] 2. Thermal anomaly: subdivided into local overheating and rapid temperature rise.

[0082] 3. Electrical anomaly: subdivided into three-phase current imbalance and abnormal back electromotive force.

[0083] 4. Acoustic anomaly: subdivided into electromagnetic whistling and mechanical friction sound.

[0084] (3) Fault mode class:

[0085] The root causes leading to the phenomenon are defined, including:

[0086] 1. Electromagnetic fault: such as inter-turn short circuit, magnetic steel demagnetization, and insulation breakdown.

[0087] 2. Mechanical fault: such as bearing pitting, air gap eccentricity, dynamic imbalance, and loose footing.

[0088] (4) Diagnosis and detection class:

[0089] The technical means for discovering faults are defined, including current characteristic analysis method, vibration spectrum analysis method, and infrared thermal imaging detection.

[0090] (5) Maintenance decision class:

[0091] Define the specific measures to solve the fault, including replacing parts, insulation repair (including rewinding and impregnation), precision adjustment (including dynamic balance correction and centering).

[0092] S12. Define the semantic relationship between object attributes.

[0093] In order to endow the knowledge graph with reasoning ability, this sub-step defines the multiple logical relationships between entities, and sets the definition domain and value domain of the relationship:

[0094] 1. Composition relationship: define "contains components" to connect devices and components, such as connecting permanent magnet synchronous motor and stator system, etc. This relationship is set to have transitivity, that is, if the motor contains the stator system, and the stator system contains the stator winding, then it can be automatically inferred that the motor contains the stator winding.

[0095] 2. Causal relationship: the definition domain of "cause" is fault mode, and the value domain is fault phenomenon. For example: (inter-turn short circuit, causes, local overheating). By …… caused is defined as the inverse relationship of "cause".

[0096] 3. Diagnosis relationship: define "detected by" to connect fault mode and detection method.

[0097] 4. Position relationship: define "located in" to connect fault mode and device components. For example: (pitting, located in, bearing inner ring).

[0098] 5. Governance relationship: define "solved by" to connect fault mode and maintenance measures.

[0099] S13. Obtain structured data and unstructured data and define data attributes.

[0100] Obtain structured data in the design and manufacturing stage and unstructured data generated in the operation and maintenance stage. In order to fuse structured data and unstructured data, this sub-step defines detailed numerical attributes for specific classes:

[0101] 1. For the fault mode class, define the following attributes:

[0102] Risk priority number (RPN value): data type is integer, data source is mapped from FMEA table, used to represent the theoretical risk level.

[0103] Severity (S): data type is integer (1-10), data source is mapped from FMEA table.

[0104] Actual occurrence frequency: data type is integer, data source is the statistical aggregation of operation log, initial value is 0, and it is dynamically accumulated with system running.

[0105] Confidence: data type is float (0.0-1.0), which represents the reliability of the fault knowledge.

[0106] 2. For the maintenance decision class, the following attributes are defined:

[0107] Average cost: data type is float.

[0108] Average time consumption: data type is float, unit is hour.

[0109] S14. Set logical constraints and axioms.

[0110] In order to ensure the consistency of the knowledge graph and prevent logical conflicts, the embodiment presets mutual exclusion constraints in the ontology: set the device components, fault phenomena and fault modes to be mutually exclusive. That is, the system limits that an entity cannot belong to both the "component" class and the "fault phenomenon" class at the same time, thereby avoiding logical confusion during knowledge extraction.

[0111] S2. Dual-channel knowledge extraction.

[0112] Step S2 aims to convert the multi-source heterogeneous data (structured data and unstructured data) obtained in step S1 into triple knowledge that can be understood by a computer and conforms to the ontology specification. In view of the heterogeneity of the data sources, this step S2 designs a "dual-channel" extraction mechanism, and the specific implementation process is as follows:

[0113] S21. Unstructured text knowledge extraction.

[0114] For unstructured text data such as maintenance work orders, fault logs and expert diagnosis reports, the embodiment constructs a BERT-BiLSTM-CRF model and performs entity recognition through the BERT-BiLSTM-CRF model, and performs relation extraction combined with an attention mechanism, as shown in Figure 2 . Specifically:

[0115] S211. Data vectorization and encoding (BERT layer): input the cleaned operation and maintenance text sequence into the pre-trained BERT model (the BERT model is one of the layers of the BERT-BiLSTM-CRF model, also known as the BERT layer). The BERT layer first converts the input text into a superimposed form of word vectors, position vectors and paragraph vectors. Then, through the self-attention mechanism (Self-Attention) of the multi-layer Transformer encoder, the text is scanned bidirectionally. The BERT layer can dynamically adjust the semantic representation of the word vector according to the context. For example, in the texts "rotor imbalance leads to vibration" and "rotor core loosening", the BERT layer can capture the subtle semantic differences of "rotor" in different contexts, solving the polysemy problem of traditional models.

[0116] S212. Context feature extraction (BiLSTM layer): input the word vector sequence containing deep semantic information output by the BERT layer into a bidirectional long short-term memory network (BiLSTM), wherein the BiLSTM layer includes a forward LSTM and a backward LSTM. The forward LSTM processes the sequence from left to right, capturing the context information (for example, after the appearance of “inter-tie short circuit”, “overheating” is more likely to appear subsequently); the backward LSTM processes from right to left, capturing the context information. Finally, the hidden layer state vectors of the two directions are spliced to form a feature vector containing complete context dependency.

[0117] S213. Entity label decoding (CRF layer): input the feature vector output by the BiLSTM layer into a conditional random field (CRF) layer for sequence labeling decoding. The CRF layer learns a label transition probability matrix to impose logical constraints on the prediction results. For example, in the BIO labeling system, the label I-fault mode (intermediate word in mechanism) must immediately follow B-fault mode (first word in mechanism) or I-fault mode, and cannot appear after O (non-entity).

[0118] For example, the input text is: “The stator winding of the motor is caused by the blockage of the cooling water channel to abnormally increase the temperature”, and the output sequence is: the CRF decodes the optimal label sequence, and identifies the entities: [equipment component: stator winding], [fault mode: blockage of cooling water channel], and [fault phenomenon: abnormal temperature rise].

[0119] S214. Semantic relation extraction layer: after identifying the entity, a classification model based on an attention mechanism is used to determine the relationship type between the entities. Specifically, the sentence containing the entity pair is input into the model, and the attention mechanism automatically gives higher weights to trigger words (such as “because”, “cause”, “result in”). The classification model based on the attention mechanism calculates that the probability of the entity pair (blockage of cooling water channel, abnormal temperature rise) belonging to the relationship type “cause” is 98%, thereby generating the triple knowledge: (blockage of cooling water channel, cause, abnormal temperature rise).

[0120] S22. Structured data knowledge extraction.

[0121] For structured data, since it has a standard column structure itself, this step uses an extract, transform, and load (ETL) mapper to convert table rows into knowledge graph subgraphs through preset rules.

[0122] S221. Field-Entity Mapping Rule: Map the failure mode column content (e.g., "bearing inner race peeling") in structured data to the failure mode entity in ontology, and map the failure effect column content (e.g., "high frequency noise") to the failure phenomenon entity in ontology; map the component name column content (e.g., "drive end bearing") to the equipment component entity in ontology.

[0123] S222. Attribute Extraction Rule: Extract the quantitative evaluation parameters in structured data as numerical attributes of nodes for storage, extract the S, O, and D columns of numerical values, calculate the risk priority number (RPN) = S x O x D, and encapsulate the above numerical values as attribute dictionaries and mount them to the corresponding failure mode nodes.

[0124] S223. Relationship Construction Rule: According to the logical structure of FMEA, automatically instantiate the following relationships: construct a (failure mode, cause, failure phenomenon) triple; construct a (failure mode, located in, equipment component) triple.

[0125] The following is an example of input data behavior: {Component: Rotor, Failure Mode: Magnetic Steel Demagnetization, Failure Effect: Reduced Back EMF, S: 9, O: 3, D: 5}.

[0126] The output triple is: 1. (Magnetic Steel Demagnetization, Cause, Reduced Back EMF) 2. (Magnetic Steel Demagnetization, Located in, Rotor).

[0127] The output attribute is: Magnetic Steel Demagnetization, RPN = 135.

[0128] S3. Knowledge Fusion.

[0129] This step aims to solve the problems of entity name diversity, attribute description redundancy, and data source conflict in the "instance layer original triples" generated in step S2. Through entity alignment, attribute fusion, and conflict resolution mechanisms, fragmented knowledge from different data sources is integrated into a unified high-quality, unambiguous knowledge graph.

[0130] The specific steps are as follows:

[0131] S31. Entity Alignment Based on Semantic Similarity: Since there are a large number of non-standard terms (e.g., "stator winding" is expressed as "coil", "winding wire") in unstructured text data, and standard terms are used in structured data, entity alignment is needed to eliminate semantic ambiguity. The specific operations include:

[0132] 1. Vector representation: Use the BERT model trained in step S21 to extract the semantic vector representation of all entity names in the original triple. Let the two entity names to be aligned be (from maintenance log) and (structured data), whose corresponding vectors are and .

[0133] 2. Similarity calculation: the cosine similarity formula is used to calculate the semantic closeness of the two:

[0134] (1)

[0135] 3. Threshold determination and merging: set the alignment threshold to (preferably 0.85 in this embodiment).

[0136] If , it is determined that and refer to the same object. The system will perform a merging operation: keep the standard name of the symbol ontology specification (usually in FMEA), store as an "alias" attribute of the entity, and redirect all relationships originally pointing to to .

[0137] For example, the similarity between "stator coil" (from the entity in the maintenance log) and "stator winding" (from the entity in the structured data) is 0.92, so the system automatically replaces all "stator coil" nodes in the knowledge graph with "stator winding", eliminating redundant nodes in the knowledge graph.

[0138] S32. Attribute fusion based on "master-slave strategy".

[0139] For the same fault mode entity, structured data provides theoretical design attributes, while maintenance logs provide field statistical attributes. This embodiment uses a "master-slave fusion strategy" to construct a full-dimensional fault portrait. The specific operations include:

[0140] 1. Anchor the main data source: the nodes generated by structured data (FMEA data) are the main data source, and their risk priority number (RPN), severity (S), occurrence (O), and detectability (D) are retained as priori theoretical attributes, and they are given a high confidence mark.

[0141] 2. Mount the slave data source: the nodes generated by unstructured data (maintenance logs) are the slave data, and the number of occurrences of the fault mode in the historical time window is counted to generate the actual occurrence frequency (Frequency) attribute; the average maintenance time (MTTR) attribute is generated by counting the average time of the maintenance work order corresponding to the fault.

[0142] 3. Attribute merging: mount the statistical attributes of the above "slave data source" to the entity node determined by the "main data source".

[0143] For example, FMEA data: {entity: bearing pitting, RPN: 160}, log data: {entity: bearing spalling, frequency: 56 times, MTTR: 4.5 hours}, fused node: {entity: bearing pitting, alias: bearing spalling, RPN: 160, frequency: 56, MTTR: 4.5}.

[0144] S33. Multi-source data conflict resolution.

[0145] When different data sources have logical conflicts in describing the same attribute, the preset confidence priority rules are used for processing. The specific rules are as follows:

[0146] 1. Theoretical parameter conflict: If the "speculative risk level" extracted from unstructured data (maintenance log text) is inconsistent with the "severity S" in structured data (FMEA), the structured data (FMEA) is used as the reference (because FMEA represents design specifications).

[0147] 2. Measured parameter conflict: If the difference between the "estimated maintenance time" in structured data (FMEA) and the "average maintenance time" in unstructured data (maintenance log text) exceeds 20%, the unstructured data (maintenance log text) is used as the reference (because the log represents the real situation on site), and the attribute is marked as "to be corrected" for subsequent feedback to the design department.

[0148] 3. Artificial intervention interface: For automatic fusion results with a confidence level below the threshold (such as 0.6 in this embodiment), the system marks them as "to be audited" and pushes them to the expert audit interface for manual confirmation before writing them into the official knowledge graph.

[0149] S34. Constructing a fault knowledge graph.

[0150] The triplets after cleaning in steps S31-S33 are batch imported into a graph database (such as Neo4j). Specifically:

[0151] 1. Node creation: Create nodes for fault modes, fault phenomena, device components, etc., and write in the fused attributes.

[0152] 2. Relationship creation: Create edges such as cause, location, and contain, and add weight attributes to the edges (initial weight is set to 1.0).

[0153] 3. Index construction: Create full-text indexes for the "fault phenomenon" and "fault mode" node name attributes to speed up subsequent retrieval and reasoning speed.

[0154] S4. Intelligent fault reasoning and dynamic updating based on knowledge graph.

[0155] Step S4 aims to respond to the user's diagnosis request and provide accurate fault location using the constructed unified knowledge graph; at the same time, the newly added maintenance records are used to correct the graph weight in reverse, and the dynamic updating of the knowledge base is realized. The specific steps are as follows:

[0156] S41. Intelligent fault reasoning based on "theoretical (i.e. theoretical risk priority number, RPN value) and practical (i.e. actual occurrence frequency)" double weighting algorithm: when the operation and maintenance personnel input the natural language description of the fault phenomenon, the system is no longer a simple keyword matching, but a probability reasoning based on the knowledge graph path. The specific reasoning logic is as follows:

[0157] 1. Semantic analysis and anchor positioning: use the BERT-BiLSTM-CRF model trained in S21 to analyze the user's input query statement (for example: "There is a clear whistling sound at the front cover of the motor during operation, and the temperature rises quickly"). The system identifies the fault phenomenon entities [whistling] and [temperature rise], and locates the corresponding fault phenomenon nodes in the graph as the starting point of reasoning.

[0158] 2. Multi-hop path tracing: taking the located fault phenomenon node as the starting point, performing reverse breadth-first search along the edges of the "caused by" relationship, finding all reachable fault mode nodes, and forming a candidate fault set .

[0159] 3. Comprehensive recommendation score calculation: in order to solve the ambiguity problem that "one phenomenon may correspond to multiple reasons", this embodiment calculates the recommendation score of each candidate fault reason as follows:

[0160] (2)

[0161] In formula (2), is the risk priority number of the fault in FMEA, representing the most dangerous possibility in theory; represents the maximum value in the risk priority number; is the actual occurrence frequency of the fault in the knowledge graph, representing the most possible possibility in reality; represents the comprehensive of all fault frequencies; is the adjustment coefficient, with a value range of 0-1.

[0162] 4. Sorting and output: sort the candidate fault set by comprehensive recommendation score from high to low, and output the most likely fault reason and its corresponding maintenance measures and detection methods.

[0163] ​For example, for the "whistling" phenomenon, the knowledge graph finds two causes: "bearing cage fracture" (high RPN, low frequency) and "uneven air gap" (medium RPN, high frequency). After calculating the score, the system preferentially recommends "uneven air gap" and prompts to check the rotor eccentricity, which is consistent with the actual working condition of motor aging.

[0164] S42. Dynamic updating of the knowledge graph based on closed-loop feedback.

[0165] The traditional knowledge base is fixed after construction, while the present embodiment constructs a closed-loop updating mechanism that interfaces with the maintenance management system (CMMS), enabling the knowledge graph to have a dynamic updating mechanism.

[0166] The specific updating process is as follows:

[0167] 1. Event listening: The system background listens to the work order state flow of the maintenance management system (CMMS) in real time. When it detects that the work order state has changed to "closed" and contains the "confirmed cause" field, the updating process is triggered.

[0168] 2. Real-time weight correction: including analyzing the confirmed cause in the work order (for example, "lubricating grease dryness"). Search for the corresponding fault mode node in the knowledge graph. Perform atomic operation: Node.Frequency = Node.Frequency + 1. Effect: With the accumulation of this fault frequency, its weight component will gradually increase in the reasoning formula of S41. This means that if a certain type of fault has recently broken out, the system will preferentially recommend that fault when diagnosing, achieving self-adaptation to changes in working conditions.

[0169] 3. New knowledge discovery and storage: If the entity (such as the new type of component "ceramic bearing") or relationship extracted from the new work order does not exist in the existing graph, and the comprehensive confidence of the model extraction is higher than the preset new knowledge threshold, then the system automatically creates new nodes and edges in the candidate area of the knowledge graph. The formula for calculating the comprehensive confidence is as follows:

[0170] (3)

[0171] In formula (3), is the average probability value of the entity label sequence output by the BERT-BiLSTM-CRF model (i.e., the decoding score normalization value of the CRF layer), is the relationship category normalized exponential function probability value output by the classification model based on the attention mechanism, is the weight coefficient (the present embodiment preferably takes 0.5).

[0172] In the present embodiment, the new knowledge threshold is set to 0.95, i.e., when ​If so, the system automatically creates a new node and a new edge in the candidate area of the knowledge graph, and assigns the comprehensive confidence value calculated above to the "confidence" data attribute of the new node defined in step S13 as the initial confidence record of the knowledge point; then a trigger is sent to the expert review process to inform the field experts to review, and after the review is passed, the candidate knowledge is formally incorporated into the main knowledge graph, completing the expansion of the knowledge base, thereby realizing the dynamic updating of the knowledge graph.

[0173] In order to further verify the effectiveness, scientificity and progressiveness of the motor-oriented fault mode knowledge graph construction and updating method proposed in the present application compared with the prior art, the present embodiment selects the full life cycle operation and maintenance data of a permanent magnet synchronous motor (PMSM) of a certain motor manufacturing enterprise to perform full-process verification. The data includes structured data in the design and manufacturing stage (such as FMEA document data) and unstructured data generated in the operation and maintenance stage (such as historical maintenance work order records).

[0174] In order to intuitively show the complete processing flow of the present application from multi-source heterogeneous data to intelligent diagnosis results, the following selects representative sample data from the data, and focuses on detailing the specific implementation process of the "double-channel knowledge extraction (step S2)", "knowledge fusion (step S3)" and "double-weighted intelligent reasoning (step S4)" steps in the method of the present application.

[0175] Part of the structured design data (FMEA) sample selected in the present embodiment is shown in Table 1, and part of the unstructured operation and maintenance data (maintenance work order) sample selected is shown in Table 2.

[0176] Table 1 Part of the structured design data sample

[0177]

[0178] Table 2 Part of the unstructured operation and maintenance data sample

[0179]

[0180] Based on the sample data in Tables 1 and 2 above, the specific execution logic of the method of the present embodiment is as follows:

[0181] 1. Double-channel knowledge extraction:

[0182] For the stator winding inter-turn short circuit record in Table 1, the system automatically instantiates the fault mode node "stator inter-turn short circuit" using the ETL mapper, and extracts the RPN value (135) as its prior attribute.

[0183] ​For the repair order in Table 2, the system uses a pre-trained BERT-BiLSTM-CRF model to perform sequence labeling on the text, for example, identifying the entity "large current fluctuation" (fault phenomenon) and "coil turn-to-turn short circuit" (fault mode), and extracting the relationship to construct a (coil turn-to-turn short circuit, cause, large current fluctuation) triple.

[0184] 2. Knowledge fusion:

[0185] Entity alignment: The system calculates the semantic similarity and identifies that "stator winding" in FMEA and "coil" in the order refer to the same component, and merges "coil turn-to-turn short circuit" to the standard node "stator turn-to-turn short circuit".

[0186] Attribute fusion: The system traverses all orders in the past three years and finds that "stator turn-to-turn short circuit" has occurred 142 times, "bearing pitting" has occurred 85 times, "cooling water channel blockage" has occurred 56 times, and "permanent magnet demagnetization" has occurred 8 times. The system mounts these frequency data to the corresponding nodes of the knowledge graph.

[0187] 3. Double-weighted intelligent reasoning:

[0188] When the user inputs the query "motor current abnormality", the system calculates the comprehensive recommendation score of each candidate fault based on formula (2). In the calculation process of this embodiment, in order to ensure the comparability of numerical values in different dimensions, the parameters are normalized as follows:

[0189] (1) Normalization of theoretical risk denominator (i.e. the maximum value of risk priority number, ): set to the FMEA theoretical maximum value of 1000 (i.e. severity (S) 10 x occurrence (O) 10 x detection (D) 10).

[0190] (2) Normalization of actual occurrence frequency denominator (i.e. the comprehensive of all fault frequencies, ): According to statistics, under the current query condition, the actual occurrence frequency of all candidate faults (including the top 4 in Table 3 and other long-tail low-frequency faults) is 532 times.

[0191] (3) Adjustment coefficient ( ): set , i.e. the comprehensive score is composed of 40% of the theoretical risk part and 60% of the actual frequency part, and , indicating that more weight is given to the actual data on site.

[0192] Based on the above settings, the calculation results of the comprehensive recommendation scores of each fault mode are shown in Table 3.

[0193] Table 3 Calculation results of comprehensive recommendation scores of each fault mode

[0194]

[0195] The calculation process is described as follows by taking "stator inter-turn short circuit" as an example:

[0196] (1) Theoretical part calculation: divide the RPN value (135) by the maximum value (1000) to obtain a normalized value of 0.135; multiply by the weight coefficient 0.4 to obtain a theoretical score component of 0.054.

[0197] (2) Actual part calculation: divide the actual frequency (142) by the total frequency of candidate faults (532) to obtain a frequency ratio of about 0.267; multiply by the weight coefficient 0.6 to obtain an actual score component of 0.160.

[0198] (3) Comprehensive summation: add the two parts 0.054 + 0.160 to obtain a final comprehensive recommendation score of 0.214.

[0199] The calculation process is the same as above, and will not be repeated here. The calculation results are shown in Table 3. Although the theoretical RPN (160) of "bearing pitting" is higher than that of "stator inter-turn short circuit" (135), the latter has a higher actual occurrence frequency in the field, and after double weighting, the system correctly ranks "stator inter-turn short circuit" in the first recommended position, which is consistent with the field experience of "high-frequency fault priority investigation", and also takes into account the attention to high-risk faults.

[0200] In order to verify the technical advantages of the BERT-BiLSTM-CRF model adopted in the present application in processing complex professional texts in the field of electric machines, the present embodiment compares it with the traditional BiLSTM-CRF model (without pre-training layer) and the IDCNN-CRF model. The precision (Precision), recall (Recall) and F1 value (F1-Score) are used as evaluation indexes, and the experimental results are shown in Table 4 and Figure 3

[0201] Table 4 Comparison of entity recognition performance of different knowledge extraction models

[0202]

[0203] As can be seen from the data in Table 4, the F1 value of the BERT-BiLSTM-CRF model adopted in the present application is improved by about 12-15 percentage points compared with the traditional baseline model, which proves that the introduction of the BERT pre-training layer can effectively capture the deep semantic features of professional terms such as "stator inter-turn short circuit" and "tooth slot torque", and significantly improve the extraction quality of unstructured data.

[0204] ​To verify the diagnostic accuracy of the double-weighting algorithm based on the "theoretical risk priority number and actual occurrence frequency" proposed in the present application, 200 historical real fault cases are selected for backtracking test in the present embodiment, and the TOP-3 diagnostic accuracy (i.e. the proportion of real fault causes appearing in the top 3 recommended results) of three different recommendation strategies is compared, as shown in Table 5 and Figure 4 Fig. 5, wherein strategy A is a strategy based only on the theoretical risk priority number (RPN), strategy B is a strategy based only on the actual occurrence frequency, and strategy C is a strategy of the present application using the double-weighting algorithm based on the "theoretical risk priority number and actual occurrence frequency".

[0205] Table 5 Comparison of TOP-3 diagnostic accuracy of different recommendation strategies

[0206]

[0207] The experimental data show that the double-weighting algorithm of the present application improves the TOP-3 diagnostic accuracy to 89.8%, proving that the deep integration of the theoretical norms of the design end and the practical experience of the operation and maintenance end can effectively overcome the limitations of a single data source and significantly improve the intelligent level and practical value of motor fault diagnosis.

[0208] The method of the present embodiment successfully breaks down the barriers between the structured data generated in the design and manufacturing stage and the unstructured data generated in the operation and maintenance stage by constructing a standardized motor fault domain ontology model and using the double-channel mechanism of BERT-BiLSTM-CRF deep learning model and FMEA rule mapping. Using the reasoning algorithm based on the double-weighting of "theoretical risk priority number and actual occurrence frequency", not only can complex faults such as "stator inter-turn short circuit" and "bearing pitting" be accurately located, but also the diagnostic strategy can be dynamically adjusted according to the field conditions. Through closed-loop linkage with the maintenance management system, the present application realizes incremental adaptive updating of the knowledge graph, solves the problems of easy aging and high maintenance cost of the knowledge base of the traditional expert system, and has significant application value in the whole life cycle management of industrial motors.

[0209] II. Embodiment 2

[0210] The present embodiment discloses a motor-oriented fault mode knowledge graph construction and updating system, comprising:

[0211] The motor fault domain ontology model construction module is configured to define a core concept class hierarchy for representing motor fault knowledge, define object attributes and semantic relationships, obtain structured data and unstructured data and define data attributes, and then set logical constraints and axioms.

[0212] The double-channel knowledge extraction module is configured to extract knowledge from unstructured data by using a deep learning model and extract knowledge from structured data by using a mapping method based on the motor fault domain ontology model, and generate original triple knowledge.

[0213] The knowledge fusion module is configured to perform entity alignment, attribute fusion and conflict resolution on the original triple knowledge, integrate knowledge from different data sources to construct a high-quality and unambiguous knowledge graph, and store the knowledge graph in a graph database.

[0214] The intelligent fault reasoning and dynamic updating module is configured to respond to a diagnosis request of a user by using the constructed knowledge graph, implement intelligent reasoning and sorting from a fault phenomenon to a root cause based on a preset double-weighting algorithm, and if a new component or a new fault relationship that does not exist in the knowledge graph is included in the work order and the extraction confidence is higher than a preset new knowledge threshold, incorporate the new component or the new fault relationship into the main knowledge graph to implement dynamic updating of the knowledge graph.

[0215] For the system embodiment, since it basically corresponds to the method embodiment, the implementation process of the functions and roles of each module in the system is specifically described in the implementation process of the corresponding steps in the method described in Embodiment 1, and will not be repeated here.

[0216] The above only describes the basic principles and preferred embodiments of the present application, and those skilled in the art can make many changes and improvements based on the above description, and these changes and improvements should belong to the protection scope of the present application.

Claims

1. A method for constructing and updating a fault mode knowledge graph for motors, characterized in that, Including the following steps: Constructing an ontology model for the motor fault domain: Define a hierarchy of core concept classes to represent motor fault knowledge, define object attributes and semantic relationships, acquire structured and unstructured data and define data attributes, and then set logical constraints and axioms; Dual-channel knowledge extraction: Based on the aforementioned motor fault domain ontology model, a deep learning model is used to extract knowledge from unstructured data, and a mapping method is used to extract knowledge from structured data to generate original triplet knowledge. Knowledge fusion: The original triple knowledge is subjected to entity alignment, attribute fusion and conflict resolution. Knowledge from different data sources is integrated to construct a high-quality, unambiguous knowledge graph and stored in a graph database. Intelligent fault reasoning and dynamic updating based on knowledge graph: Utilizing the constructed knowledge graph, in response to user diagnostic requests, intelligent reasoning and ranking from fault phenomena to root causes are achieved based on a preset dual-weighted algorithm; at the same time, if the work order contains new components or new fault relationships that do not exist in the knowledge graph and the extraction confidence is higher than the preset new knowledge threshold, it is incorporated into the main knowledge graph to achieve dynamic updating of the knowledge graph.

2. The method according to claim 1, characterized in that, The core concept class hierarchy is a top-down five-dimensional concept tree structure, with the five dimensions being equipment and component class, fault phenomenon class, fault mode class, diagnosis and detection class, and maintenance decision class; the object attributes and semantic relationships include at least the compositional relationship connecting equipment and components, the causal relationship connecting fault modes and fault phenomena, the diagnostic relationship connecting fault modes and detection methods, the positional relationship connecting fault modes and equipment components, and the governance relationship connecting fault modes and maintenance measures; the data attributes include at least attributes for the fault mode class and attributes for the maintenance decision class.

3. The method according to claim 1, characterized in that, The method of extracting knowledge from unstructured data using a deep learning model specifically includes: A BERT-BiLSTM-CRF model is constructed to perform entity recognition on unstructured data. The BERT-BiLSTM-CRF model includes a BERT layer for converting unstructured data into context-related vectors, a BiLSTM layer for bidirectional extraction of sequence context features, and a CRF layer for annotating and decoding entity label sequences. After identifying entities, an attention-based classification model is used to extract semantic relationships between entities, thereby generating triplet knowledge.

4. The method according to claim 1, characterized in that, The knowledge extraction from structured data using a mapping method specifically includes: Based on the preset field-entity mapping rules, the field content in the structured data is mapped to the corresponding entity in the motor fault domain ontology model; Based on attribute extraction rules, the Risk Priority Number (RPN) in the structured data is extracted and calculated as an attribute of the failure mode node. Based on relational construction rules, (fault mode, cause, fault phenomenon) triples and (fault mode, location, device component) triples are automatically instantiated according to the logical structure of structured data.

5. The method according to claim 3, characterized in that, The entity alignment specifically includes: The semantic vector representation of all entity names in the original triples is extracted using the BERT-BiLSTM-CRF model; Calculate and determine the semantic similarity of entities. If the similarity exceeds the preset alignment threshold, merge them, retain the standard name of the entity, and record the alias.

6. The method according to claim 1, characterized in that, The attribute fusion adopts a master-slave strategy, specifically including: The nodes generated from structured data serve as the primary data source, preserving prior theoretical attributes, which include at least risk priority, severity, occurrence, and detectability. The nodes generated from unstructured data are used as the data source to statistically analyze the actual frequency of failures and the average repair time. The statistical attributes from the data source will be attached to the entity nodes determined by the main data source to achieve attribute fusion.

7. The method according to claim 1, characterized in that, The conflict resolution specifically includes: In case of conflicting theoretical parameters, structured data shall prevail. In case of conflicting measured parameters, unstructured data shall prevail. For fused data with a confidence level below the preset threshold, the data confirmed by manual review shall prevail.

8. The method according to claim 1, characterized in that, The process of utilizing the constructed knowledge graph to respond to user diagnostic requests and achieving intelligent reasoning and ranking from fault symptoms to root causes based on a preset dual-weighted algorithm specifically includes: Using the trained deep learning model, the query statement entered by the user is parsed, and the corresponding fault phenomenon node is located in the knowledge graph; Starting from the fault phenomenon node, a reverse breadth-first search is performed along the causal relationship to obtain a candidate fault set; The comprehensive recommendation score for each candidate fault is calculated using a pre-set dual weighted algorithm of "theoretical risk priority number and actual occurrence frequency". Sort the candidate fault set from high to low according to the comprehensive recommendation score, and output the fault cause and its corresponding maintenance measures and detection methods.

9. The method according to claim 1, characterized in that, If a work order contains new components or new fault relationships that do not exist in the knowledge graph and the extraction confidence level is higher than a preset new knowledge threshold, it is then incorporated into the main knowledge graph to achieve dynamic updates of the knowledge graph. Specifically, this includes: Monitor the work order status of the maintenance management system in real time, and trigger the update process when the work order is completed and the cause is diagnosed; If the cause of diagnosis in the work order exists in the existing knowledge graph, then the corresponding fault mode node is retrieved in the knowledge graph, and the actual occurrence frequency attribute of the fault mode node is cumulatively corrected. If the entity or relation extracted from the new work order does not exist in the existing knowledge graph and the overall confidence level is higher than the preset new knowledge threshold, the corresponding new node and new edge will be automatically created in the candidate area, and the expert review process will be triggered. After the review is approved, it will be incorporated into the main knowledge graph to realize the dynamic update of the knowledge graph.

10. A fault mode knowledge graph construction and updating system for motors, characterized in that, include: The module for building the ontology model in the field of motor faults is configured to define the core concept class hierarchy used to represent knowledge of motor faults, define object attributes and semantic relationships, acquire structured and unstructured data and define data attributes, and then set logical constraints and axioms. The dual-channel knowledge extraction module is configured to extract knowledge from unstructured data using a deep learning model and from structured data using a mapping method based on the ontology model of the motor fault domain, generating original triplet knowledge. The knowledge fusion module is configured to perform entity alignment, attribute fusion, and conflict resolution on the original triple knowledge, and to integrate knowledge from different data sources to construct a high-quality, unambiguous knowledge graph and store it in a graph database. The intelligent fault reasoning and dynamic update module is configured to utilize the constructed knowledge graph to respond to user diagnostic requests and achieve intelligent reasoning and ranking from fault phenomena to root causes based on a preset dual-weighted algorithm. At the same time, if the work order contains new components or new fault relationships that do not exist in the knowledge graph and the extraction confidence is higher than the preset new knowledge threshold, it is incorporated into the main knowledge graph to achieve dynamic updates of the knowledge graph.

Images