Device fault and maintenance knowledge intelligent association method based on large model

Abstract

The invention discloses an equipment fault and maintenance knowledge intelligent association method based on a large model, and relates to the technical field of equipment intelligent operation and maintenance and industrial artificial intelligence, and the method comprises the steps: carrying out the structural analysis of historical fault work order data and maintenance record data of target equipment, and constructing a fault-maintenance knowledge pair data set; performing semantic coding on the fault description text and the maintenance scheme text by using a pre-trained large model, generating a corresponding semantic vector, and constructing a fault-maintenance semantic association map; and performing semantic coding on the real-time fault description, screening candidate fault nodes based on semantic similarity, calculating the semantic matching degree of the maintenance scheme in the semantic association map, and outputting a target maintenance scheme sequence after matching degree sorting. According to the invention, the intelligent level of equipment fault maintenance and the reuse efficiency of historical maintenance knowledge are improved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent equipment operation and maintenance and industrial artificial intelligence technology, and in particular to an intelligent association method for equipment fault and maintenance knowledge based on a large model. Background Technology

[0002] With the continuous development of industrial internet, intelligent manufacturing, and equipment lifecycle management technologies, equipment operation and maintenance models are gradually evolving from traditional reactive maintenance and periodic inspections to predictive maintenance and intelligent repair centered on data-driven, knowledge-assisted, and intelligent decision-making. In complex industrial scenarios such as power, rail transportation, petrochemicals, and high-end equipment manufacturing, equipment operating environments are highly variable, structures are complex, and fault types exhibit high heterogeneity. Fault diagnosis and repair processes often rely on a large amount of historical work orders, repair records, and expert experience. While existing equipment management systems can store and query fault work orders and repair records, they primarily rely on structured fields or keyword searches, making it difficult to perform deep semantic understanding and correlation modeling of fault descriptions and repair solutions in natural language form. This results in persistent problems such as "difficulty in reusing repair experience for similar faults," "low utilization of historical knowledge," and "reliance on manual experience for repair decisions." Especially in the emergency fault handling of complex equipment, how to quickly and accurately match the most valuable repair solutions from massive amounts of historical fault and repair data has become a key technical bottleneck restricting the improvement of intelligent equipment operation and maintenance levels.

[0003] CN119690051A discloses a method and system for intelligent industrial fault prediction and maintenance based on a large model. This scheme collects equipment operation data and extracts features, utilizes a pre-trained large model to predict equipment faults, and generates dynamic maintenance plans based on this, thereby reducing equipment operation risks and improving production efficiency. This technology focuses on equipment condition prediction and maintenance strategy optimization, emphasizing the capabilities of large models in time series modeling and fault trend prediction. However, its core processing object is still mainly numerical operational data, without systematically semantically modeling historical fault work orders and maintenance texts, nor constructing a computable semantic relationship between fault descriptions and maintenance plans. Therefore, it is difficult to achieve refined plan matching and reuse based on historical maintenance knowledge.

[0004] CN120894009A proposes a method, system, and storage medium for equipment fault diagnosis and repair based on a large model. This approach inputs equipment fault descriptions, operational data, and operational images into a pre-trained large model to identify the causes of target faults and generate priority troubleshooting steps, thereby assisting manual fault diagnosis and repair. While this method improves the collaboration and efficiency of fault diagnosis and repair to some extent, its technical approach mainly revolves around the reasoning process of "single fault – single repair," focusing on fault cause identification driven by current multimodal information. It lacks systematic knowledge extraction and long-term accumulation of historical faults and repair records, and it does not form a semantic-level knowledge association network for similar fault scenarios, making it difficult to support cross-time and cross-equipment repair experience transfer and ranking recommendations.

[0005] In summary, existing equipment operation and maintenance technologies based on large models generally suffer from problems such as insufficient utilization of historical fault and maintenance knowledge, lack of stable semantic association mechanism between fault description and maintenance plan, and limited interpretability and reusability of maintenance plan recommendation results. This invention provides an intelligent association method for equipment fault and maintenance knowledge based on large models, which effectively solves the problems of fragmented maintenance knowledge, low matching efficiency, and difficulty in reusing experience in existing technologies, thereby improving the intelligence and knowledge level of equipment fault handling. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the specification abstract and the title of the invention, to avoid obscuring the purpose of this section, the specification abstract, and the title of the invention. Such simplifications or omissions shall not be used to limit the scope of the invention.

[0007] Given that existing equipment fault diagnosis and repair technologies often store historical fault work orders and repair records in a scattered manner in text form, lacking a unified semantic modeling and association mechanism, which leads to difficulties in reusing repair experience in similar fault scenarios, low efficiency in matching repair solutions, and reliance on manual experience for results, this invention is proposed.

[0008] Therefore, the problem to be solved by this invention is how to use a large model to perform unified semantic encoding and association modeling of equipment historical fault descriptions and maintenance plans, so as to achieve accurate matching of fault semantics and intelligent sorting output of maintenance plans in real-time fault scenarios, thereby improving the accuracy, efficiency and knowledge reuse capability of equipment fault maintenance decisions.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for intelligently associating equipment fault and maintenance knowledge based on a large model, which includes, Collect historical fault work order data and maintenance record data of the target equipment, perform structured parsing on the historical fault work order data and the maintenance record data, and construct a fault-maintenance knowledge pair dataset; The fault-maintenance knowledge is input into the fault description text and maintenance plan text of the dataset and semantically encoded into a pre-trained large model to generate a fault semantic vector set and a maintenance semantic vector set, respectively, and a fault-maintenance semantic association graph is constructed. Obtain the real-time fault description text to be matched, and use the big model to perform semantic encoding on the real-time fault description text to generate a real-time fault semantic vector. Calculate the semantic similarity value between the real-time fault semantic vector and each vector in the fault semantic vector set, and filter fault nodes with semantic similarity values ​​greater than a preset threshold to form a candidate fault node set; Based on the fault-maintenance semantic association graph, the maintenance semantic vector associated with each node in the candidate fault node set is obtained, the semantic matching score between the real-time fault semantic vector and each maintenance semantic vector is calculated, and the target maintenance scheme sequence is output in descending order according to the semantic matching score.

[0010] Secondly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of the above-described intelligent association method for equipment fault and maintenance knowledge based on a large model.

[0011] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the above-described intelligent association method for equipment fault and maintenance knowledge based on a large model.

[0012] Compared with existing technologies, the beneficial effects of this invention are as follows: By collecting historical fault work order data and maintenance record data of the target equipment and performing structured parsing to construct a fault-maintenance knowledge pair dataset, the scattered unstructured historical maintenance experience is transformed into computable structured knowledge assets, avoiding the problem of low knowledge utilization caused by relying on manual experience retrieval in traditional methods, thereby realizing the systematic accumulation and reuse of enterprise maintenance knowledge; By using a pre-trained large model to perform deep semantic encoding on fault description text and maintenance plan text to generate fault semantic vector sets and maintenance semantic vector sets, the limitations of traditional keyword matching methods that can only capture literal similarity are overcome, and the deep semantic features and contextual relationships of fault descriptions can be understood, so that fault descriptions with different expressions but similar semantics can be accurately identified, significantly improving the recall and accuracy of fault matching; By constructing a fault-maintenance semantic association graph and establishing bidirectional association edges with fault component identifiers as association indices, the structured relationship between fault phenomena and maintenance plans is realized. By mapping semantic relationships, the system can quickly locate relevant maintenance knowledge based on the graph topology. Compared with traditional linear retrieval methods, the graph-based knowledge organization supports multi-hop reasoning and association discovery, significantly improving the efficiency of knowledge retrieval and the accuracy of association recommendation. By performing L2 norm normalization on the real-time fault semantic vector and performing a vector inner product operation with the candidate maintenance semantic vector to obtain the initial matching degree score, and then combining it with the semantic similarity value of the candidate fault node for weighted fusion calculation to obtain the final semantic matching degree score, a comprehensive evaluation of multi-dimensional semantic information is achieved. It considers both the direct semantic correlation between real-time faults and historical maintenance plans and incorporates the confidence information of fault node matching. The dual semantic verification mechanism effectively reduces the risk of mismatch and ensures that the recommended maintenance plan has higher reliability and applicability. Finally, the target maintenance plan sequence is output in descending order of semantic matching degree score, which significantly shortens the fault diagnosis and maintenance decision time, improves equipment maintenance efficiency, and reduces maintenance costs caused by misjudgment. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of a method for intelligently associating equipment fault and maintenance knowledge based on a large model. Detailed Implementation

[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0015] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0016] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0017] As mentioned in the background section above, XXXXX. To address the aforementioned problems, this invention provides a method for intelligently associating equipment faults and maintenance knowledge based on a large model.

[0018] Reference Figure 1 , Figure 1 This is a flowchart illustrating an intelligent association method for equipment fault and maintenance knowledge based on a large model, according to an embodiment of the present invention. Figure 1 As shown, a method for intelligently associating equipment fault and maintenance knowledge based on a large model includes: S1: Collect historical fault work order data and maintenance record data of the target equipment, perform structured parsing on the historical fault work order data and maintenance record data, and construct a fault-maintenance knowledge pair dataset; S1.1: Collect historical fault work order data from the operation and maintenance management system of the target equipment, and at the same time collect maintenance record data corresponding to the historical fault work order data from the operation and maintenance management system. It should be noted that historical fault work order data includes fields for fault occurrence time, fault phenomenon description, fault equipment number, and fault handling status; maintenance record data includes fields for maintenance time, maintenance personnel, maintenance operation content, and replacement parts information.

[0019] S1.2: Perform text preprocessing on historical fault work order data and maintenance record data to remove invalid characters, redundant symbols and blank fields, and obtain cleaned fault work order data and cleaned maintenance record data. S1.3: The regular expression matching method is used to perform structured parsing on the fault work order data and the maintenance record data after cleaning. The fault description text is extracted from the fault phenomenon description field, the fault component identifier is extracted from the fault equipment number field, and the maintenance plan text is extracted from the maintenance operation content field. S1.3.1: Construct regular expression matching rules for the fault phenomenon description field, where the regular expression matching rules include fault type keyword patterns, fault location keyword patterns, and fault symptom keyword patterns; S1.3.2: Apply regular expression matching rules to the fault phenomenon description field in the cleaned fault work order data, locate text fragments in the fault phenomenon description field that match the fault type keyword pattern, fault location keyword pattern, and fault symptom keyword pattern, and concatenate and integrate the located text fragments to obtain the fault description text. S1.3.3: Construct regular expression matching rules for the faulty equipment number field, where the regular expression matching rules include equipment type encoding pattern, equipment serial number pattern, and equipment location encoding pattern; S1.3.4: Apply regular expression matching rules to the fault equipment number field in the cleaned fault work order data, extract the coded strings in the fault equipment number field that conform to the equipment type coding pattern, equipment serial number pattern and equipment location coding pattern, and perform standardized format conversion on the extracted coded strings to obtain the fault component identifier; S1.3.5: Construct regular expression matching rules for the maintenance operation content field, where the regular expression matching rules include maintenance action keyword patterns, maintenance object keyword patterns, and maintenance parameter keyword patterns; S1.3.6: Apply regular expression matching rules to the maintenance operation content field in the cleaned maintenance record data, locate text fragments in the maintenance operation content field that match the maintenance action keyword pattern, maintenance object keyword pattern, and maintenance parameter keyword pattern, and arrange and combine the located text fragments according to the execution order of the maintenance operations to obtain the maintenance plan text. S1.3.7: Perform format validation on the fault description text, fault component identifier, and maintenance plan text respectively. Verify whether the character length of the fault description text is greater than the preset minimum length threshold, verify whether the fault component identifier conforms to the preset encoding format specification, and verify whether the maintenance plan text contains complete maintenance operation steps. Output the fault description text, fault component identifier, and maintenance plan text that pass the format validation as the structured parsing result. S1.4: Using the faulty component identifier as the association key, pair the fault description text with the maintenance plan text one by one to generate fault-maintenance knowledge pairs; S1.4.1: Based on the structured parsing results, extract the fault component identifier and fault description text corresponding to each cleaning fault work order data, and establish the first mapping relationship table between the fault component identifier and the fault description text; S1.4.2: Extract the fault component identifier and maintenance plan text corresponding to each cleaning and maintenance record data based on the structured parsing results, and establish a second mapping relationship table between the fault component identifier and the maintenance plan text; S1.4.3: Traverse each record in the first mapping table and obtain the faulty component identifier of the current record as the current association key; S1.4.4: Use the current association key to perform an index query in the second mapping table to retrieve the maintenance plan text record that matches the current association key; S1.4.5: Determine the number of search results for the index query; Specifically, if the number of search results is zero, the current record is marked as having no matching record and is skipped; if the number of search results is one, the fault description text corresponding to the current association key is directly paired with the retrieved maintenance plan text; if the number of search results is greater than one, nearest neighbor matching is performed based on the time difference between the maintenance time field and the fault occurrence time field, and the maintenance plan text with the smallest time difference is selected for pairing.

[0020] S1.4.6: Encapsulate the successfully paired fault description text, fault component identifier, and repair plan text into a fault-repair knowledge pair record; It should be noted that the fault-repair knowledge record includes a fault description text field, a fault component identification field, and a repair plan text field.

[0021] S1.4.7: Repeat steps S1.4.3 to S1.4.6 until all records in the first mapping table have been traversed. Then, aggregate the generated fault-maintenance knowledge pairs to form a fault-maintenance knowledge pair set. S1.5: Perform integrity verification on fault-maintenance knowledge pairs, remove invalid records where the fault description text or maintenance plan text in the fault-maintenance knowledge pairs is empty, classify and store the fault-maintenance knowledge pairs that pass the verification according to the fault type, and construct a fault-maintenance knowledge pair dataset. S1.5.1: Traverse each fault-maintenance knowledge pair record in the fault-maintenance knowledge pair set, and sequentially read the fault description text field, fault component identifier field, and maintenance plan text field in the current fault-maintenance knowledge pair record; S1.5.2: Perform a non-empty check on the fault description text field to determine whether the character length of the fault description text field is greater than zero. If the character length is equal to zero, mark the current fault-maintenance knowledge pair record as a first-type invalid record. S1.5.3: Perform a non-empty check on the maintenance plan text field to determine whether the character length of the maintenance plan text field is greater than zero. If the character length is equal to zero, mark the current fault-maintenance knowledge pair record as a second type of invalid record. S1.5.4: Perform format validity validation on the fault component identification field to determine whether the fault component identification field conforms to the preset equipment coding standard. If the fault component identification field does not conform to the preset equipment coding standard, mark the current fault-maintenance knowledge record as a third type of invalid record. S1.5.5: Remove the fault-maintenance knowledge pairs marked as Category I invalid records, Category II invalid records, and Category III invalid records from the fault-maintenance knowledge pair set to obtain the fault-maintenance knowledge pair set that has passed the verification. S1.5.6: Construct a keyword dictionary for fault types; It should be noted that the fault type keyword dictionary includes a variety of preset fault types and their corresponding feature keyword lists; the preset fault types include electrical fault types, mechanical fault types, control fault types and communication fault types.

[0022] S1.5.7: Traverse and verify each fault-maintenance knowledge pair record in the fault-maintenance knowledge pair set, extract the fault description text in the current fault-maintenance knowledge pair record, match the fault description text with the feature keyword list in the fault type keyword dictionary, and count the number of keywords in the fault description text that match each feature keyword list. S1.5.8: Determine the fault type label of the current fault-maintenance knowledge pair record based on the number of keywords. Take the preset fault type with the largest number of keywords as the fault type label of the current fault-maintenance knowledge pair record. If there are multiple preset fault types with the same number of keywords, mark the current fault-maintenance knowledge pair record as a composite fault type. S1.5.9: Based on the fault type label, group the records of each fault-maintenance knowledge pair in the set of fault-maintenance knowledge pairs that have passed the verification, and group the records of fault-maintenance knowledge pairs with the same fault type label into the same fault type group, generating multiple fault type subsets; S1.5.10: Create an index directory for multiple fault type subsets according to fault type labels, encapsulate and store the multiple fault type subsets and index directory, and construct a fault-maintenance knowledge pair dataset; It should be noted that the index directory records the mapping relationship between the storage locations of each fault type label and the corresponding fault type subset.

[0023] In an optional embodiment, suppose that during routine operation and maintenance of a substation of a power company, a disconnector switchgear of model "GIS-220kV-01" (fault device number "EQP-GIS-220kV-01-A phase") malfunctions. Maintenance personnel report a real-time fault description via the equipment monitoring terminal: "The A phase disconnector switch exhibits an incomplete closing phenomenon during operation, accompanied by abnormal noise, and the local temperature rises to 85°C." Based on this real-time fault description, historical faults and maintenance knowledge will be automatically linked to recommend maintenance solutions to the maintenance personnel. First, historical fault work order data and maintenance record data for this equipment will be collected from the operation and maintenance management system. For example, a historical fault work order might show: "Fault phenomenon: Poor contact after closing the A phase disconnector switch, abnormal temperature rise to 80°C; Fault device number: EQP-GIS-220kV-01-A phase." The corresponding maintenance record data would be: "Maintenance operation: Disassemble the A phase contact, clean the contact surface and apply conductive grease, readjust the closing position parameters to the standard value; Replacement parts: None." By parsing with regular expressions, the fault description text "faulty closing, temperature rise" is extracted from the fault phenomenon description field; "EQP-GIS-220kV-01-A phase" is extracted from the equipment number field as the fault component identifier; and "disassembling contacts, cleaning contact surfaces, adjusting closing parameters" is extracted from the maintenance operation content field as the maintenance plan text. Using the fault component identifier as the association key, this information is paired to generate a fault-maintenance knowledge pair, which is then stored in the dataset after integrity verification.

[0024] S2: Input the fault description text and maintenance plan text of the fault-maintenance knowledge dataset into the pre-trained large model for semantic encoding, generate fault semantic vector set and maintenance semantic vector set respectively, and construct fault-maintenance semantic association graph with fault component identifier as association index; S2.1: Load the pre-trained large model, which adopts an encoder model based on the Transformer architecture. The encoder model includes multiple layers of self-attention mechanism layers and feedforward neural network layers. S2.1.1: Construct the input embedding layer of the encoder model based on the Transformer architecture; Furthermore, the input embedding layer includes a lexical embedding matrix and a positional embedding matrix; the dimension of the lexical embedding matrix is ​​the product of the vocabulary size and the hidden layer dimension; the dimension of the positional embedding matrix is ​​the product of the maximum sequence length and the hidden layer dimension.

[0025] S2.1.2: Construct a multi-layer self-attention mechanism layer for the encoder model, and set up a multi-head attention calculation module in the self-attention mechanism layer; Furthermore, each self-attention mechanism layer includes a query weight matrix, a key weight matrix, and a value weight matrix; the dimensions of the query weight matrix, key weight matrix, and value weight matrix are all products of the hidden layer dimension and the attention head dimension; the multi-head attention calculation module performs matrix multiplication operations on the input vector with the query weight matrix, key weight matrix, and value weight matrix respectively to obtain the query vector, key vector, and value vector, performs a dot product operation on the query vector and key vector and divides it by a scaling factor, performs softmax normalization, and then performs a weighted sum with the value vector to obtain the attention output vector.

[0026] The preferred output matrix of the self-attention mechanism layer is formulated as follows: ; in, This is the output matrix of the self-attention mechanism layer. For querying the matrix, Let K be the permutation matrix of the key matrix. For value matrices, The attention head dimension has a value of 64. This is the adaptive scaling factor, with a value of 0.1. It is the Frobenius norm. This is for outputting the projection matrix.

[0027] It should be noted that the output matrix of the self-attention mechanism layer... The value range is a real number matrix with a dimension equal to the hidden layer dimension of 768. The matrix elements take values ​​in the range of (-∞, +∞). After subsequent layer normalization processing, the element values ​​will be standardized.

[0028] S2.1.3: After each self-attention mechanism layer, a feedforward neural network layer is constructed, wherein the feedforward neural network layer includes a first linear transformation layer, an activation function layer, and a second linear transformation layer; S2.1.4: Add a residual connection module and a layer normalization module after each self-attention mechanism layer and feedforward neural network layer, respectively; It should be noted that the first linear transformation layer maps the attention output vector from the hidden layer dimension to the intermediate layer dimension; the activation function layer uses the GELU activation function to perform a nonlinear transformation on the mapping result; the second linear transformation layer maps the nonlinear transformation result from the intermediate layer dimension back to the hidden layer dimension; the residual connection module adds the input vector and output vector of the current layer element by element; and the layer normalization module performs mean and variance normalization on the addition result.

[0029] S2.1.5: Stack and connect the input embedding layer, multi-layer self-attention mechanism layer, multi-layer feedforward neural network layer, residual connection module and layer normalization module in sequence to construct a complete encoder model structure; Preferably, the encoder model has 12 layers, the hidden layer dimension is 768, and the number of attention heads is 12.

[0030] S2.1.6: The encoder model is pre-trained using a large-scale general corpus; Specifically, the pre-training process employs a masked language model task, randomly masking 15% of the words in the input sequence. The encoder model is then trained to predict the masked words based on contextual information, and all weight parameters of the encoder model are updated through backpropagation.

[0031] S2.1.7: Use a corpus of equipment failure domains to fine-tune the pre-trained encoder model for domain adaptation; Furthermore, the equipment failure domain corpus includes equipment failure description text, maintenance operation text, and equipment technical manual text. Through domain-adaptive fine-tuning, the encoder model learns the professional terminology and semantic expression patterns of the equipment failure domain.

[0032] S2.1.8: Serialize and store the weight parameters of the encoder model after domain-adaptive fine-tuning to generate a pre-trained model weight file. Load the pre-trained model weight file into the encoder model to obtain the pre-trained large model.

[0033] S2.2: Read each fault-maintenance knowledge pair sequentially from the fault-maintenance knowledge pair dataset, and obtain the fault description text, maintenance plan text, and fault component identifier of the fault-maintenance knowledge pair; S2.3: Perform word segmentation on the fault description text, convert the fault description text into a word sequence, and add a special classification tag at the beginning of the word sequence to generate a fault text word sequence; S2.3.1: Load a word segmenter that matches the pre-trained large language model. The word segmenter includes a word list file and a word segmentation rule configuration file. The word list file records the mapping relationship between all words and word indexes. S2.3.2: Read the fault description text of the current fault-maintenance knowledge pair record from the fault-maintenance knowledge pair dataset, perform character-level preprocessing on the fault description text, convert uppercase letters in the fault description text to lowercase letters, and remove leading and trailing whitespace characters from the fault description text. S2.3.3: The preprocessed fault description text is segmented using a byte-pair encoding algorithm, which splits the fault description text into multiple sub-word units and generates a fault description sub-word sequence. S2.3.4: Based on the mapping relationship between lexical units and lexical indices in the lexical file, convert each sub-word unit in the fault description sub-word sequence into the corresponding lexical index value, and generate the fault description lexical index sequence; S2.3.5: Insert the term index value corresponding to the special classification tag at the beginning of the fault description term index sequence, where the special classification tag is the [CLS] tag, and the term index value of the [CLS] tag is obtained from the term file; S2.3.6: Insert the term index value corresponding to the sequence end marker at the end of the fault description term index sequence after inserting the special classification marker, where the sequence end marker is the [SEP] marker, to obtain the complete fault description term index sequence; S2.3.7: Determine whether the length of the complete fault description term index sequence exceeds the preset maximum sequence length threshold. If it exceeds the preset maximum sequence length threshold, truncate the complete fault description term index sequence, retaining the first preset maximum sequence length threshold of term index values. If it does not exceed the preset maximum sequence length threshold, pad the complete fault description term index sequence by adding the term index values ​​corresponding to the padding markers at the end to the preset maximum sequence length threshold, thus obtaining the fault text term sequence. It should be noted that the preset maximum sequence length threshold is determined based on the design parameters of the pre-trained large model's architecture and the statistical characteristics of text length in the equipment fault domain. Specifically, this threshold must ensure that the vast majority of fault descriptions and repair solutions can be fully encoded, while avoiding excessive consumption of computational resources due to excessively long sequences. Typically, this threshold is set with reference to the maximum positional encoding length during model pre-training (e.g., 512 tokens) and combined with the distribution analysis of text length in historical data (e.g., ensuring that more than 95% of text samples do not require truncation).

[0034] S2.4: Input the fault text word sequence into the pre-trained large model, obtain the hidden state vector at the position corresponding to the special classification label, and use this hidden state vector as the fault semantic vector; S2.4.1: Based on the index value of each word in the fault text word sequence, query the corresponding word embedding vector in the word embedding matrix to generate a fault text word embedding vector sequence; S2.4.2: Based on the position index of each word in the fault text word sequence, query the corresponding position embedding vector in the position embedding matrix to generate a fault text position embedding vector sequence; S2.4.3: Add the fault text word embedding vector sequence and the fault text position embedding vector sequence position by position to obtain the fault text input embedding vector sequence; S2.4.4: Input the fault text into the first self-attention mechanism layer of the large model, and calculate the attention weights between the embedded vectors at each position through the multi-head attention calculation module to obtain the first-layer attention output vector sequence.

[0035] S2.4.5: After the first layer attention output vector sequence is processed by the residual connection module and the layer normalization module, it is input into the first layer feedforward neural network layer. After being processed by the first linear transformation layer, the activation function layer and the second linear transformation layer in sequence, the first layer feedforward output vector sequence is obtained. S2.4.6: Input the first layer feedforward output vector sequence into the second to twelfth layers of the encoder model in sequence, and repeat the self-attention mechanism calculation and feedforward neural network calculation to obtain the final layer output vector sequence; S2.4.7: Extract the hidden state vector corresponding to the first position from the final layer output vector sequence, where the first position corresponds to the position of the special classification label [CLS], and use the hidden state vector corresponding to the first position as the fault semantic vector; It should be noted that the final layer output vector sequence contains a hidden state vector of the same length as the fault text word sequence; the dimension of the fault semantic vector is the same as the dimension of the hidden layer.

[0036] S2.5: Perform word segmentation on the maintenance plan text to generate a maintenance text word sequence. Input the maintenance text word sequence into the pre-trained large model to obtain the hidden state vectors at the corresponding positions of special classification labels, and use the hidden state vectors as maintenance semantic vectors. S2.5.1: Read the maintenance plan text of the current fault-maintenance knowledge pair record from the fault-maintenance knowledge pair dataset, perform character-level preprocessing on the maintenance plan text, convert uppercase letters in the maintenance plan text to lowercase letters, and remove whitespace characters at the beginning and end of the maintenance plan text. S2.5.2: The preprocessed maintenance plan text is segmented using a byte-pair encoding algorithm, which splits the maintenance plan text into multiple sub-word units and generates a maintenance plan sub-word sequence. S2.5.3: Based on the mapping relationship between lexical units and lexical indexes in the lexical file, convert each sub-word unit in the sub-word sequence of the maintenance plan into the corresponding lexical index value, and generate the lexical index sequence of the maintenance plan; S2.5.4: Insert the term index value corresponding to the special classification marker [CLS] at the beginning of the maintenance plan term index sequence, and insert the term index value corresponding to the sequence end marker [SEP] at the end of the maintenance plan term index sequence to obtain the complete maintenance plan term index sequence; S2.5.5: The length of the complete maintenance plan term index sequence is adjusted. If the length of the complete maintenance plan term index sequence exceeds the preset maximum sequence length threshold, it is truncated. If it does not exceed the threshold, it is padded to obtain the maintenance text term sequence. S2.5.6: Input the inspection text word sequence into the input embedding layer of the encoder model, query the corresponding word embedding vector and position embedding vector according to the word index value of each word in the inspection text word sequence, and add the word embedding vector and position embedding vector to obtain the inspection text input embedding vector sequence. S2.5.7: Input the inspection text into the embedded vector sequence and then input it into the first to twelfth layers of the encoder model. After feature extraction processing by the self-attention mechanism layer and the feedforward neural network layer of each layer, the final layer output vector sequence of the inspection text is obtained. S2.5.8: Extract the hidden state vector corresponding to the first position from the final output vector sequence of the inspection text, and use the hidden state vector corresponding to the first position as the inspection semantic vector; It should be noted that the first position corresponds to the position of the special classification marker [CLS]; the dimension of the maintenance semantic vector is the same as the dimension of the fault semantic vector.

[0037] S2.6: Traverse all fault-maintenance knowledge pairs in the fault-maintenance knowledge pair dataset, aggregate the generated fault semantic vectors into a fault semantic vector set, and aggregate the generated maintenance semantic vectors into a maintenance semantic vector set.

[0038] S2.7: Using the faulty component identifier as the association index, establish bidirectional association edges between each fault semantic vector in the fault semantic vector set and the corresponding maintenance semantic vector in the maintenance semantic vector set, and generate a fault-maintenance association edge set. S2.8: Use the faulty component identifier as a graph node, the fault semantic vector and the maintenance semantic vector as attribute vectors of the graph node, and the set of fault-maintenance associated edges as graph connection relationships to construct a fault-maintenance semantic association graph.

[0039] In an optional embodiment, fault-maintenance knowledge is input into a pre-trained large model (such as a Transformer-based encoder model). For example, the fault description text "faulty closing, temperature rise" is converted into a 768-dimensional fault semantic vector, and the maintenance plan text "disassemble contacts, clean contact surfaces" is converted into a corresponding maintenance semantic vector. The system uses the fault component identifier "EQP-GIS-220kV-01-A phase" as the association index to establish association edges between faults and maintenance at the semantic level, ultimately constructing a semantic association graph containing multiple fault nodes and maintenance nodes, where each node has an attribute vector to facilitate subsequent semantic matching.

[0040] S3: Obtain the real-time fault description text to be matched, use the large model to perform semantic encoding on the real-time fault description text, and generate a real-time fault semantic vector. S3.1: Obtain the real-time fault description text to be matched through the equipment monitoring terminal or operation and maintenance work order system; S3.2: Perform text preprocessing on the real-time fault description text to obtain the cleaned real-time fault description text; It should be noted that the real-time fault description text includes fault phenomenon description information, fault occurrence time information, and faulty device number information; invalid characters, redundant symbols, and blank content in the real-time fault description text are removed.

[0041] S3.3: Perform text standardization processing on the cleaned real-time fault description text, and convert the synonyms and abbreviations in the cleaned real-time fault description text into standard terminology to obtain standardized real-time fault description text; S3.4: Perform word segmentation on the standardized real-time fault description text to generate a real-time fault text word sequence, and input the real-time fault text word sequence into the pre-trained large model to obtain the hidden state vector; Furthermore, the normalized real-time fault description text is converted into a word sequence, and a special classification label is added to the beginning of the word sequence to generate a real-time fault text word sequence; through feature extraction of multiple layers of self-attention mechanism layer and feedforward neural network layer, the hidden state vector at the corresponding position of the special classification label is obtained; S3.5: Perform L2 norm normalization on the hidden state vector and normalize the magnitude of the hidden state vector to a unit length to obtain the real-time fault semantic vector.

[0042] In an optional embodiment, the real-time fault description text "A-phase disconnect switch not closing properly, accompanied by abnormal noise, local temperature rise to 85°C" is obtained, and preprocessed and normalized (e.g., "A-phase" is unified as "A-phase", and "85°C" is converted to the standard term "high temperature"); word segmentation is used to generate a word sequence, which is input into a pre-trained large model, and the hidden state vector at the [CLS] marker position is extracted as the real-time fault semantic vector, and L2 norm normalization is performed to make its modulus unit length.

[0043] S4: Calculate the semantic similarity value between the real-time fault semantic vector and each vector in the fault semantic vector set, and filter out fault nodes with semantic similarity values ​​greater than a preset threshold to form a candidate fault node set. S4.1: Load the fault semantic vector set from the fault-maintenance semantic association graph, and obtain each fault semantic vector and its corresponding fault component identifier in the fault semantic vector set; S4.2: Perform L2 norm normalization on each fault semantic vector in the fault semantic vector set, normalize the magnitude of each fault semantic vector to a unit length, and obtain a normalized fault semantic vector set. Furthermore, the specific formula for the normalized fault semantic vector is as follows: ; in, This is the normalized fault semantic vector. This is the original fault semantic vector. The semantic vector dimension has a value of 768. Let f be the i-th component of vector f. This is the peak suppression coefficient, with a value of 0.01.

[0044] It should be noted that the normalized fault semantic vector The range is approximately a unit-length vector, i.e. The range of each component is [-1, 1], and the normalized vector is located near the d-dimensional unit hypersphere.

[0045] S4.3: Traverse each normalized fault semantic vector in the set of normalized fault semantic vectors, calculate the vector inner product between the real-time fault semantic vector and each normalized fault semantic vector, and use the vector inner product as the semantic similarity value. The specific formula is as follows: ; in, Real-time fault semantic vector With historical fault semantic vector The semantic similarity value between them The semantic vector dimension has a value of 768. Real-time fault semantic vector The i-th component, Historical fault semantic vector The i-th component, This is the distance penalty factor, with a value of 0.1. Real-time fault semantic vector With historical fault semantic vector The L2 norm of the vector difference between them.

[0046] It should be noted that semantic similarity values The value range is [0,1], when the semantic similarity value When the value is close to 1, real-time faults and historical faults have a high semantic similarity; when the semantic similarity value is close to 1, real-time faults and historical faults have a high semantic similarity. When the value is close to 0, it indicates that the semantic difference between the two is large. A preset threshold is set based on historical fault matching statistics. The preset threshold is a floating-point value between 0.65 and 0.85, which is used to filter candidate fault nodes.

[0047] S4.4: Compare the semantic similarity value with a preset threshold, and filter out fault semantic vectors with semantic similarity values ​​greater than the preset threshold to obtain a preliminary set of filtered fault semantic vectors.

[0048] Specifically, based on the fault device number information corresponding to the real-time fault semantic vector, the device type consistency is checked with the fault component identifier corresponding to each fault semantic vector in the preliminary set of fault semantic vectors. Fault semantic vectors with inconsistent device types are removed to obtain the set of fault semantic vectors after verification.

[0049] S4.4.1: Traverse each semantic similarity value in the semantic similarity value sequence, and read the current semantic similarity value and its corresponding normalized fault semantic vector in turn.

[0050] S4.4.2: Compare the current semantic similarity value with the preset threshold to determine the relationship between the current semantic similarity value and the preset threshold; Furthermore, when the current semantic similarity value is greater than the preset threshold, the normalized fault semantic vector corresponding to the current semantic similarity value is marked as a high similarity vector, and the high similarity vector is added to the initial screening set of fault semantic vectors; when the current semantic similarity value is equal to the preset threshold, the normalized fault semantic vector corresponding to the current semantic similarity value is marked as a critical similarity vector, and the critical similarity vector is added to the set of fault semantic vectors to be reviewed; when the current semantic similarity value is less than the preset threshold, the normalized fault semantic vector corresponding to the current semantic similarity value is marked as a low similarity vector, and the low similarity vector is skipped without processing.

[0051] S4.4.3: Repeat steps S4.4.1 to S4.4.2 until all semantic similarity values ​​in the semantic similarity value sequence have been traversed; S4.4.4: Count the number of fault semantic vectors in the initial screening set of fault semantic vectors, and determine whether the number of fault semantic vectors meets the preset minimum matching requirement; Furthermore, when the number of fault semantic vectors is greater than or equal to the preset minimum matching number, the initial set of fault semantic vectors is kept unchanged, and the process proceeds to step S4.4.6; when the number of fault semantic vectors is less than the preset minimum matching number and the set of fault semantic vectors to be reviewed is not empty, the critical similarity vectors in the set of fault semantic vectors to be reviewed are merged into the initial set of fault semantic vectors, and the process proceeds to step S4.4.6; when the number of fault semantic vectors is less than the preset minimum matching number and the set of fault semantic vectors to be reviewed is empty, the threshold dynamic adjustment operation is performed, and the process proceeds to step S4.4.5.

[0052] S4.4.5: Adjust the preset threshold by decreasing it stepwise according to the preset step size to obtain the adjusted preset threshold. The preset step size is 0.05. Determine whether the adjusted preset threshold is less than the preset minimum threshold lower limit. Specifically, when the adjusted preset threshold is greater than or equal to the preset minimum threshold lower limit, the adjusted preset threshold is used as the new preset threshold, and the filtering operation is re-executed in step S4.4.1; when the adjusted preset threshold is less than the preset minimum threshold lower limit, the threshold adjustment is stopped, the current preliminary filtering fault semantic vector set is output as the final filtering result, and a low matching degree alarm message is generated.

[0053] S4.4.6: Sort each fault semantic vector in the preliminary set of fault semantic vectors in descending order according to the corresponding semantic similarity value to obtain the sorted preliminary set of fault semantic vectors.

[0054] S4.5: Take the fault component identifiers corresponding to each fault semantic vector in the verified fault semantic vector set as candidate fault nodes, aggregate them to form a candidate fault node set, and record the semantic similarity value corresponding to each candidate fault node in the candidate fault node set.

[0055] In an optional embodiment, the semantic similarity value between the real-time fault semantic vector and each vector in the fault semantic vector set is calculated. For example, the similarity between the fault semantic vector and the historical fault "poor contact after closing, abnormal temperature rise" is 0.82, while the similarity with other unrelated faults (such as "insulator breakage") is less than 0.5. A preset threshold of 0.70 is set, and fault semantic vectors with similarity greater than 0.70 (such as vectors corresponding to 0.82 and 0.75) are filtered out. Inconsistent vectors are eliminated according to the consistency verification of equipment type (such as ensuring that they are all "disconnecting switch" type equipment) to form a candidate fault node set.

[0056] S5: Based on the fault-maintenance semantic association graph, obtain the maintenance semantic vector associated with each node in the candidate fault node set, calculate the semantic matching score between the real-time fault semantic vector and each maintenance semantic vector, sort them in descending order of semantic matching score, and output the target maintenance scheme sequence. S5.1: Traverse each candidate fault node in the candidate fault node set, and based on the fault component identifier corresponding to each candidate fault node, query the fault-maintenance association edge associated with the fault component identifier in the fault-maintenance semantic association graph; S5.2: Traverse the graph along the fault-maintenance association edge to obtain the maintenance semantic vector connected to the candidate fault node; S5.3: Aggregate the maintenance semantic vectors and their corresponding faulty component identifiers to form a candidate maintenance semantic vector set; S5.4: Perform L2 norm normalization on each maintenance semantic vector in the candidate maintenance semantic vector set, calculate the L2 norm value of each maintenance semantic vector, and obtain the normalized maintenance semantic vector set. S5.4: Traverse each normalized maintenance semantic vector in the set of normalized maintenance semantic vectors, and perform a vector inner product operation between the real-time fault semantic vector and each normalized maintenance semantic vector to obtain the initial matching degree score; S5.5: From the mapping table between each candidate fault node and its corresponding semantic similarity value in the candidate fault node set, obtain the semantic similarity value of the candidate fault node corresponding to the normalized maintenance semantic vector. Then, perform a weighted fusion calculation between the semantic similarity value and the initial matching score to obtain the semantic matching score. The specific formula is as follows: ; in, The semantic matching score is the score. For real-time fault semantic vectors, This represents the candidate maintenance semantic vector. This represents the semantic similarity value of the corresponding candidate faulty nodes. The semantic vector dimension has a value of 768. Real-time fault semantic vector The i-th component, Let be the i-th component of the candidate maintenance semantic vector e. The similarity fusion weight is set to 0.6. The base weight is set to 0.4. This is the slope parameter of the Sigmoid function, with a value of 5.

[0057] It should be noted that the semantic matching score The value range is (0,1), when the semantic matching score is... When the score is close to 1, the maintenance plan is highly matched with the real-time fault. If the result is less than the preset output threshold of 0.7, an insufficient matching alarm message will be generated.

[0058] S5.6: Repeat steps S5.4 to S5.5 until all normalized maintenance semantic vectors in the normalized maintenance semantic vector set have been traversed, generate a semantic matching score sequence, and establish an association record between each semantic matching score in the semantic matching score sequence and the corresponding maintenance semantic vector. S5.7: Sort the semantic matching scores in descending order to obtain the semantic matching score sequence, and determine whether the highest semantic matching score in the semantic matching score sequence meets the preset output threshold requirement, and output the target maintenance plan sequence. Furthermore, when the highest semantic matching score is greater than or equal to the preset output threshold, the corresponding maintenance plan text is extracted from the fault-maintenance knowledge pair dataset based on the maintenance semantic vector corresponding to each semantic matching score in the semantic matching score sequence, and a target maintenance plan sequence is generated; when the highest semantic matching score is less than the preset output threshold, an insufficient matching alarm message is generated, and the maintenance plan text corresponding to the top preset number of maintenance semantic vectors in the semantic matching score sequence is output as a reference maintenance plan sequence.

[0059] S5.8: Associate and encapsulate the target maintenance plan sequence and its corresponding semantic matching score to generate maintenance plan recommendation results.

[0060] It should be noted that the maintenance plan recommendation results include a maintenance plan text field, a semantic matching score field, and a faulty component identifier field. The maintenance plan recommendation results are output to the operation and maintenance work order system or equipment monitoring terminal, so that operation and maintenance personnel can select a suitable maintenance plan based on the semantic matching score.

[0061] In an optional embodiment, based on the fault-maintenance semantic association graph, maintenance semantic vectors associated with candidate fault nodes are obtained (e.g., the maintenance semantic vector description corresponding to the fault node is "disassemble contacts, clean contact surfaces"). The semantic matching score between the real-time fault semantic vector and these maintenance semantic vectors is calculated and weighted by combining them with the semantic similarity value of the candidate fault nodes. For example, the matching scores are 0.88, 0.79, and 0.70, respectively; sorted in descending order, if the highest score of 0.88 is greater than the preset output threshold of 0.80, the corresponding maintenance plan text is extracted from the knowledge pair dataset to generate a target maintenance plan sequence (S5.7), such as "Step 1: De-energize and disassemble phase A contacts; Step 2: Clean contact surfaces and apply conductive grease; Step 3: Recalibrate closing position parameters." Finally, the maintenance plan recommendation results (including plan text, matching score, and fault component identifier) ​​are output to the maintenance work order system to assist maintenance personnel in making quick decisions.

[0062] In summary, this invention constructs a fault-maintenance knowledge pair dataset by collecting historical fault work order data and maintenance record data of target equipment and performing structured parsing. This transforms scattered, unstructured historical maintenance experience into computable, structured knowledge assets, avoiding the low knowledge utilization rate caused by relying on manual experience retrieval in traditional methods. This achieves the systematic accumulation and reuse of enterprise maintenance knowledge. By utilizing a pre-trained large model to perform deep semantic encoding on fault description text and maintenance plan text to generate fault semantic vector sets and maintenance semantic vector sets, this invention overcomes the limitation of traditional keyword matching methods that can only capture literal similarity. It can understand the deep semantic features and contextual relationships of fault descriptions, enabling accurate identification of fault descriptions that are semantically similar but have different expressions, significantly improving the recall and accuracy of fault matching. By constructing a fault-maintenance semantic association graph and establishing bidirectional association edges using fault component identifiers as association indices, a structured semantic mapping between fault phenomena and maintenance plans is achieved. The system leverages graph topology to quickly locate relevant maintenance knowledge. Compared to traditional linear retrieval methods, graph-based knowledge organization supports multi-hop reasoning and association discovery, significantly improving the efficiency of knowledge retrieval and the accuracy of association recommendations. By normalizing the real-time fault semantic vector using the L2 norm and performing a vector inner product operation with candidate maintenance semantic vectors to obtain an initial matching score, and then combining this with the semantic similarity values ​​of candidate fault nodes for weighted fusion calculation to obtain the final semantic matching score, a comprehensive evaluation of multi-dimensional semantic information is achieved. This considers both the direct semantic relevance between real-time faults and historical maintenance plans, and incorporates the confidence information of fault node matching. The dual semantic verification mechanism effectively reduces the risk of mismatches, ensuring that the recommended maintenance plans have higher reliability and applicability. Finally, the target maintenance plan sequence is output in descending order of semantic matching scores, significantly shortening fault diagnosis and maintenance decision-making time, improving equipment maintenance efficiency, and reducing maintenance costs caused by misjudgments.

[0063] This embodiment also provides a computer device applicable to the intelligent association method of equipment fault and maintenance knowledge based on a large model, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent association method of equipment fault and maintenance knowledge based on a large model as proposed in the above embodiment.

[0064] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0065] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the intelligent association method for equipment fault and maintenance knowledge based on a large model, as proposed in the above embodiments.

[0066] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0067] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligently associating equipment fault and maintenance knowledge based on a large model, characterized in that: include, Collect historical fault work order data and maintenance record data of the target equipment, perform structured parsing on the historical fault work order data and the maintenance record data, and construct a fault-maintenance knowledge pair dataset; The fault-maintenance knowledge is input into the fault description text and maintenance plan text of the dataset and semantically encoded into a pre-trained large model to generate a fault semantic vector set and a maintenance semantic vector set, respectively, and a fault-maintenance semantic association graph is constructed. Obtain the real-time fault description text to be matched, and use the big model to perform semantic encoding on the real-time fault description text to generate a real-time fault semantic vector. Calculate the semantic similarity value between the real-time fault semantic vector and each vector in the fault semantic vector set, and filter fault nodes with semantic similarity values ​​greater than a preset threshold to form a candidate fault node set; Based on the fault-maintenance semantic association graph, the maintenance semantic vector associated with each node in the candidate fault node set is obtained, the semantic matching score between the real-time fault semantic vector and each maintenance semantic vector is calculated, and the target maintenance scheme sequence is output in descending order according to the semantic matching score.

2. The intelligent association method for equipment fault and maintenance knowledge based on a large model as described in claim 1, characterized in that: The method for outputting the target maintenance plan sequence is as follows: The L2 norm normalization process is performed on each maintenance semantic vector in the candidate maintenance semantic vector set, and the L2 norm value of each maintenance semantic vector is calculated to obtain the normalized maintenance semantic vector set. Traverse each normalized maintenance semantic vector in the set of normalized maintenance semantic vectors, and perform a vector inner product operation between the real-time fault semantic vector and each normalized maintenance semantic vector to obtain the initial matching degree score; From the mapping relationship table between each candidate fault node and its corresponding semantic similarity value in the candidate fault node set, obtain the semantic similarity value of the candidate fault node corresponding to the normalized maintenance semantic vector, and perform weighted fusion calculation on the semantic similarity value and the initial matching score to obtain the semantic matching score. Sort the semantic matching scores in descending order to obtain a semantic matching score sequence, and determine whether the highest semantic matching score in the semantic matching score sequence meets the preset output threshold requirement, and output the target maintenance plan sequence. The target maintenance plan sequence and its corresponding semantic matching score are associated and encapsulated to generate maintenance plan recommendation results.

3. The intelligent association method for equipment fault and maintenance knowledge based on a large model as described in claim 2, characterized in that: Determining whether the highest semantic matching score in the semantic matching score sequence meets a preset output threshold requirement includes: When the highest semantic matching score is greater than or equal to the preset output threshold, the corresponding maintenance plan text is extracted from the fault-maintenance knowledge pair dataset according to the maintenance semantic vector corresponding to each semantic matching score in the semantic matching score sequence, and the target maintenance plan sequence is generated. When the highest semantic matching score is less than the preset output threshold, an alarm message for insufficient matching score is generated, and the maintenance plan text corresponding to the top preset number of maintenance semantic vectors in the semantic matching score sequence is output as a reference maintenance plan sequence.

4. The intelligent association method for equipment fault and maintenance knowledge based on a large model as described in claim 2, characterized in that: The method for obtaining the candidate maintenance semantic vector set is as follows: Traverse each candidate fault node in the candidate fault node set, and based on the fault component identifier corresponding to each candidate fault node, query the fault-maintenance association edge associated with the fault component identifier in the fault-maintenance semantic association graph; Perform a graph traversal along the fault-maintenance association edges to obtain the maintenance semantic vectors connected to the candidate fault nodes; The maintenance semantic vectors and their corresponding faulty component identifiers are aggregated to form a candidate maintenance semantic vector set.

5. The intelligent association method for equipment fault and maintenance knowledge based on a large model as described in claim 2, characterized in that: The method for generating the real-time fault semantic vector is as follows: Obtain real-time fault description text through equipment monitoring terminals or maintenance work order systems; The real-time fault description text is preprocessed to obtain a cleaned real-time fault description text. The cleaned real-time fault description text is subjected to text normalization processing, and the synonyms and abbreviations in the cleaned real-time fault description text are converted into standard terminology to obtain the normalized real-time fault description text. The standardized real-time fault description text is segmented to generate a real-time fault text word sequence, and the real-time fault text word sequence is input into a pre-trained large model to obtain a hidden state vector. The hidden state vector is normalized using the L2 norm, and its magnitude is normalized to a unit length to obtain the real-time fault semantic vector.

6. The intelligent association method for equipment fault and maintenance knowledge based on a large model as described in claim 4, characterized in that: The method for constructing the fault-maintenance semantic association graph is as follows: Using the faulty component identifier as the association index, establish bidirectional association edges between each fault semantic vector in the fault semantic vector set and the corresponding maintenance semantic vector in the maintenance semantic vector set, and generate a fault-maintenance association edge set. The faulty component identifier is used as a graph node, the fault semantic vector and the maintenance semantic vector are used as attribute vectors of the graph node, and the set of fault-maintenance associated edges is used as the graph connection relationship to construct a fault-maintenance semantic association graph.

7. The intelligent association method for equipment fault and maintenance knowledge based on a large model as described in claim 6, characterized in that: The method for generating the fault semantic vector set and the maintenance semantic vector set is as follows: Read each fault-maintenance knowledge pair sequentially from the fault-maintenance knowledge pair dataset, and obtain the fault description text, maintenance plan text, and fault component identifier of the fault-maintenance knowledge pair. The fault description text is segmented into words, converted into a word sequence, and a special classification marker is added to the beginning of the word sequence to generate a fault text word sequence. The fault text word sequence is input into a pre-trained large model to obtain the hidden state vector at the corresponding position of the special classification label, and this hidden state vector is used as the fault semantic vector. The maintenance plan text is segmented to generate a maintenance text word sequence. The maintenance text word sequence is input into a pre-trained large model to obtain the hidden state vector at the corresponding position of the special classification label. The hidden state vector is then used as the maintenance semantic vector. Traverse all fault-maintenance knowledge pairs in the fault-maintenance knowledge pair dataset, aggregate the generated fault semantic vectors into a fault semantic vector set, and aggregate the generated maintenance semantic vectors into a maintenance semantic vector set.

8. The intelligent association method for equipment fault and maintenance knowledge based on a large model as described in claim 7, characterized in that: The pre-trained large model adopts an encoder model based on the Transformer architecture; the encoder model includes multiple layers of self-attention mechanism layers and feedforward neural network layers.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent association method for equipment fault and maintenance knowledge based on a large model as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent association method for equipment fault and maintenance knowledge based on a large model as described in any one of claims 1 to 8.

Images