Dynamic equipment fault progressive diagnosis method based on knowledge enhancement large model

Abstract

The invention discloses a dynamic equipment fault progressive diagnosis method based on a knowledge enhancement large model, and relates to the technical field of equipment intelligent fault diagnosis and operation and maintenance, and the diagnosis method comprises the following specific steps: S100, building a domain knowledge graph: based on historical fault data and equipment manual information, carrying out data arrangement and labeling, and carrying out mapping on the domain knowledge graph; the method comprises the following steps: constructing a knowledge graph which takes equipment key parts, sensor measuring points, typical fault modes, fault symptoms, reason mechanisms and maintenance measures as nodes, associates the parts with the fault modes and takes a causal relationship between the fault symptoms and potential reasons as edges, and defining knowledge expression specifications; according to the method, the knowledge graph and the large-scale pre-training language model are fused, an intelligent diagnosis engine with equipment domain knowledge is constructed, the accuracy of equipment fault diagnosis is remarkably improved, the introduction of the knowledge graph enables the diagnosis process to refer to rich domain knowledge and historical experience, and the diagnosis efficiency is improved. And the defects of a pure data driving model in the face of rare or complex faults are made up.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent fault diagnosis and operation and maintenance of equipment, in particular to a dynamic equipment fault progressive diagnosis method based on a knowledge-enhanced large model. BACKGROUND

[0002] With the continuous improvement of industrial automation and intelligence, the operation and maintenance of various complex equipment has become a key link to ensure production efficiency and safety. Equipment failure not only causes production interruption, but also may cause safety accidents and huge economic losses. Therefore, developing efficient and accurate equipment fault diagnosis technology is of great significance to improve the reliability and efficiency of industrial production. In recent years, with the rapid development of artificial intelligence technology, especially the wide application of deep learning technology, new ideas and solutions have been provided for equipment fault diagnosis. However, traditional equipment fault diagnosis methods still have many limitations in the face of complex and variable equipment operating environment and diversified fault modes, and it is difficult to meet the needs of actual production.

[0003] Traditional equipment fault diagnosis methods mainly include expert system-based fault diagnosis methods and deep learning-based data-driven diagnosis methods. Although the expert system-based fault diagnosis method has certain field specificity, the diagnosis rules are fixed and difficult to maintain. Once new or complex fault modes occur in the equipment, the diagnosis accuracy will be significantly reduced. At the same time, the maintenance and update of expert knowledge require a lot of manual intervention, which leads to the inability to timely expand and update the system knowledge. The deep learning-based data-driven diagnosis method trains a large amount of historical data to achieve fault detection and diagnosis through pattern recognition, but it is strongly dependent on data size, quality and type. Once there is a lack of sufficient training data for a specific fault, the diagnosis performance will decrease significantly. In addition, deep learning models generally have black box characteristics and lack explainability, making it difficult to clearly give the fault diagnosis reasoning process and judgment basis, reducing the credibility in actual engineering application. At the same time, most of these models are static models, which are difficult to perform further information feedback and progressive analysis based on preliminary diagnosis results, especially in complex and multi-cause fault scenarios.

[0004] Therefore, developing a dynamic equipment fault progressive diagnosis method based on a knowledge-enhanced large model not only improves the accuracy and efficiency of equipment fault diagnosis, but also enhances the explainability and dynamic adaptability of the diagnosis process, providing more reliable and efficient technical support for industrial production and equipment operation and maintenance. SUMMARY

[0005] The purpose of the present application is to make up for the shortcomings of the prior art, and provide a dynamic equipment fault progressive diagnosis method based on a knowledge enhanced large model, which comprises the following steps:

[0006] The present application provides the following technical solutions to solve the above technical problems: a dynamic equipment fault progressive diagnosis method based on a knowledge enhanced large model, which comprises the following steps:

[0007] S100, constructing a field knowledge graph: based on historical fault data and equipment manual information, after data arrangement and labeling, a knowledge graph is constructed, which comprises nodes of equipment key components, sensor measuring points, typical fault modes, fault symptoms, cause mechanisms and maintenance measures, and edges of component-fault mode association and fault symptom-potential cause causal relationship, and a knowledge expression specification is defined;

[0008] S200, constructing a knowledge enhanced large model: based on a large pre-trained language model, the knowledge graph content is integrated into the large model to construct a knowledge enhanced large model, the model is fine-tuned and trained by using fault diagnosis field question-answer pairs and accident case data, and is optimized by contrast learning to construct the knowledge enhanced large model;

[0009] S300, real-time state monitoring and feature extraction: the target equipment is monitored in real time by a plurality of source sensors, the running condition data is collected and preprocessed, and the state features of the equipment at the current time are extracted by a multi-scale feature fusion algorithm;

[0010] S400, progressive multi-round diagnosis reasoning: additional information is obtained by the large model according to the requirements proposed by the preliminary diagnosis, the additional information and the equipment state feature description are input into the knowledge enhanced large model, fault reasoning is performed, the confidence is updated by an iterative formula, whether the diagnosis termination condition is met is judged according to the confidence entropy and the maximum confidence, if not, additional information is obtained to continue reasoning, until a clear conclusion is obtained;

[0011] S500, diagnosis result output and knowledge graph update: when the model confidence reaches the threshold or is narrowed to a single fault cause, output the diagnosis result containing fault positioning, type, cause analysis and maintenance suggestion and disposal suggestion, and feed back the new data and conclusion in the diagnosis process to the knowledge base to update the knowledge graph relationship weight through the dynamic knowledge update factor.

[0012] Further, in the S100 of constructing the field knowledge graph, the large pre-training language model refers to LLaMA pre-trained by large-scale text data, which has natural language understanding and generation capabilities, can be used as a basis, and can be integrated with knowledge graph content and fine-tuned with fault diagnosis field data to adapt to the language model of the equipment fault diagnosis task.

[0013] Further, in the S100 of constructing the field knowledge graph, the specific steps of constructing the knowledge graph are:

[0014] Collect historical fault data and equipment manual information of the target equipment, clean, classify and arrange the data, and remove redundant or incorrect information;

[0015] Extract key entities from the preprocessed data and define them as knowledge graph nodes, including: equipment key components, sensor measuring points, typical fault modes, fault symptoms, cause mechanisms and maintenance measures;

[0016] Determine the association relationship between entities and define it as the edge of the knowledge graph, including: the association relationship between components and fault modes and the causal relationship between fault symptoms and potential causes;

[0017] Based on the co-occurrence number of entities in historical fault data, the annotated causal confidence and timeliness, the relationship weight is calculated by the knowledge graph entity relationship weight calculation formula;

[0018] Define knowledge in a standardized semantic format, clearly define entity attributes and relationship weights, and form a knowledge expression specification that can be referenced by large models;

[0019] Graph storage and management: store the constructed knowledge graph in Neo4j, and establish indexes for nodes and edges.

[0020] Further, in the S100 of constructing the field knowledge graph, the relationship weight is calculated by the knowledge graph entity relationship weight calculation formula, and the set of entity nodes in the knowledge graph is defined as: The calculation formula of the edge relationship weight between and is: wherein, is the edge relationship weight between the entity nodes and , and It is an entity node in the knowledge graph, representing key components of equipment, sensor measurement points, and typical failure modes. and is the index identifier, It is the loop variable in the sum operation, count For historical failures and The number of co-occurrences of conf is the causal confidence level annotated by the experts, is the dynamic decay factor, the total number of entity nodes in the knowledge graph, and the formula is: ,age The time between the last update of the relationship and the current time. is the attenuation coefficient.

[0021] Furthermore, in the step S200, the knowledge enhancement model is constructed by the attention mechanism function Attn( ) The knowledge graph entity With the large model hidden state For cross-modal fusion, the formula is: ,in, Big model at the moment The hidden state vector of Updated hidden state vector after integrating knowledge graph entities, is an entity node in the knowledge graph, is the index identifier, is the knowledge weight coefficient, Attn( ) is the attention mechanism function, calculating and The association weight of Multilayer perceptron is used to perform nonlinear transformation on the fused features and optimize the model through contrastive learning. The formula is: ,in, is the loss value of contrastive learning, The hidden state of the large model after integrating knowledge, Is the knowledge graph except Other entities, is the cosine similarity, is the temperature parameter, is the real fault entity, Indicates the moment, is an index variable.

[0022] Further, the S300, the sensor types used in real-time state monitoring and feature extraction are: vibration sensor, using speed sensor to monitor the vibration signal of the equipment during operation; temperature sensor: using infrared temperature sensor to measure the temperature of the equipment components; pressure sensor: using absolute pressure sensor to collect pressure data in the equipment or pipeline; current / voltage sensor: using current transformer to monitor the working current and voltage of electrical components; flow sensor: using turbine flowmeter to collect state data of lubrication system and cooling system.

[0023] Further, the S300, the state features of the equipment at the current time are extracted by a multi-scale feature fusion algorithm in real-time state monitoring and feature extraction. The multi-scale feature extraction is performed on the time series data collected by the sensor at time t, wherein, is the time series data collected by the sensor at time t, represents the time t, represents feature splicing, WT( ) is wavelet transform, ( ) is self-attention mechanism, ( ) is convolutional neural network, used to extract local features of signals, and the feature fusion is performed through a learnable weight matrix The formula is: , is the original feature vector after fusion, is the final feature vector after linear transformation, is a learnable weight matrix for feature transformation, is a bias term to adjust the feature offset.

[0024] Further, the S400, the update of confidence in progressive multi-round diagnosis reasoning, the hypothesis set of the mth round of diagnosis is , and the confidence vector of the mth round of diagnosis is , represents the total number of fault hypotheses in the hypothesis set of the mth round of diagnosis, and is updated by an iterative formula, the formula is: wherein, is a diagnosis weight matrix for adjusting the direction of confidence update, represents element-wise multiplication, is the hypothesis associated knowledge graph subgraph, is an attention mechanism for calculating the relevance of state features and knowledge graph subgraph , ​is the original feature vector after fusion, is a subgraph of the knowledge graph, is the first hypothesized fault in the first round of diagnosis, is the confidence offset introduced by the new information, is the iteration coefficient, is the activation function that converts numerical values into probability distributions, represents the iteration round of diagnosis, is the index identifier.

[0025] Further, the S400, the determination of the diagnosis termination condition in the progressive multi-round diagnosis reasoning, sets the confidence entropy as , measures the uncertainty of the confidence distribution, when or , terminate the iteration, otherwise obtain new data to generate continue reasoning, wherein is the index identifier, represents the iteration round of diagnosis, is the first hypothesized confidence in the first round of diagnosis, is the number of fault categories in the hypothesis set, is the confidence entropy threshold, when the confidence entropy is less than or equal to , it is considered that the diagnosis result is determined enough, is the maximum confidence value in the first round of diagnosis, is the maximum confidence threshold, when the maximum confidence is greater than or equal to , it is considered that the credibility of a single fault hypothesis is high enough.

[0026] Further, the S500, the diagnosis result output and the knowledge graph update automatically update the knowledge graph relationship weight through the knowledge graph dynamic update algorithm in the knowledge graph update, design a dynamic knowledge update factor , the calculation formula is: , wherein, is the relationship weight between the entity and before updating, is the relationship weight increment caused by the new case, is the relationship weight between the entity and after updating, is the dynamic knowledge update factor, which determines the influence degree of the new case on the weight, and the calculation formula is: , represents the time, and novelty For case novelty score, calculate the KL divergence between case features and historical data distribution: novelty , is the equipment state feature vector of the current case, is the newly diagnosed case, is the adjustment parameter, which controls the sensitivity of the case novelty to the update factor.

[0027] Compared with the prior art, the dynamic equipment fault progressive diagnosis method based on the knowledge enhanced large model has the following beneficial effects:

[0028] Firstly, the intelligent diagnosis engine with equipment domain knowledge is constructed by fusing the knowledge graph and the large-scale pre-training language model, which significantly improves the accuracy of equipment fault diagnosis. The introduction of the knowledge graph enables the diagnosis process to refer to rich domain knowledge and historical experience, making up for the shortcomings of pure data-driven models in the face of rare or complex faults. At the same time, the powerful reasoning ability of the large model combined with the logical relationship network of the knowledge graph realizes deep causal reasoning and hypothesis testing, ensuring the intelligence and precision of the diagnosis results. This innovative way of knowledge and data fusion not only improves the accuracy of diagnosis, but also enhances the self-adaptability and intelligence level of the system.

[0029] Secondly, by introducing the knowledge graph and citing knowledge basis in the diagnosis report in detail, the invention can give clear and transparent diagnosis conclusions and reasoning processes, enhancing the user's trust in the diagnosis results. In addition, real-time data stream analysis and diagnosis are supported, which can respond in a timely manner when the equipment state changes, realizing online monitoring of dynamic running equipment. At the same time, through the updating mechanism of knowledge and model, the invention can continuously learn new knowledge and optimize the diagnosis ability, ensuring the adaptability to new fault patterns. This dynamic adaptability and continuous optimization capability make the invention have stronger application value and competitiveness in complex industrial environments.

[0030] Other advantages, objects and features of the present invention will be in part apparent and in part pointed out below in the specification, as some of them will be observed by those versed in the art upon examination of the text below, or will be learned from the practice of the invention. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any inventive labor.

[0032] Figure 1 A flowchart of a dynamic equipment fault progressive diagnosis method based on a knowledge-enhanced large model;

[0033] Figure 2 A framework diagram of a dynamic equipment fault progressive diagnosis method based on a knowledge-enhanced large model. DETAILED DESCRIPTION

[0034] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined inventive purposes, the specific embodiments, structures, features and effects thereof according to the present application are described in detail below in combination with the drawings and preferred embodiments.

[0035] Embodiment one:

[0036] A wind power equipment gearbox fault diagnosis scene embodiment.

[0037] In the operation process of a 2MW wind turbine generator of a certain wind farm, the gearbox as the core transmission component is long-term under complex working conditions of high speed and heavy load, and is extremely prone to gear meshing wear and bearing lubrication problems. Once the gearbox is abnormal, it will not only cause the unit to shut down, but also may cause a chain damage, causing huge economic losses. In order to realize early detection and accurate diagnosis of faults, the dynamic equipment fault progressive diagnosis method based on the knowledge-enhanced large model of the present application is applied to real-time monitoring and intelligent analysis of the running state of the gearbox.

[0038] S100, constructing a field knowledge graph: collecting historical fault reports, maintenance records and equipment manuals of the model wind power gearbox in the past 5 years, which contain gear wear, bearing failure and lubricating oil leakage fault cases, as shown in Figure 2 After cleaning and arranging the data, eliminating repeated or invalid information, classifying and archiving according to fault types, symptoms, causes and maintenance measures, extracting high-speed shaft, planetary gear, bearing components, vibration abnormalities, temperature sudden rise, noise increase symptoms, and corresponding fault modes, cause mechanisms and maintenance strategies as knowledge graph nodes, through analyzing the accompanying appearance frequency of each entity in the historical data, the degree of causal association annotated by experts and time influence factors, using the knowledge graph entity relationship weight calculation formula to determine the association strength between nodes, setting the entity node set in the knowledge graph as: , the edge relationship weight The calculation formula is: , wherein is the edge relationship weight between the entity nodes and in the knowledge graph, , and are entity nodes in the knowledge graph, representing equipment key components, sensor measuring points, typical fault modes, and is the index identifier, is the loop variable in the summation operation, count is the co-occurrence number of and , conf is the expert-labeled causal confidence, is the dynamic decay factor, the total number of entity nodes in the knowledge graph, and the formula is: , age is the time length from the latest relationship update time to the current time, is the decay coefficient, such as the association weight of high-speed shaft and bearing failure, and the causal weight of vibration anomaly and gear wear. After defining the knowledge in the standardized semantic format, it is stored in the Neo4j graph database and indexed to form a structured gearbox failure knowledge network.

[0039] S200, build a knowledge-enhanced large model. Based on the LLaMA pre-training language model with natural language understanding capability, the entity information in the knowledge graph (such as the feature description of bearing failure and maintenance cases) is cross-modal fused with the model hidden layer state through the attention mechanism function to build a knowledge-enhanced large model, and the formula is: , wherein, is the hidden layer state vector of the large model at time , and is the updated hidden layer state vector after fusing the knowledge graph entity, is the entity node in the knowledge graph, is the index identifier, is the knowledge weight coefficient, Attn( ) is the attention mechanism function, and the association weight of and is calculated, is a multi-layer perceptron used for nonlinear transformation of the fused features, so that the model can learn the association logic of components and failures in the graph. The model is fine-tuned using fault diagnosis question-answer pairs in the wind power field (such as "What causes the gearbox vibration anomaly?") and historical accident cases (such as an event of bearing burnout due to lubricating oil pollution in a certain wind farm), so that it masters the reasoning logic of gearbox failure. Then, the model is optimized through contrastive learning, and the formula is: , wherein, is the loss value of contrastive learning, is the hidden layer state of the large model after fusing the knowledge, is other entities in the knowledge graph except , is the cosine similarity, is the temperature parameter, For real fault entities, denotes the time, is an index variable, enhancing its ability to distinguish similar faults (such as gear wear and bearing failure), and building a knowledge-enhanced diagnostic model that adapts to the wind power scene.

[0040] S300, real-time state monitoring and feature extraction, install vibration sensors at the gearbox body and bearing seat positions to collect real-time vibration signals; install infrared temperature sensors at the bearing end cover and gear meshing position to monitor component temperature changes; connect pressure sensors and flow sensors to the lubricating oil pipeline to obtain lubrication system parameters; at the same time, monitor the motor input and output current through the current transformer; after denoising and filtering preprocessing, use multi-scale feature fusion algorithm containing wavelet transform, convolutional neural network and long short-term memory network to extract mutation features of vibration signals and abnormal trend state features of temperature curves from time domain and frequency domain. The formula is: where, is the time series data collected by the sensor at time , denotes the time, denotes feature splicing, WT( ) is wavelet transform, ( ) is self-attention mechanism, ( ) is a convolutional neural network used to extract local features of signals through a learnable weight matrix for feature fusion. The formula is: , is the original feature vector after fusion, is the final feature vector after linear transformation, is a learnable weight matrix used for feature transformation, is a bias term to adjust the feature offset.

[0041] S400, progressive multi-round diagnostic reasoning:

[0042] S410, progressive information acquisition: according to the needs raised by the preliminary diagnosis, the large model obtains additional information to narrow down the diagnosis range. For example, the system retrieves other sensor data, historical data or performs specific detection operations according to the model prompt; or the operation and maintenance personnel can check the site situation according to the clues provided by the model in human-computer interaction. The knowledge graph provides guidance at this stage, such as suggesting "check the temperature of a certain component" or "obtain vibration spectrum analysis", so as to obtain additional evidence that helps to distinguish different faults.

[0043] S420, a new round of diagnostic reasoning: The additional information acquired in S410 is fed into the diagnostic engine along with the original data, and the knowledge-enhanced large model performs a new round of reasoning and analysis. With this additional evidence, the model can eliminate some previous hypotheses, more accurately evaluate remaining candidates, and potentially propose new, more specific hypotheses. The results are then reassessed. If uncertainty remains, further feedback is provided to obtain more information and proceed to the next round. This iterative process forms a progressive diagnostic process until the model outputs a clear, high-confidence fault diagnosis conclusion.

[0044] The acquired additional information and the fused feature vectors are input into the knowledge-enhanced large model, and the multiple fault hypotheses initially generated by the model, such as "bearing wear", "gear meshing failure", and "lubricant contamination", are eliminated to obtain a set of eliminated fault hypotheses, which are then assigned initial confidence levels. The confidence levels of each hypothesis are updated through an iterative formula. Suppose the hypothesis set for the mth round of diagnosis is , the confidence vector of the mth round of diagnosis is , It represents the total number of fault hypotheses in the mth round of diagnostic hypothesis set and is updated by the iterative formula: ,in, is the diagnostic weight matrix, which is used to adjust the direction of confidence update. represents element-wise multiplication, Assumption The associated knowledge graph subgraph, Is the attention mechanism, calculating state features and knowledge graph subgraphs The correlation degree, is the original eigenvector after fusion, is a knowledge graph subgraph, It is The first round of diagnosis A hypothetical failure, The confidence offset introduced by the new information, is the iteration coefficient, Activation function, which converts numerical values ​​into probability distribution, represents the diagnostic iteration round, is an index identifier. In this process, the model adjusts the reasoning direction based on the diagnostic weight matrix, uses the attention mechanism to calculate the correlation between the current feature and each fault subgraph in the knowledge graph (such as the vibration characteristics and temperature threshold corresponding to bearing wear), and combines the confidence offset introduced by the newly added sensor data (such as the lubricating oil sampling analysis results) to gradually correct the credibility of each hypothesis. or ,in is the belief entropy, is the index identifier, represents the diagnostic iteration round, The maximum confidence value in the first round of diagnosis, The confidence of the first hypothesis in the diagnosis, is the number of fault categories in the hypothesis set, is the confidence entropy threshold, when the confidence entropy is less than or equal to , it is considered that the diagnosis result is sufficient to determine, The maximum confidence value in the first round of diagnosis, The maximum confidence threshold, when the maximum confidence is greater than or equal to , it is considered that the credibility of a single fault hypothesis is high enough (indicating that the uncertainty of the diagnosis result is low), or the maximum confidence of a certain fault hypothesis exceeds the threshold, the reasoning is terminated; otherwise, continue to obtain additional information for refined analysis of the vibration spectrum for multiple iterations.

[0045] The core content of this step is: according to the needs put forward by the preliminary diagnosis, the large model obtains additional information to narrow the diagnosis range. Due to the addition of additional evidence, the model can exclude some previous hypotheses, more accurately evaluate the remaining candidates, and possibly propose more specific new hypotheses.

[0046] S500, diagnosis result output and knowledge graph update: if the model determines that the confidence of "bearing wear" reaches the termination condition, output the diagnosis result: the fault location is the high-speed shaft bearing, the reason is that the metal particle deposition in the lubricating oil causes abrasive wear, the maintenance suggestion is to replace the bearing and clean the lubrication system, and at the same time, the new correlation data between "lubricating oil metal particle concentration and bearing wear" found in this diagnosis is fed back to the knowledge base. The influence of the new case on the relationship weight of the knowledge graph is calculated through the dynamic knowledge update factor, and the dynamic knowledge update factor is designed, and the calculation formula is: , wherein is the relationship weight between the entity and before updating, is the relationship weight increment caused by the new case, is the relationship weight between the entity and after updating, is the dynamic knowledge update factor, which determines the influence degree of the new case on the weight, and the calculation formula is: , represents the moment, and novelty is the case novelty score, which is calculated by comparing the KL divergence between the case features and the historical data distribution: novelty , is the equipment state feature vector of the current case, is the new diagnosis case, is a regulation parameter, which controls the sensitivity of the update factor to the novelty of the case, such as enhancing the association weight of “lubricant contamination” and “bearing wear” to make the atlas continue to iterate and optimize, and adapt to new fault modes.

[0047] The embodiment realizes progressive reasoning of the gearbox fault by constructing a knowledge graph containing key components, fault modes and causal relationships of the gearbox, and integrating it into the LLaMA pre-training model to form a knowledge-enhanced large model, combining multi-sensor real-time monitoring and multi-scale feature fusion algorithm. Through the entity relationship weight calculation formula of the knowledge graph, the attention mechanism fusion algorithm and the dynamic knowledge update factor core mechanism, the method combines domain knowledge with data-driven reasoning. It not only accurately locates the current fault (such as high-speed shaft bearing wear) based on historical fault data, but also iteratively optimizes the knowledge graph through new cases to improve the diagnosis ability of new faults, providing an efficient and reliable technical solution for intelligent operation and maintenance of wind power equipment.

[0048] Embodiment Two:

[0049] Petroleum chemical centrifugal pump fault diagnosis scene embodiment.

[0050] The centrifugal pump in the crude distillation unit of a certain petrochemical enterprise undertakes the key task of transporting high-temperature crude oil, and its operating state directly affects the material balance and production efficiency of the distillation tower. In recent years, the pump has shown phenomena such as flow fluctuation and abnormal temperature rise of the bearing. Considering the continuity and danger of petrochemical production, it is necessary to quickly and accurately diagnose the fault cause. The application of the diagnostic method, combined with the operating data and domain knowledge of the centrifugal pump, realizes accurate positioning and analysis of the fault.

[0051] S100, construct a domain knowledge graph: collect the fault records, repair reports and equipment technical manuals of the centrifugal pump in the past three years, including typical fault cases such as impeller wear, mechanical seal leakage and bearing overheating. After cleaning the data to remove abnormal records caused by human operation errors, classify and organize them according to fault types, extract key components such as impeller, pump shaft, mechanical seal and bearing, as well as signs such as flow drop, pressure fluctuation and vibration intensification, and corresponding fault modes, causes (such as medium corrosion, cavitation) and maintenance measures as knowledge graph nodes. Through analyzing the co-occurrence frequency of entities in historical data (such as the number of times mechanical seal leakage and pressure fluctuation occur together), expert-labeled causal confidence (such as the impact of cavitation on impeller wear), and time decay factor, the edge weight is determined using the knowledge graph entity relationship weight calculation formula: such as strengthening the association weight of “medium containing particles” and “impeller wear”, the knowledge is stored in the Neo4j database in a standardized format after being defined, and an efficient retrieval graph structure is established, as shown in Figure 1 .

[0052] S200, in the construction of knowledge-enhanced large models, based on the LLaMA pre-trained language model, the feature description of mechanical seal failure and the entity information of impeller wear repair case in the knowledge graph are fused into the model hidden layer through the attention mechanism function to construct the knowledge-enhanced large model, the formula is: , the model is fine-tuned using petrochemical field fault diagnosis question and answer data (such as "possible reasons for insufficient flow of centrifugal pump") and actual accident cases (such as an event of perforation of impeller due to medium cavitation in a device), the formula is: , so that it learns the fault logic of petrochemical centrifugal pump, optimizes the model through contrastive learning, and distinguishes the faults with similar symptoms of "bearing overheating" and "mechanical seal leakage", and optimizes the knowledge-enhanced model for petrochemical scene.

[0053] S300, real-time state monitoring and feature extraction: install vibration sensors on the pump body bearing seat to monitor running vibration; install temperature sensors in the mechanical seal cavity and bearing box to collect temperature data; install pressure sensors and turbine flow meters on the inlet and outlet pipelines to obtain pressure and flow parameters; monitor motor current through current transformers; after noise reduction preprocessing, use multi-scale feature fusion algorithm combining wavelet transform and convolutional neural network to extract abnormal frequency components of vibration signals, mutation points of temperature curve, and fluctuation mode features of flow pressure, the formula is: , and then generate a feature vector that comprehensively represents the pump running state through a feature fusion algorithm, the formula is: .

[0054] S400, in the progressive multi-round diagnosis reasoning, the feature vector is input into the model, and the fault hypotheses of "mechanical seal wear", "impeller cavitation" and "bearing oil shortage" are initially generated and assigned with confidence, and the confidence is updated through iteration formula, the formula is: , the model adjusts the reasoning direction according to the diagnosis weight matrix, matches the current features with the correlation degree of each fault subgraph (such as pressure fluctuation features and temperature threshold of mechanical seal wear) in the knowledge graph through attention mechanism, and combines the confidence offset introduced by medium component analysis information to gradually narrow down the fault range, when or , stop reasoning; otherwise, continue to obtain additional data such as vibration spectrum analysis and lubricating oil viscosity detection for multi-round diagnosis.

[0055] S500, diagnosis result output and knowledge graph update, if the model determines that "mechanical seal wear" is the main fault, the output result is: fault location is the mechanical seal of pump shaft, the reason is that corrosive components in the medium cause damage to the seal surface, the maintenance suggestion is to replace the seal assembly and filter the medium, at the same time, the new data of "correlation between medium corrosion and mechanical seal life" found in this diagnosis is fed back to the knowledge base, the influence of new cases on relationship weight is calculated through dynamic knowledge update factor, the formula is: As the correlation weight of enhancing "medium corrosion" and "mechanical seal failure" is increased, the knowledge graph can adapt to new fault challenges brought by changes in petrochemical media.

[0056] In summary, in the fault diagnosis scene of the petrochemical centrifugal pump, the embodiment realizes the precise diagnosis of the fault under the complex working condition by constructing the professional knowledge graph (covering the fault correlation of the impeller and the mechanical seal part), combining the knowledge enhanced large model and the multi-source sensor data fusion. The method uses the entity relationship weight calculation formula of the knowledge graph to quantify the correlation strength between the components and the fault, extracts the comprehensive features of the parameters containing vibration and temperature through the multi-scale feature fusion algorithm, and ensures the diagnosis accuracy through the confidence entropy and maximum confidence determination mechanism in the progressive reasoning. In addition, the dynamic knowledge update factor enables the model to adaptively update the knowledge graph according to the changes in petrochemical media (such as the increase of corrosion), and improves the diagnosis adaptability to the petrochemical industry specific faults (such as medium cavitation and mechanical seal corrosion), which provides technical support for ensuring the continuity of petrochemical production.

[0057] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the present application, and any equivalent embodiments with equivalent changes are equivalent to the above embodiments. Any modification, equivalent change and modification of the above embodiments according to the technical essence of the present application are still within the scope of the present application.

Claims

1. A dynamic device fault progressive diagnosis method based on a knowledge-enhanced large model, characterized in that, The specific steps of the diagnosis method are: S100, constructing a field knowledge graph: based on historical fault data and equipment manual information, after data cleaning and labeling, a knowledge graph is constructed, which contains device key components, sensor measuring points, typical fault modes, fault symptoms, cause mechanisms and maintenance measures as nodes, and the association between components and fault modes and the causal relationship between fault symptoms and potential causes as edges, and the knowledge expression specification is defined; S200, constructing a knowledge-enhanced large model: based on a large pre-trained language model, the knowledge graph content is integrated into the large model to construct a knowledge-enhanced large model, the model is fine-tuned and trained using fault diagnosis domain question and answer pairs and accident case data, and is optimized through contrastive learning to construct a knowledge-enhanced large model; S300, real-time state monitoring and feature extraction: real-time state monitoring of the target equipment is performed through multiple source sensors, operating condition data is collected and preprocessed, and the state features of the equipment at the current time are extracted through a multi-scale feature fusion algorithm; S400, progressive multi-round diagnosis reasoning: additional information is obtained and input into the knowledge-enhanced large model according to the requirements raised by the preliminary diagnosis, the additional information and the device state feature description are input into the knowledge-enhanced large model, fault reasoning is performed, the confidence is updated through iteration formula, and whether the diagnosis termination condition is met is judged according to the confidence entropy and the maximum confidence, if not, additional information is obtained and the reasoning is continued until a clear conclusion is reached; S500, diagnosis result output and knowledge graph update: when the model confidence reaches a threshold or is narrowed to a single fault cause, the diagnosis result and disposal suggestion containing fault positioning, type, cause analysis and maintenance suggestion are output, and the new data and conclusion in the diagnosis process are fed back to the knowledge base, and the knowledge graph relationship weight is updated through a dynamic knowledge update factor. 2.The knowledge-enhanced large model-based dynamic device fault progressive diagnosis method according to claim 1, characterized in that, In the S100, the large pre-trained language model refers to LLaMA pre-trained on a large amount of text data, which has natural language understanding and generation capabilities, can be used as a basis, and can be integrated with knowledge graph content and fine-tuned with fault diagnosis domain data to adapt to the language model of the equipment fault diagnosis task. 3.The knowledge-enhanced large model-based dynamic device fault progressive diagnosis method according to claim 1, characterized in that, The specific steps of constructing the knowledge graph in S100 include: (1) Collecting historical fault data and equipment manual information of the target equipment, cleaning, classifying and arranging the data, and removing redundant or incorrect information; (2) Extracting key entities from the preprocessed data and defining them as knowledge graph nodes, including: device key components, sensor measuring points, typical fault modes, fault symptoms, cause mechanisms and maintenance measures; (3) Determine the association between entities and define it as the edge of the knowledge graph, including: the association between components and fault modes and the causal relationship between fault symptoms and potential causes; (4) Based on the co-occurrence number of entities in historical fault data, the annotated causal confidence and timeliness, the relationship weight is calculated through the knowledge graph entity relationship weight calculation formula; (5) Define knowledge in a standardized semantic format, clearly define entity attributes and relationship weights, and form a knowledge expression specification that can be referenced by the large model; (6) Atlas storage and management: the completed knowledge graph is stored in Neo4j, and the index of nodes and edges is established.

4. The knowledge-enhanced large model-based dynamic device fault progressive diagnosis method according to claim 3, characterized in that, The relationship weight is calculated by the knowledge graph entity relationship weight calculation formula, including: Define the set of entity nodes in the knowledge graph as: , Edge relationship weight The calculation formula is: , wherein, is the edge relation weight between entity nodes and in the knowledge graph, , and are entity nodes in the knowledge graph, representing equipment key components, sensor measuring points, typical failure modes, and are index identifiers, is a loop variable in summation operation, count is the co-occurrence number of and in historical failures, conf is the expert-annotated causal confidence, is a dynamic damping factor, the total number of entity nodes in the knowledge graph, The formula is: , age is the time length from the latest update time of the relationship to the current time, is the decay coefficient. 5.The knowledge-enhanced large model-based dynamic device fault progressive diagnosis method according to claim 1, wherein, The S200, constructing the knowledge enhanced large model, constructs the knowledge enhanced large model through an attention mechanism function Attn( ) to perform cross-modal fusion between knowledge graph entities and large model hidden layer states , and the formula is: , wherein, The large model is in time The hidden layer state vector, The updated hidden layer state vector after fusing the knowledge graph entity, Is an entity node in the knowledge graph, Is an index identifier, Is a knowledge weight coefficient, Attn ) is an attention mechanism function, which calculates The association weight of and Multi-layer perceptron, used for nonlinear transformation of fused features, model optimization through contrastive learning, formula: , wherein, is a loss value of contrastive learning, is a hidden layer state of a large model after fusing knowledge, is an entity other than in a knowledge graph, is a cosine similarity, is a temperature parameter, is a real fault entity, represents a moment, is an index variable. 6.The knowledge-enhanced large model-based dynamic device fault progressive diagnosis method according to claim 1, wherein, The sensor types used in the real-time state monitoring and feature extraction of the S300 are: Vibration sensor: speed sensor is used to monitor the vibration signal of the equipment during operation; Temperature sensor: infrared temperature sensor is used to measure the temperature of the equipment components; Pressure sensor: absolute pressure sensor is used to collect pressure data inside the equipment or in the pipeline; Current / voltage sensor: current transformer is used to monitor the working current and voltage of electrical components; Flow sensor: turbine flowmeter is used to collect state data of the lubrication system and cooling system. 7.The knowledge-enhanced large model based dynamic device fault progressive diagnosis method according to claim 1, wherein, In the real-time state monitoring and feature extraction, the multi-scale feature fusion algorithm is used to extract the state features of the equipment at the current time, and the time sequence data of the sensor is multi-scale feature extraction is performed, The calculation formula is: , wherein, is the time series data collected by the sensor at time , represents the time, represents the feature concatenation, WT( ) is the wavelet transform, ( ) is the self-attention mechanism, ) is a convolutional neural network used to extract local features of the signal through a learnable weight matrix The formula for feature fusion is: ,​ is the original feature vector after fusion, is the final feature vector after linear transformation, is the learnable weight matrix for feature transformation, is the bias term to adjust the feature offset. 8.The knowledge-enhanced large model based dynamic device fault progressive diagnosis method according to claim 1, characterized in that, The S400 updates the confidence in the progressive multi-round diagnosis reasoning. Assuming that a hypothesis set of the mth round of diagnosis is , and a confidence vector of the mth round of diagnosis is , , wherein the total number of fault hypotheses in the hypothesis set of the mth round of diagnosis is represented, and the confidence is updated through an iterative formula, The formula is: , wherein, is a diagnostic weight matrix, used to adjust the direction of confidence update, denotes element-wise multiplication, is a hypothesis associated knowledge graph subgraph, is an attention mechanism, calculating the state feature and the relevance to the knowledge graph subgraph , is the fused original feature vector, is the knowledge graph subgraph, is the hypothesis fault of the round of diagnosis, is the confidence offset introduced by the new information, is the iteration coefficient, is an activation function, converting numerical values into probability distribution, represents the iteration round of diagnosis, is an index identifier. 9.The knowledge-enhanced large model based dynamic device fault progressive diagnosis method according to claim 1, wherein, The S400, the determination of the diagnosis termination condition in the progressive multi-round diagnosis reasoning, sets the confidence entropy as , measures the uncertainty of the confidence distribution, when or , terminate the iteration, otherwise get the new data to generate Continuing the reasoning, is the index identifier, represents the diagnostic iteration round, No. In the first round of diagnosis The confidence level of a hypothesis, is the number of fault categories in the hypothesis set, Is the credibility entropy threshold, when the credibility entropy is less than or equal to The diagnosis is considered sufficiently certain when No. The maximum confidence value in the round diagnosis, Maximum confidence threshold, when the maximum confidence is greater than or equal to When , the credibility of the single fault hypothesis is considered high enough. 10.The knowledge-enhanced large model based dynamic device fault progressive diagnosis method according to claim 1, wherein, The S500, diagnosis result output and knowledge graph update through knowledge graph dynamic update algorithm can automatically update the knowledge graph relationship weight, and a dynamic knowledge update factor is designed , The calculation formula is: , wherein, pre-update entity with the relationship weight, the relationship weight increment caused by the new case, post-update entity with the relationship weight, is a dynamic knowledge update factor, determining the influence degree of the new case on the weight, and the calculation formula is: , represents the time, novelty is the case novelty score, which is calculated by comparing the KL divergence of the case features and the historical data distribution: novelty , is the equipment state feature vector of the current case, is the newly diagnosed case, is a regulation parameter, controlling the sensitivity of the update factor to the case novelty.

Images