Device anomaly root cause diagnosis methods, equipment, and media based on large models and intelligent agents.
By collaborating with a multi-agent system and a large model, and combining time-series large models and multi-scale feature analysis, the accuracy and efficiency of root cause diagnosis of power generation equipment anomalies are solved, enabling rapid and accurate root cause localization and fault handling, and adapting to the equipment monitoring needs of complex environments.
Patent Information
- Application Number
- CN202611449208.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies suffer from insufficient accuracy, low fault handling efficiency, and poor adaptability and scalability in the diagnosis of root causes of power generation equipment anomalies. In particular, they are difficult to quickly and accurately locate the root causes of anomalies when faced with complex data and diverse data types.
By employing a multi-agent system and a large model working together, and through data-driven and knowledge-guided reasoning mechanisms, the complex root cause diagnosis task is decomposed into multiple sub-tasks, which are handled by different agents. Combining device network topology relationships, historical fault cases, and domain knowledge, the root cause diagnosis is carried out collaboratively using a time-series large model and multi-scale feature analysis.
It enables accurate and rapid diagnosis of the root causes of equipment anomalies, reduces false positives and false negatives, improves fault handling efficiency, reduces operation and maintenance costs, adapts to different equipment and environments, and has adaptability and scalability.
Smart Images

Figure CN123114832A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of equipment fault diagnosis technology, specifically to a method, device, and medium for diagnosing the root causes of equipment anomalies based on a large model and intelligent agents. Background Technology
[0002] In the operation of power generation equipment, such as thermal power generating units, abnormal situations frequently occur. Traditional methods for diagnosing the root causes of equipment anomalies mainly rely on human experience, relying on reviewing equipment operation logs, monitoring indicators, and other data, combined with the knowledge and experience of technical personnel, to determine the cause of the anomaly. However, with the continuous increase in equipment complexity, the amount of data generated by the equipment is exponentially increasing, and the data types are diverse, including structured data (such as performance indicators) and unstructured data (such as log text). Manual analysis is not only inefficient but also prone to omissions and misjudgments, making it difficult to quickly and accurately locate the root cause of the anomaly.
[0003] While existing machine learning-based anomaly detection methods have improved efficiency to some extent, they typically only detect whether a device is malfunctioning, offering limited ability to diagnose the root causes of anomalies. They often lack a deep understanding and reasoning ability regarding complex data relationships, failing to extract the key factors leading to anomalies from massive datasets. Summary of the Invention
[0004] This invention provides a method, device, and medium for diagnosing the root causes of equipment anomalies based on a large model and intelligent agents. By collaborating with a large model through a multi-agent system, the complex root cause diagnosis task is decomposed into multiple sub-tasks, which are then processed by different intelligent agents. Combined with an inference mechanism, this addresses the problems of insufficient accuracy in diagnosing the root causes of equipment anomalies, low fault handling efficiency, and poor adaptability and scalability in the prior art.
[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A method for diagnosing the root causes of device anomalies based on large models and intelligent agents, which involves performing the following steps via computer equipment: S1. Collect log data, performance index data and historical event records during the operation of the power generation equipment, and perform cleaning, noise reduction and annotation preprocessing on the log data, performance index data and historical event records to obtain preprocessed equipment data for anomaly detection in large models. S2. Use a finely tuned and trained large model to perform anomaly detection on the preprocessed device data, identify whether the device is abnormal, and record the abnormal data when an anomaly occurs. The abnormal data includes the anomaly type and the degree of anomaly, and is used for root cause diagnosis by the root cause diagnosis agent in the multi-agent system. S3. By using the root cause diagnosis agent in the multi-agent system, combined with the device network topology, historical fault cases and domain knowledge, root cause diagnosis is performed on abnormal data to obtain root cause diagnosis results. The root cause diagnosis agent adopts a reasoning mechanism that integrates data-driven, time-series large model and knowledge guidance, and coordinates time-series pattern recognition and multi-scale feature analysis to achieve root cause diagnosis, thereby diagnosing the root cause of equipment abnormalities. S4. Present the root cause diagnosis results in a visual manner and provide processing suggestions for equipment maintenance and fault handling. Optimize the large model and agent decision rules based on user feedback to assist maintenance personnel in completing fault diagnosis and fault handling.
[0006] Preferably, the specific steps of S1 are as follows: S11. Use a data acquisition tool to collect the log data, performance index data and historical event records from the device sensors, monitoring system and log files in real time, and clean them to remove noisy data and duplicate data; S12. Parse and vectorize the log data to convert unstructured log text into structured data representation; standardize the performance indicator data to ensure that the performance indicator data have the same scale. S13. Add root cause labels to the fault events in the historical event records to form historical fault data for fine-tuning training of large models.
[0007] Preferably, the specific steps of S2 are as follows: S21. Select a pre-trained large model, wherein the pre-trained large model is a GPT series model or a BERT model; S22. Fine-tune the pre-trained large model using historical fault data, which includes normal data, abnormal data and corresponding fault root cause labels, so that the pre-trained large model learns specific fault modes and characteristics of the equipment. S23. Use the finely tuned and trained large model to perform anomaly detection on the preprocessed device data, and output whether the device is abnormal, as well as the type and degree of the anomaly. The output of the anomaly detection is determined by the following formula:
[0008] In the formula, X is the preprocessed sequence of device data input; Let be the hidden state vector extracted by the large model for the last time step of sequence X, where t is the time step index; W and b are the weight matrix and bias vector of the classification layer, respectively; y is the anomaly type random variable; and k is the anomaly type index. This is a normalized exponential function used to map the output vector of the classification layer to a probability distribution; It is a conditional probability function; This represents the probability that the device belongs to the k-th type of anomaly given the input sequence X; the category with the highest probability is taken as the anomaly type, and this probability value is taken as the degree of anomaly.
[0009] Preferably, the specific steps of S3 are as follows: S31. Obtain the device network topology from the device configuration management system or network monitoring system, wherein the network topology includes the connection relationship and dependency relationship between devices; S32. Based on the network topology, identify the abnormally related nodes with dependencies from the abnormal nodes; S33. Combining historical failure cases and domain knowledge with the abnormal associated nodes, and based on the abnormal type and degree, use reasoning algorithms and rule engines to perform root cause diagnosis and determine the root cause.
[0010] Preferably, the specific steps of S33 are as follows: S331. Utilize a large-scale time-series model to capture the long-period dependence and nonlinear characteristics in the equipment's time-series data, and perform parameter prediction. The prediction formula is as follows:
[0011] In the formula, L is the length of the time window; The input timing data is of length L. For large time-series model functions; These are model parameters; This is the predicted value for the next moment; This refers to the physical mechanism constraint term corresponding to the anomaly type index kk; These are the constraint weighting coefficients; S332. The time-series data is decomposed using multi-scale feature analysis, and the decomposition formula is as follows:
[0012] In the formula, m is the scale index; M is the total number of scales; Let t be the component of time at the m-th scale; extract statistical features and frequency domain features for each scale component, and perform historical trend root cause detection and micro-feature enhancement identification; S333. The prediction residuals of the time-series large model are fused with the anomaly scores obtained from the multi-scale feature analysis, and weighted using the degree of anomaly to obtain a comprehensive anomaly warning score. The fusion formula is as follows:
[0013] In the formula, The comprehensive anomaly warning score for time t; and For weighting; It is the absolute value symbol; Let be the weight of the m-th scale; Let m be the anomaly detection function at the m-th scale; This represents the prediction value of the large time series model at time t.
[0014] Preferably, the specific steps of S33 also include: S334. Construct a time-series-text cross-modal fusion framework, and perform similarity matching between the fault feature vector extracted from the time-series large model and the text-vectorized expert experience vector. The similarity calculation formula is as follows:
[0015] In the formula, This is the fault feature vector extracted from the time series large model, where the subscript f represents the fault feature; To transform unstructured expert experience into a semantic vector using text vector technology, the superscript k represents the anomaly type index, and the subscript e represents the expert experience. The superscript T in the vector indicates the transpose of the vector; and Representing vectors respectively and The Euclidean norm; The cosine similarity is used; when the cosine similarity exceeds a preset threshold, the corresponding expert experience root cause is used as a candidate root cause.
[0016] Preferably, the specific steps of S33 also include: S335. A critical point detection mechanism is introduced for transfer learning, transferring source domain unit fault features to target domain units; transfer learning is achieved by minimizing the difference in feature distribution between the source and target domains, with the optimization objective being:
[0017] In the formula, Indicates To find the minimum value of the optimized variable; Represents the mathematical expectation; This indicates that the random variable follows the corresponding data distribution; (z,c) represents the input sample z sampled from the data distribution and its corresponding fault label c; Data distribution for source domain units; J represents the data distribution of the target domain units; J is the fault prediction loss function. The output prediction of the time-series large model function for the input sample z; This represents the maximum mean difference between the feature distributions of the source and target domains. This refers to the transfer regularization coefficient; when a critical change in the distribution of data in the target domain is detected, transfer learning is automatically initiated, and the model parameters are updated. .
[0018] Preferably, the specific steps of S4 are as follows: S41. The result presentation agent receives the root cause diagnosis results, generates a root cause diagnosis report and visualization charts, and provides processing suggestions based on the root cause diagnosis results. S42. Establish a user feedback channel to collect user evaluations of the accuracy and completeness of the root cause diagnosis results, as well as suggestions for improvement. S43. Based on user feedback and new equipment failure data, optimize the parameters of the large model and the agent decision rules. The optimization objective is to minimize the sum of prediction error and user feedback loss.
[0019] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.
[0020] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.
[0021] As can be seen from the above technical solution compared with the prior art, the present invention has the following beneficial effects: 1. This invention utilizes a multi-agent system and a large model to work collaboratively, comprehensively considering equipment performance indicators, log data, network topology relationships, and historical fault cases. It employs a reasoning mechanism that integrates data-driven, time-series large models, and knowledge guidance, combining time-series pattern recognition with multi-scale feature analysis. This enables more accurate location of the root cause of equipment anomalies, reducing misjudgments and missed judgments. Simultaneously, it automates anomaly detection and root cause diagnosis, quickly providing maintenance personnel with handling suggestions, shortening fault handling time, and improving equipment availability and production efficiency. 2. This invention reduces the time and effort spent on manual analysis and troubleshooting, and proactively identifies and resolves potential equipment problems, preventing equipment failures from escalating and reducing enterprise operation and maintenance costs and production losses caused by equipment failures. At the same time, the system can continuously optimize the analysis model and intelligent agent decision rules based on new data and user feedback, adapting to different types of equipment and complex and ever-changing operating environments. Furthermore, it can easily expand the functions and number of intelligent agents to meet the ever-growing needs for equipment monitoring and analysis. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the method steps in an embodiment of the present invention; Figure 2 This is a schematic diagram of the logical structure of the root cause diagnosis intelligent agent in an embodiment of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.
[0024] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but should not be used to limit the scope of the present invention.
[0025] This invention provides a method for diagnosing the root causes of device anomalies based on large models and intelligent agents, such as... Figure 1 As shown, perform the following steps using a computer device: S1. Collect log data, performance index data and historical event records during the operation of the power generation equipment, and perform cleaning, noise reduction and annotation preprocessing on the log data, performance index data and historical event records to obtain preprocessed equipment data for anomaly detection in large models. S2. Use a finely tuned and trained large model to perform anomaly detection on the preprocessed device data, identify whether the device is abnormal, and record the abnormal data when an anomaly occurs. The abnormal data includes the anomaly type and the degree of anomaly, and is used for root cause diagnosis by the root cause diagnosis agent in the multi-agent system. S3. By using the root cause diagnosis agent in the multi-agent system, combined with the device network topology, historical fault cases and domain knowledge, root cause diagnosis is performed on abnormal data to obtain root cause diagnosis results. The root cause diagnosis agent adopts a reasoning mechanism that integrates data-driven, time-series large model and knowledge guidance, and coordinates time-series pattern recognition and multi-scale feature analysis to achieve root cause diagnosis, thereby diagnosing the root cause of equipment abnormalities. S4. Present the root cause diagnosis results in a visual manner and provide processing suggestions for equipment maintenance and fault handling. Optimize the large model and agent decision rules based on user feedback to assist maintenance personnel in completing fault diagnosis and fault handling.
[0026] Example: This embodiment uses a computer device to perform root cause diagnosis of equipment anomalies in an industrial site as an example. In this industrial site, the equipment runs continuously and generates log data, performance index data, and historical event records. When an anomaly occurs in the equipment, the computer device executes the method described in this application to complete the entire process from data acquisition, anomaly detection, root cause diagnosis to result presentation and feedback optimization. S11. Use a data acquisition tool to collect the log data, performance index data and historical event records from the device sensors, monitoring system and log files in real time, and clean them to remove noisy data and duplicate data.
[0027] For example, the data acquisition tool can acquire performance index data such as temperature, vibration, and current from the device sensors at a acquisition frequency of, for example, once per second; acquire device operating status data from the monitoring system; and acquire alarm logs, operation logs, and event records from log files. During the cleaning process, a sliding window statistical method can be used to remove noisy data that exceeds a reasonable range, and a hash deduplication method can be used to remove duplicate data.
[0028] S12. Parse and vectorize the log data to convert unstructured log text into structured data representation; standardize the performance indicator data to ensure that the performance indicator data have the same scale.
[0029] For example, regular expressions can be used to parse the log text, extracting fields such as timestamp, device identifier, log level, event type, and event description. Then, a text vectorization model can be used to convert the event description into a structured vector of, for example, 128 dimensions. For performance index data, Z-score standardization can be used to ensure that performance index data of different dimensions have the same scale.
[0030] S13. Add root cause labels to the fault events in the historical event records to form historical fault data for fine-tuning training of large models.
[0031] For example, confirmed fault events can be extracted from historical event records, and labels can be added according to the fault root cause category, such as "sensor drift", "communication interruption", "overload", "component aging", etc., and normal data, abnormal data and corresponding fault root cause labels can be combined to form historical fault data.
[0032] By collecting, cleaning, denoising, vectorizing, and standardizing the data, high-quality, structured, and scale-uniform device data can be provided for subsequent large-scale model anomaly detection and multi-agent root cause diagnosis, reducing the interference of noise and duplicate data on the analysis results and improving the accuracy of subsequent anomaly identification and root cause localization.
[0033] S21. Select a pre-trained large model, wherein the pre-trained large model is a GPT series model or a BERT model.
[0034] For example, based on the data type and computing resources of the equipment anomaly analysis task, you can choose the GPT series model or the BERT model as the base model; if the log text accounts for a high proportion, the BERT model can be chosen first; if strong sequence generation and reasoning capabilities are required, the GPT series model can be chosen first.
[0035] S22. Fine-tune the pre-trained large model using historical fault data, which includes normal data, abnormal data, and corresponding fault root cause labels, so that the pre-trained large model learns specific fault modes and characteristics of the equipment.
[0036] For example, historical fault data can be divided into training, validation, and test sets, and the pre-trained large model can be fine-tuned using, for example, a cross-entropy loss function, so that the model learns the data patterns of the device under normal and different abnormal states.
[0037] S23. Use the finely tuned and trained large model to perform anomaly detection on the preprocessed device data, and output whether the device is abnormal, as well as the type and degree of the anomaly. The output of the anomaly detection is determined by the following formula:
[0038] In the formula, X is the preprocessed sequence of device data input; Let be the hidden state vector extracted by the large model for the last time step of sequence X, where t is the time step index; W and b are the weight matrix and bias vector of the classification layer, respectively; y is the anomaly type random variable; and k is the anomaly type index. This is a normalized exponential function used to map the output vector of the classification layer to a probability distribution; It is a conditional probability function; This represents the probability that the device belongs to the k-th type of anomaly given the input sequence X; the category with the highest probability is taken as the anomaly type, and this probability value is taken as the degree of anomaly.
[0039] By analyzing the preprocessed data using a finely tuned large model, it is possible to quickly identify whether the device is abnormal, as well as the type and degree of the abnormality. The abnormal data is then passed to the subsequent root cause diagnosis agent, thereby providing a clear abnormal object and abnormal range for root cause diagnosis, and improving the efficiency and accuracy of abnormality detection.
[0040] S31, such as Figure 2 As shown, the device network topology relationship is obtained from the device configuration management system or network monitoring system. The network topology relationship includes the connection relationship and dependency relationship between devices.
[0041] For example, the connection relationships between servers, databases, switches, and sensors can be obtained from the device configuration management system, and the dependencies between devices can be obtained from the network monitoring system, such as a server depending on a database, or a data acquisition terminal depending on a switch.
[0042] S32. Based on the network topology, identify abnormally related nodes with dependencies from the abnormal nodes.
[0043] For example, if the anomaly detection agent determines that a server is malfunctioning and that server depends on a database, it will further check whether the database also has abnormal indicators. If the database has indicators such as increased response latency and abnormal connection count, it will be identified as an abnormal associated node.
[0044] S33. Combining historical fault cases and domain knowledge, the abnormal associated nodes are used to perform root cause diagnosis using reasoning algorithms and rule engines to determine the root cause.
[0045] For example, real-time data of abnormally associated nodes, historical failure cases, and domain knowledge can be input into the root cause diagnosis agent. The rule engine can match rules such as "database connection pool exhaustion leads to abnormal server response" to initially identify candidate root causes. Further, by combining time series large models and multi-scale feature analysis, the candidate root causes can be verified and ranked to finally determine the root cause.
[0046] Furthermore, the specific steps of S33 are as follows: S331. Utilize a large-scale time-series model to capture the long-period dependence and nonlinear characteristics in the equipment's time-series data, and perform parameter prediction. The prediction formula is as follows:
[0047] In the formula, L is the length of the time window; The input timing data is of length L. For large time-series model functions; These are model parameters; This is the predicted value for the next moment; This refers to the physical mechanism constraint term corresponding to the anomaly type index kk; These are the constraint weighting coefficients; S332. The time-series data is decomposed using multi-scale feature analysis, and the decomposition formula is as follows:
[0048] In the formula, m is the scale index; M is the total number of scales; Let t be the component of time at the m-th scale; extract statistical features and frequency domain features for each scale component, and perform historical trend root cause detection and micro-feature enhancement identification; S333. The prediction residuals of the time-series large model are fused with the anomaly scores obtained from the multi-scale feature analysis, and weighted using the degree of anomaly to obtain a comprehensive anomaly warning score. The fusion formula is as follows:
[0049] In the formula, The comprehensive anomaly warning score for time t; and For weighting; It is the absolute value symbol; Let be the weight of the m-th scale; Let m be the anomaly detection function at the m-th scale; This represents the prediction value of the large time series model at time t.
[0050] S334. Construct a time-series-text cross-modal fusion framework, and perform similarity matching between the fault feature vector extracted from the time-series large model and the text-vectorized expert experience vector. The similarity calculation formula is as follows:
[0051] In the formula, This is the fault feature vector extracted from the time series large model, where the subscript f represents the fault feature; To transform unstructured expert experience into a semantic vector using text vector technology, the superscript k represents the anomaly type index, and the subscript e represents the expert experience. The superscript T in the vector indicates the transpose of the vector; and Representing vectors respectively and The Euclidean norm; The cosine similarity is used; when the cosine similarity exceeds a preset threshold, the corresponding expert experience root cause is used as a candidate root cause.
[0052] S335. A critical point detection mechanism is introduced for transfer learning, transferring source domain unit fault features to target domain units; transfer learning is achieved by minimizing the difference in feature distribution between the source and target domains, with the optimization objective being:
[0053] In the formula, Indicates To find the minimum value of the optimized variable; Represents the mathematical expectation; This indicates that the random variable follows the corresponding data distribution; (z,c) represents the input sample z sampled from the data distribution and its corresponding fault label c; Data distribution for source domain units; J represents the data distribution of the target domain units; J is the fault prediction loss function. The output prediction of the time-series large model function for the input sample z; This represents the maximum mean difference between the feature distributions of the source and target domains. This refers to the transfer regularization coefficient; when a critical change in the distribution of data in the target domain is detected, transfer learning is automatically initiated, and the model parameters are updated. .
[0054] By integrating data-driven approaches, large-scale temporal models, and knowledge guidance into a multi-agent system, and by coordinating temporal pattern recognition and multi-scale feature analysis, the system can pinpoint root causes from multiple perspectives, including network topology, historical cases, domain knowledge, temporal prediction residuals, cross-modal expert experience, and transfer learning. This reduces misjudgments and omissions caused by analysis of a single data source and improves the comprehensiveness and accuracy of root cause diagnosis.
[0055] S41. The result presentation agent receives the root cause diagnosis results, generates a root cause diagnosis report and visualization charts, and provides processing suggestions based on the root cause diagnosis results.
[0056] For example, the results presentation agent can generate a root cause diagnosis report that includes the time of the anomaly, the type of anomaly, the degree of the anomaly, the nodes associated with the anomaly, the root cause, the confidence level, and the handling suggestions. It can also generate visualization charts such as network topology anomaly propagation diagrams, time series indicator change diagrams, and root cause ranking diagrams. The handling suggestions can include "check the database connection pool configuration", "replace the sensor", "adjust the load distribution strategy", etc.
[0057] S42. Establish a user feedback channel to collect user evaluations of the accuracy and completeness of the root cause diagnosis results, as well as suggestions for improvement. For example, an accuracy score, a completeness score, and a text feedback input box can be set in the results presentation interface, allowing users to evaluate the root cause diagnosis results and submit improvement suggestions.
[0058] S43. Based on user feedback and new equipment failure data, optimize the parameters of the large model and the agent decision rules. The optimization objective is to minimize the sum of prediction error and user feedback loss.
[0059] For example, if users report that the root cause of a certain type of fault is not accurately located, the corresponding data can be added to the training set, and the parameters of the large model or the weights of the agent rules can be adjusted; if new equipment fault data shows that a certain rule has a high false alarm rate, the priority of that rule can be reduced or the rule conditions can be modified.
[0060] Visual presentations and processing suggestions help users quickly understand the root causes of anomalies and take maintenance measures. By continuously optimizing large model parameters and agent decision-making rules through user feedback and new fault data, the solution can adapt to different types of equipment and complex and ever-changing operating environments, improving the system's adaptability, scalability, and long-term analysis accuracy.
[0061] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.
[0062] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.
[0063] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the device anomaly root cause diagnosis methods based on large models and intelligent agents in the above embodiments.
[0064] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.
[0065] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another.
[0066] For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media.
[0067] The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid state disks (SSDs)).
[0068] It should be noted that in this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.
[0069] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0070] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0071] The embodiments of the present invention are given for the purposes of illustration and description. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A method for diagnosing the root causes of device anomalies based on large models and intelligent agents, characterized in that, Perform the following steps using a computer device: S1. Collect log data, performance index data and historical event records during the operation of the power generation equipment, and perform cleaning, noise reduction and annotation preprocessing on the log data, performance index data and historical event records to obtain preprocessed equipment data for anomaly detection in large models. S2. Use a finely tuned and trained large model to perform anomaly detection on the preprocessed device data, identify whether the device is abnormal, and record the abnormal data when an anomaly occurs. The abnormal data includes the anomaly type and the degree of anomaly, and is used for root cause diagnosis by the root cause diagnosis agent in the multi-agent system. S3. By using the root cause diagnosis agent in the multi-agent system, combined with the device network topology, historical fault cases and domain knowledge, root cause diagnosis is performed on abnormal data to obtain root cause diagnosis results. The root cause diagnosis agent adopts a reasoning mechanism that integrates data-driven, time-series large model and knowledge guidance, and coordinates time-series pattern recognition and multi-scale feature analysis to achieve root cause diagnosis, thereby diagnosing the root cause of equipment abnormalities. S4. Present the root cause diagnosis results in a visual manner and provide processing suggestions for equipment maintenance and fault handling. Optimize the large model and agent decision rules based on user feedback to assist maintenance personnel in completing fault diagnosis and fault handling.
2. The device anomaly root cause diagnosis method based on large models and intelligent agents as described in claim 1, characterized in that: The specific steps of S1 are as follows: S11. Use a data acquisition tool to collect the log data, performance index data and historical event records from the device sensors, monitoring system and log files in real time, and clean them to remove noisy data and duplicate data; S12. Parse and vectorize the log data to convert unstructured log text into structured data representation; Standardize the performance index data to ensure that the performance index data have the same scale; S13. Add root cause labels to the fault events in the historical event records to form historical fault data for fine-tuning training of large models.
3. The device anomaly root cause diagnosis method based on large models and intelligent agents as described in claim 1, characterized in that: The specific steps of S2 are as follows: S21. Select a pre-trained large model, wherein the pre-trained large model is a GPT series model or a BERT model; S22. Fine-tune the pre-trained large model using historical fault data, which includes normal data, abnormal data and corresponding fault root cause labels, so that the pre-trained large model learns specific fault modes and characteristics of the equipment. S23. Use the finely tuned and trained large model to perform anomaly detection on the preprocessed device data, and output whether the device is abnormal, as well as the type and degree of the anomaly. The output of the anomaly detection is determined by the following formula: In the formula, X is the preprocessed sequence of device data input; Let be the hidden state vector extracted by the large model for the last time step of sequence X, where t is the time step index; W and b are the weight matrix and bias vector of the classification layer, respectively; y is the anomaly type random variable; and k is the anomaly type index. This is a normalized exponential function used to map the output vector of the classification layer to a probability distribution; It is a conditional probability function; This represents the probability that the device belongs to the k-th type of anomaly given the input sequence X; the category with the highest probability is taken as the anomaly type, and this probability value is taken as the degree of anomaly.
4. The device anomaly root cause diagnosis method based on large models and intelligent agents as described in claim 3, characterized in that: The specific steps of S3 are as follows: S31. Obtain the device network topology from the device configuration management system or network monitoring system, wherein the network topology includes the connection relationship and dependency relationship between devices; S32. Based on the network topology, identify the abnormally related nodes with dependencies from the abnormal nodes; S33. Combining historical failure cases and domain knowledge with the abnormal associated nodes, and based on the abnormal type and degree, use reasoning algorithms and rule engines to perform root cause diagnosis and determine the root cause.
5. The device anomaly root cause diagnosis method based on large models and intelligent agents as described in claim 4, characterized in that: The specific steps of S33 are as follows: S331. Utilize a large-scale time-series model to capture the long-period dependence and nonlinear characteristics in the equipment's time-series data, and perform parameter prediction. The prediction formula is as follows: In the formula, L is the length of the time window; The input timing data is of length L. For large time-series model functions; These are model parameters; This is the predicted value for the next moment; This refers to the physical mechanism constraint term corresponding to the anomaly type index kk; These are the constraint weighting coefficients; S332. The time-series data is decomposed using multi-scale feature analysis, and the decomposition formula is as follows: In the formula, m is the scale index; M is the total number of scales; Let t be the component of time t at the m-th scale; Statistical and frequency domain features are extracted for each scale component, and historical trend root cause detection and micro-feature enhancement identification are performed. S333. The prediction residuals of the time-series large model are fused with the anomaly scores obtained from the multi-scale feature analysis, and weighted using the degree of anomaly to obtain a comprehensive anomaly warning score. The fusion formula is as follows: In the formula, The comprehensive anomaly warning score for time t; and For fusion weights; It is the absolute value symbol; Let be the weight of the m-th scale; Let m be the anomaly detection function at the m-th scale; This represents the prediction value of the large time series model at time t.
6. The device anomaly root cause diagnosis method based on large models and intelligent agents as described in claim 5, characterized in that: The specific steps of S33 also include: S334. Construct a time-series-text cross-modal fusion framework, and perform similarity matching between the fault feature vector extracted from the time-series large model and the text-vectorized expert experience vector. The similarity calculation formula is as follows: In the formula, This is the fault feature vector extracted from the time series large model, where the subscript f represents the fault feature; To transform unstructured expert experience into a semantic vector using text vector technology, the superscript k represents the anomaly type index, and the subscript e represents the expert experience. The superscript T in the vector indicates the transpose of the vector; and Representing vectors respectively and The Euclidean norm; The cosine similarity is used; when the cosine similarity exceeds a preset threshold, the corresponding expert experience root cause is used as a candidate root cause.
7. The device anomaly root cause diagnosis method based on large models and intelligent agents as described in claim 5, characterized in that: The specific steps of S33 also include: S335. A critical point detection mechanism is introduced for transfer learning, transferring source domain unit fault features to target domain units; transfer learning is achieved by minimizing the difference in feature distribution between the source and target domains, with the optimization objective being: In the formula, Indicated by To find the minimum value of the optimized variable; Represents the mathematical expectation; This indicates that the random variable follows the corresponding data distribution; (z,c) represents the input sample z sampled from the data distribution and its corresponding fault label c; Data distribution for source domain units; J represents the data distribution of the target domain units; J is the fault prediction loss function. The output prediction of the time-series large model function for the input sample z; This represents the maximum mean difference between the feature distributions of the source and target domains. This refers to the transfer regularization coefficient; when a critical change in the distribution of data in the target domain is detected, transfer learning is automatically initiated, and the model parameters are updated. .
8. The device anomaly root cause diagnosis method based on large models and intelligent agents as described in claim 1, characterized in that: The specific steps of S4 are as follows: S41. The result presentation agent receives the root cause diagnosis results, generates a root cause diagnosis report and visualization charts, and provides processing suggestions based on the root cause diagnosis results. S42. Establish a user feedback channel to collect user evaluations of the accuracy and completeness of the root cause diagnosis results, as well as suggestions for improvement. S43. Based on user feedback and new equipment failure data, optimize the parameters of the large model and the agent decision rules. The optimization objective is to minimize the sum of prediction error and user feedback loss.
9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 8.
10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 8.
