A method and system for servicing a turn machine

Abstract

The present application relates to the technical field of test and repair of transfer equipment, and particularly relates to a transfer equipment repair method and system, and the present application proposes the following scheme: firstly, a knowledge extraction network is configured to extract a repair knowledge sequence; operation data is collected and processed through data mining to generate an entity label five tuple; secondly, the entity label and the repair knowledge sequence are constructed into a repair knowledge graph for repair of the transfer equipment. The knowledge extraction network comprises an input layer, a feature extraction layer and an output layer, and knowledge sequence optimization and network model updating. The operation data processing comprises preprocessing and data classification. The construction of the repair knowledge graph comprises preprocessing, word embedding model conversion, bidirectional gate recurrent unit extraction of global features, algorithm extraction of association rules and calculation and integration of knowledge nodes, thereby improving the intelligent level of test and repair decision of the transfer equipment.

Description

Technical Field

[0001] This invention relates to the field of testing and maintenance technology for rotating equipment, and in particular to a method and system for the maintenance of rotating equipment. Background Technology

[0002] With the continuous advancement of industrial technology, switching equipment has become increasingly diverse and complex. Personnel in this field often need to retrieve a large amount of textual information, such as equipment manuals and maintenance guides, during their daily operation and maintenance work, only obtaining a small portion of it to guide their work. The internet has seen a surge in resources, with general search engines and professional forums becoming channels for switching equipment maintenance personnel to find information. However, these channels typically provide search results containing numerous web links mixed with irrelevant information, requiring users to expend considerable time and effort to filter them. With the rise and maturity of knowledge graph technology, more and more fields, such as medicine and law, are adopting knowledge graphs to organize information and use them as a data foundation to build efficient intelligent search and consultation applications. The switching equipment maintenance field can learn from this approach to create accurate, fast, and efficient intelligent search and question-and-answer systems.

[0003] For example, Chinese patent CN110196157B relates to a diagnostic method based on a remote analysis system for holographic diagnosis of rotating equipment, belonging to the field of fault diagnosis. This invention includes a basic information module, a vibration information module, a lubrication recording module, an electrical operation module, a maintenance data module, and an operating parameter module. The diagnostic method is as follows: (i) After the rotating equipment starts and enters normal operation, the relevant measuring points connected to the rotating equipment transmit equipment information to the system and present various characteristic quantities; (ii) By sequentially clicking on the basic information module, vibration information module, lubrication recording module, electrical operation module, maintenance data module, and operating parameter module on the interface, information measured by different detection methods will be displayed; (iii) Vibration information and characteristics of the rotating equipment in various directions, structural characteristics and fault characteristic prompts of the foundation support, and information such as the installation method of the rotating equipment are measured through the basic information module.

[0004] All of the above patents suffer from the problems described in the background: traditional maintenance methods for rotating equipment heavily rely on expert experience, resulting in low efficiency; the variety of rotating equipment makes data retrieval difficult; and data-driven diagnosis lacks interpretability. A precise, rapid, and comprehensive maintenance method for rotating equipment is a practical necessity in the current development of industrial production technology. To address these issues, this application designs a maintenance method and system for rotating equipment. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a method and system for the maintenance of transfer equipment. First, a knowledge extraction network is configured to extract maintenance knowledge sequences. Operational data is collected and processed through data mining to generate entity label quintuples. Second, the entity labels and maintenance knowledge sequences are used to construct a maintenance knowledge graph for the maintenance of transfer equipment. The knowledge extraction network includes an input layer, a feature extraction layer, and an output layer, as well as knowledge sequence optimization and network model updates. Operational data processing includes preprocessing and data classification. The construction of the maintenance knowledge graph includes preprocessing, word embedding model transformation, bidirectional gated recurrent unit extraction of global features, algorithmic extraction of association rules, and computational integration of knowledge nodes.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for overhauling rotating equipment, the method comprising:

[0008] Construct an equipment maintenance knowledge base and extract maintenance knowledge sequences through a maintenance knowledge extraction network;

[0009] Collect operational data from the transfer equipment, process the operational data through data mining, and generate five-tuple entity labels for the transfer equipment.

[0010] The five-tuple of entity tags of the transfer equipment and the maintenance knowledge sequence are used to construct a maintenance knowledge graph of the transfer equipment, and the transfer equipment is maintained according to the maintenance knowledge graph of the transfer equipment.

[0011] The maintenance knowledge extraction network includes:

[0012] The input layer maps the relevant knowledge of the transfer equipment and the fault cases through a pre-trained word vector model, transforms them into a word vector matrix, and calculates the knowledge feature vector based on the word vector matrix.

[0013] The feature extraction layer extracts semantic features from the knowledge feature vector through a bidirectional long short-term memory neural network;

[0014] The output layer scores semantic features, calculates a label score matrix, extracts the optimal sequence based on the label score matrix, decodes the optimal sequence, and calculates the maintenance knowledge sequence.

[0015] The input layer also includes a self-attention mechanism, which is used to calculate an attention matrix based on the word vector matrix and adjust the word vector matrix based on the attention matrix. The bidirectional long short-term memory neural network consists of a forward LSTM and a backward LSTM, which respectively acquire forward and backward information in the knowledge feature vector word sequence, and merge the forward and backward information into the output at the same time to obtain semantic features.

[0016] The output layer also includes a Conditional Random Field (CRF) and a Viterbi algorithm. The CRF is used to globally evaluate the scoring sequence of the semantic feature sequence and calculate the transition score matrix. The Viterbi algorithm calculates the label sequence with the highest score based on dynamic programming to generate the optimal sequence.

[0017] The formula for calculating the optimal sequence is:

[0018]

[0019] Among them, B q Let Score{·} denote the optimal sequence, Score{·} denote the Viterbi algorithm, j denote the index of a single label in the semantic feature sequence, M denote the total number of labels in the semantic feature sequence, D denote the transition score matrix, and y j Let y represent the j-th label in the semantic feature sequence. j+1 Let Q represent the (j+1)th label in the semantic feature sequence, and let Q represent the label score matrix. Let p represent matrix multiplication, and p represent conditional probability.

[0020] The maintenance knowledge extraction network also includes knowledge sequence optimization and network model updating.

[0021] The knowledge sequence optimization configuration is set in the output layer, including filtering redundant information, checking semantic consistency, and dynamic optimization mechanism;

[0022] The network model update configuration is configured in the output layer, including model retraining, model performance monitoring, and model version management.

[0023] The data mining process for processing operational data includes operational data preprocessing and data classification.

[0024] The operation data preprocessing uses a programming language to statistically analyze the operation data and obtain data on the classification of the transfer equipment.

[0025] The data classification extracts entity labels from the transfer equipment classification data using a clustering algorithm, generating a five-tuple of transfer equipment entity labels.

[0026] The entity labels extracted from the transfer equipment classification data using a clustering algorithm include:

[0027] The DBSCAN clustering algorithm is used to perform preliminary clustering of the transfer equipment classification data to obtain core points, and the number of cluster centers is determined by the number of core points.

[0028] The kernel points after DBSCAN clustering are used as the initial center clusters of the K-means clustering algorithm, and the Euclidean distance from the transfer equipment classification data to each initial center cluster is calculated.

[0029] Clustering is performed based on Euclidean distance, and the average clustering error of the initial central clusters is calculated. The central clusters are then updated based on the average clustering error.

[0030] Determine if the convergence condition is met. If the convergence condition is not met, continue to update the central cluster. If the convergence condition is met, output the entity label of the transfer equipment classification item data.

[0031] The construction of the knowledge graph for the maintenance of the transfer equipment includes:

[0032] Preprocess the five-tuple of entity tags of the transfer equipment and the maintenance knowledge sequence to obtain text information;

[0033] The text information is converted into text vectors using a word embedding model, and the text vectors are encoded by combining text position information to obtain tag vectors and knowledge vectors.

[0034] Global features of tag vectors and knowledge vectors are extracted based on bidirectional gated recurrent units to obtain deep semantic information of tag vectors and knowledge vectors;

[0035] The association rules between the deep semantic information are extracted using the Apriori algorithm. The label vector and knowledge vector are matched according to the association rules, and integrated knowledge nodes are calculated.

[0036] The bidirectional gated loop unit includes a forward GRU and a backward GRU;

[0037] The vectors are processed by the forward GRU to gradually generate the forward hidden state of each vector in the sequence;

[0038] The backward GRU is used to process the vectors to generate the backward hidden state of each vector;

[0039] The forward hidden state and the backward hidden state are concatenated to form a complete vector feature representation. The fused features are then compressed using max pooling to calculate the global features.

[0040] The acquisition of deep semantic information of label vectors and knowledge vectors also includes convolutional neural networks and multi-layer aggregation. The convolutional neural network calculates the semantic features of global features at different scales through convolutional pooling kernels. The multi-layer aggregation aggregates the semantic features at different scales to calculate the deep semantic information of label vectors and knowledge vectors.

[0041] The integrated knowledge nodes include:

[0042] The Apriori algorithm is used to mine deep semantic information and extract association rules.

[0043] Map association rules to association rule vectors;

[0044] The association rule vectors are accumulated, and the association rule vectors are used to determine whether they are associated with the tag vectors and knowledge vectors through an indicator function. The overall association contribution value is calculated, and the influence coefficient of the association rule part on the final integrated knowledge node is calculated by combining the association rule vector weight coefficient.

[0045] Calculate the interaction vector between the knowledge vector and the tag vector;

[0046] The knowledge vector, tag vector, and interaction vector are combined using weighting coefficients, and the knowledge nodes are integrated using the influence coefficients.

[0047] The calculation formula for the integrated knowledge node is as follows:

[0048]

[0049] Among them, K i Let F represent the integrated knowledge node, γ represent the weight coefficient of the association rule vector, s represent the individual number of the association rule vector, S represent the total number of association rule vectors, and F represent the total number of association rule vectors. s Let T represent the s-th association rule vector, α represent the label vector weight coefficient, β represent the knowledge vector weight coefficient, and T... i Let S represent the i-th label vector. i Let F represent the i-th knowledge vector. i δ represents the interaction vector between the i-th label vector and the i-th knowledge vector. is The indicator function is δ, which represents the expression for the positive correlation between the i-th label vector and the i-th knowledge vector and the s-th association rule vector. is =1, when the i-th label vector and the i-th knowledge vector are negatively correlated or uncorrelated with the s-th association rule vector, δ is =0, ⊙ represents the Hadamard product.

[0050] The construction of the knowledge graph for the maintenance of the transfer equipment also includes:

[0051] Create a graph database and design its data schema;

[0052] The integrated knowledge nodes and the association rules are imported into the graph database to create a knowledge graph for the maintenance of transfer equipment.

[0053] The maintenance of the rotating equipment includes fault prediction, fault warning, maintenance plan development, and updated diagrams.

[0054] The fault prediction includes:

[0055] The real-time operating data of the transfer equipment is monitored by sensors, and the real-time operating data is mapped to the maintenance knowledge graph of the transfer equipment, and the knowledge graph mapping data is calculated.

[0056] A fault prediction network is constructed, and the operating data and the knowledge graph mapping data are used as input parameters of the fault prediction network. The fault prediction network is used to train the input parameters and output the potential fault modes and potential fault rates of the switching equipment.

[0057] The fault prediction network includes:

[0058] The input layer includes a convolutional module and a memory module. The convolutional module extracts real-time features based on the running data, and the memory module extracts semantic features based on the knowledge graph mapping data.

[0059] The feature fusion layer is used to fuse the real-time features and the semantic features using a normalized embedded Gaussian function to calculate the fused features;

[0060] The output layer includes a first activation function module and a second activation function module. The first activation function module calculates the confidence probability of the fused features using a Softmax classifier and takes the fault category corresponding to the highest confidence probability as the potential fault mode of the transfer equipment. The second activation function module calculates the fault probability of the fused features using a ReLU activation function and determines whether the convergence condition is met. If the convergence condition is not met, the fault probability is backpropagated using a loss function. If the convergence condition is met, the fault probability is output as the potential fault rate.

[0061] A system for overhauling equipment, the system comprising a data acquisition module, a knowledge base construction module, a data processing module, a knowledge graph construction module, and an overhaul module;

[0062] The data acquisition module is used to collect relevant knowledge, fault cases, and operational data of the transfer equipment;

[0063] The knowledge base construction module is used to build an equipment maintenance knowledge base based on relevant knowledge and fault cases of the transfer equipment.

[0064] The data processing module is configured with a data processing strategy, which is used to calculate the maintenance knowledge sequence and the five-tuple of entity tags of the switching equipment based on the equipment maintenance knowledge base and the operating data.

[0065] The knowledge graph construction module is configured with a node integration strategy, which is used to merge the maintenance knowledge sequence and the five-tuple of entity tags of the transfer equipment.

[0066] The maintenance module is used to maintain the rotating equipment based on the maintenance knowledge graph of the rotating equipment. The maintenance module is configured with a fault prediction strategy, which is used to predict the potential fault modes and potential fault rates of the rotating equipment.

[0067] The data processing module includes a knowledge sequence extraction unit and a tag quintuple extraction unit;

[0068] The knowledge sequence extraction unit is used to extract maintenance knowledge sequences through the maintenance knowledge extraction network.

[0069] The tag quintuple extraction unit is used to process the operating data through data mining to generate quintuple tags for the transfer equipment entity;

[0070] The data processing strategy includes knowledge extraction logic and tag classification logic. The knowledge extraction logic is configured within the knowledge sequence extraction unit, and the tag classification logic is configured within the tag quintuple extraction unit.

[0071] The knowledge graph construction module includes a node preprocessing unit, a node integration unit, and a graph construction unit;

[0072] The node preprocessing unit is used to preprocess the maintenance knowledge sequence and the equipment entity label quintuple to generate label vectors and knowledge vectors;

[0073] The node integration unit is used to fuse tag vectors and knowledge vectors to form a comprehensive knowledge node;

[0074] The graph construction unit is used to import the fused knowledge nodes into the graph database to create a knowledge graph for the maintenance of transfer equipment.

[0075] The node integration strategy includes vector encoding logic and knowledge fusion logic. The vector encoding logic is configured within the node preprocessing unit, and the knowledge fusion logic is configured within the node integration unit.

[0076] The maintenance module includes a fault prediction unit, a maintenance planning unit, and a feedback recording unit.

[0077] The fault prediction unit monitors equipment operation data in real time through sensors and maps the data to a knowledge graph to predict potential fault modes and failure rates.

[0078] The maintenance planning unit formulates a detailed maintenance plan based on the fault prediction results, including maintenance time arrangement, task allocation and resource scheduling;

[0079] The feedback recording unit is used to record the operation steps, problems found and solutions during the maintenance process, and update the feedback information to the knowledge graph;

[0080] The fault prediction strategy includes fault prediction logic, which is configured within the fault prediction unit.

[0081] Compared with the prior art, the beneficial effects of the present invention are:

[0082] This invention integrates and organizes a large amount of equipment maintenance-related information through knowledge graph technology, including troubleshooting steps, parts replacement guidelines, maintenance records, etc. By using knowledge graphs, operators can find the information they need more quickly and execute maintenance steps accurately, thereby reducing error rates and improving equipment reliability. Attached Figure Description

[0083] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0084] Figure 1 This is a flowchart illustrating a method for overhauling a rotating equipment according to Embodiment 1 of the present invention.

[0085] Figure 2 This is a diagram of the knowledge extraction network structure in Embodiment 1 of the present invention;

[0086] Figure 3 This is a flowchart of the optimal sequence extraction process in Embodiment 1 of the present invention;

[0087] Figure 4 This is a flowchart of entity label extraction in Embodiment 1 of the present invention;

[0088] Figure 5 This is a flowchart illustrating the construction of a knowledge graph for the maintenance of transfer equipment in Embodiment 1 of the present invention.

[0089] Figure 6 This is a flowchart of the deep semantic information extraction process in Embodiment 1 of the present invention;

[0090] Figure 7 This is a flowchart illustrating the integrated knowledge node calculation process in Embodiment 1 of the present invention.

[0091] Figure 8 This is a fault prediction network structure diagram of Embodiment 1 of the present invention;

[0092] Figure 9 This is a module diagram of a rotating equipment maintenance system according to Embodiment 2 of the present invention. Detailed Implementation

[0093] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0094] Example 1

[0095] Please see Figure 1 The present invention provides an embodiment of a method for overhauling rotating equipment, the specific steps of which are as follows:

[0096] S1: Collect relevant knowledge and fault cases of the rotating equipment to build an equipment maintenance knowledge base;

[0097] S2: Configure a knowledge extraction network to process the equipment maintenance knowledge base and extract maintenance knowledge sequences.

[0098] S3: Obtain raw data of the transfer equipment, perform statistics on the raw data of the transfer equipment using a programming language, and obtain the data of the transfer equipment category items;

[0099] S4: Process the data of the transfer equipment classification items through data mining to obtain the five-tuple of entity labels for transfer equipment;

[0100] S5: Construct a maintenance knowledge graph of the transfer equipment by combining the five-tuple of entity tags of the transfer equipment and the maintenance knowledge sequence;

[0101] S6: Monitor equipment operation data in real time through sensors and map the data to the maintenance knowledge graph of the rotating equipment to predict the potential failure modes and failure rates of the rotating equipment;

[0102] S7: Based on the potential failure modes and failure rates, formulate a maintenance plan, obtain feedback information according to the implementation process of the maintenance plan, and update the feedback information to the maintenance knowledge graph of the rotating equipment.

[0103] The knowledge related to the rotating equipment includes collecting relevant operation manuals, maintenance guidelines, and technical specifications. The fault cases include fault descriptions, cause analyses, and solutions. The operation manuals include equipment installation guidelines, operating procedures, and maintenance precautions, recording the operating requirements and standards of the rotating equipment under different working conditions. The maintenance guidelines provide daily maintenance steps and precautions for the equipment, covering lubrication, tightening, cleaning, and replacement of vulnerable parts. The technical specifications include the equipment's design specifications, technical standards, operating parameters (such as temperature range, pressure range, and speed range), and performance indicators.

[0104] The equipment maintenance knowledge base provides comprehensive knowledge support for the entire maintenance process, ensuring the effectiveness and efficiency of equipment operation, maintenance, and troubleshooting. The construction of the equipment maintenance knowledge base includes:

[0105] S1.1: Based on the relevant knowledge of the transfer equipment, determine the field and scope of the equipment body;

[0106] S1.2: Based on the domain and scope of the device ontology, determine the keywords and key terms of the domain, and build a device ontology knowledge model from top to bottom. The device ontology knowledge model includes a top-level ontology, a concept subtree, and sub-layer instances.

[0107] S1.3: Add the relevant knowledge and fault cases of the transfer equipment as knowledge features to the sub-layer instance. Perform logical verification on the specific equipment ontology knowledge model based on expert knowledge to determine whether it meets the ontology construction principles of the equipment ontology domain. If it does, instantiate the equipment ontology knowledge model. If it does not, continue to add sub-layer instances.

[0108] S1.4: Input the instantiated device ontology knowledge model into the device maintenance knowledge base.

[0109] The principles for constructing device ontology are derived from Gruber's five basic principles: accuracy, consistency, extensibility, minimum ontology commitment, and coding bias. Accuracy means that the ontology should accurately convey the meaning of the terms it contains. Consistency means that the definitions of the ontology are consistent; the initial definition of the ontology must be logically consistent with the results derived from it, without any contradictions. Extensibility means that in the construction of the ontology, not only should industry-consensus terminology be used, but the scope of use of concepts should also be clearly defined to ensure that the ontology can be easily extended without external interference. As knowledge of the device evolves and improves, the constructed device ontology should maintain extensibility. Minimum ontology commitment means reducing the constraints on the target objects in the construction of the ontology. In the process of constructing the device ontology, the constraints on terms should be minimized as much as possible to ensure the diversity of the constructed ontology. Coding bias means that concepts need to be introduced at the knowledge level, rather than relying on specific symbolic level encoding.

[0110] The knowledge extraction network includes:

[0111] The input layer maps the relevant knowledge of the transfer equipment and the fault cases through a pre-trained word vector model, and calculates the knowledge feature vector;

[0112] The feature extraction layer extracts semantic features from the knowledge feature vector through a bidirectional long short-term memory neural network;

[0113] The output layer scores semantic features, extracts the optimal sequence, decodes the optimal sequence, and calculates the maintenance knowledge sequence.

[0114] Please see Figure 2The knowledge extraction network structure diagram of this embodiment of the invention is shown below. This network uses a multi-layer neural network structure, including an input layer, a feature extraction layer, and an output layer, to perform knowledge mapping, feature extraction, scoring, and optimal sequence derivation, ultimately generating a maintenance knowledge sequence. The specific knowledge extraction network is as follows:

[0115] The input layer is primarily responsible for mapping equipment maintenance knowledge and fault cases into a high-dimensional vector space, providing a foundation for subsequent feature extraction. It maps equipment maintenance knowledge and fault cases of the transfer equipment entities using a pre-trained word vector model, transforming them into a word vector matrix. This matrix is ​​then converted into the query vector H, key vector K, and value vector V of the input parameters, represented by word embeddings in the text. An attention matrix is ​​calculated using a self-attention mechanism, the formula for which is as follows:

[0116]

[0117] Among them, M A Let represent the attention matrix, Softmax(·) represent the activation function, T represent the matrix transpose, and d represent the linear dimension;

[0118] The word vector matrix is ​​adjusted based on the attention matrix and continuously updated during network training to optimize the explicit knowledge features of the word vector matrix. The knowledge feature vector is calculated based on the weight coefficient matrix, and the formula for calculating the knowledge feature vector is as follows:

[0119]

[0120] Among them, V c ω represents the knowledge feature vector. q ω k and ω v Represents the weight coefficient matrix;

[0121] The main function of the feature extraction layer is to extract deep semantic features from the knowledge feature vector and obtain contextual information from the time series. This layer includes a bidirectional long short-term memory network, which consists of a forward LSTM and a backward LSTM to obtain forward and backward information from the word sequence of the knowledge feature vector, respectively. The forward LSTM processes the input sequence from left to right to generate forward semantic information at each time step, and the backward LSTM processes the input sequence from right to left to generate backward semantic information at each time step. The forward and backward information are then merged into the output at the same time step to obtain semantic features.

[0122] The output of the bidirectional long short-term memory network contains contextual information for each position in the input sequence, enabling the network to understand the complex relationships between words in the sequence. Through the feature extraction layer, the temporal dependencies in the knowledge feature vector are extracted, providing semantic context support for subsequent sequence scoring and optimal sequence derivation.

[0123] Please see Figure 3 The flowchart of the optimal sequence extraction in this embodiment of the invention is as follows: The output layer first scores the semantic features to obtain the score sequence of the semantic feature sequence. The scoring process is implemented through a linear transformation and a softmax function to generate the score of the label in each semantic feature and calculate the label score matrix.

[0124] Then, the scoring sequence of the semantic feature sequence is globally evaluated by the conditional random field. The relationship between adjacent labels in the sequence is considered to determine whether the transition score matrix satisfies global optimality. If it does not satisfy global optimality, the transition score matrix is ​​recalculated. If it satisfies global optimality, the label score matrix is ​​adjusted by the transition score matrix to make the label score matrix more consistent with the sequence dependency pattern.

[0125] The optimal sequence is obtained using the Viterbi algorithm based on adjacent labels. The Viterbi algorithm calculates the score of all possible paths from the beginning of the sequence to the current time step, records the best path at each step (i.e., the label sequence with the highest score), and after calculating the score of the best path at the last step, backtracking is performed to progressively determine the optimal label at each time step, generating the optimal sequence. The optimal sequence is then decoded to obtain the maintenance knowledge sequence. This step mainly maps the optimal label sequence back to the original maintenance knowledge. The formula for calculating the optimal sequence is:

[0126]

[0127] Among them, B q Let Score{·} denote the optimal sequence, Score{·} denote the Viterbi algorithm, j denote the index of a single label in the semantic feature sequence, M denote the total number of labels in the semantic feature sequence, D denote the transition score matrix, and y j Let y represent the j-th label in the semantic feature sequence. j+1 Let Q represent the (j+1)th label in the semantic feature sequence, and let Q represent the label score matrix. This represents matrix multiplication, where p represents conditional probability;

[0128] By using global evaluation of random conditional fields and the Viterbi algorithm, the output layer does not only select the highest-scoring label at each time step, but also considers the dependencies of the entire sequence to ensure that the generated label sequence is globally optimal.

[0129] The knowledge extraction network also includes knowledge sequence optimization and network model updating;

[0130] In the knowledge extraction network, the generated maintenance knowledge sequence, after initial extraction, may still contain redundant information, inaccurate labels, or sequence arrangements requiring further optimization. To improve the quality and practicality of the final output maintenance knowledge sequence, knowledge sequence optimization is also required. This optimization is configured in the output layer and includes filtering redundant information, checking semantic consistency, and a dynamic optimization mechanism. Filtering redundant information uses rule-based reasoning and frequency-based pruning algorithms to identify frequently occurring repetitive information in multiple cases. Contextual relevance analysis determines whether redundant information is related to the current equipment status and fault type, retaining relevant information and eliminating irrelevant information to ensure the simplicity and efficiency of the maintenance knowledge sequence. Checking semantic consistency uses a sequence labeling model to perform semantic consistency checks on the optimized knowledge sequence. Contextual relationship analysis evaluates the logical relationships between knowledge points to ensure the knowledge sequence is operable in practical applications. For detected semantic inconsistencies, automatic correction is performed using an expert system or predefined rules to ensure the semantic coherence and consistency of all knowledge points in the maintenance knowledge sequence. The dynamic optimization mechanism includes online learning methods to update and optimize the knowledge sequence in real time based on equipment data and fault cases.

[0131] As the equipment operating environment changes, fault modes evolve, and new knowledge accumulates, the knowledge extraction network model also needs to be updated regularly to ensure the system can continuously provide accurate maintenance knowledge. This network model update includes model retraining, model performance monitoring, and model version management. Model retraining involves collecting the latest equipment operating data, maintenance records, and fault cases to construct an updated training set. Incremental learning is then used to add the updated training set to the existing model for training, updating network weights, integrating new and old knowledge, and improving the model's adaptability. Model performance monitoring is used to set performance metrics and periodically update the model. The accuracy of the network model in extracting data from transfer equipment is evaluated, and the model's performance in fault prediction and knowledge extraction is analyzed in real time. When performance degradation is detected, the model update process is automatically triggered. Model diagnostic tools are applied to analyze the causes of model errors in specific scenarios and guide model optimization and retraining. The model version management is used to record version information for each model update, including update content, training dataset, and performance metrics. When the new version of the model performs poorly in practical applications, it can be quickly rolled back to the previous stable version. Multiple model versions are tested under different equipment or fault scenarios, and the best-performing model version is selected as the network model.

[0132] The raw data of the transfer equipment includes sensor data, operation logs, and historical fault records;

[0133] The statistical analysis of the raw data of the transfer equipment using a programming language includes semantic conversion, unified formatting, word segmentation, and word frequency statistics. The data of the transfer equipment classification items includes equipment name, parameter name, and abstract conceptual nouns.

[0134] The specific steps for processing the data on the classification items of transfer equipment through data mining are as follows:

[0135] S3.1: Perform data cleaning on the data of the transfer equipment category, including removing duplicate data, normalizing data, and removing blank remarks data;

[0136] S3.2: Based on the cleaned data of the transfer equipment category items, construct a category profile for the transfer equipment. Extract entity labels from the transfer equipment category item data using a clustering algorithm. Represent the entity labels with knowledge O = {O1, O2, O3, O4, O5}, where O represents the transfer equipment entity label quintuple, O1 represents the concept set of the transfer equipment entity label quintuple, O2 represents the set of hierarchical relationships of the transfer equipment entity label quintuple, O3 represents the characteristics of the transfer equipment entity label quintuple, O4 represents the rule set, and O5 represents the description of the transfer equipment entity label quintuple.

[0137] Please see Figure 4 The entity label extraction flowchart of this invention shows the specific steps for extracting entity labels for transfer equipment classification data using a clustering algorithm as follows:

[0138] S3.2.1: Perform preliminary clustering of the transfer equipment classification data according to the DBSCAN clustering algorithm, obtain kernel points, and determine the number of cluster centers based on the number of kernel points;

[0139] S3.2.2: Use the kernel points after DBSCAN clustering as the initial center clusters of the K-means clustering algorithm, and calculate the Euclidean distance from the transfer equipment classification data to each initial center cluster;

[0140] S3.2.3: Perform clustering based on Euclidean distance and calculate the average clustering error of the initial centroid clusters; update the centroid clusters based on the average clustering error.

[0141] S3.2.4: Repeat step S3.2.3 until all initial central clusters are updated. Determine if the convergence condition is met. If the convergence condition is not met, continue updating the central clusters. If the convergence condition is met, output the entity labels of the transfer equipment classification data.

[0142] Specifically, by combining DBSCAN and K-means clustering algorithms, cluster analysis is performed on the classification data of transfer equipment. Through iterative optimization of the central clusters, more accurate clustering results are obtained, leading to a better understanding of the operating status of the transfer equipment and contributing to improved system efficiency and reduced energy consumption. DBSCAN is a density-based clustering algorithm used to identify closely connected data points in space and divide them into clusters. DBSCAN can be used for preliminary clustering to obtain kernel points. A kernel point is a core region containing at least a specified number of data points within a given radius. Since each kernel point may represent a cluster, the number of kernel points can be used to initially determine the number of central clusters.

[0143] The kernel points obtained from the initial clustering using DBSCAN are used as the initial centroids for the K-means clustering algorithm, providing an initial cluster center location. Euclidean distance is used for cluster partitioning; by calculating the Euclidean distance from each initial centroid to the data point of the transfer equipment category, each piece of operational information is assigned to the nearest centroid. The cluster mean error is the average of the Euclidean distances from all included data points to the centroid. The cluster mean error can be used to update the initial centroids to improve the clustering effect. The position of the centroids can be adjusted based on the calculated cluster mean error, thus updating the centroids. A convergence condition is defined to determine whether the updated centroids meet the condition, thus determining whether to end the iteration. If the convergence condition is met, the clustering result is output.

[0144] Please see Figure 5 This invention provides a flowchart for constructing a knowledge graph for the maintenance of transfer equipment. The purpose of constructing this knowledge graph is to integrate the entity tags of the equipment and maintenance knowledge in a structured manner, forming a knowledge graph that can be efficiently retrieved and applied. The construction of the knowledge graph includes:

[0145] S5.1: Preprocess the five-tuple of entity tags of the transfer equipment and the maintenance knowledge sequence to obtain text information, which includes tag information and knowledge information.

[0146] The preprocessing includes:

[0147] S5.1.1: Perform data cleaning on the equipment entity tag quintuple and maintenance knowledge sequence, removing null values, duplicate data and noisy data;

[0148] S5.1.2: Implement standardization processing to unify data of different formats into a consistent format and standardize the use of terminology to ensure data interoperability between different modules;

[0149] S5.1.3: Use a word segmentation tool to segment the text and generate a basic vocabulary list.

[0150] S5.1.4: Perform part-of-speech tagging to identify device nouns, operation verbs, fault nouns, etc. in the text, and prepare for subsequent word embedding and vectorization.

[0151] S5.2: The text information is converted into a text vector through a word embedding model, and the text vector is encoded in combination with the text position information to obtain a tag vector and a knowledge vector, thus converting the text information into a numerical vector representation for easy computer processing;

[0152] A pre-trained word embedding model is applied to transform the preprocessed text information into text vectors. Each word is mapped to a fixed-dimensional vector through the word embedding model, preserving the semantic information of the word.

[0153] By encoding the text with positional information and combining it with the position of words in the sentence, a more context-sensitive vector representation can be generated. Positional information encoding can be performed using either relative or absolute positional encoding.

[0154] Please see Figure 6 A flowchart of deep semantic information extraction according to an embodiment of the present invention;

[0155] S5.3: Extract global features of tag vectors and knowledge vectors based on bidirectional gated loop units, obtain deep semantic information of tag vectors and knowledge vectors, capture deep semantic information of each tag and knowledge point, and ensure that in practical applications, the knowledge graph can accurately reflect the status of the equipment and its associated maintenance knowledge.

[0156] Bidirectional gated recurrent units (Bi-GRUs) are neural network architectures that extend the traditional unidirectional GRU, enabling simultaneous processing of forward and backward information in sequential data. GRUs are gated recurrent neural networks that primarily control information flow through update and reset gates, thus mitigating the vanishing gradient problem encountered by traditional RNNs in long sequences. Bi-GRU combines two GRUs: one processes the sequence from left to right, and the other from right to left, allowing the network to obtain more complete contextual information about the input data.

[0157] First, the vectors are processed using a forward GRU to generate a forward hidden state for each vector in the sequence. The forward hidden state not only contains information about the current vector but also accumulates information about all vectors preceding it. Then, the vectors are processed using a backward GRU to generate a backward hidden state for each vector. The backward hidden state starts from the end of the sequence and accumulates information about the current vector and all subsequent vectors. The forward and backward hidden states are concatenated to form a complete vector feature representation. The fused features are compressed using a max pooling operation to calculate global features. The pooling operation can retain key information in the features while reducing redundant data.

[0158] After extracting global features, further extracting deep semantic information from label vectors and knowledge vectors is a crucial step in constructing a high-quality knowledge graph. By combining this with convolutional neural networks (CNNs), which can better capture local features and patterns, high-order features of local regions in the sequence can be extracted.

[0159] A convolution kernel is applied to the global feature representation of each label vector or knowledge vector. Convolution is performed using a sliding window. The size of the convolution kernel can be adjusted according to the specific global features, usually between 3 and 5, to capture local patterns between neighboring vectors and generate feature maps. These feature maps reflect the local features of the input sequence. Through these feature maps, the system can better understand the semantic information in the local context. Local pooling is performed on the convolved feature maps, and then global pooling is performed in a larger context window to calculate semantic features at different scales.

[0160] By employing hierarchical aggregation techniques to combine semantic features at different scales, the system computes deep semantic information from label vectors and knowledge vectors. By integrating low-level detailed features with high-level abstract features, the system can generate a more expressive feature vector. This hierarchical integration ensures that nodes in the knowledge graph possess both rich detail and semantic integrity.

[0161] S5.4: Extract the association rules between the deep semantic information according to the Apriori algorithm, match the tag vector and knowledge vector according to the association rules, and calculate and integrate the knowledge nodes.

[0162] Please see Figure 7 The present invention integrates a knowledge node calculation flowchart. The specific steps of S5.4 in the present invention are as follows:

[0163] S5.4.1: Process deep semantic information and mine potential association rules between equipment tags and maintenance knowledge using the Apriori algorithm. The Apriori algorithm is a classic association rule mining algorithm. It discovers frequently occurring itemsets from the dataset through iteration and generates meaningful association rules, which represent the potential relationship between equipment tags and maintenance knowledge.

[0164] Association rules can be represented as form S A →K A S A It is a specific tag in the tag set, K A It is a specific piece of knowledge in a knowledge set, representing a specific tag S in a tag set. A When it appears, a certain knowledge K A It also has a high probability of occurring;

[0165] S5.4.2: Map association rules to association rule vectors. The association rule vectors represent a fixed pattern related to the label vectors and knowledge vectors. Vectorization enables these rules to be further processed and calculated by the model, making the application of association rules more flexible and efficient. By converting rules into vectors, matrix operations and vector operations can be performed efficiently, thereby accelerating the processing flow.

[0166] S5.4.3: The association rule vectors are accumulated to comprehensively consider the relationship between all possible tags and knowledge. During the accumulation process, an indicator function is used to determine whether the association rule vectors are associated with the current tag vectors and knowledge vectors, thereby calculating the overall contribution of these association rule vectors to the integrated knowledge nodes.

[0167] S5.4.4: Combine the association rule vector weight coefficient to calculate the influence coefficient of the association rule part on the final integrated knowledge node, integrate the influence of multiple association rules, improve the expressiveness of the knowledge node, and ensure that the generated node can accurately reflect the complex relationship between equipment status and maintenance knowledge;

[0168] S5.4.5: Combine the knowledge vector and the label vector by dot product to calculate the interaction vector between the knowledge vector and the label vector. The interaction vector can capture more complex semantic information, such as how a specific device state affects the implementation of a specific operation. The introduction of the interaction vector enhances the semantic understanding ability of the system, so that the knowledge node not only contains simple labels and knowledge information, but also contains the dynamic relationship between them.

[0169] S5.4.6: Combine knowledge vectors, tag vectors, and interaction vectors through weighting coefficients, integrate their interrelationships, and calculate and integrate knowledge nodes through the influence coefficients.

[0170] Through the detailed steps described above, useful association rules can be extracted from deep semantic information. These rules are then combined with label vectors and knowledge vectors to generate integrated knowledge nodes. This process not only improves the expressiveness of the knowledge graph but also enhances the system's understanding of equipment status and maintenance procedures, ultimately providing technical support for achieving intelligent equipment management and efficient maintenance.

[0171] The calculation formula for the integrated knowledge node is as follows:

[0172]

[0173] Among them, K i Let F represent the integrated knowledge node, γ represent the weight coefficient of the association rule vector, s represent the individual number of the association rule vector, S represent the total number of association rule vectors, and F represent the total number of association rule vectors. s Let T represent the s-th association rule vector, α represent the label vector weight coefficient, β represent the knowledge vector weight coefficient, and T... i Let S represent the i-th label vector. i Let F represent the i-th knowledge vector. i δ represents the interaction vector between the i-th label vector and the i-th knowledge vector. is The indicator function is δ, which represents the expression for the positive correlation between the i-th label vector and the i-th knowledge vector and the s-th association rule vector. is =1, when the i-th label vector and the i-th knowledge vector are negatively correlated or uncorrelated with the s-th association rule vector, δ is =0, ⊙ represents the Hadamard product.

[0174] The construction of the knowledge graph for the maintenance of the transfer equipment also includes:

[0175] S5.5: Create a graph database and design the data schema of the graph database. Graph databases are different from traditional relational databases. They store data in the form of nodes and edges, which are more suitable for representing the relationships between entities. The graph database is created using software Neo4j, ArangoDB and Amazon Neptune. The data schema design includes node type definition, edge type definition and attribute definition.

[0176] S5.6: Import the integrated knowledge nodes and the association rules into the graph database to create a knowledge graph for the maintenance of transfer equipment.

[0177] The node type definition includes:

[0178] Equipment node: Represents the transfer equipment entity, containing basic equipment information such as equipment ID, equipment name, category, and operating status;

[0179] Knowledge nodes: These represent maintenance knowledge related to the equipment, including operating procedures, maintenance guidelines, solutions, and other information.

[0180] Fault node: Represents the fault information of the equipment, including fault type, occurrence time, fault description, cause analysis, etc.

[0181] Tag node: Represents the tag information of the device, such as device category tag, operating status tag, maintenance record tag, etc.

[0182] The edge type definition includes:

[0183] Association: The association between device nodes and knowledge nodes, used to represent possible maintenance steps or fault solutions for a specific device in a certain state;

[0184] Impact Relationship: The impact relationship between faulty nodes and equipment nodes, used to represent the impact of a fault on the operating status of the equipment;

[0185] Dependency: The dependency relationship between knowledge nodes is used to indicate that a certain maintenance step depends on other steps or knowledge points.

[0186] The attribute definition includes:

[0187] Each node and edge can have several attributes. For example, the attributes of a device node may include manufacturer, installation date, and service life; the attributes of a knowledge node may include step description, tool requirements, and execution time.

[0188] With proper attribute settings, the system can efficiently store and retrieve devices and related information in the graph database;

[0189] After importing nodes and edges, a preliminary knowledge graph structure is created using a graph database, forming a network of nodes and relationships. This network represents the interrelationships between devices, faults, tags, and knowledge points. To make the knowledge graph more accurate and refined, the relationships between nodes can be further processed. For example, based on the actual usage of the devices and historical data, the association weights between devices and knowledge nodes can be dynamically adjusted to reflect the actual performance of the devices in different environments.

[0190] The specific steps of S6 are as follows:

[0191] S6.1: Various types of sensors are installed on the rotating equipment to monitor its operating status. These sensors may include temperature sensors, vibration sensors, pressure sensors, current sensors, etc. These sensors continuously collect the equipment's operating data;

[0192] S6.2: Transform operational data into high-dimensional feature vectors through feature extraction algorithms. For example, for vibration data, extract time-domain features (such as mean and variance) and frequency-domain features (such as spectral peak value and frequency band energy). These feature vectors can better describe the operating status of the equipment and provide a basis for subsequent fault prediction.

[0193] S6.3: The high-dimensional feature vector is mapped to nodes in the knowledge graph using predefined mapping rules or a trained mapping model. The mapping rules can be based on historical data analysis or trained using a machine learning model. The mapped data points find corresponding fault modes or related maintenance knowledge in the knowledge graph, thereby enriching the real-time information of the graph nodes.

[0194] S6.4: Combine high-dimensional feature vectors with nodes in the knowledge graph to calculate the knowledge graph mapping data;

[0195] S6.5: Construct a fault prediction network, using the operating data and the knowledge graph mapping data as input parameters of the fault prediction network, train the input parameters through the fault prediction network, and output the potential fault modes and potential fault rates of the switching equipment.

[0196] Please see Figure 8 The fault prediction network structure diagram of this invention is shown below. The fault prediction network is a complex neural network structure used to predict potential equipment failure modes and failure rates from real-time operational data acquired from sensors and information extracted from knowledge graphs. The network consists of three parts: an input layer, a feature fusion layer, and an output layer. Each layer performs a specific function to ensure accurate fault prediction. The fault prediction network includes:

[0197] The input layer includes a convolutional module and a memory module. The convolutional module extracts real-time features based on the operational data. The convolutional module contains a convolutional neural network, which processes the operational features using convolutional kernels and pooling kernels to effectively extract complex features from the real-time operational data, especially nonlinear relationships that are difficult to capture using simple statistical methods. These features provide high-quality input to the fault prediction model, ensuring that the model can identify potential fault modes. The memory module extracts semantic features based on the knowledge graph mapping data. The memory module receives data mapped from the knowledge graph. This data includes the equipment's historical operating status, fault modes, and related maintenance knowledge. The memory module can extract semantic features from this data. These features contain the equipment's behavioral patterns under different conditions and the relationship between these patterns and fault occurrence. It can remember long-term dependency information and control which information needs to be retained and which can be discarded through memory units and forget gates, capturing key semantic information in complex operational state changes.

[0198] The feature fusion layer is responsible for fusing real-time features extracted from the input layer with semantic features to form a unified feature vector for subsequent prediction. This layer achieves effective feature fusion through normalization and embedded Gaussian functions. Before fusing features, features from different sources need to be normalized. Since real-time features and semantic features may have different dimensions and numerical ranges, normalization ensures that these features have the same influence during fusion. The real-time features and semantic features are fused through normalization and embedded Gaussian functions. Features are weighted using a Gaussian distribution to calculate the fused features. The Gaussian function can smoothly fuse different features and ensure that the transition in the feature space is continuous and reasonable. This smooth fusion method helps avoid sudden changes or discontinuities between features, improving the stability of the model. The fused feature vector contains both dynamic features from real-time running data and static and historical information from the knowledge graph. This comprehensive representation can provide a more comprehensive information foundation for fault prediction.

[0199] The output layer includes a first activation function module and a second activation function module. The first activation function module calculates the confidence probability of the fused features using a Softmax classifier and takes the fault category corresponding to the highest confidence probability as the potential fault mode of the transfer equipment. The second activation function module calculates the fault probability of the fused features using a ReLU activation function and determines whether the convergence condition is met. If the convergence condition is not met, the fault probability is backpropagated using a loss function. If the convergence condition is met, the fault probability is output as the potential fault rate.

[0200] The fault warning system generates warning information based on the fault mode recognition results and pushes it to maintenance personnel, indicating the possible faults and their severity.

[0201] The formulation of the maintenance plan includes:

[0202] Based on the equipment's operating status and production plan, schedule maintenance time reasonably to minimize production downtime.

[0203] Based on the equipment's fault characteristics and maintenance knowledge sequence, specific maintenance tasks are assigned, and maintenance steps and required tools are determined.

[0204] Based on the maintenance tasks and equipment status, dispatch the necessary maintenance personnel, tools and spare parts to ensure the smooth progress of the maintenance work;

[0205] The updated knowledge graph is used to establish a maintenance feedback mechanism. Through feedback from maintenance personnel's actual operations, the knowledge graph is optimized and updated. Specifically, this includes:

[0206] Feedback collection: Collect feedback from maintenance personnel via mobile devices or computers, including the rationality of maintenance procedures and the effectiveness of solutions;

[0207] Data analysis: Analyze the collected feedback data to identify shortcomings and areas for improvement in the knowledge graph;

[0208] Knowledge Updates: Based on feedback, update the maintenance knowledge and fault cases in the knowledge graph to ensure the timeliness and accuracy of the knowledge graph.

[0209] Example 2

[0210] Please see Figure 9 The present invention provides an embodiment of a rotating equipment maintenance system, the system comprising a data acquisition module, a knowledge base construction module, a data processing module, a knowledge graph construction module, and a maintenance module;

[0211] The data acquisition module is used to collect relevant knowledge, fault cases, and operational data of the transfer equipment;

[0212] The knowledge base construction module is used to build an equipment maintenance knowledge base based on relevant knowledge and fault cases of the transfer equipment.

[0213] The data processing module is configured with a data processing strategy, which is used to calculate the maintenance knowledge sequence and the five-tuple of entity tags of the switching equipment based on the equipment maintenance knowledge base and the operating data.

[0214] The knowledge graph construction module is configured with a node integration strategy, which is used to merge the maintenance knowledge sequence and the five-tuple of entity tags of the transfer equipment.

[0215] The maintenance module is used to maintain the rotating equipment based on the maintenance knowledge graph of the rotating equipment. The maintenance module is configured with a fault prediction strategy, which is used to predict the potential fault modes and potential fault rates of the rotating equipment.

[0216] The data processing module includes a knowledge sequence extraction unit and a tag quintuple extraction unit;

[0217] The knowledge sequence extraction unit is used to extract maintenance knowledge sequences through the maintenance knowledge extraction network.

[0218] The tag quintuple extraction unit is used to process the operating data through data mining to generate quintuple tags for the transfer equipment entity;

[0219] The data processing strategy includes knowledge extraction logic and tag classification logic. The knowledge extraction logic is used to extract maintenance knowledge sequences from the equipment maintenance knowledge base, and the tag classification logic is used to extract tag quintuples. The knowledge extraction logic is configured within the knowledge sequence extraction unit, and the tag classification logic is configured within the tag quintuple extraction unit.

[0220] The knowledge graph construction module includes a node preprocessing unit, a node integration unit, and a graph construction unit;

[0221] The node preprocessing unit is used to preprocess the maintenance knowledge sequence and the equipment entity label quintuple to generate label vectors and knowledge vectors;

[0222] The node integration unit is used to fuse tag vectors and knowledge vectors to form a comprehensive knowledge node;

[0223] The graph construction unit is used to import the fused knowledge nodes into the graph database to create a knowledge graph for the maintenance of transfer equipment.

[0224] The node integration strategy includes vector encoding logic and knowledge fusion logic. The vector encoding logic is used to convert the text information into text vectors through a word embedding model, and encode the text vectors in combination with text position information to obtain tag vectors and knowledge vectors. The knowledge fusion logic is used to match the tag vectors and knowledge vectors according to association rules to calculate and integrate knowledge nodes. The vector encoding logic is configured in the node preprocessing unit, and the knowledge fusion logic is configured in the node integration unit.

[0225] The maintenance module includes a fault prediction unit, a maintenance planning unit, and a feedback recording unit.

[0226] The fault prediction unit monitors equipment operation data in real time through sensors and maps the data to a knowledge graph to predict potential fault modes and failure rates.

[0227] The maintenance planning unit formulates a detailed maintenance plan based on the fault prediction results, including maintenance time arrangement, task allocation and resource scheduling;

[0228] The feedback recording unit is used to record the operation steps, problems found and solutions during the maintenance process, and update the feedback information to the knowledge graph;

[0229] The fault prediction strategy includes fault prediction logic, which is used to predict the fault status of the transfer equipment, and the fault prediction logic is configured within the fault prediction unit.

[0230] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for overhauling rotating equipment, characterized in that, The method includes: Construct an equipment maintenance knowledge base and extract maintenance knowledge sequences through a maintenance knowledge extraction network; Collect operational data from the transfer equipment, process the operational data through data mining, and generate five-tuple entity labels for the transfer equipment. The five-tuple of entity tags of the transfer equipment and the maintenance knowledge sequence are used to construct a maintenance knowledge graph of the transfer equipment, and the transfer equipment is maintained according to the maintenance knowledge graph of the transfer equipment. The maintenance knowledge extraction network includes: The input layer maps the relevant knowledge and fault cases of the transfer equipment to a word vector matrix through a pre-trained word vector model, and calculates the knowledge feature vector based on the word vector matrix. The feature extraction layer extracts semantic features from the knowledge feature vector through a bidirectional long short-term memory neural network; The output layer scores semantic features, calculates a label score matrix, extracts the optimal sequence based on the label score matrix, decodes the optimal sequence, and calculates the maintenance knowledge sequence. The output layer also includes a conditional random field (CRF) and a Viterbi algorithm. The CRF is used to globally evaluate the scored sequence of the semantic feature sequence and calculate the transition score matrix. The Viterbi algorithm calculates the label sequence with the highest score based on dynamic programming to generate the optimal sequence. The formula for calculating the optimal sequence is: ， in, Let Score{•} denote the optimal sequence, Score{•} denote the Viterbi algorithm, j denote the index of a single label in the semantic feature sequence, M denote the total number of labels in the semantic feature sequence, and D denote the transition score matrix. This represents the j-th label in the semantic feature sequence. Let Q represent the (j+1)th label in the semantic feature sequence, and let Q represent the label score matrix. This represents matrix multiplication, where p represents conditional probability; The construction of the knowledge graph for the maintenance of the transfer equipment includes: Preprocess the five-tuple of entity tags of the transfer equipment and the maintenance knowledge sequence to obtain text information; The text information is converted into text vectors using a word embedding model, and the text vectors are encoded by combining text position information to obtain tag vectors and knowledge vectors. Global features of tag vectors and knowledge vectors are extracted based on bidirectional gated recurrent units to obtain deep semantic information of tag vectors and knowledge vectors; The association rules between the deep semantic information are extracted using the Apriori algorithm. The label vector and knowledge vector are matched according to the association rules, and integrated knowledge nodes are calculated.

2. The method for overhauling a rotating equipment according to claim 1, characterized in that, The input layer also includes a self-attention mechanism, which is used to calculate an attention matrix based on the word vector matrix and adjust the word vector matrix based on the attention matrix. The bidirectional long short-term memory neural network consists of a forward LSTM and a backward LSTM, which respectively acquire forward and backward information in the knowledge feature vector word sequence, and merge the forward and backward information into the output at the same time to obtain semantic features.

3. The method for overhauling a rotating equipment according to claim 1, characterized in that, The maintenance knowledge extraction network also includes knowledge sequence optimization and network model updating. The knowledge sequence optimization configuration is set in the output layer, including filtering redundant information, checking semantic consistency, and dynamic optimization mechanism; The network model update configuration is configured in the output layer, including model retraining, model performance monitoring, and model version management.

4. The method for overhauling a rotating equipment according to claim 1, characterized in that, The data mining process for processing operational data includes operational data preprocessing and data classification. The operation data preprocessing uses a programming language to statistically analyze the operation data and obtain data on the classification of the transfer equipment. The data classification extracts entity labels from the transfer equipment classification data using a clustering algorithm, generating a five-tuple of transfer equipment entity labels.

5. The method for overhauling a rotating equipment according to claim 4, characterized in that, The entity labels extracted from the transfer equipment classification data using a clustering algorithm include: The DBSCAN clustering algorithm is used to perform preliminary clustering of the transfer equipment classification data to obtain core points, and the number of cluster centers is determined by the number of core points. The kernel points after DBSCAN clustering are used as the initial center clusters of the K-means clustering algorithm, and the Euclidean distance from the transfer equipment classification data to each initial center cluster is calculated. Clustering is performed based on Euclidean distance, and the average clustering error of the initial central clusters is calculated. The central clusters are then updated based on the average clustering error. Determine if the convergence condition is met. If the convergence condition is not met, continue to update the central cluster. If the convergence condition is met, output the entity label of the transfer equipment classification item data.

6. The method for overhauling a rotating equipment according to claim 1, characterized in that, The bidirectional gated loop unit includes a forward GRU and a backward GRU; The vectors are processed by the forward GRU to gradually generate the forward hidden state of each vector in the sequence; The backward GRU is used to process the vectors to generate the backward hidden state of each vector; The forward hidden state and the backward hidden state are concatenated to form a complete vector feature representation. The fused features are then compressed using max pooling to calculate the global features.

7. The method for overhauling a rotating equipment according to claim 6, characterized in that, The acquisition of deep semantic information of label vectors and knowledge vectors also includes convolutional neural networks and multi-layer aggregation. The convolutional neural network calculates the semantic features of global features at different scales through convolutional pooling kernels. The multi-layer aggregation aggregates the semantic features at different scales to calculate the deep semantic information of label vectors and knowledge vectors.

8. The method for overhauling a rotating equipment according to claim 1, characterized in that, The integrated knowledge nodes include: The Apriori algorithm is used to mine deep semantic information and extract association rules. Map association rules to association rule vectors; The association rule vectors are accumulated, and the association rule vectors are used to determine whether they are associated with the tag vectors and knowledge vectors through an indicator function. The overall association contribution value is calculated, and the influence coefficient of the association rule part on the final integrated knowledge node is calculated by combining the association rule vector weight coefficient. Calculate the interaction vector between the knowledge vector and the tag vector; The knowledge vector, tag vector, and interaction vector are combined using weighting coefficients, and the knowledge nodes are integrated using the influence coefficients.

9. The method for overhauling a rotating equipment according to claim 8, characterized in that, The calculation formula for the integrated knowledge node is as follows: ， in, This indicates the integration of knowledge nodes. This represents the weight coefficient of the association rule vector, s represents the individual index of the association rule vector, and S represents the total number of association rule vectors. This represents the vector of the s-th association rule. This represents the label vector weight coefficient. This represents the weight coefficient of the knowledge vector. This represents the i-th label vector. This represents the i-th knowledge vector. This represents the interaction vector between the i-th label vector and the i-th knowledge vector. This indicates the indicator function, which states that when the i-th label vector and the i-th knowledge vector are positively correlated with the s-th association rule vector, When the i-th label vector and the i-th knowledge vector are negatively correlated or uncorrelated with the s-th association rule vector, , It represents the Hadamardi (or Hadama) stack.

10. The method for overhauling a rotating equipment according to claim 1, characterized in that, The construction of the knowledge graph for the maintenance of the transfer equipment also includes: Create a graph database and design its data schema; The integrated knowledge nodes and the association rules are imported into the graph database to create a knowledge graph for the maintenance of transfer equipment.

11. The method for overhauling a rotating equipment according to claim 1, characterized in that, The maintenance of the rotating equipment includes fault prediction, fault warning, maintenance plan development, and updated diagrams.

12. The method for overhauling a rotating equipment according to claim 11, characterized in that, The fault prediction includes: The real-time operating data of the transfer equipment is monitored by sensors, and the real-time operating data is mapped to the maintenance knowledge graph of the transfer equipment, and the knowledge graph mapping data is calculated. A fault prediction network is constructed, and the operating data and the knowledge graph mapping data are used as input parameters of the fault prediction network. The fault prediction network is used to train the input parameters and output the potential fault modes and potential fault rates of the switching equipment.

13. The method for overhauling a rotating equipment according to claim 12, characterized in that, The fault prediction network includes: The input layer includes a convolutional module and a memory module. The convolutional module extracts real-time features based on the running data, and the memory module extracts semantic features based on the knowledge graph mapping data. The feature fusion layer is used to fuse the real-time features and the semantic features using a normalized embedded Gaussian function to calculate the fused features; The output layer includes a first activation function module and a second activation function module. The first activation function module calculates the confidence probability of the fused features using a Softmax classifier and takes the fault category corresponding to the highest confidence probability as the potential fault mode of the transfer equipment. The second activation function module calculates the fault probability of the fused features using a ReLU activation function and determines whether the convergence condition is met. If the convergence condition is not met, the fault probability is backpropagated using a loss function. If the convergence condition is met, the fault probability is output as the potential fault rate.

14. A rotating equipment maintenance system, implemented based on a rotating equipment maintenance method as described in any one of claims 1-13, characterized in that, The system includes a data acquisition module, a knowledge base construction module, a data processing module, a knowledge graph construction module, and a maintenance module. The data acquisition module is used to collect relevant knowledge, fault cases, and operational data of the transfer equipment; The knowledge base construction module is used to build an equipment maintenance knowledge base based on relevant knowledge and fault cases of the transfer equipment. The data processing module is configured with a data processing strategy, which is used to calculate the maintenance knowledge sequence and the five-tuple of entity tags of the switching equipment based on the equipment maintenance knowledge base and the operating data. The knowledge graph construction module is configured with a node integration strategy, which is used to merge the maintenance knowledge sequence and the five-tuple of entity tags of the transfer equipment. The maintenance module is used to maintain the rotating equipment based on the maintenance knowledge graph of the rotating equipment. The maintenance module is configured with a fault prediction strategy, which is used to predict the potential fault modes and potential fault rates of the rotating equipment.

15. The maintenance system for a rotating equipment according to claim 14, characterized in that, The data processing module includes a knowledge sequence extraction unit and a tag quintuple extraction unit; The knowledge sequence extraction unit is used to extract maintenance knowledge sequences through the maintenance knowledge extraction network. The tag quintuple extraction unit is used to process the operating data through data mining to generate quintuple tags for the transfer equipment entity; The data processing strategy includes knowledge extraction logic and tag classification logic. The knowledge extraction logic is configured within the knowledge sequence extraction unit, and the tag classification logic is configured within the tag quintuple extraction unit.

16. The maintenance system for a rotating equipment according to claim 14, characterized in that, The knowledge graph construction module includes a node preprocessing unit, a node integration unit, and a graph construction unit; The node preprocessing unit is used to preprocess the maintenance knowledge sequence and the equipment entity label quintuple to generate label vectors and knowledge vectors; The node integration unit is used to fuse tag vectors and knowledge vectors to form a comprehensive knowledge node; The graph construction unit is used to import the fused knowledge nodes into the graph database to create a knowledge graph for the maintenance of transfer equipment. The node integration strategy includes vector encoding logic and knowledge fusion logic. The vector encoding logic is configured within the node preprocessing unit, and the knowledge fusion logic is configured within the node integration unit.

17. The maintenance system for a rotating equipment according to claim 14, characterized in that, The maintenance module includes a fault prediction unit, a maintenance planning unit, and a feedback recording unit. The fault prediction unit monitors equipment operation data in real time through sensors and maps the data to a knowledge graph to predict potential fault modes and failure rates. The maintenance planning unit formulates a detailed maintenance plan based on the fault prediction results, including maintenance time arrangement, task allocation and resource scheduling; The feedback recording unit is used to record the operation steps, problems found and solutions during the maintenance process, and update the feedback information to the knowledge graph; The fault prediction strategy includes fault prediction logic, which is configured within the fault prediction unit.

Images