Large Model-Based Fault Detection Method and System

Abstract

The present invention discloses a fault detection method and system based on a large model, belonging to the technical field of fault detection. Specifically, it includes: collecting multi-source fault data from an industrial system and performing preprocessing to extract fault features to form a data set; loading a pre-trained large language model and fine-tuning it according to the fault feature data set; using the fine-tuned model and the fault feature data set to construct a fault detection model, training and evaluating it in combination with an adaptive optimization strategy to obtain a trained model; inputting the real-time collected fault data into the model to output preliminary fault detection results, and further obtaining fault type, location, and probability information through a multi-level fault analysis method, while feeding back the results to optimize the model. The present invention improves the accuracy and efficiency of fault detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault detection, specifically a fault detection method and system based on a large model. Background Art

[0002] Fault detection methods collect and analyze various data during the operation of a system, and use specific algorithms or models to determine whether there are abnormalities or faults in the system. The purpose is to detect and locate faults in a timely manner so as to take corresponding measures for repair, thereby avoiding or reducing the impact and losses caused by faults to the system. Traditional fault detection methods, such as rule-based methods, statistics-based methods, and machine learning-based methods, although can meet the fault detection requirements of specific fields to a certain extent, often have problems such as insufficient generalization ability, high model complexity, and limited processing ability for large-scale data.

[0003] For example, the Chinese patent application with the publication number CN118069400A discloses a fault detection method, system and electronic device, including: obtaining target data corresponding to a target device, the target device being a device to be fault-detected, the target data including multiple sub-data, and each sub-data being data for fault identification in different data sources corresponding to the target device; inputting the target data into a fault analysis model so that the fault analysis model obtains a target fault detection result according to the target data, the target fault detection result including the fault type and fault cause of the target device, and the fault analysis model being a model trained based on historical fault data. Using the method of a large model effectively reduces the human input in fault detection. By analyzing the target data corresponding to the target device, which includes multiple sub-data for fault identification, compared with the method of identifying faults only from a single data bus or diagnostic data, the accuracy and comprehensiveness of the fault detection result are improved.

[0004] For example, the Chinese patent with the authorization announcement number CN115629930B discloses a fault detection method, device, equipment and storage medium based on a DSP system, including: obtaining a fault signal; inputting the fault signal into a preset fault detection model, and based on the fault detection model, classifying the fault signal and diagnosing the fault information of the classified fault signal to obtain fault information; wherein, the fault detection model is obtained by iteratively training a model to be trained based on a fault signal sample, the category weight information of the fault signal sample, and the fault information label of the fault signal sample. This technical solution classifies and diagnoses the fault information of the fault signal that appears in the signal processing system of the DSP processor based on a pre-trained fault detection model, and quickly and accurately outputs the fault information, without the need for developers or users to manually debug a large number of problem locations, improving the efficiency of fault detection of the DSP system.

[0005] The above prior arts all have the following problems: In CN118069400A, although it mentions inputting target data into a fault analysis model to obtain a fault detection result, it does not mention real-time requirements and feedback mechanisms; it only mentions fault types and causes, but does not mention fault location and fault probability information; in CN115629930B, it mainly focuses on fault signals inside the DSP system and classifies and diagnoses based on these signals; the detail and accuracy of the fault detection result are limited; although it mentions quickly and accurately outputting fault information, it may not discuss real-time requirements and feedback mechanisms in detail. Summary of the Invention

[0006] In view of the deficiencies of the prior art, the present invention proposes a fault detection method and system based on a large model, which collects multi-source fault data from an industrial system and preprocesses it, extracts fault features to form a data set; loads a pre-trained large language model and fine-tunes it according to the fault feature data set; uses the fine-tuned model and the fault feature data set to construct a fault detection model, combines an adaptive optimization strategy for training and evaluation, and obtains a trained model; inputs the real-time collected fault data into the model, outputs a preliminary fault detection result, and further obtains fault type, location, and probability information through a multi-level fault analysis method, and at the same time feeds back the result to optimize the model. The present invention improves the accuracy and efficiency of fault detection.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A fault detection method based on a large model, comprising:

[0009] Step S1: Collect multi-source fault data from an industrial system, preprocess it, and at the same time, extract fault features from the preprocessed multi-source fault data to obtain a fault feature data set;

[0010] Step S2: Load a pre-trained large language model and fine-tune it according to the fault feature data set;

[0011] Step S3: Use the fine-tuned large language model and the fault feature data set to construct a fault detection model, combine an adaptive optimization strategy, train and evaluate the fault detection model, and optimize the fault detection model according to the evaluation result to obtain a trained fault detection model;

[0012] Step S4: Input the real-time collected fault data into the trained fault detection model. The fault detection model outputs the preliminary fault detection result according to the input data. Combine the multi-level fault analysis method to conduct multi-level analysis on the preliminary fault detection result, obtain the fault type, fault location and fault probability information, and feedback the real-time fault detection result and the result of multi-level analysis to the fault detection model for optimization;

[0013] The specific steps of the said step S4 include:

[0014] S4.1: Obtain the real-time fault data and conduct preprocessing to obtain the preprocessed real-time fault data , where, represents the Mth preprocessed real-time fault data, and M represents the number of preprocessed real-time fault data;

[0015] S4.2: Input into the trained fault detection model. The fault detection model outputs the preliminary fault detection result according to the input data;

[0016] S4.3: Combine the multi-level fault analysis method to conduct in-depth analysis on the preliminary fault detection result, and through the knowledge graph and expert system, combine the characteristic information in the real-time fault data and historical fault cases to judge the fault type, fault location and fault probability information;

[0017] S4.4: Feedback the real-time fault detection result and the result of multi-level analysis to the fault detection model, and use the reinforcement learning algorithm to conduct online update according to the feedback result;

[0018] The specific steps of the said S4.3 include:

[0019] S4.31: Obtain and the preliminary fault detection result;

[0020] S4.32: Extract fault knowledge from historical fault cases, expert experience and equipment manuals, and use the knowledge graph method to represent the extracted fault knowledge as a graphical knowledge structure. Among them, the fault type, fault location and fault probability information are represented as nodes, and the relationships between them are represented as edges;

[0021] S4.33: Input the preliminary fault detection result into the pre-loaded expert system. The expert system matches and reasons the fault detection result according to the fault knowledge in the knowledge graph to obtain the fault type, fault location and fault probability information;

[0022] S4.34: According to the reasoning result of the expert system, combine to judge the specific type and location of the fault;

[0023] S4.35: Analyze historical failure cases using machine learning algorithms to obtain the probability distribution of failure occurrences, and evaluate the probability of failure based on the current operating state of the device, the specific type and location of the failure.

[0024] The specific steps of S4.4 are as follows:

[0025] S4.41: Integrate the real-time failure detection results and the multi-level analysis results to generate a failure detection report.

[0026] S4.42: Load the pre-built reinforcement learning model, input the failure detection report as the feedback result into the reinforcement learning model, and design the reward function and state transition rules. Use the reinforcement learning strategy to train the reinforcement learning model so that the reinforcement learning model can learn and adapt to new failure modes, and obtain the trained reinforcement learning model. The formula is:

[0027] ;

[0028] where, represents the quality function at time t+1 under the state z, action d, failure feature f, and context information g. represents the quality function at time t under the state z, action d, failure feature f, and context information g. represents the learning rate. represents the reward immediately obtained after taking action d. represents the discount factor. represents the maximum value of the quality function under the next state, action, failure feature, and context information at time t. , , , respectively represent the next state, action, failure feature, and context information. represents the feature weight. represents the feature function. represents the constraint weight. represents the constraint function.

[0029] S4.43: According to the training results of the reinforcement learning model, adjust the parameters of the failure detection model, and evaluate the accuracy of the adjusted failure detection model. At the same time, according to the evaluation results, further optimize and adjust the failure detection model.

[0030] If the evaluation result is not satisfactory, return to step S4.42 to retrain the reinforcement learning model and update the failure detection model again.

[0031] If the evaluation result meets the requirements, output the result and end the process.

[0032] Specifically, the specific formula of the adaptive optimization strategy in step S3 is as follows:

[0033] ;

[0034] where t represents the current moment, and respectively represent the first-order moment estimate and the second-order moment estimate of the fault detection model at time t, represents the fault detection model parameters at time t + 1, represents the fault detection model parameters at time t, represents the step size for updating the control parameters, and represent the exponential decay rate of the moment estimate at time t, and b represents a constant.

[0035] The fault detection system based on the large model includes: a data processing module, a model fine-tuning module, a fault detection module, and a fault analysis module;

[0036] The data processing module is used to collect multi-source fault data from the industrial system and perform preprocessing and feature extraction;

[0037] The model fine-tuning module is used to fine-tune the pre-trained large language model using the fault feature dataset;

[0038] The fault detection module is used to construct a fault detection model using the fine-tuned large language model and the fault feature dataset, and perform training and evaluation, and optimize the model according to the evaluation results;

[0039] The fault analysis module is used to input the real-time collected fault data into the trained fault detection model, output the preliminary fault detection results, and perform multi-level analysis in combination with the multi-level fault analysis method.

[0040] Specifically, the fault analysis module includes: a real-time detection unit, a multi-level analysis unit, and a feedback optimization unit;

[0041] The real-time detection unit is used to input the real-time collected fault data into the trained fault detection model and output the preliminary fault detection results;

[0042] The multi-level analysis unit is used to perform multi-level analysis on the preliminary fault detection results;

[0043] The feedback optimization unit is used to feedback the real-time detection results and the results of the multi-level analysis to the fault detection model for optimization.

[0044] An electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of a fault detection method based on a large model are implemented.

[0045] A computer-readable storage medium stores computer instructions, and when the computer instructions run, the steps of a fault detection method based on a large model are executed.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. The present invention proposes a fault detection system based on a large model and optimizes and improves the architecture, operation steps, and processes. The system has the advantages of simple processes, low investment and operation costs, and low production work costs.

[0048] 2. The present invention proposes a fault detection method based on a large model. By collecting multi-source fault data from an industrial system, preprocessing and feature extraction are performed, and combined with a pre-trained large language model for fine-tuning, a fault detection model is constructed and optimized. This process not only improves the accuracy and efficiency of fault detection but also enables the model to better adapt to the actual situation of the industrial system, reducing the probability of false alarms and missed alarms.

[0049] 3. The present invention proposes a fault detection method based on a large model. By inputting the real-time collected fault data into the trained fault detection model, preliminary fault detection results can be quickly output; through a multi-level fault analysis method to further analyze the results, detailed fault type, location, and probability information can be obtained, providing strong support for the rapid location and repair of faults; at the same time, feedback of the real-time detection results and analysis results to the model for optimization can continuously improve the detection ability and accuracy of the model, improving the stability of the industrial system. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the fault detection method based on a large model of the present invention;

[0051] Figure 2 It is a principle flowchart of the fault detection method based on a large model of the present invention;

[0052] Figure 3 It is a flowchart for implementing real-time fault detection results in the fault detection method based on a large model of the present invention;

[0053] Figure 4 It is an architecture diagram of the fault detection system based on a large model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] Embodiment 1

[0055] Please refer to Figures 1-3, an embodiment provided by the present invention: A fault detection method based on a large model, comprising the following steps:

[0056] Step S1: Collect multi-source fault data from the industrial system and perform preprocessing. At the same time, extract fault features from the preprocessed multi-source fault data to obtain a fault feature dataset;

[0057] Further, the specific steps of step S1 include:

[0058] S1.1: Determine the types of fault data to be collected, such as vibration signals, temperature signals, and pressure signals, and install sensors at various key parts of the industrial system to collect fault data in real time. At the same time, ensure the accuracy and reliability of the sensors, as well as the real-time and continuity of data collection;

[0059] S1.2: Clean the collected raw data, remove noise, outliers, and duplicate data, and perform formatting processing on the data to convert it into a data format suitable for data mining algorithms;

[0060] S1.3: Use the decision tree classification algorithm to mine the preprocessed multi-source fault data and extract feature information related to faults, such as fault frequency, fault amplitude, and fault type. Among them, the decision tree classification algorithm is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here;

[0061] S1.4: Integrate the extracted fault feature information into a fault feature dataset for subsequent fault detection and diagnosis.

[0062] Step S2: Load the pre-trained large language model and fine-tune it according to the fault feature dataset;

[0063] Further, the specific steps of step S2 include:

[0064] S2.1: Select a pre-trained model: Select a model suitable for the current task from existing pre-trained large language models. These models have usually been trained on large-scale text data and have certain language understanding and generation capabilities. In the present invention, the ERNIE large model is selected. Among them, the ERNIE large model is a pre-trained model developed by Baidu, which enhances the representation ability of the model by introducing entity and semantic information.

[0065] S2.2: Load the pre-trained large language model: Load the selected pre-trained large language model into memory or a computing device for subsequent fine-tuning tasks;

[0066] S2.3: Prepare the fault feature dataset;

[0067] S2.4: Model Fine-tuning: Use the fault feature dataset to fine-tune the pre-trained large language model to better adapt to the current fault detection task. During the fine-tuning process, hyperparameters such as the parameters and learning rate of the large language model can be adjusted to optimize the performance of the large language model in the fault detection task. The parameter optimization formula for the large language model is:

[0068] ;

[0069] where represents the loss function, represents the probability that the large language model predicts the label under the given input features and parameters . represents the label of the i-th fault feature data, that is, the fault type or category, represents the i-th fault feature data, represents the large language model parameters, and N represents the number of samples in the fault feature dataset.

[0070] It should be noted that the goal of the fine-tuning process is to minimize the loss function by adjusting the parameters , thereby improving the accuracy of the large language model in the fault detection task.

[0071] Step S3: Use the fine-tuned large language model and the fault feature dataset to build a fault detection model, combine an adaptive optimization strategy, train and evaluate the fault detection model, and optimize the fault detection model according to the evaluation results to obtain a trained fault detection model;

[0072] Furthermore, the specific steps of step S3 include:

[0073] S3.1: Build a fault detection model framework and integrate an adaptive optimization strategy;

[0074] S3.2: Divide the fault feature dataset into a training set and a test set in a ratio of 7:3. Use the training set in the fault feature dataset to train the fault detection model, calculate the loss through forward propagation, and update the parameters of the fault detection model through backpropagation;

[0075] S3.3: According to the evaluation results on the validation set, such as accuracy, judge whether the performance of the fault detection model meets the requirements. If the performance of the fault detection model is not good, perform fault detection model optimization, including adjusting the model architecture, increasing training data, adjusting the parameters of the optimization algorithm, and repeating the process of training, evaluation, and optimization until the performance of the fault detection model reaches the preset level;

[0076] S3.4: When the fault detection model shows good performance on the validation set, it is considered that the fault detection model has been trained, and the parameters and structure of the fault detection model are saved for subsequent deployment and use in the actual environment.

[0077] Step S4: Input the real-time collected fault data into the trained fault detection model. The fault detection model outputs preliminary fault detection results based on the input data. Combining with the multi-level fault analysis method, the preliminary fault detection results are analyzed at multiple levels to obtain fault type, fault location, and fault probability information, and the real-time fault detection results and the results of multi-level analysis are fed back to the fault detection model for optimization.

[0078] The specific formula of the adaptive optimization strategy in S3 is:

[0079] ;

[0080] where t represents the current moment, and respectively represent the first-order moment estimate and the second-order moment estimate of the fault detection model at time t, represents the fault detection model parameters at time t + 1, represents the fault detection model parameters at time t, represents the step size for controlling parameter update, and represent the exponential decay rate of the moment estimate at time t, and b represents a constant.

[0081] The specific steps of S4 include:

[0082] S4.1: Obtain real-time fault data and perform preprocessing to obtain preprocessed real-time fault data , where, represents the Mth preprocessed real-time fault data, and M represents the number of preprocessed real-time fault data;

[0083] S4.2: Input into the trained fault detection model, and the fault detection model outputs preliminary fault detection results according to the input data;

[0084] Furthermore, the specific steps of S4.2 include:

[0085] S4.21: Obtain , and load the trained fault detection model to ensure that the model file is complete and not damaged, and at the same time the model parameters have been optimized to the best state;

[0086] S4.22: Use the normalization method to Perform format conversion to convert it into the input format of the fault detection model, and input the converted into the fault detection model;

[0087] The specific process of data conversion includes:

[0088] (1) Standardize or normalize continuous numerical features so that all fault feature data are on the same scale and eliminate the influence of dimensions;

[0089] (2) Fill in missing numerical values using the mean, median or mode;

[0090] (3) Complete data type conversion. Exemplarily, convert the integer type to the floating-point type.

[0091] S4.23: The fault detection model calculates the input converted extracts features and applies the learned rules to identify potential faults;

[0092] S4.24: Output the preliminary fault detection results. The fault detection results are usually presented in numerical, label or text form, depending on the design and requirements of the system.

[0093] S4.3: Combine the multi-level fault analysis method to deeply analyze the preliminary fault detection results, and through the knowledge graph and expert system, combine the feature information in the real-time fault data and historical fault cases to judge the fault type, fault location and fault probability information;

[0094] S4.4: Feed back the real-time fault detection results and multi-level analysis results to the fault detection model, and use the reinforcement learning algorithm to perform online update according to the feedback results.

[0095] The specific steps of S4.3 include:

[0096] S4.31: Obtain and the preliminary fault detection results;

[0097] S4.32: Extract fault knowledge from historical fault cases, expert experience, and equipment manuals, and use the knowledge graph method to represent the extracted fault knowledge as a graphical knowledge structure. Among them, the fault type, fault location, and fault probability information are represented as nodes, and the relationships between them are represented as edges. At the same time, the knowledge graph method is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here;

[0098] S4.33: Input the preliminary fault detection results into the pre-loaded expert system. The expert system matches and infers the fault detection results according to the fault knowledge in the knowledge graph to obtain the fault type, fault location and fault probability information;

[0099] Further, the specific steps of S4.33 include:

[0100] (1) Obtain preliminary fault detection results, which usually manifest as a series of fault signals, and preprocess the collected fault signals, including operations such as filtering, denoising, and normalization, to improve the accuracy and reliability of the signals;

[0101] (2) Input the preprocessed fault signals into a pre-loaded expert system. The expert system is a computer program based on artificial intelligence technology, which contains a large amount of fault knowledge and reasoning rules;

[0102] (3) The expert system uses the fault knowledge in the knowledge graph to match the input fault signals. The knowledge graph is a graph structure containing a large number of entities, relationships, and attributes. Among them, the entities can be devices, components, or fault types, etc., the relationships can be connection relationships, causal relationships, etc., and the attributes can be the model and parameters of the device;

[0103] (4) According to the matching results, the expert system uses the built-in reasoning mechanism for fault reasoning. The reasoning process adopts the case-based reasoning method, and the case-based reasoning method is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here;

[0104] (5) After reasoning and analysis, the expert system outputs fault type, fault location, and fault probability information, which can help users quickly locate and solve faults.

[0105] S4.34: According to the reasoning results of the expert system, combined with , judge the specific type and location of the fault;

[0106] S4.35: Use machine learning algorithms to analyze historical fault cases, obtain the probability distribution of fault occurrence, and evaluate the probability of fault occurrence according to the current operating state of the device, the specific type and location of the fault.

[0107] Further, the specific steps of S4.35 include:

[0108] (1) Obtain and historical fault case data, and perform preprocessing to generate raw data;

[0109] (2) Adopt statistical analysis methods to extract features from the raw data to obtain fault features;

[0110] (3) Load the pre-built machine learning model based on decision trees, and use the obtained fault features to train and evaluate the machine learning model based on decision trees. Among them, the training and evaluation of the machine learning model are the prior art in this field and not the creative solution of this application, so they will not be elaborated here;

[0111] (4) Input the real-time fault data into the trained machine learning model based on decision trees. The machine learning model based on decision trees will output the probability distribution of the occurrence of faults. According to the probability distribution, the possibility of the current device having a fault can be evaluated.

[0112] The specific steps of S4.4 include:

[0113] S4.41: Integrate the real-time fault detection results and the multi-level analysis results to generate a fault detection report;

[0114] S4.42: Load the pre-built reinforcement learning model, use the fault detection report as the feedback result, input it into the reinforcement learning model, and design the reward function and state transition rules. Use the reinforcement learning strategy to train the reinforcement learning model so that the reinforcement learning model can learn and adapt to new fault patterns to obtain the trained reinforcement learning model. The formula is:

[0115] ;

[0116] Among them, represents the quality function at time t + 1 under the state z, action d, fault feature f, and context information g, represents the quality function at time t under the state z, action d, fault feature f, and context information g, represents the learning rate, represents the reward immediately obtained after taking action d, represents the discount factor, represents the maximum value of the quality function under the next state, action, fault feature, and context information at time t, 、 、 、 respectively represent the next state, action, fault feature, and context information, represents the feature weight, represents the feature function, which is used to represent the importance of the fault feature, represents the constraint weight, represents the constraint function, which is used to represent the resource limit and safety requirement conditions;

[0117] In the present invention, the feature function , for emphasizing the importance of certain specific features. Exemplarily, in fault detection, some fault features may be more important than others, which can be reflected by to reflect this, the feature weight to control the influence of the feature function. And the introduction of the constraint function , for ensuring that certain constraint conditions are met when taking actions. Exemplarily, in an industrial environment, some actions may be restricted by resource limitations or safety requirements, which can be represented by to represent these constraints, the constraint weight to control the influence of the constraint function.

[0118] S4.43: According to the training results of the reinforcement learning model, adjust the parameters of the fault detection model, and evaluate the accuracy of the adjusted fault detection model. At the same time, according to the evaluation results, further optimize and adjust the fault detection model. Among them, the accuracy evaluation is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here;

[0119] If the evaluation result is not satisfactory, return to step S4.42 to retrain the reinforcement learning model and update the fault detection model again;

[0120] If the evaluation result meets the requirements, output the result and end the process.

[0121] Embodiment 2

[0122] Please refer to Figure 4 , another embodiment provided by the present invention: A fault detection system based on a large model, including:

[0123] A data processing module, a model fine-tuning module, a fault detection module, and a fault analysis module;

[0124] The data processing module is used to collect multi-source fault data from the industrial system and perform preprocessing and feature extraction. Among them, it includes collecting fault data from various sensors and log file sources, and the preprocessing includes cleaning, formatting, and normalizing to improve the data quality;

[0125] The model fine-tuning module is used to fine-tune the pre-trained large language model with the fault feature dataset to make it more suitable for the fault detection task;

[0126] The fault detection module is used to construct a fault detection model using the fine-tuned large language model and the fault feature dataset, and perform training and evaluation, and optimize the model according to the evaluation results;

[0127] The fault analysis module is used to input the real-time collected fault data into the trained fault detection model, output the preliminary fault detection result, and perform multi-level analysis in combination with the multi-level fault analysis method.

[0128] The fault detection module includes: a model construction unit, an evaluation unit, and a model optimization unit;

[0129] The model construction unit is used to construct a fault detection model by combining the fine-tuned large language model and the fault feature dataset to form a preliminary model capable of detecting faults;

[0130] The training and evaluation unit is used to train the fault detection model and evaluate its performance, such as accuracy and recall, to understand the performance of the fault detection model in the fault detection task;

[0131] The model optimization unit is used to optimize the fault detection model according to the evaluation results, such as adjusting the model structure and parameters, to improve the detection performance and generalization ability of the fault detection model.

[0132] The fault analysis module includes: a real-time detection unit, a multi-level analysis unit, and a feedback optimization unit;

[0133] The real-time detection unit is used to input the real-time collected fault data into the trained fault detection model and output the preliminary fault detection results to achieve real-time fault detection;

[0134] The multi-level analysis unit is used to perform multi-level analysis on the preliminary fault detection results, such as analysis based on fault type, fault location, and fault probability, to provide more detailed and accurate fault information, which helps to quickly locate and solve faults;

[0135] The feedback optimization unit is used to feedback the real-time detection results and the results of multi-level analysis to the fault detection model for optimization, to achieve continuous learning and optimization of the model, and improve the adaptability and accuracy of the model.

[0136] Embodiment 3

[0137] An electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the steps of the fault detection method based on a large model. For details, refer to the above method embodiments and will not be elaborated here.

[0138] A computer-readable storage medium stores computer instructions. When the computer instructions run, they execute the steps of the fault detection method based on a large model. Among them, the storage medium can be a volatile or non-volatile computer-readable storage medium.

[0139] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make changes, modifications, substitutions, and variations to the above embodiments without departing from the spirit and scope of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A fault detection method based on a large model, characterized in that: include: Step S1: Collect multi-source fault data from the industrial system and perform preprocessing. At the same time, extract fault features from the preprocessed multi-source fault data to obtain a fault feature data set; Step S2: Load the pre-trained large language model and fine-tune it according to the fault feature dataset; Step S3: construct a fault detection model using the fine-tuned large language model and the fault feature data set, train and evaluate the fault detection model in combination with the adaptive optimization strategy, and optimize the fault detection model according to the evaluation results to obtain a trained fault detection model; Step S4: input the real-time collected fault data into the trained fault detection model, and the fault detection model outputs preliminary fault detection results according to the input data. Combined with the multi-level fault analysis method, the preliminary fault detection results are subjected to multi-level analysis to obtain the fault type, fault location and fault probability information, and the real-time fault detection results and the results of the multi-level analysis are fed back to the fault detection model for optimization; The specific steps of step S4 include: S4.1: Obtain real-time fault data and perform preprocessing to obtain preprocessed real-time fault data ,in, represents the Mth preprocessed real-time fault data, and M represents the number of preprocessed real-time fault data; S4.2: The trained fault detection model is input, and the fault detection model outputs preliminary fault detection results based on the input data; S4.3: Combine multi-level fault analysis methods to conduct in-depth analysis of preliminary fault detection results, and use knowledge graphs and expert systems to combine feature information in real-time fault data and historical fault cases to determine fault type, fault location, and fault probability information; S4.4: Feedback the real-time fault detection results and multi-level analysis results to the fault detection model, and use the reinforcement learning algorithm to perform online updates based on the feedback results; The specific steps of S4.3 include: S4.31: Acquisition and preliminary fault detection results; S4.32: Extract fault knowledge from historical fault cases, expert experience, and equipment manuals, and use the knowledge graph method to represent the extracted fault knowledge as a graphical knowledge structure, where the fault type, fault location, and fault probability information are represented as nodes, and the relationship between them is represented as edges; S4.33: Input the preliminary fault detection results into the preloaded expert system. The expert system matches and infers the fault detection results according to the fault knowledge in the knowledge graph to obtain the fault type, fault location and fault probability information; S4.34: Based on the reasoning results of the expert system, combined with , determine the specific type and location of the fault; S4.35: Use machine learning algorithms to analyze historical failure cases to obtain the probability distribution of failures, and evaluate the probability of failures based on the current operating status of the equipment, the specific type and location of the failure; The specific steps of S4.4 include: S4.41: Integrate the real-time fault detection results and multi-level analysis results to generate a fault detection report; S4.42: Load the pre-built reinforcement learning model, use the fault detection report as the feedback result, input it into the reinforcement learning model, design the reward function and state transition rules, use the reinforcement learning strategy to train the reinforcement learning model, so that the reinforcement learning model can learn and adapt to the new fault mode, and obtain the trained reinforcement learning model. The formula is: ； in, represents the quality function under state z, action d, fault feature f and context information g at time t+1, represents the quality function under state z, action d, fault feature f and context information g at time t, represents the learning rate, represents the reward obtained immediately after taking action d, represents the discount factor, represents the maximum value of the quality function under the next state, action, fault characteristics and context information at time t, , , , Respectively represent the next state, action, fault characteristics and context information, represents the feature weight, represents the characteristic function, represents the constraint weight, represents the constraint function; S4.43: According to the training results of the reinforcement learning model, adjust the parameters of the fault detection model, and evaluate the accuracy of the adjusted fault detection model. At the same time, according to the evaluation results, further optimize and adjust the fault detection model; If the evaluation result is not satisfactory, return to step S4.42, retrain the reinforcement learning model, and update the fault detection model again; If the evaluation result meets the requirements, the result is output and the process ends.

2. The large model-based fault detection method according to claim 1, characterized in that: The specific formula of the adaptive optimization strategy in step S3 is: ； Among them, t represents the current time, and They represent the first-order moment estimation and second-order moment estimation of the fault detection model at time t, respectively. represents the fault detection model parameters at time t+1, represents the fault detection model parameters at time t, represents the step size of the control parameter update, and represents the exponential decay rate of the moment estimate at time t, and b represents a constant.

3. A large model-based fault detection system, which is used to implement the large model-based fault detection method according to any one of claims 1 to 2, characterized in that: include: Data processing module, model fine-tuning module, fault detection module, fault analysis module; The data processing module is used to collect multi-source fault data from the industrial system and perform preprocessing and feature extraction; The model fine-tuning module is used to fine-tune the pre-trained large language model using the fault feature dataset; The fault detection module is used to construct a fault detection model using the fine-tuned large language model and the fault feature data set, and to perform training and evaluation, and optimize the model according to the evaluation results; The fault analysis module is used to input the real-time collected fault data into the trained fault detection model, output the preliminary fault detection results, and perform multi-level analysis in combination with the multi-level fault analysis method.

4. The large model-based fault detection system according to claim 3, characterized in that: The fault analysis module includes: a real-time detection unit, a multi-level analysis unit, and a feedback optimization unit; The real-time detection unit is used to input the real-time collected fault data into the trained fault detection model and output preliminary fault detection results; The multi-level analysis unit is used to perform multi-level analysis on the preliminary fault detection results; The feedback optimization unit is used to feed back the real-time detection results and the results of the multi-level analysis to the fault detection model for optimization.

5. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the large model-based fault detection method according to any one of claims 1 to 2 are implemented.

6. A computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and when the computer instructions are executed, the steps of the large model-based fault detection method described in any one of claims 1-2 are executed.

Images