Multi-round debate equipment fault reason analysis method and device based on large language model, and storage medium
Through the multi-round debate method of large language model, combined with equipment state timing data and regression model, a candidate set of equipment failure causes is generated, and multiple large language models are used to conduct multiple rounds of debate, solving the problem of difficult to deeply analyze the physical relationship of equipment failure in traditional methods, and achieving high accuracy and low cost failure cause analysis.
Patent Information
- Application Number
- CN202510537121.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-27
- Publication Date
- 2025-08-15
AI Technical Summary
In the analysis of the cause of equipment failures, traditional methods rely on deviation detection, making it difficult to deeply explore the physical relationships behind the failures. In addition, artificial intelligence-based methods are costly and have weak generalization capabilities, so they cannot effectively utilize user feedback and expert knowledge.
A multi-round debate method based on a large language model is adopted, and a time series data of equipment states is obtained, pre-processed and residual calculation is performed, and a regression model is used to generate a candidate set of failure causes, and multiple rounds of debate are conducted through multiple large language models, combining physical principles and expert knowledge to optimize the analysis of failure causes.
Improve the accuracy of equipment failure cause analysis and generalization ability to new tasks, reduce costs, and ensure the robustness and reliability of analysis results.
Smart Images
Figure CN120493201A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent equipment management technology, and in particular to a method, device and storage medium for analyzing the cause of equipment failure based on a multi-round debate of a large language model. Background Art
[0002] With the proliferation of intelligent devices and the development of IoT technology in the industrial sector, data-driven asset health management is becoming increasingly important. The widespread use of sensors generates large amounts of time-series data from industrial equipment, providing valuable information for fault detection and maintenance. Traditional fault cause analysis methods typically rely on detecting deviations or anomalies, often only capturing the surface symptoms of equipment failures. However, the underlying physical relationships behind the failures often lead to low analysis accuracy. AI-based equipment fault diagnosis methods require large amounts of labeled data to train models, which is costly and has poor generalization capabilities to new tasks.
[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention
[0004] The embodiments of the present invention provide a method, device and storage medium for analyzing the causes of multi-round debate equipment failures based on a large language model, which effectively improves the analysis accuracy and the ability to generalize to new tasks and reduces costs.
[0005] On the one hand, an embodiment of the present invention provides a method for analyzing the cause of a multi-round debate device failure based on a large language model, comprising the following steps:
[0006] Get device status time series data;
[0007] Preprocessing the device status time series data to obtain target time series data;
[0008] Calculating residuals and abnormal data based on the target time series data;
[0009] generating a candidate set of equipment failure causes using a regression model according to the residual and the target time series data;
[0010] Generate a device failure cause analysis result set using multiple large language models based on the abnormal data and the device failure cause candidate set;
[0011] Based on the equipment failure cause analysis result set, multiple rounds of debate are conducted through the multiple large language models to obtain the target equipment failure cause analysis result.
[0012] In some embodiments, preprocessing the device status time series data to obtain target time series data includes:
[0013] Denoising the device status time series data to obtain first time series data;
[0014] Normalizing the first time series data to obtain second time series data;
[0015] Missing values are filled in the second time series data to obtain the target time series data.
[0016] In some embodiments, calculating residuals and abnormal data based on the target time series data includes:
[0017] Inputting the target time series data into a time series prediction model function to obtain a predicted value;
[0018] Calculating the residual according to the predicted value and the actual observed value;
[0019] If the residual is greater than a preset abnormality threshold, the data point at the time corresponding to the residual is taken as the abnormal data.
[0020] In some embodiments, generating a candidate set of equipment failure causes using a regression model based on the residual and the target time series data includes:
[0021] Inputting the residual and the target time series data into the regression model to obtain a significance map;
[0022] Calculate the importance score corresponding to each channel in the saliency map according to the total number of channels and the embedding matrix;
[0023] sorting the plurality of importance scores to obtain an ordered score set;
[0024] According to a preset number of channels, a plurality of target importance scores are selected from the ordered score set as the candidate set of equipment failure causes, wherein the plurality of target importance scores are importance scores ranked first by the preset number of channels.
[0025] In some embodiments, generating a device failure cause analysis result set using multiple large language models based on the abnormal data and the device failure cause candidate set includes:
[0026] Initializing the multiple large language models;
[0027] Inputting the abnormal data and the candidate set of equipment failure causes into the multiple large language models, so that the multiple large language models respectively identify the failure causes and obtain corresponding equipment failure cause rankings and reasons;
[0028] The plurality of equipment failure cause rankings and reasons are combined to obtain the equipment failure cause analysis result set.
[0029] In some embodiments, the performing of multiple rounds of debates using the multiple large language models based on the device failure cause analysis result set to obtain a target device failure cause analysis result includes:
[0030] Using the multiple large language models, the ranking and reason of each device failure cause in the device failure cause analysis result set are evaluated to obtain corresponding evaluation opinions;
[0031] According to the plurality of evaluation opinions, the equipment failure cause analysis result set is updated, and the opinion consistency is calculated according to the equipment failure cause analysis result set until the opinion consistency is greater than a preset consistency threshold, thereby obtaining the target equipment failure cause analysis result.
[0032] In some embodiments, calculating the importance score corresponding to each channel in the saliency map according to the total number of channels and the embedding matrix includes:
[0033] According to the total number of channels and the embedding matrix, the importance score corresponding to each channel in the saliency map is calculated using the importance score calculation formula, which is:
[0034]
[0035] Where R t,i is the importance score of the i-th channel at the t-th moment, d is the total number of channels, U t,i,j is the embedding matrix.
[0036] On the other hand, an embodiment of the present invention provides a device for analyzing the cause of a multi-round debate device failure based on a large language model, comprising:
[0037] The first module is used to obtain device status time series data;
[0038] The second module is used to preprocess the device status time series data to obtain target time series data;
[0039] The third module is used to calculate residuals and abnormal data based on the target time series data;
[0040] A fourth module is configured to generate a candidate set of equipment failure causes using a regression model based on the residual and the target time series data;
[0041] A fifth module is configured to generate a device failure cause analysis result set using multiple large language models based on the abnormal data and the device failure cause candidate set;
[0042] The sixth module is used to conduct multiple rounds of debates using the multiple large language models based on the equipment failure cause analysis result set to obtain the target equipment failure cause analysis result.
[0043] In another aspect, an embodiment of the present invention provides a computer device, comprising:
[0044] at least one processor;
[0045] at least one memory for storing at least one program;
[0046] When the at least one program is executed by the at least one processor, the at least one processor implements the method.
[0047] On the other hand, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described above is implemented.
[0048] The beneficial effects of the present invention are as follows:
[0049] The embodiment of the present invention first obtains device status time series data, preprocesses the device status time series data to obtain target time series data, then calculates residuals and abnormal data based on the target time series data, and uses a regression model to generate a candidate set of device failure causes based on the residuals and the target time series data. Then, based on the abnormal data and the candidate set of device failure causes, a device failure cause analysis result set is generated through multiple large language models. Finally, based on the device failure cause analysis result set, multiple rounds of debate are performed through multiple large language models to obtain target device failure cause analysis results, thereby enabling device failure cause analysis to be implemented through multiple large language models, thereby improving analysis accuracy and generalization ability for new tasks and reducing costs.
[0050] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the structures particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.
[0052] Figure 1 This is a flow chart of a method for analyzing the cause of a multi-round debate device failure based on a large language model according to an embodiment of the present invention;
[0053] Figure 2A schematic diagram of an overall process for generating the root cause of a device failure according to an embodiment of the present invention;
[0054] Figure 3 This is a schematic diagram of the structure of a device for analyzing the cause of a multi-round debate device failure based on a large language model according to an embodiment of the present invention;
[0055] Figure 4 The figure is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0056] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present application. They are merely examples of devices and methods consistent with some aspects of the embodiments of the present application as detailed in the appended claims.
[0057] It will be understood that the terms "first", "second", etc. used in this application may be used herein to describe various concepts, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the words "if" and "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".
[0058] The terms "at least one", "plurality", "each", "any", etc. used in this application include "at least one", "two" or more, "plurality" or "each", "any" or "any one", "each" or "any one" as used herein.
[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.
[0060] Before explaining the embodiments of the present application in detail, some of the nouns and terms involved in the embodiments of the present application are first explained. The nouns and terms involved in the embodiments of the present application are subject to the following explanations.
[0061] Large Language Models (LLMs) are deep learning models trained using large amounts of text data, enabling them to generate natural language text or understand the meaning of text. These models can be trained on massive datasets to provide in-depth knowledge and language production on a variety of topics. Their core idea is to learn the patterns and structure of natural language through large-scale unsupervised training, simulating, to a certain extent, the human language cognition and generation process.
[0062] With the increasing prevalence of intelligent devices and the development of the Internet of Things (IoT) in the industrial sector, data-driven asset health management (AHM) is becoming increasingly important. The widespread use of sensors generates massive amounts of time-series data from industrial equipment, providing valuable information for fault detection and maintenance. However, traditional methods for analyzing the cause of equipment failure typically rely on detecting deviations or anomalies, often only capturing the surface symptoms of equipment failures while failing to delve deeper into the underlying root causes. Although machine learning (ML) and deep learning (DL) methods have been applied to industrial fault diagnosis, existing technologies still suffer from poor interpretability, low accuracy, and insufficient integration of physical laws and expert experience when dealing with complex and volatile industrial environments. This limits their practical application. Current fault diagnosis systems often lack sufficient reliability and accuracy, especially in identifying root causes. Many data-driven models can only infer potential fault causes based on residuals in time-series data, but this approach often ignores the underlying physical relationships, resulting in unreliable model results. Furthermore, existing technologies struggle to effectively incorporate domain expert experience and physical principles into the analysis process, compromising the authenticity and precision of the results and resulting in low accuracy.
[0063] AI-based equipment fault diagnosis relies on AI technologies, such as deep learning and machine learning models, to analyze equipment operating data to identify potential faults. This approach first collects equipment operating data through sensors, cleans it, and normalizes it. Key features, such as vibration, temperature, and pressure, are then extracted from this preprocessed data. Multiple candidate codes are generated for the same task sampling. Machine learning algorithms (such as neural networks and support vector machines) are then used to train the extracted features to build a fault diagnosis model. Finally, the trained model is fed with real-time monitoring data for fault prediction and diagnosis. However, model training requires a large amount of labeled data, which is costly to acquire and annotate. Complex machine learning models lack transparency, making their decision-making processes difficult to explain. The model may perform poorly for unseen fault types and have poor generalization capabilities to new tasks.
[0064] An industrial fault diagnosis method based on knowledge graphs and large language models builds a knowledge graph of equipment and faults, combined with the natural language processing capabilities of large language models. This method can extract key information from historical fault reports, helping operations and maintenance personnel quickly locate fault causes. This method first collects domain expert knowledge and historical fault data to construct a knowledge graph of equipment, faults, and repairs. It then performs natural language processing on the collected fault descriptions to extract key information. The large language model is then trained on this processed data to learn how to generate fault diagnoses and repair recommendations. Finally, new fault descriptions are input, and the model generates possible fault causes and repair recommendations. However, building a high-quality knowledge graph requires the involvement of a large number of domain experts, which is costly and time-consuming. Training a large language model requires significant computing resources and is complex. Furthermore, general-purpose large language models may lack domain-specific expertise, resulting in inaccurate diagnostic results.
[0065] Existing methods rely too heavily on the data costs of large numbers of high-quality code examples and fail to effectively leverage knowledge from user feedback. Due to the scarcity of high-quality code data and the high cost of annotation, large-scale supervised learning approaches are not practical. Research is needed to explore strategies for leveraging user natural language feedback to improve model performance on small datasets, reducing reliance on data size and improving sample efficiency. Existing methods are unable to effectively incorporate user interaction feedback into model training, making it difficult to continuously improve model generation quality. Most methods only use feedback for online adjustments, failing to fully incorporate knowledge gained from interactions into the model.
[0066] In light of this, the embodiments of the present invention utilize multiple large language models to analyze the causes of equipment failures and conduct multiple rounds of debate and evaluation to determine the final cause of equipment failures. This improves the accuracy and reliability of fault diagnosis. It not only processes complex data but also integrates physical principles and expert knowledge to provide more robust and reliable root cause analysis. The language model can be effectively trained using user feedback, enabling the model to actively learn user-guided knowledge, thereby achieving continuously enhanced generative capabilities.
[0067] The method for analyzing the cause of failure of a multi-round debate device based on a large language model provided in an embodiment of the present application relates to the field of intelligent device management technology. The method for analyzing the cause of failure of a multi-round debate device based on a large language model provided in an embodiment of the present application can be applied to a terminal, can be applied to a server, or can be software running in a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, and a car terminal, etc., but is not limited to this; the server side can be configured as an independent physical server, or can be configured as a server cluster or distributed system composed of multiple physical servers, and can also be configured as a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application that implements the method for analyzing the cause of failure of a multi-round debate device based on a large language model, etc., but is not limited to the above forms.
[0068] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.
[0069] The following is a detailed explanation of the embodiments of the present application with reference to the accompanying drawings:
[0070] Figure 1 This is an optional flowchart of the method for analyzing the cause of a multi-round debate device failure based on a large language model provided in an embodiment of the present application. Figure 1 The method may include but is not limited to steps S101 to S106.
[0071] Step S101: Acquire device status time series data;
[0072] Step S102: pre-process the device status time series data to obtain target time series data;
[0073] Step S103: Calculate residuals and abnormal data based on the target time series data;
[0074] Step S104: Generate a candidate set of equipment failure causes using a regression model based on the residual and target time series data;
[0075] Step S105: Generate a device failure cause analysis result set using multiple large language models based on the abnormal data and the device failure cause candidate set;
[0076] Step S106: Based on the equipment failure cause analysis result set, multiple rounds of debate are conducted through multiple large language models to obtain the target equipment failure cause analysis result.
[0077] Steps S101 to S106 shown in the embodiment of the present application implement equipment failure cause analysis, improve analysis accuracy and generalization capabilities for new tasks, and reduce costs.
[0078] In step S101 of some embodiments, the device status time series data can be obtained through sensors of the industrial equipment. The device status time series data can also be obtained through other means, not limited thereto. The device status time series data is a high-frequency time series data, which can include pressure, temperature, or flow, etc.
[0079] In some embodiments, in step S102, preprocessing the device status time series data to obtain target time series data may include but is not limited to the following steps:
[0080] Denoising the device status time series data to obtain first time series data;
[0081] Normalizing the first time series data to obtain second time series data;
[0082] Fill missing values in the second time series data to obtain target time series data.
[0083] In some embodiments, the device status time series data can be first denoised to obtain first time series data, then the first time series data can be standardized to obtain second time series data, and then the second time series data can be filled with missing values to obtain target time series data to ensure the quality of the input data.
[0084] In some embodiments, in step S103, calculating residuals and abnormal data based on target time series data may include but is not limited to the following steps:
[0085] Input the target time series data into the time series prediction model function to obtain the predicted value;
[0086] Calculate the residuals based on the predicted values and the actual observed values;
[0087] If the residual is greater than the preset abnormal threshold, the data point at the time corresponding to the residual is regarded as abnormal data.
[0088] In some embodiments, the target time series data may be first input into the time series prediction model function to obtain a predicted value, wherein the calculation formula of the predicted value is: Where, is the model's predicted value for time t+1, which is the result of inferring the future state based on past time series data. f(*) represents the model function used for time series prediction, such as LSTM, CNN or other deep learning models. This function receives historical data within a time window as input and outputs the predicted value for the next moment, X t-l+1:t The target time series data input within the time window is the time series data from time t-l+1 to time t, where l represents the length of the time window. This data is used to predict the state at the next moment. It is understandable that a prediction model based on time series data (such as LSTM, CNN, etc.) can be used to predict the normal operation state of the device. This model is trained using historical time series data to predict the device state at future moments. The residual is then calculated based on the predicted value and the actual observed value. The residual calculation formula is: Where R t+1 is the residual at time t+1, X t+1 is the actual observation value at time t+1, that is, the real data point. The residual between the predicted value and the actual observation value can be calculated to detect whether there is an anomaly. If the residual is greater than the preset anomaly threshold τ, the data point at the time corresponding to the residual is regarded as an anomaly. For example, if R t+1 If the threshold τ is exceeded, which is set by statistical methods or experience, then the system is considered to have an abnormality at time t+1, which can be expressed mathematically as: Where Anomaly is the anomaly detection function, and τ is the preset anomaly threshold. Large residuals indicate equipment failure or other anomalies. High residuals can be used to determine whether the actual observations may contain faults. These anomaly periods can be marked and used as input for subsequent analysis.
[0089] In some embodiments, in step S104, generating a candidate set of equipment failure causes using a regression model based on the residual and target time series data may include but is not limited to the following steps:
[0090] Input the residual and target time series data into the regression model to obtain the significance map;
[0091] Calculate the importance score of each channel in the saliency map based on the total number of channels and the embedding matrix;
[0092] Sort multiple importance scores to obtain an ordered score set;
[0093] According to the preset number of channels, multiple target importance scores are selected from the ordered score set as a candidate set of equipment failure causes, and the multiple target importance scores are importance scores ranked first by the preset number of channels.
[0094] In some embodiments, the residuals in the regression prediction model are used on the detected abnormal data to identify the root cause of the abnormality. The expression of the prediction residual is: Where g(*) is the regression model. The residual and target time series data can be first input into the regression model to obtain a saliency map, highlighting the key features that may be related to the anomaly. The saliency map can calculate the contribution of each input channel to the high residual by regressing the residual and backpropagating the gradient (or similar methods). By processing the output of the regression model, the key channels that affect the anomaly are identified. Then, based on the total number of channels and the embedding matrix, the importance score calculation formula is used to calculate the importance score corresponding to each channel in the saliency map. The importance score calculation formula is: Where R t,i is the importance score of the i-th channel at the t-th moment, d is the total number of channels, used to calculate the contribution of each channel, U t,i,j is the embedding matrix extracted by the neural network. It can be understood that the saliency map is a matrix R, which includes multiple elements R t,i . The importance score provides a preliminary clue for locating the root cause of the equipment failure. Then, multiple importance scores are sorted to obtain an ordered score set, and finally, based on the preset number of channels, multiple target importance scores are selected from the ordered score set as the candidate set of equipment failure causes, where the multiple target importance scores are the importance scores of the preset number of channels ranked first. It can be understood that the candidate set of equipment failure causes is composed of channels with higher importance scores. For example, if the score of a channel is significantly higher than that of other channels, it will be given priority as the root cause of the potential equipment failure.
[0095] In some embodiments, in step S105, generating a device failure cause analysis result set using multiple large language models based on the abnormal data and the device failure cause candidate set may include but is not limited to the following steps:
[0096] Initialize multiple large language models;
[0097] Input abnormal data and candidate sets of equipment failure causes into multiple large language models, so that the multiple large language models can identify the failure causes respectively and obtain the corresponding equipment failure cause ranking and reasons;
[0098] The causes of multiple equipment failures are ranked and their reasons are combined to obtain an equipment failure cause analysis result set.
[0099] In some embodiments, multiple large language models can be initialized first, and then the abnormal data and the candidate set of equipment failure causes are input into the multiple large language models, so that the multiple large language models can perform fault cause identification respectively, obtain the corresponding equipment failure cause ranking and reasons, and then combine the multiple equipment failure cause rankings and reasons to obtain the equipment failure cause analysis result set. For example, based on the abnormal data and the candidate set of equipment failure causes, the high residual channels (such as temperature and pressure anomalies) detected in the abnormal data and the top k channels screened out from the saliency map of the equipment failure cause candidate set can be input into a pre-trained large-scale language model (LLM) (such as GPT-3, GPT-4, etc.). The LLM is used to analyze the correlation between the candidate causes and the system symptoms to help identify the true root cause. Multiple LLMs can be used for fault cause identification, and each LLM independently generates its own equipment failure cause ranking and reasons based on the same input. The specific number of LLMs can be adjusted according to task requirements or resource constraints (such as 3, 4 or more), and the effect is verified through experiments.
[0100] In some embodiments, in step S106, based on the device failure cause analysis result set, multiple rounds of debate are conducted using multiple large language models to obtain the target device failure cause analysis result, which may include but is not limited to the following steps:
[0101] Multiple large language models are used to evaluate the ranking and reasons of each equipment failure cause in the equipment failure cause analysis result set, and corresponding evaluation opinions are obtained;
[0102] Based on multiple evaluation opinions, the equipment failure cause analysis result set is updated, and the opinion consistency is calculated based on the equipment failure cause analysis result set until the opinion consistency is greater than the preset consistency threshold, and the target equipment failure cause analysis result is obtained.
[0103] In some embodiments, multiple large language models can first evaluate the ranking and justification of each device failure cause in the device failure cause analysis result set to obtain corresponding evaluation opinions. For example, in the second and subsequent rounds, each LLM needs to evaluate the opinions of other LLMs (i.e., the ranking and justification of the device failure causes) and adjust its own judgment based on these opinions. The device failure cause analysis result set is then updated based on the multiple evaluation opinions, and the consensus is calculated based on the device failure cause analysis result set until the consensus exceeds a preset consensus threshold, resulting in the target device failure cause analysis result. Each LLM will reconsider its initial position in different rounds until a consensus is ultimately reached. At the end of each round, a "clerk" role can be used to ensure the consistency of the opinions of all LLMs. Only if a majority of LLMs agree on a particular ranking will it be used as the final root cause ranking, resulting in the target device failure cause analysis result. It is understood that the "clerk" role is an independent system module or rule engine, not one of the LLMs participating in the debate. It can be an automated script or a lightweight LLM that only checks whether the voting results meet the termination conditions. The LLMs of this embodiment will evaluate and correct each other in multiple rounds and eventually reach a consensus, thereby reducing the deviation that may be caused by a single LLM. Each LLM provides its analysis results and ensures the accuracy of the consensus through a self-exclusion voting mechanism. Furthermore, this embodiment uses multiple rounds of debate and voting mechanisms to ultimately select the most likely root cause and sort them by priority to obtain the target equipment failure cause analysis results. This result combines physical principles and the experience of domain experts, making the root cause of equipment failure more credible and explainable. Combining the analysis method of LLM debate and data-driven model, the accuracy of the root cause analysis of equipment failure is significantly improved, especially in the application of complex industrial systems, and can correctly distinguish between symptom signals and the root cause of equipment failure.
[0104] In some embodiments, the overall process for generating the root cause of a device failure is as follows: Figure 2 As shown in the figure, data collection and preprocessing can be performed first, and then anomaly detection can be performed on the preprocessed data to obtain abnormal data, and residual analysis and potential fault channel identification can be performed to obtain prediction residuals. The abnormal data and prediction residuals are then input into the regression model to generate a significance map, and the importance score of each channel is analyzed from the significance map. The score is input into the large language model for fault cause identification to obtain the ranking and reasons of equipment failure causes. Finally, through multiple rounds of LLM debates and self-exclusion voting, the ranking and reasons of equipment failure causes of multiple LLMs can reach a consensus, or the degree of consensus is greater than the preset consistency threshold, and the final ranking and reasons of equipment failure causes are obtained as the target equipment failure cause analysis results.
[0105] In some embodiments, this embodiment, by combining a data-driven deep learning model with a large-scale language model (LLM), can introduce more domain knowledge and physical principles into the fault cause analysis, making the root cause analysis of equipment failures more accurate. This combination not only avoids the limitations of a single data-driven model in traditional methods, but also optimizes the output of the model through the LLM, ensuring the high accuracy of the analysis results. This embodiment adopts a multi-round self-exclusion voting mechanism. Through debate and correction between multiple LLMs, it eliminates the deviations that may be caused by a single model or language model, making the root cause analysis results of equipment failures more stable and reliable. Each round of debate and feedback further strengthens the reliability of the final results, making it meet the actual industrial application needs. This embodiment, by combining a saliency map with an LLM, can accurately identify potential root causes related to anomalies, especially in multi-device collaborative fault analysis, and can explore the complex causal relationships between devices, so that the system can more accurately locate the source of the problem when facing complex, nonlinear failure modes.
[0106] In some embodiments, equipment failure cause analysis can also be performed through other methods, including ensemble learning and graph neural networks. In root cause analysis based on ensemble learning, ensemble learning methods improve analysis accuracy and stability by combining the prediction results of multiple machine learning models. In root cause analysis, multiple models (such as LSTM, random forest, support vector machine, etc.) can be used for fault detection and root cause identification, and the outputs of each model are integrated into the final result using weighted voting or stacking strategies. Ensemble learning methods can reduce the bias that may be introduced by a single model while improving the reliability of analysis through the complementarity between models. The advantage of ensemble learning is that it can enhance system stability and avoid the bias and overfitting problems caused by reliance on a single model. It is particularly suitable for industrial fault detection scenarios with large data volumes and high complexity. In root cause analysis based on graph neural networks (GNNs), GNNs are suitable for handling complex inter-device relationships and can effectively model industrial equipment and the propagation of faults between them. In GNNs, industrial equipment and their states are represented as nodes and edges in a graph, with interactions between devices serving as edge features. Through graph convolutional networks (GCNs), GNNs can capture dynamic dependencies between devices and analyze the mutual impact of different devices during failures. The advantage of graph neural networks is that they can handle complex causal relationships in multi-device failures. They are particularly suitable for collaborative failure analysis of multiple devices and provide more accurate root cause identification.
[0107] In some embodiments, this embodiment combines a traditional data-driven deep root cause analysis (RCA) model with a pre-trained large-scale language model (LLM) to form a framework that combines data-driven analysis and LLM, which can give full play to the advantages of data-driven models in processing complex time series data, and at the same time integrate the knowledge and physical principles of domain experts through LLM to optimize the fault analysis results. This framework can effectively improve the accuracy and interpretability of the root cause analysis of equipment failures, and avoid the limitations of relying solely on data-driven models. It has high technical innovation and application value. At the same time, in order to further improve the reliability of the analysis results, this embodiment proposes an innovative multi-round debate strategy, in which multiple LLMs participate in the analysis and ranking of the root causes, and each LLM evaluates, corrects and optimizes the judgment of the root causes in multiple rounds of debate. Through the self-exclusion voting mechanism, LLM can adjust according to the opinions of other LLMs to avoid directly bypassing the LLM system, and achieve the same purpose by simplifying the analysis process, thereby ensuring that the final analysis results are more accurate and stable.
[0108] The beneficial effects of implementing the embodiments of the present invention include: the embodiments of the present invention first obtain device status time series data, preprocess the device status time series data to obtain target time series data, then calculate residuals and abnormal data based on the target time series data, and generate a candidate set of device failure causes using a regression model based on the residuals and the target time series data, and then generate a device failure cause analysis result set based on the abnormal data and the device failure cause candidate set through multiple large language models, and finally, based on the device failure cause analysis result set, conduct multiple rounds of debate through multiple large language models to obtain the target device failure cause analysis result, so that the device failure cause analysis can be realized through multiple large language models, thereby improving the analysis accuracy and the generalization ability of new tasks and reducing costs.
[0109] like Figure 3 As shown, an embodiment of the present invention further provides a device for analyzing the cause of a multi-round debate device failure based on a large language model, comprising:
[0110] The first module 801 is used to obtain device status time series data;
[0111] The second module 802 is used to pre-process the device status time series data to obtain target time series data;
[0112] The third module 803 is used to calculate residuals and abnormal data based on the target time series data;
[0113] The fourth module 804 is used to generate a candidate set of equipment failure causes using a regression model based on the residual and target time series data;
[0114] The fifth module 805 is configured to generate a device failure cause analysis result set using multiple large language models based on the abnormal data and the device failure cause candidate set;
[0115] The sixth module 806 is used to conduct multiple rounds of debates using multiple large language models based on the equipment failure cause analysis result set to obtain the target equipment failure cause analysis result.
[0116] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0117] like Figure 4 As shown, an embodiment of the present invention further provides a computer device, including:
[0118] at least one processor 901;
[0119] At least one memory 902, configured to store at least one program;
[0120] When at least one program is executed by at least one processor, the at least one processor implements Figure 1 The method shown.
[0121] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0122] The embodiment of the present invention further provides a computer-readable storage medium, which stores a computer program, which is executed by a processor to implement Figure 1 The method shown.
[0123] The contents of the above method embodiments are all applicable to the present storage medium embodiment. The functions specifically implemented by the present storage medium embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0124] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and essence of the present invention should be within the scope of the present invention.
Claims
1. A multi-round debate device failure cause analysis method based on a large language model, characterized in that: The following steps are involved: Get device status time series data; Preprocessing the device status time series data to obtain target time series data; Calculating residuals and abnormal data based on the target time series data; generating a candidate set of equipment failure causes using a regression model according to the residual and the target time series data; Generate a device failure cause analysis result set using multiple large language models based on the abnormal data and the device failure cause candidate set; Based on the equipment failure cause analysis result set, multiple rounds of debate are conducted through the multiple large language models to obtain the target equipment failure cause analysis result.
2. The method according to claim 1, characterized in that The preprocessing of the device status time series data to obtain target time series data includes: Denoising the device status time series data to obtain first time series data; Normalizing the first time series data to obtain second time series data; Missing values are filled in the second time series data to obtain the target time series data.
3. The method according to claim 1, characterized in that The calculating of residuals and abnormal data according to the target time series data includes: Inputting the target time series data into a time series prediction model function to obtain a predicted value; Calculating the residual according to the predicted value and the actual observed value; If the residual is greater than a preset abnormality threshold, the data point at the time corresponding to the residual is taken as the abnormal data.
4. The method according to claim 1, wherein Generating a candidate set of equipment failure causes using a regression model based on the residual and the target time series data includes: Inputting the residual and the target time series data into the regression model to obtain a significance map; Calculate the importance score corresponding to each channel in the saliency map according to the total number of channels and the embedding matrix; sorting the plurality of importance scores to obtain an ordered score set; According to a preset number of channels, a plurality of target importance scores are selected from the ordered score set as the candidate set of equipment failure causes, wherein the plurality of target importance scores are importance scores ranked first by the preset number of channels.
5. The method according to claim 1, wherein Generating a device failure cause analysis result set using multiple large language models based on the abnormal data and the device failure cause candidate set includes: Initializing the multiple large language models; Inputting the abnormal data and the candidate set of equipment failure causes into the multiple large language models, so that the multiple large language models respectively identify the failure causes and obtain corresponding equipment failure cause rankings and reasons; The plurality of equipment failure cause rankings and reasons are combined to obtain the equipment failure cause analysis result set.
6. The method according to claim 1, characterized in that The method of performing multiple rounds of debates based on the device failure cause analysis result set using the multiple large language models to obtain a target device failure cause analysis result includes: Using the multiple large language models, the ranking and reason of each device failure cause in the device failure cause analysis result set are evaluated to obtain corresponding evaluation opinions; According to the plurality of evaluation opinions, the equipment failure cause analysis result set is updated, and the opinion consistency is calculated according to the equipment failure cause analysis result set until the opinion consistency is greater than a preset consistency threshold, thereby obtaining the target equipment failure cause analysis result.
7. The method according to claim 4, characterized in that Calculating the importance score corresponding to each channel in the saliency map according to the total number of channels and the embedding matrix includes: According to the total number of channels and the embedding matrix, the importance score corresponding to each channel in the saliency map is calculated using the importance score calculation formula, which is: Where R t,i is the importance score of the i-th channel at the t-th moment, d is the total number of channels, U t,i,j is the embedding matrix.
8. A multi-round debate equipment failure cause analysis device based on a large language model, characterized in that: include: The first module is used to obtain device status time series data; The second module is used to preprocess the device status time series data to obtain target time series data; The third module is used to calculate residuals and abnormal data based on the target time series data; A fourth module is configured to generate a candidate set of equipment failure causes using a regression model based on the residual and the target time series data; A fifth module is configured to generate a device failure cause analysis result set using multiple large language models based on the abnormal data and the device failure cause candidate set; The sixth module is used to conduct multiple rounds of debates using the multiple large language models based on the equipment failure cause analysis result set to obtain the target equipment failure cause analysis result.
9. A computer device, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.