A method for air compressor fault diagnosis combining knowledge graph and multimodal large model

Abstract

This invention discloses a fault diagnosis method for air compressors that combines knowledge graphs and multimodal large models, belonging to the field of industrial fault diagnosis technology. The method includes: receiving air compressor vibration signals, extracting data features, and drawing images; converting numerical and image features into text descriptions using a trained data model and a multimodal large model, respectively, and performing terminology alignment and illusion filtering with a pre-built domain knowledge graph to generate a standardized semantic set; fusing multimodal features using a source-balanced weighting method based on macro-average logic; constructing a contextual hint engineering based on GraphRAG according to the pre-built domain knowledge graph to obtain a weighted triplet list; and using a large language model to perform diagnosis according to a "metacognitive diagnostic reasoning" process, outputting the fault type and customized maintenance suggestions. This invention effectively improves diagnostic accuracy and interpretability and suppresses model illusions by deeply integrating knowledge graph constraints and the reasoning capabilities of large models.

Description

Technical Field

[0001] This invention belongs to the field of industrial fault diagnosis technology, specifically involving an interpretable intelligent fault diagnosis method for air compressors that combines a Large Language Model (LLM) and a Domain Knowledge Graph (KG), which is particularly suitable for industrial scenarios of air compressor fault diagnosis. Background Technology

[0002] Air compressors are core equipment in industrial production, providing compressed air and are widely used in machinery manufacturing, chemical industry, building materials, and other fields. As a type of pressure-bearing special equipment, the operation of air compressors involves risks such as high pressure, high temperature, and high-speed mechanical movement. Unplanned shutdowns can lead to high maintenance costs, production interruptions, and safety hazards. However, in the field of air compressors, existing diagnostic methods such as human experience, machine learning, and direct application of general large models for fault diagnosis still face significant bottlenecks in practical engineering applications.

[0003] 1) The subjective nature of manual experience-based diagnosis and its difficulty in large-scale reuse: Traditional air compressor fault diagnosis mainly relies on the personal experience of engineers, making it difficult to form a standardized, structured and reusable diagnostic system; at the same time, as equipment models are updated and operating conditions change, experience rules are prone to becoming invalid, maintenance costs are high, and it is difficult to support large-scale industrial applications.

[0004] (2) Machine learning methods are highly dependent on data and lack interpretability: methods such as SVM, RF, CNN, and LSTM rely on a large amount of high-quality labeled data. However, there are common problems of imbalanced samples and scarce fault samples in industrial sites, which leads to insufficient generalization ability of the model under new working conditions and rare fault scenarios. At the same time, these models are mostly "black boxes" and it is difficult to provide reasoning basis that conforms to engineering mechanisms. It is difficult for maintenance personnel to verify and trust their diagnostic results.

[0005] (3) General large language model lacks domain constraints and industrial data perception capabilities: Although general LLM has powerful natural language understanding and reasoning capabilities, it lacks knowledge constraints in the air compressor domain and is prone to "illusion" type erroneous conclusions; at the same time, it is difficult to directly integrate field time-series sensor data, and cannot form an end-to-end state perception and diagnosis closed loop, making it difficult to meet the requirements of industrial applications.

[0006] In summary, there is an urgent need for an intelligent fault diagnosis system for air compressors that deeply integrates the structured semantics of knowledge graphs with the semantic reasoning capabilities of large models, can input time-series data, and possesses interpretability and self-evolution capabilities, in order to meet the actual needs of large-scale, complex, and multi-source heterogeneous diagnosis in industrial sites. Summary of the Invention

[0007] The purpose of this invention is to overcome the deficiencies in the prior art and to provide a method for diagnosing air compressor faults that combines knowledge graphs and multimodal large models.

[0008] The specific technical solution adopted in this invention is as follows:

[0009] In a first aspect, the present invention provides a method for diagnosing air compressor faults that combines knowledge graphs and multimodal large models, as detailed below:

[0010] S1: Receive the vibration signal from the target air compressor measuring point, preprocess it, calculate the numerical characteristic parameters, and draw the corresponding image of the vibration signal;

[0011] S2: Use the trained data model to convert the set of numerical feature parameters obtained in S1 into a first text feature description. The multimodal large model is invoked to convert the image obtained in S1 into a second text feature description. The large language model is invoked to compare the two text feature descriptions with a pre-built knowledge graph of the air compressor domain. The standard terms in the text are aligned, and the vector similarity calculation method is used for illusion detection and filtering to generate a standardized set of keywords;

[0012] S3: Using a source balance weighting method based on macro-average logic, weights are calculated for the two sets of standardized keywords obtained in S2, and a fused text description set with weight labels is generated.

[0013] S4: Based on the fused text description set described in S3, query the knowledge graph described in S2 to generate a list of weighted triples;

[0014] S5: Construct diagnostic prompt words containing the numerical feature parameters described in S1 and the weighted triple list described in S4, guide the large language model to perform step-by-step reasoning according to the preset metacognitive reasoning chain, and output the most likely fault type and diagnostic confidence.

[0015] S6: Based on the fault type determined in S5, query the standard fault causes and maintenance suggestions from the knowledge graph described in S2; call the large language model again, combine the characteristics of the target air compressor with the reasoning process in S5, and rewrite the standard fault causes and maintenance suggestions in a customized manner to generate the final fault diagnosis report.

[0016] Preferably, in S2, the knowledge graph of the air compressor field The specific construction method is as follows:

[0017] S21: Collect and integrate multi-source data such as relevant teaching materials, maintenance manuals, historical fault records, and expert experience documents for air compressor equipment. Utilize Named Entity Recognition (NER) and Relation Recognition (RE) technologies to automatically construct a domain knowledge graph for air compressors. And store it in Neo4j. It is a set of nodes, covering entities of five body types: physical structure of air compressor, failure mode, failure characteristics, failure cause and maintenance suggestions; This is a set of relations, describing the logical relationships between the aforementioned entities; It is a collection of attributes that records the specific parameters and descriptions of an entity.

[0018] Preferably, in step S2, the method for semantic transformation using the trained data model is as follows:

[0019] S22-1: Training a fully connected neural network (MLP) classification model. Obtain a training set of known fault modes with expert-annotated text descriptions, and calculate the corresponding feature data set according to S1. Annotate the semantic features corresponding to the training set, and use the Meta-Learning Training Method (MAML) to input the training set into the fully connected neural network (MLP) classification model for training, obtaining a fine-tunable data-to-semantic model.

[0020] S22-2: Obtain the measurement point data of the air compressor under test, and calculate the feature dataset using S1. Use a trained fully connected neural network (MLP) classification model to transform the data into semantic features, obtaining the first textual feature description of the data under test. .

[0021] Preferably, in step S2, the method for calling the multimodal large model for semantic transformation is as follows:

[0022] S23-1: Invoke the multimodal large model and convert the image drawn in S1 into Base64 encoding as visual input. Combine the various feature data sets calculated by S1, guide the model to perform cross-modal semantic extraction through preset prompt word templates, and output a preliminary semantic representation of the image features. .

[0023] S23-2: Invoke the Large Language Model (LLM) to perform feature cross-validation between the preliminary text results from image recognition and the feature data set calculated in S1. If a significant contradiction occurs, a low confidence warning is triggered, and this situation is added as a new prompt word and merged into the preset prompt words. The LLM is then invoked again to complete the operation in S23-1. This process continues until no significant contradictions occur during feature cross-validation, at which point the final image semantic description (i.e., the second text feature description) is output. .

[0024] Preferably, in step S2, the method for aligning the text feature description with standard terminology is as follows:

[0025] S24-1: From graph databases The system dynamically extracts all existing "fault characteristic" entities, groups them according to node attributes, and constructs a standard terminology library. .

[0026] S24-2: Will and The input is fed into a large language model (LLM), using pre-defined prompt word templates, explicitly requiring the LLM to play an expert role and extract from a standard terminology database. Select with and Matched keywords, natural language description analysis, and final output. and and their respective standard terms.

[0027] Preferably, the method for hallucination detection and filtering in S2 is as follows:

[0028] S25: Using the Sentence-BERT embedding model to... and The standardized keyword mappings output by the LLM are high-dimensional vectors. By calculating the cosine similarity between the two mappings, if the similarity is below a preset threshold, the system automatically determines that the mapping is a phantom, thus rejecting the mapping, deleting the corresponding mapping group, and finally outputting a standardized keyword JSON list. and .

[0029] Preferably, S3 is as follows:

[0030] S3-1: Source Marking and Counting. Receives a set of text descriptions from numerical features. Record the number of text descriptions ; Receive a set of text descriptions from the image. Record the number of text descriptions .

[0031] S3-2: Calculation of source balance factor. Determine the effective number of modes. (If both channels have output, then) ). Calculate the uniform weights for numerical text descriptions: ; Calculate the uniform weights for image-text descriptions: .

[0032] Preferably, in step S5, the metacognitive reasoning chain is specifically as follows:

[0033] S5-1: Constructing Diagnostic Hints The set of feature parameters calculated by S1 and the KG triples with weight labels are then used. and Convert to structured text and add prompts.

[0034] S5-2: Metacognitive Chain-Based Reasoning. When the system invokes LLM, it loads a pre-defined prompt template to guide the model to reason according to five predefined metacognitive steps:

[0035] 1) Text parsing: Analyze the quantitative indicators of features and compare them with the matching relationships returned by the knowledge graph.

[0036] 2) Forming a preliminary judgment: Calculate the weighted matching score of each potential fault.

[0037] 3) Preliminary evaluation: Self-reflection to determine if there are any conflicts.

[0038] 4) Make the final decision: Determine the most likely rotor state and provide detailed judgment criteria.

[0039] 5) Assess confidence level: Based on factors such as feature matching degree and knowledge coverage, give a confidence score of 0-100%.

[0040] Secondly, this invention provides an air compressor fault diagnosis system that combines knowledge graphs and multimodal large models, including:

[0041] The signal processing and feature extraction module is used to receive the vibration signal from the target air compressor measuring point, calculate the numerical feature parameters after preprocessing, and draw the image corresponding to the vibration signal.

[0042] The cross-modal semantic mapping module is used to convert the set of numerical feature parameters obtained from the signal processing and feature extraction modules into a first text feature description using a trained data model. The multimodal large model is invoked to convert the image obtained from the signal processing and feature extraction modules into a second text feature description. The large language model is invoked to compare the two text feature descriptions with a pre-built knowledge graph of the air compressor domain. The standard terms in the text are aligned, and the vector similarity calculation method is used for illusion detection and filtering to generate a standardized set of keywords;

[0043] The adaptive feature fusion module is used to calculate weights for the two standardized keyword sets finally obtained by the cross-modal semantic mapping module using a source balance weighting method based on macro-average logic, and generate a fused text description set with weight labels.

[0044] The GraphRAG-based metacognitive diagnostic reasoning module is used to query the knowledge graph of the cross-modal semantic mapping module based on the fused text description set of the adaptive feature fusion module, and generate a list of weighted triples; it is also used to construct diagnostic prompt words that include the numerical feature parameters of the signal processing and feature extraction module and the list of weighted triples, guide the large language model to perform step-by-step reasoning according to the preset metacognitive reasoning chain, and output the most likely fault type and diagnostic confidence.

[0045] The knowledge-enhanced report generation module is used to query standard fault causes and maintenance suggestions from the knowledge graph described by the cross-modal semantic mapping module based on the fault type determined by the GraphRAG-based metacognitive diagnostic reasoning module; then, it calls the large language model again, combining the target air compressor features with the reasoning process of the GraphRAG-based metacognitive diagnostic reasoning module, to customize and rewrite the standard fault causes and maintenance suggestions, generating the final fault diagnosis report.

[0046] Thirdly, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can realize the intelligent fault diagnosis method for air compressors based on multimodal fusion and knowledge graph as described in any of the first aspects.

[0047] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the intelligent fault diagnosis method for air compressors based on multimodal fusion and knowledge graph as described in any of the first aspects.

[0048] Fifthly, the present invention provides a computer electronic device, including a memory and a processor; the memory is used to store a computer program; the processor is used to implement, when executing the computer program, the intelligent fault diagnosis method for air compressors based on multimodal fusion and knowledge graph as described in any of the first aspects.

[0049] Compared with the prior art, the present invention has the following advantages:

[0050] 1) More accurate and reliable diagnostic results: By using the dual path of processing numerical features through MLP model and processing images through multimodal large model, and combining the source balance weighting method based on macro average logic to fuse multimodal features, the problem of insufficient information in single modality data (only numerical or only image) under complex working conditions is effectively made up for.

[0051] 2) Significantly suppress large model hallucinations: By introducing knowledge graphs as underlying fact constraints, similarity verification and hallucination detection are performed during the semantic transformation stage, ensuring that the diagnostic opinions generated by LLM are supported by real physical expert knowledge.

[0052] 3) The reasoning process is interpretable: Given a knowledge graph, metacognitive reasoning chain technology is used to make the diagnostic process no longer a black box decision, but to output a reasoning path according to the logic of "observation-association-hypothesis-decision", which is convenient for on-site engineers to review.

[0053] 4) Strong knowledge loop and scalability: The system uses Neo4j to store expert experience and can easily expand fault types and maintenance strategies by updating CSV data files, achieving deep decoupling between diagnostic models and domain knowledge. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the overall process architecture of the method of the present invention.

[0055] Figure 2 This refers to the ontology layer structure of the knowledge graph constructed in the method of this invention.

[0056] Figure 3 This is an architecture diagram of an intelligent fault diagnosis system for air compressors based on knowledge graphs and multimodal large models, as described in the present invention. Detailed Implementation

[0057] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in various embodiments of the present invention can be combined accordingly without mutual conflict.

[0058] like Figure 1 As shown, the method includes: receiving air compressor data signals, performing data preprocessing, calculating data features and drawing images; converting numerical features and image features into semantic descriptions through a data model and a multimodal large model, respectively, and performing standardization alignment and fact verification based on a knowledge graph; using a source balance weighting method based on macro-average logic to fuse multimodal features; constructing a contextual prompting engineering based on GraphRAG, and constructing prompt words to guide the large language model to follow the metacognitive reasoning chain for fault diagnosis; and finally generating a customized fault report based on the diagnosis results by querying information from the knowledge graph.

[0059] The method of the present invention specifically includes the following steps:

[0060] S1, Signal Processing and Feature Extraction:

[0061] The vibration signal from the target air compressor measuring point is received, preprocessed, and then the numerical characteristic parameters are calculated. The corresponding image of the vibration signal is then plotted.

[0062] In practical use, the specific steps are as follows:

[0063] The system receives the target air compressor measurement point location (e.g., horizontal direction of the drive end bearing) input by the user and clearly informs the user that vibration signal data of the corresponding measurement point needs to be uploaded, along with the required sampling parameters (e.g., sampling frequency and duration). After receiving the raw vibration signal uploaded by the user, the system first performs data preprocessing, including denoising, detrending, and normalization. Subsequently, based on the preprocessed data, it calculates the set of time-domain characteristic parameters. (Including root mean square value, peak value, kurtosis, waveform factor, etc.) and the set of frequency domain characteristic parameters (This includes the amplitude of each harmonic and the energy percentage of characteristic frequencies obtained through Fast Fourier Transform). Simultaneously, the system automatically generates data images reflecting the equipment status, including time-domain waveforms. Frequency domain spectrum and axis trajectory diagram This provides input for subsequent visual semantic analysis.

[0064] S2, Cross-modal semantic mapping and knowledge alignment:

[0065] The numerical feature parameter set obtained from S1 is converted into a first text feature description using the trained data model. The multimodal large model is invoked to convert the image obtained in S1 into a second text feature description. The large language model is invoked to compare the two text feature descriptions with a pre-built knowledge graph of the air compressor domain. The standard terms in the text are aligned, and the vector similarity calculation method is used for illusion detection and filtering to generate a standardized set of keywords;

[0066] This step aims to transform the numerical values ​​and images obtained in S1 into standardized semantic descriptions that can interact with the knowledge graph. Its core innovation lies in introducing the knowledge graph as a standard library, and through a multi-stage alignment and filtering mechanism, effectively suppressing the "illusion" problem of general models in specialized fields.

[0067] In a preferred embodiment of the present invention, the domain knowledge graph construction method in this step is as follows:

[0068] S21, Domain Knowledge Graph Construction, such as Figure 2 As shown:

[0069] This process involves collecting and integrating heterogeneous text data from multiple sources related to air compressor equipment, including textbooks, maintenance manuals, historical maintenance work orders, and expert experience documents. Named Entity Recognition (NER) technology is used to extract entities from the text (e.g., "male / female rotor," "intake valve," "imbalance," "blockage"), and Relation Extraction (RE) technology is used to identify relationships between entities (e.g., "cause," "manifests as"). The extracted entities, relationships, and attributes are then structured and stored in the Neo4j graph database, forming a domain knowledge graph containing five types of ontologies: physical structure, failure modes, failure characteristics, failure causes, and maintenance suggestions. .

[0070] in, This is a collection of physical nodes, covering at least five categories of the air compressor: physical structure (such as "male and female rotors", "intake valve", "oil pump", "oil separator core", etc.), failure modes (such as "rotor misalignment", "intake valve leakage", "oil filter blockage", "oil separator core damage", etc.), failure characteristics (such as "obvious double frequency components", "time domain signal drops too quickly after unloading", "irregular shaft trajectory", etc.), failure causes (such as "excessive pipeline stress", "intake valve core or seat wear", "excessive lubricating oil replacement cycle", "excessive residual impurities in the oil circuit"), and maintenance recommendations (such as "check and eliminate stress in connecting pipelines", "check valve core and seat wear", "strictly replace oil filter on time", "monitor oil separator core pressure difference", etc.). This is a set of relationships that describes the logical connections between the entities mentioned above (such as "behaves as", "leads to", "solution is" etc.). It is a collection of attributes that records the specific parameters and descriptions of an entity.

[0071] In a preferred embodiment of the present invention, the semantic transformation method based on the data model in this step is specifically as follows:

[0072] S22-1, Model Training:

[0073] Obtain a labeled training dataset containing multiple sets of vibration signals under known fault conditions and their corresponding expert text descriptions (e.g., "first harmonic amplitude dominates," "amplitude increases with increasing rotational speed"). Calculate the set of feature parameters for these signals using the S1 method. Train a fully connected neural network (MLP) classification model using the Meta-Learning Approach (MAML), which learns the mapping from feature parameter vectors to textual semantic descriptions, resulting in a fine-tunable base model.

[0074] S22-2, Feature Transformation:

[0075] For the air compressor data to be diagnosed, the set of feature parameters calculated by S1 is input into the trained MLP model. The model outputs the first textual feature description of this set of feature data. .

[0076] In a preferred embodiment of the present invention, the semantic transformation method based on the multimodal large model in this step is specifically as follows:

[0077] S23-1, Preliminary identification:

[0078] Calling a multimodal large model (such as Zhipu GLM-4V) will generate the time-domain waveform plotted by S1. Frequency domain spectrum and axis trajectory diagram After the image is converted to Base64 encoding, it serves as visual input. Simultaneously, the feature data set calculated by S1 (e.g., "1st octave amplitude: 5.6 mm / s") is used as contextual information. The model is guided through analysis using preset prompt templates (e.g., "You are an air compressor fault diagnosis expert. Please analyze this vibration spectrum diagram and, based on the provided feature data, describe the main frequency components, noise level, etc. shown in the diagram..."), outputting a preliminary semantic description of the image features. .

[0079] S23-2, Cross-validation and Iterative Correction:

[0080] Image semantic description output by multimodal large model The image description (e.g., "1X frequency component dominates") is compared with the numerical features calculated by S1 (e.g., "1st harmonic amplitude accounts for 65%). If there is a significant contradiction (e.g., the image description is "1X frequency component dominates," but the amplitude of the 1X frequency in the numerical features is lower than the amplitude of the 2X frequency), a low-confidence warning is triggered. The system merges this contradictory information (including the image description, contradictory numerical features, and warning information) into the existing prompts as new prompts and calls the multimodal large model again for recognition. This process is iterated until there is no significant contradiction between the image description and the numerical features, and the final reliable image semantic description (i.e., the second text feature description) is output. .

[0081] In a preferred embodiment of the present invention, the standard alignment method of semantic features to the knowledge graph in this step is as follows:

[0082] S24-1, Constructing a standard terminology database:

[0083] From Neo4j knowledge graph Extract all "fault characteristic" class entities and their attributes to construct a standard terminology library. (For example: {"high noise level", "large axial vibration", "existence of sideband frequency"}).

[0084] S24-2, LLM Guide Alignment:

[0085] Describe the text features obtained above and image semantic description The input is fed into the Large Language Model (LLM). It uses preset prompt word templates (such as "You are an air compressor fault diagnosis expert, please refer to the standard terminology library"). The model is guided to perform analysis by selecting one or more standard keywords that best match the input semantic description, and outputs formatted alignment results.

[0086] In a preferred embodiment of the present invention, the hallucination detection and filtering method in this step is specifically as follows:

[0087] S25, Hallucination Detection and Filtering:

[0088] Using the Sentence-BERT text embedding model, the original text is embedded into a text. , The standardized keywords corresponding to the LLM output are mapped to high-dimensional vectors. The cosine similarity of each pair of "original description vector - standardized keyword vector" is calculated. If the similarity is below a preset threshold (e.g., 0.5), the alignment is considered to be "illusion," and the system automatically discards the mapping result to ensure that the final output of the standardized keyword list is based on reliable semantic associations.

[0089] S3, Multimodal fusion based on source equilibrium weighting method of macro-average logic:

[0090] A source balance weighting method based on macro-average logic is adopted to calculate the weights for the two sets of standardized keywords obtained in S2, and generate a fused text description set with weight labels.

[0091] In this step, the system does not simply concatenate the results of the two modalities. Instead, it employs a source balancing strategy to eliminate quantitative bias, effectively addressing the problem of quantitative imbalance bias and ensuring equal power between the textual descriptions of numerical features and the textual descriptions of image modalities at the logical reasoning level. The system calculates a standardized weight label for the semantic descriptions from both numerical and image features. This method follows the macro-averaging principle: that is, in the final fusion decision, each information source (modality) has an equal status, and the sum of its contribution weights remains constant in the system, unaffected by fluctuations in the quantity of textual descriptions produced by that source. The weight calculation follows... ,in As weight, For the number of information sources, This represents the number of text descriptions for that information source. Finally, the system outputs a normalized list of text descriptions with weighted score labels.

[0092] S4, GraphRAG-based context-hint engineering build:

[0093] Based on the fused text description set obtained from S3, query the knowledge graph obtained from S2 to generate a list of weighted triples.

[0094] In a preferred embodiment of the present invention, the steps are as follows:

[0095] Execute Cypher query and Within the knowledge graph, all relevant nodes and relationships are grouped and encapsulated into structured triples "fault-possession-feature," generating the final weighted KG triples. and Construct diagnostic clues based on knowledge graphs.

[0096] S5, Metacognitive Diagnostic Reasoning:

[0097] Construct diagnostic prompt words that include numerical feature parameters obtained in S1 and a list of weighted triples obtained in S4. Guide the large language model to follow the preset metacognitive reasoning chain, gradually analyze features, form and evaluate preliminary judgments, make final decisions and evaluate diagnostic confidence, and output the most likely fault type and diagnostic confidence.

[0098] In a preferred embodiment of the present invention, the multimodal data feature alignment method in this step is specifically as follows:

[0099] S51, Diagnostic prompt word construction:

[0100] Build diagnostic prompts It adopts a structured format and contains at least the following parts:

[0101] System role and task definition: The LLM is clearly required to act as a senior air compressor fault diagnosis expert.

[0102] Observational facts: Clearly list the key values ​​obtained from S1.

[0103] Knowledge Evidence Area: Lists the weighted KG triples obtained from S4.

[0104] Reasoning rules and constraints: It emphasizes that reasoning must be based on provided objective facts and knowledge evidence, and prohibits fabricating information.

[0105] Metacognitive reasoning chain instruction: This mandates that LLM learners follow these five steps to think and output the intermediate processes:

[0106] Step 1 (Text and Numerical Analysis): Analyze the numerical feature parameters obtained from S1 and the list of weighted triples obtained from S4 in the "Observation Facts Area" one by one.

[0107] Step 2 (Forming a Preliminary Judgment): Based on the analysis in Step 1, list all possible failure hypotheses. Calculate a preliminary weighted matching score for each hypothesis based on the tightness of feature matching and the weight of evidence.

[0108] Step 3 (Preliminary Evaluation): Examine whether there are contradictions or competing relationships between different hypotheses. Assess which features are shared by all hypotheses and which are the key differentiating features.

[0109] Step 4 (Make the final decision): Based on the analysis in Step 2 and Step 3, determine the one or two most likely fault types and provide detailed judgment criteria that can be traced back to the "observation fact area" and the "knowledge evidence area".

[0110] Step 5 (Assess confidence level): Based on feature coverage, consistency of evidence weights, and the presence of ambiguous information, give a confidence score of 0-100%.

[0111] S52, Model Calling and Parsing:

[0112] Calling LLM (such as GPT-4, GLM-4) for processing The system analyzes the output and extracts the final fault type, judgment criteria, and confidence level.

[0113] S6, Knowledge Enhancement Report Generation:

[0114] Based on the fault type determined in S5, standard fault causes and maintenance suggestions are queried from the knowledge graph described in S2; the large language model is invoked again, and the standard fault causes and maintenance suggestions are customized and rewritten by combining the characteristics of the target air compressor (such as compression stages, compression medium, compressor operating mode, purpose, etc., which can be selected as needed) and the reasoning process in S5, to generate the final fault diagnosis report.

[0115] In a preferred embodiment of the present invention, the steps are as follows:

[0116] Based on the fault type determined by S5, the system automatically sends a Cypher query statement (e.g., "MATCH (cause:fault cause)-[:`hasReason`]->(state:rotor state{name:'rotor imbalance'})RETURN collect(cause.name) AS standard_causes) to the Neo4j knowledge graph to retrieve the corresponding standard fault cause and standard maintenance recommendations. Subsequently, LLM is invoked again to integrate the detailed reasoning process from S5, the specific characteristic parameters of the current equipment, and the standard knowledge retrieved from the knowledge graph. LLM's task is to logically optimize and context-customize the standard knowledge, generating a final diagnostic report (including diagnostic conclusions, reasoning process, cause analysis, and maintenance recommendations) for field engineers, containing in-depth fault cause analysis and specific, actionable maintenance recommendations.

[0117] In addition, this invention also provides an air compressor fault diagnosis system that combines knowledge graphs and multimodal large models, such as... Figure 3 As shown, the system mainly includes:

[0118] The signal processing and feature extraction module is used to receive the vibration signal from the target air compressor measuring point, calculate the numerical feature parameters after preprocessing, and draw the image corresponding to the vibration signal.

[0119] The cross-modal semantic mapping module is used to convert the set of numerical feature parameters obtained from the signal processing and feature extraction modules into a first text feature description using a trained data model. The multimodal large model is invoked to convert the image obtained from the signal processing and feature extraction modules into a second text feature description. The large language model is invoked to compare the two text feature descriptions with a pre-built knowledge graph of the air compressor domain. The standard terms in the text are aligned, and the vector similarity calculation method is used for illusion detection and filtering to generate a standardized set of keywords;

[0120] The adaptive feature fusion module is used to calculate weights for the two standardized keyword sets finally obtained by the cross-modal semantic mapping module using a source balance weighting method based on macro-average logic, and generate a fused text description set with weight labels.

[0121] The GraphRAG-based metacognitive diagnostic reasoning module is used to query the knowledge graph of the cross-modal semantic mapping module based on the fused text description set of the adaptive feature fusion module, and generate a list of weighted triples; it is also used to construct diagnostic prompt words that include the numerical feature parameters of the signal processing and feature extraction module and the list of weighted triples, guide the large language model to perform step-by-step reasoning according to the preset metacognitive reasoning chain, and output the most likely fault type and diagnostic confidence.

[0122] The knowledge-enhanced report generation module is used to query standard fault causes and maintenance suggestions from the knowledge graph described by the cross-modal semantic mapping module based on the fault type determined by the GraphRAG-based metacognitive diagnostic reasoning module; then, it calls the large language model again, combining the target air compressor features with the reasoning process of the GraphRAG-based metacognitive diagnostic reasoning module, to customize and rewrite the standard fault causes and maintenance suggestions, generating the final fault diagnosis report.

[0123] In addition, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the above-mentioned intelligent fault diagnosis method for air compressors based on multimodal fusion and knowledge graph.

[0124] In addition, the present invention provides a computer-readable storage medium, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, it can realize the above-mentioned intelligent fault diagnosis method for air compressors based on multimodal fusion and knowledge graph.

[0125] In addition, the present invention provides a computer electronic device, including a memory and a processor; wherein, the memory is used to store a computer program; and the processor is used to implement the above-mentioned intelligent fault diagnosis method for air compressors based on multimodal fusion and knowledge graph when the computer program is executed.

[0126] The methods and effects of the present invention will be specifically illustrated below through examples.

[0127] Example

[0128] This embodiment provides a method for diagnosing air compressor faults by combining knowledge graphs and multimodal large models. Specifically, it takes the rotor misalignment fault diagnosis of an air compressor VT3004 in a factory as an example to demonstrate the complete closed-loop process from the original signal to the intelligent diagnostic report.

[0129] Step 1: Signal Processing and Feature Extraction (S1)

[0130] By selecting the measurement point "Non-coupling end bearing measurement point" through the system interface, the system automatically prompts the user to upload the radial (X, Y direction) vibration signal of the corresponding measurement point. Upload the radial (X, Y direction) vibration velocity signal of the non-coupling end bearing measurement point of the VT3004 air compressor, with an input sampling frequency of 25kHz and a speed of 1238 rpm. After automatic signal preprocessing, the system performs Fast Fourier Transform (FFT) analysis and automatically plots the time-domain waveform, spectrum, and shaft center trajectory. Simultaneously, the following key characteristics are calculated:

[0131] Time-domain feature set Vertically: The maximum amplitude of the time-domain signal is 0.250408, the minimum is -0.20199, the peak-to-peak value is 0.452404, the kurtosis is 3.181921, and the skewness is -0.021412, etc. Horizontally: The maximum amplitude of the time-domain signal is 0.252009, the minimum is -0.196878, the peak-to-peak value is 0.448887, the kurtosis is 3.182535, and the skewness is -0.032434, etc.

[0132] Frequency domain feature set Vertical direction: Maximum amplitude is 0.014273, total spectral energy is 5.594073, spectral width is 12499 Hz, amplitude of harmonics and sub-harmonics, signal-to-noise ratio is -39.81 dB, etc. Horizontal direction: Maximum amplitude is 0.012401, total spectral energy is 5.442578, spectral width is 12499 Hz, amplitude of harmonics and sub-harmonics, signal-to-noise ratio is -39.81 dB, etc.

[0133] Channel characteristics. Time-domain amplitude ratio: Peak amplitude ratio is 0.993648, RMS amplitude ratio is 1.027688, etc. Frequency-domain amplitude ratio: Peak amplitude ratio is 1.150924, total energy ratio is 1.027835, etc. Harmonic amplitude ratio: 0.5 frequency division amplitude ratio is 1.435973, 2nd harmonic amplitude ratio is 5.490255, etc.

[0134] Step 2: Cross-modal semantic mapping (S2)

[0135] Numerical to Semantic: Inputting a set of data features into a trained MLP model for classification, and outputting text descriptions. {Horizontal direction: "Multiple harmonics appear", "Predominantly second harmonic", "Phase difference of radial vibration is close to 180 degrees"...; Vertical direction: ...}

[0136] Image-to-text conversion: Input three drawn images into the multimodal large model GLM-4V with preset prompts, and output text descriptions. {Horizontal direction: "Exhibits good periodicity", "1st harmonic component is dominant", "Axial trajectory is banana-shaped"...; Vertical direction: ...}

[0137] Cross-validation and iterative correction: Calling a large language model (LLM) with pre-set prompt words will... Calculated with S1 , Numerical features undergo logical cross-validation. The system detects a significant contradiction between the text description and the data (the image description states "1X frequency component dominates," but the numerical features show "horizontal amplitude is 0.000517, and second harmonic amplitude is 0.000996"), triggering a low-confidence warning. The system merges the contradictory information and adds it to the original prompts, then re-invokes the multimodal large model for recognition. This process is repeated until there are no significant contradictions between the image description and the numerical features, at which point a final, reliable image semantic description is output. {Horizontal direction: "Exhibits good periodicity", "2nd harmonic component is dominant", "Axial trajectory is banana-shaped"... Vertical direction: ...}

[0138] Building a Standard Terminology Base: From the Neo4j Knowledge Graph Extract all "fault characteristic" class entities and their attributes to construct a standard terminology library. .

[0139] LLM boot alignment: LLM from Select with and Standardized descriptions with the same semantics output formatted alignment results (e.g., "2nd harmonic component dominates" → "2nd harmonic is dominant").

[0140] Hallucination Detection and Filtering: Using the Sentence-BERT text embedding model, each pair of "original description-standard keywords" is mapped to a high-dimensional vector. The cosine similarity of each pair of vectors is calculated, and if it is less than 0.5, the mapping result is automatically discarded. Finally, a list of filtered standardized keywords is output. and .

[0141] Step 3: Multimodal fusion based on source balance weighting method of macro average logic (S3)

[0142] Assumption There are 3 text descriptions. There are 5 text descriptions. Use Perform weight calculation and obtain the result. The weight of each text description is , Each text description has a weight of 0.1.

[0143] Calculated After being directly attached as an attribute to all text descriptions of the corresponding source, all text descriptions are aggregated into a unified list.

[0144] Step 4: GraphRAG-based contextual hint project building (S4)

[0145] Execute Cypher query and Within the knowledge graph, all relevant nodes and relationships are grouped and encapsulated into structured triples to generate the final KG triple description. and The corresponding weights calculated in step 3 are attached. All triples are summarized into a unified list to construct the contextual cue word project. The cue words passed to the LLM explicitly state: "The sum of the modal weights is equal. Please ignore the difference in the number of entries and deduce the logic based on the individual weights."

[0146] Step 5: Metacognitive Diagnostic Reasoning (S5)

[0147] The system constructs prompts containing all the above information, guiding the LLM to execute a metacognitive reasoning chain for fault diagnosis. The final judgment is: "Rotor misalignment" has a confidence level of 92%; "Mechanical looseness" has a confidence level of 60%. "Rotor misalignment" is selected as the final diagnostic result.

[0148] Step 6: Generate Knowledge Enhancement Report (S6)

[0149] The system queries the knowledge graph to obtain the standard causes and recommendations for "rotor misalignment". LLM combines the data characteristics and reasoning process in this case to determine the most likely cause and the most reasonable recommendation, generating a complete report.

[0150] Through the above process, this invention achieves a complete closed loop from multimodal data input to interpretable, knowledge-enhanced intelligent diagnostic report output.

[0151] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for diagnosing air compressor faults by combining knowledge graphs and multimodal large models, characterized in that, Specifically as follows: S1: Receive the vibration signal from the target air compressor measuring point, preprocess it, calculate the numerical characteristic parameters, and draw the corresponding image of the vibration signal; S2: Use the trained data model to convert the set of numerical feature parameters obtained in S1 into a first text feature description. The multimodal large model is invoked to convert the image obtained in S1 into a second text feature description. ; The large language model is invoked to compare the two text feature descriptions with a pre-built knowledge graph of the air compressor domain. The standard terms in the text are aligned, and the vector similarity calculation method is used for illusion detection and filtering to generate a standardized set of keywords; S3: Using a source balance weighting method based on macro-average logic, weights are calculated for the two sets of standardized keywords obtained in S2, and a fused text description set with weight labels is generated. S4: Based on the fused text description set described in S3, query the knowledge graph described in S2 to generate a list of weighted triples; S5: Construct diagnostic prompt words containing the numerical feature parameters described in S1 and the weighted triple list described in S4, guide the large language model to perform step-by-step reasoning according to the preset metacognitive reasoning chain, and output the most likely fault type and diagnostic confidence. S6: Based on the fault type determined in S5, query the standard fault causes and maintenance suggestions from the knowledge graph described in S2; call the large language model again, combine the characteristics of the target air compressor with the reasoning process in S5, and rewrite the standard fault causes and maintenance suggestions in a customized manner to generate the final fault diagnosis report. In S2, the method for calling the multimodal large model for semantic transformation is as follows: The image obtained in S1 is encoded and input into a multimodal large model. Combined with the set of numerical feature parameters obtained in S1, a draft of the second text feature description is generated by prompting words. The large language model is invoked to cross-validate the initial draft with the numerical feature parameter set obtained in S1. If a contradiction exists, a warning is triggered and the process is iteratively corrected until a second text feature description that matches the initial draft is output. ; S3 is specifically as follows: Let the number of valid information sources be... The number of text descriptions after fusion from numerical features is The number of text descriptions after fusion of image features is The unified weight of the fused text description formed by the numerical features is then... The unified weight of the text description after the fusion of image features is: ; The calculated weights are appended as attributes to all text descriptions of the corresponding source.

2. The air compressor fault diagnosis method combining knowledge graph and multimodal large model according to claim 1, characterized in that, In S2, the knowledge graph of the air compressor field The specific construction method is as follows: Collect multi-source data in the field of air compressors, and use named entity recognition and relation extraction technology to construct and store a knowledge graph containing five types of entities: physical structure, fault mode, fault characteristics, fault causes and maintenance suggestions.

3. The air compressor fault diagnosis method combining knowledge graph and multimodal large model according to claim 1, characterized in that, In S2, the method for semantic transformation using the trained data model is as follows: A fully connected neural network classification model is trained using a meta-learning method to learn the mapping from numerical feature parameter vectors to text feature descriptions. Then, the set of numerical feature parameters obtained in S1 is input into the trained fully connected neural network classification model to obtain the first text feature description. .

4. The air compressor fault diagnosis method combining knowledge graph and multimodal large model according to claim 1, characterized in that, In step S2, the method for aligning text feature descriptions with standard terminology is as follows: From a pre-built knowledge graph of the air compressor field Extract all fault feature entities to construct a standard terminology library; analyze the first text feature description using a large language model. With the second text feature description Select matching standard keywords from the standard terminology library.

5. The air compressor fault diagnosis method combining knowledge graph and multimodal large model according to claim 1, characterized in that, In S2, the method for hallucination detection and filtering is as follows: The text embedding model is used to map the first text feature description and the second text feature description to the standard keywords obtained after alignment, respectively, into vectors, and the cosine similarity is calculated; if the similarity is lower than the preset threshold, it is judged as a hallucination and the mapping result is filtered.

6. The air compressor fault diagnosis method combining knowledge graph and multimodal large model according to claim 1, characterized in that, In S5, the metacognitive reasoning chain is specifically as follows: Text parsing, analyzing feature parameters and linking them to knowledge graph evidence; Form a preliminary judgment, list the fault hypotheses and calculate the weighted matching score; The initial assessment is evaluated, and contradictions and key distinguishing features between hypotheses are examined. Make the final decision and determine the most likely type of failure and the basis for it; Assess confidence levels and assign a diagnostic confidence score.

7. A fault diagnosis system for air compressors that combines knowledge graphs and multimodal large models, characterized in that, include: The signal processing and feature extraction module is used to receive the vibration signal from the target air compressor measuring point, calculate the numerical feature parameters after preprocessing, and draw the image corresponding to the vibration signal. The cross-modal semantic mapping module is used to convert the set of numerical feature parameters obtained from the signal processing and feature extraction modules into a first text feature description using a trained data model. The multimodal large model is invoked to convert the image obtained from the signal processing and feature extraction modules into a second text feature description. ; The large language model is invoked to compare the two text feature descriptions with a pre-built knowledge graph of the air compressor domain. The standard terms in the text are aligned, and a vector similarity calculation method is used for illusion detection and filtering to generate a standardized keyword set; the method of calling a multimodal large model for semantic transformation is as follows: The image obtained in S1 is encoded and input into a multimodal large model. Combined with the set of numerical feature parameters obtained in S1, a draft of the second text feature description is generated by prompting words. The large language model is invoked to cross-validate the initial draft with the numerical feature parameter set obtained in S1. If a contradiction exists, a warning is triggered and the process is iteratively corrected until a second text feature description that matches the initial draft is output. ; The adaptive feature fusion module uses a source-balanced weighting method based on macro-average logic to calculate weights for the two standardized keyword sets obtained from the cross-modal semantic mapping module, generating a fused text description set with weighted labels; specifically as follows: Let the number of valid information sources be... The number of text descriptions after fusion from numerical features is The number of text descriptions after fusion of image features is The unified weight of the fused text description formed by the numerical features is then... The unified weight of the text description after the fusion of image features is: ; The calculated weights are then appended as attributes to all text descriptions from the corresponding source. The GraphRAG-based metacognitive diagnostic reasoning module is used to query the knowledge graph of the cross-modal semantic mapping module based on the fused text description set of the adaptive feature fusion module, and generate a list of weighted triples; it is also used to construct diagnostic prompt words that include the numerical feature parameters of the signal processing and feature extraction module and the list of weighted triples, guide the large language model to perform step-by-step reasoning according to the preset metacognitive reasoning chain, and output the most likely fault type and diagnostic confidence. The knowledge-enhanced report generation module is used to query standard fault causes and maintenance suggestions from the knowledge graph described by the cross-modal semantic mapping module based on the fault type determined by the GraphRAG-based metacognitive diagnostic reasoning module; then, it calls the large language model again, combining the target air compressor features with the reasoning process of the GraphRAG-based metacognitive diagnostic reasoning module, to customize and rewrite the standard fault causes and maintenance suggestions, generating the final fault diagnosis report.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can realize the air compressor fault diagnosis method that combines knowledge graph and multimodal large model as described in any one of claims 1 to 6.

Images