Power station abnormal data analysis method and system based on causal reasoning

Abstract

The invention relates to the field of intelligent electric fields, in particular to a power station abnormal data analysis method and system based on causal reasoning. The method specifically comprises the following steps: collecting multi-modal data; carrying out data preprocessing, multi-modal feature fusion and abnormal preliminary screening on the collected multi-modal data; establishing a digital twinborn model and virtual-real mapping for the fused multi-modal data, and performing causal reasoning analysis and federated learning collaborative optimization on candidate abnormal data; aiming at the abnormal condition of the multi-modal data, performing Bayesian abnormal confidence evaluation by adopting a Bayesian reasoning framework; and after receiving a final analysis result, operation and maintenance personnel confirm and correct an abnormal conclusion and optimize the digital twinborn model by adopting a feedback learning mechanism. The method has the advantages of multi-modal fusion, causal reasoning analysis innovation, federated learning cooperation and digital twinborn prediction, and the intelligent degree of electric field abnormal data analysis is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of smart electric fields, and specifically to a method and system for analyzing abnormal data from power plants based on causal reasoning. Background Technology

[0002] During operation, electrical equipment and power systems may experience abnormal phenomena such as voltage fluctuations, harmonic interference, and partial discharge. Existing methods for monitoring electric field anomalies typically rely on fixed threshold judgments or manual experience analysis, which presents several problems: fixed threshold methods are susceptible to environmental noise and sensor errors, resulting in a high false alarm rate; they are highly dependent on manual intervention, requiring maintenance personnel to troubleshoot anomalies one by one, leading to low work efficiency; existing methods lack the ability to model trends from historical data and dynamically adjust thresholds, making them difficult to adapt to complex operating environments; traditional single-modal data analysis cannot fully explore the correlations and causal relationships between multi-dimensional data; there is a lack of collaborative learning mechanisms among distributed monitoring points, hindering knowledge sharing and model optimization; furthermore, existing methods are mostly passive anomaly detection, lacking predictive analysis capabilities based on physical models; and anomaly source tracing analysis lacks interpretability and struggles to provide clear causal reasoning paths.

[0003] Therefore, there is an urgent need for an intelligent anomaly data analysis method that integrates multimodal data analysis, causal reasoning, federated learning, and digital twin technologies to improve the accuracy, interpretability, and predictive maintenance capabilities of electric field monitoring. Summary of the Invention

[0004] To address the problems existing in the prior art, the first aspect of this invention provides a method for analyzing abnormal data from power plants based on causal reasoning, specifically including the following steps: Y1. Collect multimodal data; Y2. Perform data preprocessing, multimodal feature fusion, and preliminary anomaly screening on the collected multimodal data; Y3. Establish a digital twin model and virtual-real mapping for the fused multimodal data, and perform adaptive threshold calculation based on the multimodal data; perform anomaly detection based on the adaptive threshold calculation result; if the multimodal data is normal, continue to monitor and collect the multimodal data; if the multimodal data is abnormal, perform causal reasoning analysis and federated learning collaborative optimization on the candidate abnormal data. Y4. For the aforementioned multimodal data anomalies, a Bayesian inference framework is used to perform Bayesian anomaly confidence assessment. Y5. After receiving the final analysis results, the operation and maintenance personnel use a feedback learning mechanism to confirm and correct the abnormal conclusions and optimize the digital twin model. The optimized digital twin model data is then fed back to the digital twin model and the virtual-real mapping end for model adjustment.

[0005] Preferably, the multimodal data includes voltage, current, electric field strength, temperature and humidity, vibration, sound and environmental interference data, and the multimodal feature fusion of the multimodal data adopts a multi-head attention mechanism.

[0006] Preferably, the multimodal data is uploaded to the data processing center via an edge computing gateway to form a continuous time-series dataset.

[0007] Preferably, the data preprocessing includes noise filtering, missing value imputation, data normalization, and feature extraction. Preferably, the digital twin model is based on physical constraint optimization, and the digital twin model includes a geometric model, a physical model, and a behavioral model.

[0008] Preferably, the causal reasoning analysis employs a counterfactual analysis method; the federated learning collaborative optimization employs a differential privacy-preserving aggregation algorithm.

[0009] Preferably, in the Bayesian anomaly confidence assessment, if the anomaly is a false alarm, the threshold is optimized and the optimized threshold result is fed back to the adaptive threshold calculation to adjust the adaptive threshold parameter; if it is a real anomaly, the anomaly is confirmed, and the development trend and possible consequences of the anomaly are predicted, and the final analysis result is output.

[0010] Preferably, the feedback learning mechanism includes a system operation layer, a feedback processing layer, a model optimization layer, and a knowledge update layer.

[0011] Preferably, the operation and maintenance personnel receive the system operation layer data and provide feedback evaluation results on the system operation data; the feedback processing layer collects the feedback data corresponding to each evaluation result and generates feedback labels; the model optimization layer uses parameter adjustment strategies to evaluate model performance; and based on the model performance evaluation results, the model is improved in the predictive modeling process at the knowledge update layer.

[0012] A second aspect of the present invention provides an analysis system employing the above-described causal reasoning-based abnormal data analysis method for power plant sites, specifically comprising: Multimodal data acquisition module: Various types of sensors and acquisition terminals are deployed in the electric field to collect and upload multi-dimensional and multimodal monitoring data of the electric field operation; Multimodal feature fusion module: used to extract and fuse multidimensional features in the time domain, frequency domain, and time-frequency domain to generate a comprehensive feature representation; Digital twin modeling module: used to build digital twin models of electric field equipment, realizing virtual-real mapping and state synchronization; Digital twin simulation verification module: used for simulation verification and predictive analysis of anomalies; Federated Learning Collaboration Module: Used to enable secure collaborative learning and model optimization among distributed monitoring points; Multi-level decision fusion module: used to integrate multi-source information, perform Bayesian anomaly confidence assessment, and make intelligent decisions; Predictive maintenance module: Used to generate proactive maintenance suggestions and risk warnings based on anomaly prediction results; Adaptive learning optimization module: Used to continuously optimize the parameters and performance of each module based on multi-source feedback.

[0013] Compared with the prior art, the present invention has the following significant innovative advantages: (1) This invention has the advantage of multimodal fusion. By fusing multimodal data such as electrical, environmental, vibration and sound, the accuracy and robustness of anomaly detection are significantly improved; by using a multi-head attention mechanism to adaptively fuse different modal features, cross-modal correlations are effectively mined and the limitations of single-modal analysis are reduced.

[0014] (2) This invention has innovative advantages in causal reasoning analysis. It constructs a causal graph model based on the electric field system, provides interpretable anomaly source analysis, and significantly improves the credibility and interpretability of the results; it uses counterfactual reasoning to verify the cause of anomalies, effectively distinguishes between real anomalies and occasional noise, and greatly reduces the false alarm rate; through causal chain tracing, it identifies the root cause and propagation path of anomalies, and provides a scientific basis for accurate maintenance.

[0015] (3) This invention has the advantages of federated learning collaboration. It realizes secure collaborative learning among distributed monitoring points, achieving knowledge sharing and model optimization while protecting data privacy; it adopts differential privacy technology and adaptive aggregation algorithm to ensure the dual improvement of data security and model performance; and it significantly improves the overall anomaly detection performance and generalization ability through global collaborative optimization.

[0016] (4) This invention has the advantage of digital twin prediction. It constructs a digital twin model of electric field equipment to realize the transformation from passive detection to active prediction; it provides scientific and reliable predictive maintenance suggestions based on anomaly evolution prediction based on physical constraints; and it ensures the accuracy of anomaly judgment and the effectiveness of maintenance decisions through virtual-real fusion simulation verification.

[0017] (5) The present invention significantly improves the intelligence level of electric field anomaly data analysis. It can realize multi-level and multi-dimensional intelligent anomaly analysis, greatly reduce manual dependence and improve operation and maintenance efficiency; it provides intelligent analysis reports containing causal explanations and predictive suggestions to support scientific decision-making and precise maintenance; it has adaptive learning and continuous optimization capabilities, and its performance and adaptability are continuously improved as the running time increases. Attached Figure Description

[0018] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is an overall architecture diagram of a method for analyzing abnormal data from power plants based on causal reasoning; Figure 2 This is a flowchart illustrating the specific operation of a method for analyzing abnormal data from power plants based on causal reasoning. Figure 3 This is a diagram illustrating the feedback learning mechanism architecture described in a method for analyzing abnormal data from power plants based on causal reasoning. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0020] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0021] The first aspect of this embodiment provides a method for analyzing abnormal data of power plants based on causal reasoning.

[0022] Figure 1 The overall architecture diagram of the electric field anomaly data analysis method includes a data acquisition layer, a data preprocessing layer, a core technology layer, a decision fusion layer, an intelligent analysis layer, an output display layer, and a feedback learning layer. The specific functions of each architecture layer are as follows: like Figure 1 As shown, this embodiment employs a multimodal sensor to acquire electric field data in the data acquisition layer. Specifically, the electric field data is presented as multimodal data, including voltage, current, electric field strength, temperature and humidity, vibration, sound, and environmental interference data. The acquired electric field data undergoes noise filtering, missing value imputation, data normalization, and feature extraction in the data preprocessing layer. In this embodiment, adaptive wavelet denoising is used for noise filtering, and the formula is:

[0023] in, It is a signal The result of wavelet transform, It is an inverse wavelet transform operation, and thresholded represents the coefficients after thresholding.

[0024] In this embodiment, linear interpolation is used to fill in missing values. The formula is as follows:

[0025] in, Given points coordinates For known points coordinate, For interpolation coordinate, For interpolation coordinate.

[0026] In this embodiment, Min-Max standardization is used for data normalization. For data with different dimensions, the formula is:

[0027] in, and These are the minimum and maximum values ​​of the data, respectively. This is the normalized data.

[0028] like Figure 1 As shown, further, in this embodiment, the preprocessed electric field data is transmitted to the core technology layer as multimodal data, wherein, S1. In the core technology layer, multimodal feature fusion operations are performed on multimodal data.

[0029] Specifically, the steps of the multimodal feature fusion operation are as follows: S11. Adaptive fusion of multimodal features is performed using a multi-head attention mechanism, with the following formula:

[0030] in, Let be the attention weights for the i-th modality. For the features of the i-th mode, For global feature representation, This is a learnable weight matrix.

[0031] S12. Perform cross-modal correlation modeling on multimodal data, and establish the correlation matrix formula among multimodal data as follows:

[0032] in, For the correlation between mode i and mode j, The physical distance between sensors. This is the distance attenuation parameter.

[0033] S13. Generate a dynamic threshold based on time series data in the autoregressive (AR) model, using the following formula:

[0034] in, For the data at the current moment, These are the autoregressive coefficients. Let be the order of the autoregression. This is the noise term.

[0035] Specifically, the formula for generating a dynamic threshold with a predicted trend is:

[0036] in, The mean of the predicted values, Standard deviation, This is the threshold adjustment factor. It should be noted that when data exceeds this threshold, it is marked as candidate outlier data.

[0037] S2. In the core technology layer, causal reasoning analysis is performed on the fused multimodal data. The specific steps are as follows: S21. Construct a causal graph model based on multimodal data, using multimodal data as variables to describe and monitor the causal relationships and propagation paths between data; S22. Associate candidate anomaly data with multimodal data using contextual information. Specifically, input the candidate anomaly data and multimodal data into a large language model with causal reasoning enhancement capabilities; S23. Using counterfactual reasoning, analyze "whether the anomaly would still occur if a certain factor did not exist" to verify the reliability of the anomaly causes corresponding to the candidate anomaly data. S24. Using causal chain tracing, identify the root cause and propagation path of candidate abnormal data, and generate an interpretable anomaly tracing report; S25. Combining knowledge graphs in the electric field domain, accurately classify and assess the severity of anomalies in electric field systems.

[0038] S3. In the core technology layer, a digital twin model and virtual-real mapping are established based on the fused multimodal data. The specific steps are as follows: S31. Using a digital twin model based on physical constraints for optimization, simulate and verify the detected candidate anomaly data to evaluate the authenticity and impact range of the anomaly. S32. Based on physical constraints and historical evolution patterns, predict the development trend and possible consequences of anomalies; S33. Combining the results of multimodal data and causal reasoning analysis, conduct a comprehensive anomaly reliability assessment; S34. Finally, generate an intelligent decision-making scheme that includes predictive maintenance recommendations.

[0039] The digital twin modeling process includes geometric, physical, and behavioral models. A spatial topology graph of multi-sensor data is established to describe the physical connections and influence relationships between sensors, enabling virtual-real mapping and state synchronization.

[0040] S4. In the core technology layer, federated learning is used to collaboratively optimize the fused multimodal data to achieve collaborative optimization of the global model and improve the overall anomaly detection performance of the electric field system. The specific steps are as follows: S41. Establish a federated learning framework among distributed electric field monitoring points to achieve secure sharing of digital twin model parameters; S42. Differential privacy technology is used to protect the data privacy of each monitoring point; S43. Dynamically adjust the weight contribution between each monitoring point through an adaptive aggregation algorithm; like Figure 1 As shown, furthermore, multimodal features are fused in the intelligent analysis layer to perform adaptive threshold calculation, wherein, For causal reasoning analysis, causal attribution analysis is performed. Specifically, counterfactual reasoning is used to verify the causes of anomalies. The formula is as follows:

[0041] Where Y represents the abnormal event and X represents the potential cause. This represents an intervention (removing cause X), and the causal relationship is verified by comparing the probability of abnormality before and after the intervention.

[0042] For digital twin models, anomaly detection and verification are performed based on anomaly evolution prediction using physical constraints. The formula is as follows:

[0043] in, For predicting the loss function, For physical constraints, These are the constraint weighting coefficients.

[0044] For federated learning collaborative optimization, a differential privacy-preserving aggregation algorithm is used for predictive modeling, with the following formula:

[0045] in, For the local model parameters of the k-th monitoring point, As weight, Gaussian noise was added to protect privacy.

[0046] The specific steps for adaptive threshold generation for multimodal data are as follows: X1. Perform collaborative preprocessing on the collected multimodal data. Collaborative preprocessing includes noise filtering, missing value imputation, and cross-modal normalization. X2. Extract time-domain, frequency-domain, and time-frequency-domain multi-dimensional features to construct a multimodal feature fusion matrix; X3. Utilize multi-head attention mechanism to weightedly fuse features from different modalities to generate a comprehensive feature representation; X4. Based on multimodal fusion features, an adaptive trend modeling method is used to generate dynamic thresholds, including a multivariate time series prediction model that considers the correlation between sensors. X5. Combine the physical constraints of the digital twin model to correct and verify the dynamic threshold.

[0047] like Figure 1 As shown, this embodiment further combines the multiple verification results obtained from the above causal reasoning analysis, digital twin model, and federated learning collaborative optimization. A Bayesian inference framework is used in the decision fusion layer to perform Bayesian anomaly confidence assessment, and the reliability of the anomaly data is comprehensively verified. Specifically, when the assessment determines that the anomaly is real, tiered alarm information and predictive maintenance suggestions are generated; when the assessment determines that the anomaly is a false alarm, the model parameters are updated and the detection threshold is optimized. For Bayesian anomaly confidence assessment, the formula for calculating anomaly confidence by integrating multi-source information is as follows:

[0048] Evidence includes multimodal features, causal inference results, and digital twin simulation results.

[0049] To verify the reliability of outlier data, this embodiment establishes a multiple hypothesis testing method, the formula of which is as follows:

[0050] in, This indicates the null hypothesis (no anomalies). The opposing hypothesis (an anomaly exists) is used. The mean of the pre-defined normal data is used. Bonferroni correction is employed to control the false positive rate in multiple comparisons.

[0051] Where m is the number of comparisons.

[0052] like Figure 1As shown, further, in this embodiment, the final analysis results are output through the output display layer. Specific output methods include real-time alarm system output, visualization interface output, anomaly analysis report output, anomaly trend chart output, historical event archive output, and predictive maintenance suggestions are provided so that maintenance personnel can obtain them promptly. Specifically, In the real-time alarm system output, dynamic threshold determination is performed on the real-time multimodal data. When the multimodal data exceeds the specified threshold, the real-time alarm system is triggered. The formula is:

[0053] in, For current data, This is a dynamic threshold calculated in real time.

[0054] like Figure 1 As shown, the electric field system in this embodiment further supports a manual feedback mechanism. At the feedback learning layer, maintenance personnel provide operational feedback, model optimization, and knowledge updates based on anomaly conclusions. The feedback results are used for parameter updates of multimodal feature data, knowledge enhancement in causal reasoning analysis, model optimization of the digital twin model, and collaborative learning in federated learning collaborative optimization. This gradually improves the accuracy and robustness of the electric field system in anomaly analysis.

[0055] The anomaly reported by operations and maintenance personnel is addressed by optimizing the model parameters by minimizing the loss function, as shown in the formula:

[0056] in, For the genuine labels confirmed by operations and maintenance personnel, For the model prediction results, These are the model parameters. By minimizing the loss function, the large language model with causal inference enhancement and adaptive trend modeling are optimized to improve the system's accuracy.

[0057] like Figure 2 As shown, the specific operation steps of the electric field anomaly data analysis method in this invention are as follows: Y1. Acquire multimodal data; specifically, deploy various types of sensors in the electric field, including voltage sensors, current transformers, electric field strength probes, and environmental monitoring devices (such as temperature and humidity sensors, meteorological sensors, etc.), to acquire multimodal data on the electric field operation in real time. The acquired data is uploaded to the data processing center through an edge computing gateway to form a continuous time-series dataset; Y2. Perform data preprocessing, multimodal feature fusion, and preliminary anomaly screening on the collected multimodal data; specifically, the preprocessing operations on the collected data include: Y21. Noise Filtering: Eliminate burst noise using wavelet denoising or Kalman filtering methods; Y22, Missing value imputation: Filling in missing data through interpolation or model prediction; Y23. Normalization: Normalize data of different dimensions to a comparable scale.

[0058] Y24. After data cleaning, a dynamic threshold is generated using a time series forecasting model (e.g., Autoregressive Moving Average (ARIMA) or Long Short-Term Memory (LSTM) neural network). If data exceeds the dynamic threshold range, it is marked as candidate outlier data.

[0059] Y3. Establish a digital twin model and virtual-real mapping for the fused multimodal data, establish a multi-sensor spatial topology diagram, and perform adaptive threshold calculation based on the multimodal data. Perform anomaly detection based on the adaptive threshold calculation results. If the multimodal data is normal, continue to monitor and collect multimodal data. If the multimodal data is abnormal, perform causal reasoning analysis and federated learning collaborative optimization on the candidate abnormal data. Specifically, in causal reasoning analysis, candidate abnormal data and their contextual information (including timestamps, equipment operating conditions, and data change trends at adjacent times) are input into a pre-trained large language model. The large language model performs reasoning based on historical operating data and a knowledge base to complete the following tasks: Identify the source of the anomaly (e.g., whether voltage fluctuations are caused by sudden load changes, weather interference, or equipment aging). Determine whether the anomaly is a false alarm (such as some data anomalies caused by short-term sensor jitter).

[0060] Y4. For anomalies in multimodal data, a Bayesian inference framework is used to evaluate the confidence level of Bayesian anomalies. If the anomaly is a false alarm, the threshold is optimized and the optimized threshold result is fed back into the adaptive threshold calculation to adjust the adaptive threshold parameters. If the anomaly is a real anomaly, the anomaly is confirmed, and the development trend and possible consequences of the anomaly are predicted. The final analysis results are then output. Specifically, the electric field system cross-validates the inference results of the large language model with the output of the adaptive trend modeling. For example, if the large language model determines that an anomaly is caused by a sensor malfunction, while the adaptive trend modeling shows that the relevant data shows no anomalies in multiple sensors, then the anomaly is determined to be a false alarm; conversely, if the two conclusions are consistent, the anomaly is confirmed as a real anomaly. Finally, the system automatically generates alarms or alerts. Once an anomaly is confirmed, the system outputs results, including real-time alarms (an alarm notification pops up on the maintenance personnel's terminal and is pushed to mobile devices); a visualization interface (displaying the trend curve of the anomaly data, the distribution of anomaly points, and the determination of the source in the monitoring platform); and periodic reports (generating anomaly event analysis reports periodically for subsequent traceability and management).

[0061] Y5. After receiving the final analysis results, the operations and maintenance personnel use a feedback learning mechanism to confirm, correct, and optimize the anomaly conclusions and the digital twin model. The optimized digital twin model data is then fed back to the digital twin model and the virtual-real mapping end for model adjustment. Specifically, after confirming the anomaly results, the operations and maintenance personnel can provide feedback on the system output, such as labeling it as "real anomaly" or "false alarm." This feedback data is stored and used for continuous fine-tuning of the large language model and dynamic optimization of the adaptive trend modeling method, enabling the system to continuously improve its accuracy in identifying different types of anomalies over long-term operation.

[0062] like Figure 3 As shown, the feedback learning mechanism described in this embodiment includes a system operation layer, a feedback processing layer, a model optimization layer, and a knowledge update layer.

[0063] Specifically, operations and maintenance personnel receive system operation layer data, which includes anomaly detection results output, predictive maintenance suggestions, causal analysis reports, and real-time alarm information.

[0064] Maintenance personnel evaluate the received data, including whether the maintenance suggestion is invalid, valid, marked as a false alarm, and confirmed as a genuine anomaly. At the feedback processing layer, feedback data is collected and feedback labels are generated. Feedback quality is evaluated based on these labels, and the feedback data is stored. At the model optimization layer, the stored feedback data is calculated using a loss function, and model performance is evaluated using gradient update algorithms and parameter tuning strategies. Based on the model performance evaluation results, at the knowledge update layer, predictive model improvements are implemented to integrate learning strategies, multimodal feature weights are optimized for hyperparameter tuning, adaptive threshold parameters are adjusted for model structure optimization, and causal relationship updates in causal inference analysis are performed for adaptive learning rate adjustment.

[0065] Further optimization involves monitoring and evaluating the performance of the ensemble learning strategy, and improving the effectiveness evaluation. Specifically, the learning strategy is adjusted, and the adjusted strategy is fed back to the model optimization layer for further optimization, while the current strategy is continued and fed back to the system operation layer along with the update results from the knowledge update layer.

[0066] The second aspect of this embodiment provides a system for analyzing abnormal data from power plants using a causal reasoning-based method. Specifically, it includes a multimodal data acquisition module, a multimodal feature fusion module, a digital twin modeling module, a digital twin simulation verification module, an adaptive threshold generation module, a causal reasoning analysis module, a federated learning collaboration module, a multi-level decision fusion module, a predictive maintenance module, an intelligent output module, and an adaptive learning optimization module. The operation of each module is as follows: Multimodal data acquisition module: Various types of sensors and acquisition terminals are deployed in the electric field to collect and upload multi-dimensional and multimodal monitoring data of the electric field operation, including electrical parameters, environmental parameters, vibration and sound, etc. Multimodal feature fusion module: used to extract and fuse multidimensional features in the time domain, frequency domain, and time-frequency domain to generate a comprehensive feature representation; Digital twin modeling module: used to build digital twin models of electric field equipment, realizing virtual-real mapping and state synchronization; Digital twin simulation verification module: used for simulation verification and predictive analysis of anomalies; Adaptive threshold generation module: used to generate dynamic adaptive thresholds based on multimodal fusion features and physical constraints; Causal Reasoning Analysis Module: Used to build causal graph models, perform anomaly tracing and counterfactual reasoning verification; Federated Learning Collaboration Module: Used to enable secure collaborative learning and model optimization among distributed monitoring points; Multi-level decision fusion module: used to integrate multi-source information, perform Bayesian anomaly confidence assessment, and make intelligent decisions; Predictive maintenance module: Used to generate proactive maintenance suggestions and risk warnings based on anomaly prediction results; Intelligent output module: Used to output multi-level analysis reports containing causal explanations and predictive suggestions; Adaptive learning optimization module: Used to continuously optimize the parameters and performance of each module based on multi-source feedback.

[0067] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

[0068] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0070] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0071] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for analyzing abnormal data from power plants based on causal reasoning, characterized in that, Specifically, the following steps are included: Y1. Collect multimodal data; Y2. Perform data preprocessing, multimodal feature fusion, and preliminary anomaly screening on the collected multimodal data; Y3. Establish a digital twin model and virtual-real mapping for the fused multimodal data, and perform adaptive threshold calculation based on the multimodal data; perform anomaly detection based on the adaptive threshold calculation result; if the multimodal data is normal, continue to monitor and collect the multimodal data; if the multimodal data is abnormal, perform causal reasoning analysis and federated learning collaborative optimization on the candidate abnormal data. Y4. For the aforementioned multimodal data anomalies, a Bayesian inference framework is used to perform Bayesian anomaly confidence assessment. Y5. After receiving the final analysis results, the operation and maintenance personnel use a feedback learning mechanism to confirm and correct the abnormal conclusions and optimize the digital twin model. The optimized digital twin model data is then fed back to the digital twin model and the virtual-real mapping end for model adjustment.

2. The method for analyzing abnormal data of power plants based on causal reasoning according to claim 1, characterized in that, The multimodal data includes voltage, current, electric field strength, temperature and humidity, vibration, sound and environmental interference data. The multimodal feature fusion of the multimodal data adopts a multi-head attention mechanism.

3. The method for analyzing abnormal data of power plants based on causal reasoning according to claim 2, characterized in that, The multimodal data is uploaded to the data processing center through an edge computing gateway to form a continuous time-series dataset.

4. The method for analyzing abnormal data of power plants based on causal reasoning according to claim 1, characterized in that, The data preprocessing includes noise filtering, missing value imputation, data normalization, and feature extraction.

5. The method for analyzing abnormal data of power plants based on causal reasoning according to claim 1, characterized in that, The digital twin model is based on physical constraint optimization and includes a geometric model, a physical model, and a behavioral model.

6. The method for analyzing abnormal data of power plants based on causal reasoning according to claim 5, characterized in that, The causal reasoning analysis employs a counterfactual analysis method; the federated learning collaborative optimization employs a differential privacy-preserving aggregation algorithm.

7. The method for analyzing abnormal data of power plants based on causal reasoning according to claim 1, characterized in that, In the Bayesian anomaly confidence assessment, if the anomaly is a false alarm, the threshold is optimized and the optimized threshold result is fed back to the adaptive threshold calculation to adjust the adaptive threshold parameter. If the anomaly is genuine, it is confirmed, and the development trend and possible consequences of the anomaly are predicted. Finally, the analysis results are output.

8. The method for analyzing abnormal data of power plants based on causal reasoning according to claim 1, characterized in that, The feedback learning mechanism includes a system operation layer, a feedback processing layer, a model optimization layer, and a knowledge update layer.

9. The method for analyzing abnormal data of power plants based on causal reasoning according to claim 8, characterized in that, The operation and maintenance personnel receive the system operation layer data and provide feedback evaluation results on the system operation data; the feedback processing layer collects the feedback data corresponding to each evaluation result and generates feedback tags; In the model optimization layer, a parameter tuning strategy is used to evaluate model performance. Based on the model performance evaluation results, model improvements are made in the predictive modeling process at the knowledge update layer.

10. A system employing the causal reasoning-based abnormal data analysis method for power plant as described in claims 1-9, characterized in that, Specifically, it includes: Multimodal data acquisition module: Various types of sensors and acquisition terminals are deployed in the electric field to collect and upload multi-dimensional and multimodal monitoring data of the electric field operation; Multimodal feature fusion module: used to extract and fuse multidimensional features in the time domain, frequency domain, and time-frequency domain to generate a comprehensive feature representation; Digital twin modeling module: used to build digital twin models of electric field equipment, realizing virtual-real mapping and state synchronization; Digital twin simulation verification module: used for simulation verification and predictive analysis of anomalies; Federated Learning Collaboration Module: Used to enable secure collaborative learning and model optimization among distributed monitoring points; Multi-level decision fusion module: used to integrate multi-source information, perform Bayesian anomaly confidence assessment, and make intelligent decisions; Predictive maintenance module: Used to generate proactive maintenance suggestions and risk warnings based on anomaly prediction results; Adaptive learning optimization module: Used to continuously optimize the parameters and performance of each module based on multi-source feedback.

Images