Database anomaly detection method for power plant statistical management system

Abstract

The application relates to the technical field of database anomaly detection, and provides a database anomaly detection method for a power plant statistical management system, which can perform collaborative anomaly identification and analysis on power plant statistical data; the method comprises the following steps: combining a language embedding-VAE clustering model to realize effective clustering of the power plant statistical data, thereby providing high-quality data input for a subsequent model; and improving the analysis capability for data fluctuation sources through a multiple decomposition-LSTM model, so that accurate identification and explanation of key abnormal points are realized; through the application, efficient analysis and early warning of power plant statistics can be realized, important data support and early warning capability are provided for intelligent operation and maintenance and decision support of the power plant.

Description

Technical Field

[0001] This invention relates to the field of database anomaly detection technology, and in particular to a database anomaly detection method for a power plant statistical management system. Background Technology

[0002] With the continuous improvement of information technology levels in power plants, statistical management systems have been widely applied to various aspects of power plant production operations, equipment management, and business decision-making. The databases within these systems contain a large amount of data, including power generation plans, equipment operating parameters, fuel quantity and quality data, and major and minor production indicators. The completeness and accuracy of this data directly affect the operational efficiency and safety of the power plant. However, in actual operation, due to factors such as system complexity, diverse data sources, and high data update frequency, various anomalies often exist in the system's database. Failure to detect and handle these anomalies in a timely manner can negatively impact decision support and safety management. However, existing database anomaly detection methods still have many shortcomings: First, existing methods struggle to capture such multi-scale, non-linear temporal characteristics, leading to insensitivity to identifying potential equipment anomalies such as frequent start-ups and shutdowns, and abnormal shutdowns. Second, existing methods, based on traditional detection approaches using univariate analysis or simple indicator monitoring, are unable to promptly reveal hidden systemic deviations or plan execution anomalies, reducing the intelligent response capability of scheduling. Summary of the Invention

[0003] This invention provides a database anomaly detection method for power plant statistical management systems. It can simultaneously perform in-depth anomaly analysis on power plant statistical data, enhancing the system's ability to automatically perceive, locate, and interpret potential risks. This method combines a semantic embedding-VAE clustering model with a multi-factor decomposition-LSTM model, possessing strong temporal feature learning and semantic anomaly recognition capabilities, making it suitable for intelligent anomaly detection scenarios in complex business data environments of power plants. On one hand, this invention introduces a clustering strategy combining semantic embedding and anomaly-aware VAE models. By compressing high-dimensional embedding vectors and combining them with cluster boundary modeling, it achieves effective clustering of power plant statistical data, thus providing high-quality data input for subsequent models. On the other hand, this invention constructs an LSTM model that integrates multi-factor decomposition and occlusion prediction mechanisms to decompose the original data into multiple conceptual items such as trends, periods, and residuals. It also introduces an occlusion sample and residual-driven Shapley value analysis mechanism, effectively improving the ability to analyze data fluctuation sources and achieving accurate identification and interpretation of key anomalies. Through the above collaborative modeling and analysis, this invention can efficiently output power plant anomaly detection results, providing important data support and early warning capabilities for intelligent operation and maintenance and decision support of power plants.

[0004] This invention provides a database anomaly detection method for a power plant statistical management system, the method comprising the following steps:

[0005] Step S1: Collect power plant statistical data from the power plant statistical management system database; power plant statistical data includes power generation plans, equipment operating parameters, fuel quantity and quality data, and major and minor production indicators.

[0006] Step S2: Perform data cleaning, standardization, and feature engineering on the power plant statistical data to generate preprocessed statistical data;

[0007] Feature engineering processing includes extracting moving average features, difference processing features, and time lag features from power plant statistical data;

[0008] Step S3: Construct a semantic embedding-VAE clustering model. Use the semantic embedding-VAE clustering model to cluster the preprocessed statistical data to obtain statistical clustering feature data.

[0009] Step S4: Establish an LSTM model, use the time series data multiple decomposition method and fusion occlusion prediction mechanism to optimize the LSTM model, thereby constructing a multiple decomposition-LSTM model. Process the preprocessed statistical data and statistical clustering feature data through the multiple decomposition-LSTM model to obtain the power plant anomaly detection results.

[0010] Step S5: Based on the power plant anomaly detection results, generate anomaly reports and early warning notifications to assist in decision-making.

[0011] Furthermore, step S3, the process of obtaining statistical clustering feature data, specifically includes the following steps:

[0012] Step S31: Perform word segmentation, stop word removal, and stemming on the preprocessed statistical data to generate a standardized text sequence;

[0013] Step S32: Use word embedding technology to convert the standardized text sequence into a dense vector and generate a high-dimensional embedding vector;

[0014] Step S33: Construct an anomaly-aware VAE model. A two-stage strategy is used to compress high-dimensional embedding vectors into a low-dimensional space to generate low-dimensional text vectors. The two-stage strategy includes unsupervised pre-training and anomaly-aware fine-tuning. In the anomaly-aware fine-tuning stage, a small number of labeled anomaly samples are introduced to introduce an additional penalty mechanism for the anomaly samples, which makes the anomaly-aware VAE model perform poorly when reconstructing anomaly data, thereby enhancing the separability between anomaly and normal samples.

[0015] Step S34: Construct an enhanced HDBSCAN clustering model. The enhanced HDBSCAN clustering model includes an online stage and an offline stage. In the online stage, Bubble-tree is used to perform online data summarization on low-dimensional text vectors, dynamically maintaining the compressed representation of text semantics to obtain Bubble-tree low-dimensional embedding vectors. In the offline stage, the main cluster structure is reconstructed from the Bubble-tree low-dimensional embedding vectors, the semantic distribution boundary is learned, and statistical clustering feature data is generated. Bubble-tree is an online data summarization structure for dynamic data. It adopts a tree-like hierarchical organization to cluster and compress low-dimensional text vectors. Each bubble node in the Bubble-tree represents a local semantic cluster, containing the centroid, density information, and embedding range, ultimately forming a text semantic compression tree, where each leaf node represents a Bubble-tree low-dimensional embedding vector.

[0016] Furthermore, step S4, the process of obtaining the power plant anomaly detection results, specifically includes the following steps:

[0017] Step S41: Input the preprocessed statistical data and statistical clustering feature data into the multi-factor LSTM model for training to obtain the trained LSTM model;

[0018] Step S42: Decompose the preprocessed structured data into multiple concept items, namely trend concept items, cycle concept items and residual concept items. When decomposing the cycle concept item, further refine it into two sub-concepts: seasonal fluctuation item and short-term cycle fluctuation item, and construct a more refined set of decomposed concepts to obtain the decomposed concept set.

[0019] Step S43: Perform a masking operation on each concept in the decomposed concept set to generate masked samples; calculate the prediction result after each masking based on the masked samples and the trained LSTM model to obtain the masking prediction result.

[0020] Step S44: Introduce a residual-sensitive factor to dynamically adjust the weights of the multi-concept terms and construct a residual-driven Shapley enhancement mechanism; based on the occlusion prediction results, evaluate the influence of the multi-concept terms on the prediction results of the trained LSTM model through the residual-driven Shapley enhancement mechanism to obtain SHAP contribution data.

[0021] Step S45: Based on the trained LSTM model, determine whether the current time point is an anomaly. For detected anomalies, combine the SHAP contribution data to determine the root cause of the anomaly and generate the power plant anomaly detection result.

[0022] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0023] This invention introduces a semantic embedding and anomaly perception VAE clustering model. By converting statistical data such as scheduling records and operation reports into high-dimensional semantic vectors, and combining a small number of labeled anomaly samples under an unsupervised learning framework, semantic reconstruction and clustering recognition are performed. The system can effectively identify abnormal expressions, potential risk signals and deviations from normal operation and maintenance behaviors in the text, thereby significantly improving the perception accuracy and input quality of statistical data of subsequent anomaly detection models.

[0024] This invention constructs an LSTM model that integrates multiple decomposition methods and masking prediction mechanisms, enabling deep modeling and fine-grained detection of anomalies in structured time-series data from power plants. This model not only captures long-term trends and short-term fluctuations in the data but also enhances the interpretability of anomaly detection through a residual-driven Shapley contribution mechanism. Compared to traditional methods that rely solely on error judgment, this invention can clearly identify the dominant factors behind anomalies, achieving intelligent localization of the root causes of anomalous data, improving the accuracy of anomaly detection and the reliability of decision-making, and solving the problem of existing statistical systems lacking in-depth analytical capabilities for complex dynamic data.

[0025] Through the aforementioned collaborative analysis mechanism, this invention effectively breaks down data silos and significantly enhances the power plant statistical management system's ability to integrate and process multi-source information. The organic combination of the embedding-VAE clustering model and the multi-decomposition-LSTM model in this invention not only achieves intelligent detection that comprehensively covers all types of data in the database, but also improves the response speed and visualization interpretation capabilities of data early warning, and enhances the system's auxiliary decision-making level in key business links such as operation and maintenance scheduling, anomaly investigation, and risk prevention, providing key technical support for building a highly reliable and intelligent power information management platform. Attached Figure Description

[0026] Figure 1 This invention provides a flowchart illustrating a database anomaly detection method for a power plant statistical management system;

[0027] Figure 2 This is a diagram illustrating the influence of multiple concept terms on the prediction results of the trained LSTM model in Example 5. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] Example 1, according to Figure 1This invention provides a database anomaly detection method for a power plant statistical management system, the method comprising the following steps:

[0030] Step S1: Data Acquisition: Collect power plant statistical data from the power plant statistical management system database; power plant statistical data includes power generation plans, equipment operating parameters, fuel quantity and quality data, and major and minor production indicators.

[0031] Step S2: Data preprocessing: Perform data cleaning, standardization, and feature engineering on the power plant statistical data to generate preprocessed statistical data;

[0032] Feature engineering processing includes extracting moving average features, difference processing features, and time lag features from power plant statistical data;

[0033] Step S3: Clustering Feature Extraction: Construct a semantic embedding-VAE clustering model, and use the semantic embedding-VAE clustering model to cluster the preprocessed statistical data to obtain statistical clustering feature data;

[0034] Step S4: Anomaly Detection: Establish an LSTM model, use the time series data multiple decomposition method and fusion masking prediction mechanism to optimize the LSTM model, thereby constructing a multiple decomposition-LSTM model. Process the preprocessed statistical data and statistical clustering feature data through the multiple decomposition-LSTM model to obtain the power plant anomaly detection results.

[0035] Step S5: Early Warning Decision: Based on the abnormal detection results of the power plant, generate anomaly reports and early warning notifications to assist in decision-making.

[0036] Example 2, based on Example 1, specifically includes the following steps in step S3 to obtain statistical clustering feature data:

[0037] Step S31: Perform word segmentation, stop word removal, and stemming on the preprocessed statistical data to generate a standardized text sequence;

[0038] Step S32: Use word embedding technology to convert the standardized text sequence into a dense vector and generate a high-dimensional embedding vector;

[0039] Step S33: Construct an anomaly-aware VAE model. A two-stage strategy is employed to compress high-dimensional embedding vectors into a low-dimensional space, generating low-dimensional text vectors. This two-stage strategy includes unsupervised pre-training and anomaly-aware fine-tuning. In the anomaly-aware fine-tuning stage, a small number of labeled anomaly samples are introduced, and an additional penalty mechanism is applied to these samples. This causes the anomaly-aware VAE model to perform poorly when reconstructing anomalous data, thereby enhancing the separability between anomaly and normal samples. The formula used is as follows:

[0040] Joint loss function for the anomaly detection fine-tuning stage:

[0041] ;

[0042] in, Represents the total loss function. This represents the VAE loss of normal samples. This represents the anomaly perception penalty coefficient. Input data representing anomalous samples, This represents the reconstruction output of the outlier sample; This represents the reconstruction error of outlier samples. This represents the expected value for outlier samples;

[0043] Step S34: Construct enhanced HDBSCAN clustering. Enhanced HDBSCAN clustering includes online and offline stages. In the online stage, Bubble-tree is used to perform online data summarization on low-dimensional text vectors, dynamically maintaining the compressed representation of text semantics to obtain Bubble-tree low-dimensional embedding vectors. In the offline stage, the main cluster structure is reconstructed from the Bubble-tree low-dimensional embedding vectors, the semantic distribution boundary is learned, and statistical clustering feature data is generated. Bubble-tree is an online data summarization structure for dynamic data. It adopts a tree-like hierarchical organization to cluster and compress low-dimensional text vectors. Each bubble node in the Bubble-tree represents a local semantic cluster, containing the centroid, density information, and embedding range, ultimately forming a text semantic compression tree, where each leaf node represents a Bubble-tree low-dimensional embedding vector.

[0044] Example 3, based on Example 1, specifically includes the following steps in step S3 of obtaining statistical clustering feature data:

[0045] Step E1: Perform word segmentation, stop word removal, and stemming on the preprocessed statistical data to generate a standardized text sequence;

[0046] Step E2: Use word embedding technology to transform the standardized text sequence into a dense vector, generating a high-dimensional embedding vector;

[0047] Step E3: Construct an anomaly-aware VAE model. Use a two-stage strategy of unsupervised pre-training and anomaly-aware fine-tuning to compress high-dimensional embedding vectors into low-dimensional space to generate low-dimensional text vectors. In the anomaly-aware fine-tuning stage, introduce a small number of labeled anomaly samples and introduce an additional penalty mechanism for the anomaly samples to make the anomaly-aware VAE model perform poorly when reconstructing anomaly data, thereby enhancing the separability between anomaly and normal samples.

[0048] Step E4: Use HDBSCAN clustering to process low-dimensional text vectors and generate statistical clustering feature data.

[0049] Example 4, based on Example 1, specifically includes the following steps in step S3 to obtain statistical clustering feature data:

[0050] Step Q1: Perform word segmentation, stop word removal, and stemming on the preprocessed statistical data to generate a standardized text sequence;

[0051] Step Q2: Use word embedding technology to transform the standardized text sequence into a dense vector, generating a high-dimensional embedding vector;

[0052] Step Q3: Using the VAE model, the high-dimensional embedding vector is compressed into a low-dimensional space to generate a low-dimensional text vector;

[0053] Step Q4: Use HDBSCAN clustering to process low-dimensional text vectors and generate statistical clustering feature data.

[0054] Example 5, according to Figure 2 This embodiment is based on Embodiment 2. In this embodiment, the process of obtaining the power plant anomaly detection result in step S4 specifically includes the following steps:

[0055] Step S41: Input the preprocessed statistical data and statistical clustering feature data into the multi-factor LSTM model for training to obtain the trained LSTM model;

[0056] Step S42: Decompose the preprocessed structured data into multiple concept items, namely trend concept items, cycle concept items and residual concept items. When decomposing the cycle concept item, further refine it into two sub-concepts: seasonal fluctuation item and short-term cycle fluctuation item, and construct a more refined set of decomposed concepts to obtain the decomposed concept set.

[0057] Step S43: Perform a masking operation on each concept in the decomposed concept set to generate masked samples; calculate the prediction result after each masking based on the masked samples and the trained LSTM model to obtain the masking prediction result.

[0058] Step S44: Introduce a residual-sensitive factor to dynamically adjust the weights of the multi-concept term, constructing a residual-driven Shapley enhancement mechanism; based on the occlusion prediction results, evaluate the influence of the multi-concept term on the prediction results of the trained LSTM model through the residual-driven Shapley enhancement mechanism to obtain the SHAP contribution data, using the following formula:

[0059] ;

[0060] in, Indicates a time step. Represents the concept item index, Indicates the first The concept items include trend concept items, residual concept items, seasonal fluctuation items, and short-term cyclical items; Indicates the residual sensitivity factor. At the point of time The actual observed value, Represents the set of concepts after decomposition. Indicates the removal of the first The combination of conceptual items, i.e., occlusion , Indicates a point in time Input samples, Indicates to cover up right Prediction is performed using a trained LSTM model;

[0061] ;

[0062] in, This represents SHAP contribution data. Representing a subset In cover Under the premise, from A subset selected from; Representing a subset The number of elements, express factorial, This represents the total number of elements in the concept set after decomposition. express factorial; This represents the union operation. Indicates marginal contribution;

[0063] Step S45: Based on the trained LSTM model, determine whether the current time point is an anomaly. For detected anomalies, combine the SHAP contribution data to determine the root cause of the anomaly and generate the power plant anomaly detection result.

[0064] Example 6, this example is based on Example 5. In this example, step S5: early warning decision: based on the power plant anomaly detection results, generate an anomaly report and early warning notification to assist in decision-making;

[0065] Power plant anomaly detection results:

[0066] [Abnormal Time]: 2025-04-28 14:00;

[0067] [Abnormal Indicator]: Boiler coal consumption rate (actual value 356g / kWh, predicted value 312g / kWh, deviation +14%).

[0068] [Root Causes of Abnormalities]: Short-term periodic fluctuations (contributing 48%), residual noise (contributing 30%).

[0069] [Cluster Analysis]: The current data is an isolated point in the Bubble-tree clustering, indicating an atypical operating pattern;

[0070] [Recommendation]: Check fuel quality and control parameter settings, and pay attention to repeated occurrences of abnormalities.

[0071] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention. The actual structure is not limited to this. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A database anomaly detection method for a power plant statistical management system, characterized in that: The method includes the following steps: Step S1: Collect power plant statistical data from the power plant statistical management system database; Step S2: Perform feature engineering on the power plant statistical data to generate preprocessed statistical data; Step S3: Construct a semantic embedding-VAE clustering model, and use the semantic embedding-VAE clustering model to cluster the preprocessed statistical data to obtain statistical clustering feature data; Step S4: Construct a multi-factor LSTM model, and use the multi-factor LSTM model to process the preprocessed statistical data and statistical clustering feature data to obtain the power plant anomaly detection results; In step S4, constructing the multiple decomposition-LSTM model specifically includes: establishing an LSTM model, using a time-series data multiple decomposition method and a fusion occlusion prediction mechanism, optimizing the LSTM model, and thus constructing the multiple decomposition-LSTM model. The process of obtaining the power plant anomaly detection results in step S4 specifically includes the following steps: Step S41: Input the preprocessed statistical data and statistical clustering feature data into the multi-factor LSTM model for training to obtain the trained LSTM model; Step S42: Decompose the preprocessed statistical data into multiple concept items to obtain the decomposed concept set; Step S43: Perform a masking operation on the decomposed concept set, and combine it with the trained LSTM model to generate masking prediction results; Step S44: Construct a residual-driven Shapley enhancement mechanism and combine it with the occlusion prediction results to obtain SHAP contribution data; Step S45: Based on the trained LSTM model and combined with SHAP contribution data, generate temporal anomaly detection results; Step S44 specifically includes: introducing a residual sensitivity factor to construct a residual-driven Shapley enhancement mechanism; based on the occlusion prediction results, evaluating the influence of multiple concept terms on the prediction results of the trained LSTM model through the residual-driven Shapley enhancement mechanism to obtain SHAP contribution data, using the following formula: ； in, Indicates a time step. Represents the concept item index, Indicates the first The concept items include trend concept items, residual concept items, seasonal fluctuation items, and short-term cyclical items; Indicates the residual sensitivity factor. Indicates a point in time The actual observed value, Represents the set of concepts after decomposition. Indicates the removal of the first The combination of conceptual items, i.e., occlusion , Indicates a point in time Input samples, Indicates to cover up right Prediction is performed using a trained LSTM model; ； in, This represents SHAP contribution data. Representing a subset In cover Under the premise, from A subset selected from; Representing a subset The number of elements, express factorial, This represents the total number of elements in the concept set after decomposition. express factorial; This represents the union operation. Indicates marginal contribution.

2. The database anomaly detection method for a power plant statistical management system according to claim 1, characterized in that: Feature engineering processing includes extracting moving average features, difference processing features, and time lag features from power plant statistical data.

3. The database anomaly detection method for a power plant statistical management system according to claim 1, characterized in that: The process of obtaining statistical clustering feature data in step S3 specifically includes the following steps: Step S31: Process the statistical clustering feature data to generate a standardized text sequence; Step S32: Use word embedding technology to convert the standardized text sequence into a dense vector and generate a high-dimensional embedding vector; Step S33: Construct an anomaly perception-VAE model, and use a two-stage strategy to compress high-dimensional embedding vectors into low-dimensional space to generate low-dimensional text vectors; Step S34: Construct an enhanced HDBSCAN clustering model. In the online stage, Bubble-tree is used to process low-dimensional text vectors to obtain Bubble-tree low-dimensional embedding vectors. In the offline stage, the main cluster structure is reconstructed from the Bubble-tree low-dimensional embedding vectors to generate statistical clustering feature data.

4. The database anomaly detection method for a power plant statistical management system according to claim 3, characterized in that: The two-stage strategy includes unsupervised pre-training and anomaly-aware fine-tuning.

Images