Fault detection method for operation and maintenance of electric power informatization system

Abstract

The invention relates to a fault detection method for operation and maintenance of an electric power informatization system. The method comprises the following steps: step 1, constructing and forming a time sequence data set for index monitoring data of the electric power informatization system; step 2, based on the time sequence data set formed in the step 1, carrying out adaptation on the power informatization system fault detection model based on the language model; step 3, fine tuning is carried out on the power informatization system fault detection model adapted in the step 2, and a final power informatization system fault detection model is generated; and step 4, outputting a fault detection result oriented to operation and maintenance of the electric power informatization system by adopting the electric power informatization system fault detection model finally generated in the step 3. The efficiency of automatic operation and maintenance can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power information system operation and maintenance, and relates to a fault detection method, in particular to a fault detection method for power information system operation and maintenance. Background Art

[0002] I6000 is the core support system for the information and communication operation and maintenance of the State Grid. The platform is an integrated information and communication dispatching and operation support platform that integrates "dispatching, operation, maintenance, customer service, and three lines", which can greatly improve the State Grid's power information and communication operation guarantee capabilities, information and communication business sharing and integration capabilities, and information and communication quality service capabilities. The system has the following characteristics: 1. Full life cycle management: The State Grid i6000 system can manage the entire life cycle of information equipment, including equipment procurement, warehousing, maintenance, scrapping and other links, to achieve comprehensive management of equipment information; 2. Automated operation and maintenance: The State Grid i6000 system can realize functions such as automated inspection and automatic deployment, reduce manual intervention, and improve operation and maintenance efficiency; 3. Real-time monitoring: The State Grid i6000 system can monitor the operation status of the information system in real time, discover problems in a timely manner and handle them, and ensure the stable operation of the information system; 4. Data analysis: The State Grid i6000 system can analyze the operation data of information equipment and systems, provide data support, and provide a basis for decision-making.

[0003] The traditional operation and maintenance of power information systems is completed manually by operation and maintenance personnel. During the maintenance process, there is a lot of repetitive work and the efficiency is very low. With the rapid development of business, the scale and complexity of the power information system have increased rapidly, and it is difficult to complete it relying solely on operation and maintenance personnel. Therefore, the operation and maintenance method has gradually changed. The current operation and maintenance methods are mostly automated operation and maintenance. Although automated operation and maintenance uses machine monitoring, it still requires a lot of manual participation, such as requiring operation and maintenance personnel to set thresholds and formulate rules. However, the changes in application systems, multi-dimensional scenarios, and slow changes in indicators make it difficult to maintain manually generated rules and prone to false alarms. Therefore, in order to solve this problem and make operation and maintenance fully automated without relying on humans, people thought of using AI tools.

[0004] In 2017, the concept of AIOps (Artificial Intelligence for IT Operations) was first proposed by Gartner. Based on the existing operation and maintenance data, by integrating artificial intelligence, big data and machine learning methods, it provides automatic monitoring and control of the service application environment and supporting infrastructure of large enterprises. By obtaining data from multiple sources such as applications, infrastructure, networks, clouds and existing monitoring tools, these integrated solutions are expected to provide multiple functions to minimize service interruptions and help field engineering teams quickly resolve incidents. Intelligent operation and maintenance is an emerging interdisciplinary field. The main research and application areas are divided into four aspects: fault discovery, fault location, alarm delivery, and fault recovery. Among them, key technologies and fault management systems such as Figure 1 shown.

[0005] Fault discovery collects massive amounts of operation and maintenance data, comprehensively utilizes intelligent algorithms, cloud computing and big data technologies to detect faults, monitors software and hardware indicators and system logs in real time, and promptly and accurately detects whether there are anomalies in the monitoring indicators; in the fault location part, artificial intelligence and correlation analysis methods are used to explore the correlation between system indicators and logs, and automatically or assist operation and maintenance personnel to complete fault location; alarms are compressed and clustered to avoid false alarms, and faults are intelligently distributed to downstream fault handling teams; fault recovery formulates intelligent rules, performs fault recovery work such as traffic switching, downgrading and frequency reduction according to the cause of the fault, and ensures high reliability and high availability of the system. Among them, fault detection and discovery are the basis and key to the application and development of intelligent operation and maintenance, and are of vital significance.

[0006] At present, there are several main fault detection methods used in industry and academia: fault detection methods based on traditional statistical models, fault detection methods based on integrated supervised learning, fault detection methods based on unsupervised learning and deep generative models, and fault detection methods based on pre-trained machine learning models.

[0007] The fault detection method based on traditional statistical models is one of the earliest methods studied. It assumes that data is modeled according to a specified probability distribution. If the probability of a certain data being generated from the model is lower than a certain threshold, the data is defined as abnormal, that is, abnormal data is data that cannot fit the model perfectly. Over the years, the academic community has proposed quite a number of fault detectors based on traditional statistical models, such as the adaptive Kalman filter model proposed by Knorn, Leith, et al., the non-parametric model based on kernel density estimation proposed by Heywood, Mcdermot, et al., and the wavelet analysis model proposed by Li W and Ali AGhorbani. However, these models usually have simple assumptions about time series, requiring experts to select detectors for a given time series and fine-tune the parameters of the detectors based on training data. Therefore, the fault detection method based on traditional statistical models is not applicable to complex monitoring indicator data in actual operation and maintenance scenarios.

[0008] Fault detection methods based on integrated supervised learning are proposed to avoid parameter adjustments in the above fault detection algorithms. In 2015, Yahoo proposed an extensible generic anomaly detection framework (EGADS) for large-scale time series data, and Tsinghua NetMan Laboratory proposed an integrated supervised method Opprentice. This type of algorithm uses user feedback as labels and uses the anomaly scores output by traditional detectors as features to train anomaly classifiers to perform fault detection on different types of data. Fault detection methods based on integrated supervised learning perform well on different data sets, but these methods are heavily dependent on good labels, which is usually not feasible in large-scale applications. In addition, since the algorithm needs to run multiple traditional detectors for feature extraction during detection, it will introduce a lot of computational costs, which is also a difficult problem for real-time monitoring in actual operation and maintenance.

[0009] Fault detection based on unsupervised learning and deep generative models has become increasingly popular in recent years, for example, single-class SVM and K-Means based on unsupervised learning, and Gaussian mixture models and variational autoencoders based on deep generative models. The core idea of ​​this type of algorithm is to focus on normal data rather than abnormal data. Simply put, this type of algorithm first identifies the "normal area" in the original data or some potential feature space, and then calculates the anomaly score of the data by measuring how far the observation is from the "normal area". The monitoring indicators in the operation and maintenance scenario are mainly composed of normal data, so this type of model can be easily trained even without labels. Although this type of algorithm has great advantages in model training, there is no method for calculating the anomaly score for KPI data.

[0010] Model pre-training technology is a method that uses large-scale unlabeled data to learn general language or visual representations, which can improve the performance and generalization ability of the model on downstream tasks. Model pre-training technology is usually divided into two stages: pre-training stage and fine-tuning stage. In the pre-training stage, the model performs self-supervised learning on unlabeled data to learn the intrinsic structure and rules of the data. In the fine-tuning stage, the model performs supervised learning on labeled data to learn knowledge and skills for specific tasks. Model pre-training technology can effectively solve problems such as data scarcity, transfer learning and model interpretability, and is currently a hot research direction in the field of artificial intelligence.

[0011] Therefore, in order to solve the above problems, the present invention proposes a fault detection method for power information system operation and maintenance, which can improve the efficiency of automated operation and maintenance.

[0012] After searching, no public documents of the prior art identical or similar to the present invention were found. Summary of the invention

[0013] The purpose of the present invention is to overcome the deficiencies of the prior art and to propose a fault detection method for operation and maintenance of an electric power information system, which can improve the efficiency of automated operation and maintenance.

[0014] The present invention solves the practical problem by adopting the following technical solutions:

[0015] A fault detection method for power information system operation and maintenance, comprising the following steps:

[0016] Step 1: Construct and form a time series data set for power information system indicator monitoring data;

[0017] Step 2: Based on the time series data set formed in step 1, adapt the power information system fault detection model based on the language model;

[0018] Step 3: fine-tune the power information system fault detection model adapted in step 2 to generate a final power information system fault detection model;

[0019] Step 4: Use the power information system fault detection model finally generated in step 3 to output fault detection results for power information system operation and maintenance.

[0020] Moreover, the specific steps of step 1 include:

[0021] (1) Collecting operation and maintenance monitoring indicator data such as response time indicator, error rate indicator, throughput indicator, resource utilization indicator, access volume indicator and data accuracy indicator;

[0022] (2) Preprocessing the collected operation and maintenance monitoring indicator data, including data normalization, data completion and data cleaning;

[0023] (3) Save the preprocessed data into a time series database to form a time series data set.

[0024] Moreover, the specific steps of step 2 include:

[0025] (1) Data segmentation and extraction: First, a data segmenter is designed to read the monitoring indicator time series data set from the time series database. After segmentation and extraction, a list of data substrings is generated. The length of each substring is the pre-defined sliding window size, which meets the requirements of the general language model.

[0026] (2) Data vectorization: Convert the monitoring indicator time series data substring into an embedded vector and encapsulate it into a whole with auxiliary data such as segmentation vector;

[0027] (3) Model retraining: After converting the time series data of the monitoring indicators of the information system into a format that can be understood by the general model, this data is input into the general language model and the model is retrained.

[0028] Moreover, the specific method of step (3) of step 2 to input these data into the general language model and complete the retraining of the model is:

[0029] ① Mask prediction: A fixed proportion of the data in the time series segment generated by the indicator monitoring data sequence is randomly replaced with a special [MASK] tag, and the model needs to predict these masked data;

[0030] ② Prediction of subsequent time series segments: The model will receive two time series segments as input, one is the real time series segment that will appear next, and the other is a randomly selected time series segment. The model needs to determine whether these two time series segments are continuous.

[0031] Moreover, the specific steps of step 3 include:

[0032] (1) Label generation: Generate labels automatically or manually, and label each data point in the time series to indicate whether the data at this point is abnormal.

[0033] (2) Data segmentation and extraction: Slice the input sequence according to the set length of the above information to meet the input length limit requirements of the model;

[0034] (3) Data vectorization: Convert and encapsulate input time series that the model cannot directly understand into data in an embedded format so that the model can understand it.

[0035] (4) Model structure adjustment: Add a sequence transformer layer and a fully connected layer according to the fault detection task, and finally output the fault detection results;

[0036] The sequence transformer layer is used to further model the relationship between time series in the language model output and capture the complex relationship between anomalies and the previous context. The fully connected layer is responsible for mapping the extracted features to anomaly scores to ultimately determine whether there is a fault in the sequence.

[0037] (5) After completing the above four steps, the model can be trained to complete the fine-tuning stage and generate the final power information system fault detection model.

[0038] Advantages and beneficial effects of the present invention:

[0039] 1. The present invention proposes a fault detection method for the operation and maintenance of an electric power information system. First, a set of monitoring indicators for the electric power information system is designed, and indicator monitoring data is collected from multiple electric power information systems to form a massive amount of time series data. These data are used to adapt a general language model to form a fault detection model for the electric power information system, so that the model is suitable for fault detection tasks. This step can learn rich semantic representations from a large amount of unlabeled data without the need for manual annotation operations, and the learned data representations can be migrated to downstream tasks. Next, the indicator monitoring data of the target system is used for fine-tuning, and finally a fault detection model dedicated to the target system is formed. Fine-tuning is simply fine-tuning the parameters of the model using a target data set for a specific problem or goal, thereby avoiding the problem of model overfitting caused by a small target data set.

[0040] 2. The present invention proposes a technical method for automatic fault detection in an electric power information system, which can improve the efficiency of automated operation and maintenance.

[0041] 3. The present invention utilizes the excellent architectural design of the language model to improve the effect of anomaly detection in the power information system.

[0042] 4. The present invention applies the latest language model to the field of fault detection in power information systems and improves the fault detection effect through the latest progress in machine learning technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the key technology and fault management system of the present invention;

[0044] Figure 2 It is a schematic diagram of time series fault detection of the present invention;

[0045] Figure 3 This is a system diagram of a fault detection module based on machine learning of the present invention;

[0046] Figure 4 is a process flow chart of the present invention;

[0047] Figure 5 is a process flow chart of step 2 of the present invention;

[0048] Figure 6 This is a process flow chart of step 3 of the present invention. DETAILED DESCRIPTION

[0049] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0050] A fault detection method for power information system operation and maintenance, such as Figure 4 As shown, the following steps are included:

[0051] Step 1: Construct and form a time series data set for power information system indicator monitoring data;

[0052] An effective information system monitoring mechanism is the premise and important support for ensuring the stable operation and effective management of the business. The power information system monitoring indicators are key indicators for evaluating and monitoring the activities of the information system, which helps to ensure the normal operation of the information system.

[0053] The specific steps of step 1 include:

[0054] (1) Collecting operation and maintenance monitoring indicator data such as response time indicator, error rate indicator, throughput indicator, resource utilization indicator, access volume indicator and data accuracy indicator;

[0055] (2) Preprocessing the collected operation and maintenance monitoring indicator data, including data normalization, data completion and data cleaning;

[0056] (3) Save the preprocessed data into a time series database to form a time series data set.

[0057] The working principle of step 1 is:

[0058] 1.1 Data collection;

[0059] In order to detect the failure of the information system, the present invention collects the following operation and maintenance monitoring index data through I6000 and its plug-in system:

[0060] Response time metrics

[0061] Response time is one of the important indicators for measuring the performance of information systems. It reflects the system's response speed to user requests and directly affects user experience and satisfaction. In business monitoring, the system's performance can be evaluated by monitoring and recording the system's response time. Generally speaking, the shorter the response time, the better. If the response time is too long, there may be system performance problems, which need to be adjusted or upgraded in time.

[0062] Error rate indicator

[0063] The error rate is one of the important indicators to measure the stability and reliability of information systems. It reflects the proportion of errors that occur when the system processes requests, which can help enterprises to promptly detect and resolve system failures or abnormalities. In business monitoring, the stability and reliability of the system can be evaluated by monitoring and recording the error rate of the system. Generally speaking, the lower the error rate, the better. If the error rate is too high, there may be system failures or abnormalities, which need to be checked and repaired in a timely manner.

[0064] Throughput Metrics

[0065] Throughput is one of the important indicators to measure the processing capacity of information systems. It reflects the number of requests that the system can process per unit time, which directly affects the system's concurrency and scalability. In business monitoring, the system's processing capacity can be evaluated by monitoring and recording the system's throughput. Generally speaking, the higher the throughput, the better. If the throughput is too low, there may be system bottlenecks or insufficient resources, which require timely optimization or expansion.

[0066] Resource utilization indicators

[0067] Resource utilization is one of the important indicators to measure the resource utilization of information systems. It reflects the utilization of various system resources (such as CPU, memory, disk, etc.), which can help enterprises evaluate whether the resource configuration of the system is reasonable. In business monitoring, the resource utilization of the system can be evaluated by monitoring and recording the resource utilization of the system. Generally speaking, the resource utilization should be within a reasonable range. If the resource utilization is too high, there may be system failures, affecting the normal operation of the system.

[0068] Visits metrics

[0069] The number of visits is one of the important indicators to measure the popularity of information systems and user activity. It reflects the number of requests received by the system and user access behavior, which can help enterprises understand business development trends and user needs. In business monitoring, the activity of the business and the behavioral characteristics of users can be evaluated by monitoring and recording the number of visits to the system. Generally speaking, the higher the number of visits, the better. If the number of visits is too low, there may be poor business or user loss, and timely measures need to be taken to improve business attractiveness and user experience.

[0070] Data accuracy indicators

[0071] Data accuracy is one of the important indicators for measuring the data quality of information systems. It reflects the accuracy and credibility of the system's data processing, and directly affects the accuracy and effectiveness of corporate decision-making. In business monitoring, the quality and reliability of data can be evaluated by monitoring and recording the data accuracy of the system. Generally speaking, the higher the data accuracy, the better. If the data accuracy is not high, there may be data entry errors or unreliable data sources, which need to be corrected or replaced in a timely manner.

[0072] 1.2 Data Preprocessing

[0073] The collected system monitoring indicator data needs to be preprocessed to form a time series data set for fine-tuning the machine learning model. Preprocessing mainly includes the following aspects:

[0074] Data normalization: By normalizing data, data is converted from the original range to a specific range or distribution, which can eliminate the impact of the magnitude between different data indicators, make the data more comparable and unified, and improve the training efficiency and accuracy of the model. The formula for data normalization is:

[0075]

[0076] Where x is a specific data point in the data set; X represents the data set, min(X) is the minimum value in the data set, max(X) is the maximum value in the data set; x' is the result of data normalization.

[0077] Data completion: Data completion is a technology for dealing with missing data. Missing data is a common problem in system monitoring data, which may be caused by a variety of reasons, such as data collection errors, missing data records, and measurement equipment failures. By filling in missing values, the integrity of the data set can be ensured so that analysis and model training are not affected by missing data. First, identify missing values. Data marked as NaN (Not a Number), null, or specific placeholders in the original monitoring data are considered missing values ​​and need to be completed; then, calculate the average of the 10 data before and after the missing value, and use this average value to complete the missing value;

[0078] Data cleaning: Remove useless content from monitoring data and only retain monitoring indicator data. This can improve the processing efficiency of subsequent steps and avoid the impact of useless data.

[0079] 1.3 Time Series Data Preservation

[0080] The processed data is saved in a pre-selected time series database (TSDB) to form a time series data set, which can be used in subsequent anomaly detection steps.

[0081] Step 2: Based on the time series data set formed in step 1, adapt the power information system fault detection model based on the language model;

[0082] The specific steps of step 2 include:

[0083] (1) Data segmentation and extraction: First, a data segmenter is designed to read the monitoring indicator time series data set from the time series database. After segmentation and extraction, a list of data substrings is generated. The length of each substring is the pre-defined sliding window size, which meets the requirements of the general language model.

[0084] (2) Data vectorization: Convert the monitoring indicator time series data substring into an embedded vector and encapsulate it into a whole with auxiliary data such as segmentation vector;

[0085] (3) Model retraining: After converting the time series data of the monitoring indicators of the information system into a format that can be understood by the general model, this data is input into the general language model and the model is retrained.

[0086] The specific method of step (3) of step 2 to input these data into the general language model and complete the retraining of the model is:

[0087] ① Mask prediction: A fixed proportion of the data in the time series segment generated by the indicator monitoring data sequence is randomly replaced with a special [MASK] tag, and the model needs to predict these masked data;

[0088] ② Prediction of subsequent time series segments: The model will receive two time series segments as input, one is the real time series segment that will appear next, and the other is a randomly selected time series segment. The model needs to determine whether these two time series segments are continuous.

[0089] The working principle of step 2 is:

[0090] The machine learning model structure used for anomaly detection in the power information system can adopt the structure of a general machine learning model, such as a commonly used language model, etc. However, it is necessary to use the indicator data set collected from the power information system to adapt the general model structure for training, so as to generate a fault detection model suitable for time series data. This stage is called the model adaptation stage. This stage can adjust the weights of the general model architecture to adapt it to the needs of the fault detection task.

[0091] This stage can be divided into three steps. Figure 5 As shown:

[0092] 2.1) Data segmentation and extraction: Model adaptation requires the use of a large amount of indicator monitoring data. Language models usually have certain restrictions on input. For example, the maximum string length that can be processed by the Bert model is 512, while the length of the time series data set of indicator monitoring data far exceeds this number. Therefore, it is necessary to first design a data segmenter to read the monitoring indicator time series data set from the time series database. After segmentation and extraction, a list of data substrings is generated. The length of each substring is the pre-defined sliding window size, which meets the requirements of the general model.

[0093]

[0094]

[0095] 2.2) Data vectorization: General language models usually input data in the form of word embedding vectors generated after word embedding operations and auxiliary segmentation vectors. Through word embedding operations, words can be converted into vectors that can reflect their semantic and syntactic characteristics, providing an effective numerical representation method for natural language processing tasks. In order to use general language models for information system fault detection scenarios, we also need to convert indicator monitoring data substrings into embedding vectors and encapsulate them with auxiliary data such as segmentation vectors as a whole so that they can be understood and used by general language models.

[0096]

[0097] 2.3) Model retraining

[0098] After converting the indicator monitoring data of the information system into a format that can be understood by the general model, these data can be input into the model and the model can be retrained to make the general model perform better in fault detection tasks.

[0099] At this stage, the model is trained using the monitoring indicator data in vectorized form, which mainly includes the following two steps:

[0100] Step 1: Mask prediction: A fixed proportion of the data in the time series segment generated by the indicator monitoring data sequence will be randomly replaced with a special mask mark, and the model needs to predict these masked data. This method forces the model to rely on contextual information for prediction, thereby learning rich semantic representations. The specific method is: first segment the time series segment, and then select 15% of the characters in the segment, for example, the i-th character is selected, and then: use mask replacement with a probability of 80%; use a random character replacement with a probability of 10%; and do not operate with a probability of 10%.

[0101] Step 2: Prediction of subsequent time series segments: The model needs to predict whether two given time series segments are continuous. This task aims to improve the model's ability to understand the relationship between time series segments. The model will receive two time series segments as input, one is the real next time series segment, and the other is a randomly selected time series segment. The model needs to determine whether the two time series segments are continuous.

[0102] Step 3: fine-tune the power information system fault detection model adapted in step 2 to generate a final power information system fault detection model;

[0103] The specific steps of step 3 include:

[0104] (1) Label generation: Generate labels automatically or manually, and label each data point in the time series to indicate whether the data at this point is abnormal.

[0105] (2) Data segmentation and extraction: Slice the input sequence according to the set length of the above information to meet the input length limit requirements of the model;

[0106] (3) Data vectorization: Convert and encapsulate input time series that the model cannot directly understand into data in an embedded format so that the model can understand it.

[0107] (4) Model structure adjustment: Add a sequence transformer layer and a fully connected layer according to the fault detection task, and finally output the fault detection results;

[0108] The sequence transformer layer is used to further model the relationship between time series in the language model output and capture the complex association between anomalies and the previous context. The fully connected layer is responsible for mapping the extracted features to anomaly scores, which are used to ultimately determine whether the sequence has a fault.

[0109] (5) After completing the above four steps, the model can be trained to complete the fine-tuning stage and generate the final power information system fault detection model.

[0110] The working principle of step 3 is:

[0111] After the adaptation is completed, the general language model can be used for fault detection in the power information system. However, the fault data characteristics of different power information systems are not exactly the same. In order to make the model perform better in specific system fault detection, it is also necessary to use the fault data of a specific information system to fine-tune the model and enhance the model's capabilities.

[0112] Fine-tuning a model means making small adjustments to the parameters of a more general fault detection machine learning model using the dataset of the target task. This method can make the model fit the data better and output better predictions.

[0113] The judgment of time series anomalies needs to be based on context information. If it is obviously different from the context, it is abnormal. From the perspective of actual anomaly detection applications, only the previous information can be obtained, but the following information cannot be known. Therefore, in the fine-tuning stage, the abnormal state at a certain moment is judged based on the previous information at a certain moment.

[0114] The fine-tuning phase needs to include steps such as label generation, data segmentation and extraction, data vectorization, and model structure adjustment. Figure 6 shown.

[0115] Step 1: Label generation: The present invention can automatically generate labels through algorithms or manually record labels. By generating labels, each data point in the time series is labeled to indicate whether the data at this point is abnormal.

[0116] Step 2: Data segmentation and extraction: This step is the same as the segmentation and extraction step in the model adaptation phase. The input sequence needs to be sliced ​​according to the set length of the above information to meet the input length limit of the model. The appropriate length of the data subsequence has an important impact on the model effect. If the subsequence length is too short, the sequence contains too little historical information, which is likely to cause misjudgment; if the subsequence length is too long, the sequence will contain more historical information, which will increase the accuracy of the model, but it will also cause a significant increase in the training time of the model.

[0117] Step 3: Data vectorization: This step is the same as the vectorization step in the model adaptation phase. It converts and encapsulates the input time series that the model cannot directly understand into data in an embedded format so that the model can understand it.

[0118] Step 4: Model structure adjustment: According to the fault detection task, a sequence transformer layer and a fully connected layer are added to finally output the fault detection results. The sequence transformer layer is used to further model the relationship between time series in the language model output and capture the complex association between anomalies and the previous context. The fully connected layer is responsible for mapping the extracted features to the anomaly score, which is used to finally determine whether there is a fault in the sequence. Adding the above layers can make full use of the time series information captured by the language model, extract global or local features, and thus more accurately identify the fault point or fault interval. This design further enhances the adaptability of the language model to time series characteristics and provides better results for information system fault detection.

[0119] After completing the above four steps, the model can be trained to complete the fine-tuning phase. The training algorithm is as follows:

[0120]

[0121] Step 4: Use the power information system fault detection model finally generated in step 3 to output fault detection results for power information system operation and maintenance.

[0122] The working principle of the present invention is:

[0123] During the operation of the power information system, a large amount of multi-dimensional monitoring data is generated, which contains a lot of information about the system status. The present invention introduces a pre-trained machine learning model into the field of power information system fault detection based on operation monitoring indicators, designs a corresponding fault detection model and method, and realizes online fault detection of the power information system.

[0124] like Figure 2 As shown in Figure 1, the indicator monitoring data of the power information system is a typical time series data. For example, the throughput of an information system, the CPU usage of the server running the information system, the number of visits to the Web site, etc. The time series contains a large amount of information about the observed object at a specific moment in a specific situation, so analyzing the time series has great practical value. Time series analysis is an analysis technique for studying the internal characteristics of time series. Curve fitting and parameter estimation are common methods for time series modeling. Time series analysis is widely used in application load resource forecasting, market sales volume forecasting, power plant fault warning, etc. Taking Web services as an example, usually, operation and maintenance personnel monitor various KPI indicators and analyze their corresponding time series to determine the stability of Web services. If a KPI is found to be abnormal during the time series analysis, it often means that there is a problem with its related application. The anomaly of the time series refers to the data in the time series that obviously deviates from the normal state, such as a sharp increase or decrease in data. In real applications, there are many reasons that may cause time series anomalies, such as: 1) operator errors leading to errors in data recording and storage; 2) anomalies within the system, such as anomalies caused by the effectiveness of system components; 3) anomalies caused by environmental changes in the monitored system, such as power outages in the computer room that can lead to missing monitoring data.

[0125] like Figure 3As shown, the fault detection module of the present invention has a built-in fault detection method based on machine learning, and the fault detection module will regularly receive monitoring data in the power information system such as the configured data acquisition component push. These monitoring data are a kind of time series data, which consists of an indicator name and a set of tags, and is used to describe the state of the system; after the fault detection module obtains the required monitoring indicator data, it uses a fault detection method based on machine learning to judge and conduct in-depth analysis and association of the monitoring indicator data, and detect and discover various system faults. If there is no abnormal indicator data, if the data is normal, the data will be stored locally. If there is an abnormality in the data, the abnormal data alarm is pushed to the alarm management component. After that, the alarm management module will issue an alarm according to the configuration file, and send the alarm information to the operation and maintenance personnel or the subsequent fault analysis system in a variety of configurable ways.

[0126] The innovation of the present invention lies in:

[0127] The main innovation of this invention is to use the data characteristics of the power information system and combine it with the language model to achieve fault detection. The original purpose of the language model design was mainly NLP tasks rather than fault detection.

[0128] It should be emphasized that the embodiments of the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation modes. Any other implementation modes derived by those skilled in the art based on the technical solutions of the present invention also fall within the scope of protection of the present invention.

Claims

1. A fault detection method for power information system operation and maintenance, characterized by: The following steps are involved: Step 1: Construct and form a time series data set for power information system indicator monitoring data; Step 2: Based on the time series data set formed in step 1, adapt the power information system fault detection model based on the language model; Step 3: fine-tune the power information system fault detection model adapted in step 2 to generate a final power information system fault detection model; Step 4: Use the power information system fault detection model finally generated in step 3 to output fault detection results for power information system operation and maintenance.

2. A fault detection method for power information system operation and maintenance according to claim 1, characterized in that: The specific steps of step 1 include: (1) Collecting operation and maintenance monitoring indicator data such as response time indicator, error rate indicator, throughput indicator, resource utilization indicator, access volume indicator and data accuracy indicator; (2) Preprocessing the collected operation and maintenance monitoring indicator data, including data normalization, data completion and data cleaning; (3) Save the preprocessed data into a time series database to form a time series data set.

3. A fault detection method for power information system operation and maintenance according to claim 1, characterized in that: The specific steps of step 2 include: (1) Data segmentation and extraction: First, a data segmenter is designed to read the monitoring indicator time series data set from the time series database. After segmentation and extraction, a list of data substrings is generated. The length of each substring is the pre-defined sliding window size, which meets the requirements of the general language model. (2) Data vectorization: Convert the monitoring indicator time series data substring into an embedded vector and encapsulate it into a whole with auxiliary data such as segmentation vector; (3) Model retraining: After converting the time series data of the monitoring indicators of the information system into a format that can be understood by the general model, this data is input into the general language model and the model is retrained.

4. A fault detection method for power information system operation and maintenance according to claim 3, characterized in that: The specific method of step (3) of step 2 to input these data into the general language model and complete the retraining of the model is: ① Mask prediction: A fixed proportion of the data in the time series segment generated by the indicator monitoring data sequence is randomly replaced with a special [MASK] tag, and the model needs to predict these masked data; ② Prediction of subsequent time series segments: The model will receive two time series segments as input, one is the real time series segment that will appear next, and the other is a randomly selected time series segment. The model needs to determine whether these two time series segments are continuous.

5. A fault detection method for power information system operation and maintenance according to claim 1, characterized in that: The specific steps of step 3 include: (1) Label generation: Generate labels automatically or manually, and label each data point in the time series to indicate whether the data at this point is abnormal. (2) Data segmentation and extraction: Slice the input sequence according to the set length of the above information to meet the input length limit requirements of the model; (3) Data vectorization: Convert and encapsulate input time series that the model cannot directly understand into data in an embedded format so that the model can understand it; (4) Model structure adjustment: Add a sequence transformer layer and a fully connected layer according to the fault detection task, and finally output the fault detection results; The sequence transformer layer is used to further model the relationship between time series in the language model output and capture the complex relationship between anomalies and the previous context. The fully connected layer is responsible for mapping the extracted features to anomaly scores to ultimately determine whether there is a fault in the sequence. (5) After completing the above four steps, the model can be trained to complete the fine-tuning stage and generate the final power information system fault detection model.

Images