Power transformer state evaluation and trusted data cleaning method and system

Abstract

The invention belongs to the technical field of transformer data processing, and particularly relates to a power transformer data cleaning method and system. The method mainly comprises the following steps: obtaining original data of the power transformer, carrying out abnormal data detection and correction on the original data based on physical rule filtering, outlier detection and time sequence mode analysis, extracting different types of data features from the corrected data, and carrying out cross-modal feature fusion on the extracted different types of data features; constructing a fault propagation reasoning initial model according to the data features after cross-modal feature fusion, the basic information of the power transformer and historical test data; and dynamically iterating the fault propagation reasoning initial model through a dynamic ontology updating mechanism to form a fault propagation reasoning model so as to evaluate the state of the power transformer. The space-time barrier of multi-source data is broken through, efficient integration and collaborative analysis of the data are realized, a unified data base is provided for transformer state evaluation, and the utilization rate of the multi-source data is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer data processing, and in particular relates to a method for power transformer status assessment and trusted data cleaning. Background Art

[0002] In power systems, transformers are core equipment. Accurate monitoring and data analysis of their operating status are crucial to ensuring safe and stable operation of the power system. With the rapid development of the power industry, the requirements for transformer data processing are also increasing.

[0003] Currently, the field of transformer data processing faces numerous significant challenges. The first is the existing data silo problem, resulting in inconsistent temporal and spatial data benchmarks for data acquired using different detection methods. In actual monitoring, various methods exist for transformer detection, such as online and offline testing. These methods vary in equipment characteristics, detection intervals, and data collection frequencies, making it difficult to standardize the acquired data across time and space. This makes comprehensive analysis of the transformer's overall operating status extremely difficult, making it impossible to fully and accurately grasp the transformer's real-time status, severely impacting early warning and diagnosis of transformer faults.

[0004] Secondly, traditional data cleaning methods have significant flaws. When dealing with transformer data, traditional data cleaning methods are unable to effectively distinguish between true faults and measurement errors. During transformer operation, measurement equipment may be affected by various interference factors, resulting in measurement errors. When a true fault occurs, traditional cleaning methods struggle to accurately identify the true fault signal from complex data, easily leading to misjudgments or omissions. This is extremely detrimental to the timely detection and resolution of transformer faults, potentially delaying the optimal repair opportunity and potentially causing serious power outages.

[0005] Finally, insufficient knowledge representation is a key issue. Transformer operating data contains numerous discrete parameters, and existing knowledge representation methods struggle to effectively utilize these discrete parameters to support state evolution analysis. The operating state of a transformer is a dynamically changing process, and its evolution is closely related to numerous discrete parameters. However, due to the current limitations of knowledge representation, the complex relationships between these discrete parameters cannot be effectively organized and expressed, making it difficult to accurately predict and evaluate the transformer's future operating state, and thus unable to provide strong support for equipment operation and maintenance decisions.

[0006] In summary, the existing transformer data processing technology has serious deficiencies in data silos, data cleaning, and knowledge expression. A new technical solution is urgently needed to solve these problems in order to improve the accuracy and effectiveness of transformer data processing and ensure the safe and stable operation of the power system. Summary of the Invention

[0007] The present invention provides a method for power transformer status evaluation and trustworthy data cleaning to solve the technical problems in the prior art, including data islands caused by inconsistent spatio-temporal benchmarks of multi-source data, difficulty in distinguishing real faults from measurement errors by traditional cleaning methods, and inability to effectively utilize discrete parameters for state evolution analysis in knowledge expression.

[0008] To achieve the above object, the present invention adopts the following technical solutions: A method for power transformer status evaluation and trustworthy data cleaning specifically includes the following steps: Obtain the original data of the power transformer to form a data set of basic information, historical test data, and real-time operation data; Perform abnormal data detection on the real-time operation data of the power transformer based on physical rule filtering, outlier detection, and time series pattern analysis, and perform data correction on the abnormal data based on the artificial intelligence deep learning method; Perform spatio-temporal alignment of multi-modal data on the corrected data, extract different types of data features according to the data type, and perform cross-modal feature fusion on the extracted different types of data features; Construct an initial model for fault propagation inference based on the data features after cross-modal feature fusion, the basic information of the power transformer, and the historical test data; Through the dynamic ontology update mechanism, dynamically iterate the initial model of fault propagation inference to form a fault propagation inference model for power transformer status evaluation.

[0009] The specific method for performing abnormal data detection on the real-time operation data of the power transformer is as follows: Use the Isolation Forest algorithm for outlier detection to detect outliers in the real-time operation data of the power transformer; then adopt the Dynamic Time Warping algorithm to test the morphological similarity of different data curves in the real-time operation data of the power transformer, and mark abnormal data according to the outliers and morphological similarity.

[0010] Abnormal data includes missing data, outliers, and incorrect data. For missing data and outliers, eliminate the outliers and use a conditional generative adversarial network for data filling; for incorrect data, use LSTM model transfer learning for error value correction.

[0011] The specific method for performing spatio-temporal alignment of multi-modal data on the corrected data is as follows: For heterogeneous data in the corrected data, establish a synchronization mechanism based on the equipment working condition status to reflect the equipment working condition status information, and use a time series alignment algorithm to translate and adjust the heterogeneous data according to the working condition status information.

[0012] The Fourier transform algorithm and wavelet transform algorithm are used to extract features from the data after spatio-temporal alignment of multi-modal data, and the principal component analysis algorithm and t-SNE dimensionality reduction algorithm are used to reduce redundant features in the data, completing the extraction of different types of data features.

[0013] According to different types of data features, different weights are assigned to each data feature. Based on different weights, different data features are concatenated together to construct a comprehensive feature vector. According to the comprehensive feature vector, a cross-modal attention mechanism is used to perform cross-modal feature fusion on different types of data features.

[0014] Based on the data features after cross-modal feature fusion, a causal relationship model of equipment working status and fault types is established according to the basic information of the power transformer and historical test data. Each component and parameter of the equipment is analyzed to identify how the fault propagates in the equipment after it occurs. Through data flow, information transfer and other channels, the "propagation chain" of equipment faults is found to achieve the identification of fault propagation paths and form a fault propagation knowledge graph.

[0015] A reasoning engine is formed by using a rule-based reasoning engine combined with a Bayesian network-based reasoning method. A fault propagation reasoning initial model is constructed according to the fault propagation knowledge graph and the reasoning engine.

[0016] A dynamic ontology update mechanism is adopted to perform dynamic iteration on the fault propagation reasoning initial model. Specifically: the fault propagation reasoning initial model is used to compare the real-time data with the established normal mode in the ontology, and the deviation or error index of each parameter is calculated through statistical indicators and machine learning algorithms. When these deviation or error index data exceed the preset threshold, the dynamic update process is triggered. A deep learning model is used to perform pattern recognition on the deviation or error index data exceeding the threshold to extract possible new fault features and patterns. The semantic matching algorithm is used to compare the possible new fault features and patterns with the fault classification in the existing ontology. For fault patterns with low matching degree or completely new ones, new entities and their associated relationships are generated, and the weights of relevant edges are dynamically adjusted; for existing entities and relationships, their attribute values and associated weights are recalibrated according to the real-time data feedback to ensure that the data in the fault propagation reasoning initial model can accurately reflect the current equipment status.

[0017] A power transformer status evaluation and trustworthy data cleaning system, including a data acquisition module, a data cleaning module, a data fusion module, a model construction module, and a status evaluation module; The data acquisition module is used to acquire the original data of the power transformer to form a data set of basic information, historical test data, and real-time operation data; A data cleaning module for detecting abnormal data in the original data of power transformers based on physical rule filtering, outlier detection, and time series pattern analysis, and correcting the abnormal data based on artificial intelligence deep learning methods; A data fusion module for performing spatio-temporal alignment of multi-modal data on the corrected data, extracting different types of data features according to the data type, and performing cross-modal feature fusion on the extracted different types of data features; A model construction module for constructing an initial model of fault propagation inference based on the data features after cross-modal feature fusion, the basic information of the power transformer, and historical test data; A state evaluation module for dynamically iterating the initial model of fault propagation inference through a dynamic ontology update mechanism to form a fault propagation inference model for power transformer state evaluation.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The present invention discloses a method for cleaning power transformer data. By establishing a synchronization mechanism based on the equipment operating conditions, using time series alignment algorithms such as dynamic time warping (DTW) and mutual information method, timestamp standardization and translation adjustment are performed on heterogeneous data such as oil chromatograms, vibration signals, and infrared thermal images to ensure alignment of different data sources within the same time window. Cross-modal feature fusion technologies such as weighted fusion, feature-level fusion, and cross-modal attention mechanism are used to integrate multi-dimensional data such as frequency domain features, image features, and voiceprint features into a comprehensive feature vector. This breaks the spatio-temporal barriers of multi-source data, realizes the efficient integration and collaborative analysis of data, provides a unified data base for transformer state evaluation, and significantly improves the utilization rate of multi-source data.

[0019] Furthermore, the data cleaning link adopts a multi-level abnormal detection system of "boundary constraint of the three-ratio method + isolation forest outlier detection + dynamic time warping time series analysis": the three-ratio method filters out obvious abnormal data through gas concentration thresholds, the isolation forest algorithm combines gas balance constraints and frequency domain energy distribution constraints to identify discrete points, and the dynamic time warping algorithm analyzes the morphological similarity of time series data such as load rate and oil temperature to accurately locate abnormal time series segments.

[0020] Furthermore, for the missing values and error values in the abnormal data, conditional generative adversarial network (cGAN) and LSTM model transfer learning are respectively used for repair to avoid the loss of fault signals caused by misjudgment or missed judgment in traditional methods. This technical system effectively distinguishes real fault features from measurement noise, improves the credibility of the data, and provides a reliable data basis for subsequent fault diagnosis.

[0021] Furthermore, by constructing a fault propagation knowledge graph, combining a causal relationship model and a fault propagation inference engine, in-depth mining of the complex associations between discrete parameters of equipment is achieved. The dynamic ontology update mechanism can iterate the knowledge graph based on real-time data: when the deviation of real-time data exceeds the threshold, a deep learning model is used to identify new fault patterns, new entities and their associated relationships are generated through a semantic matching algorithm, and the attributes and weights of existing entities are calibrated. This dynamically updated knowledge graph can reflect the evolution law of equipment status in real time, effectively support early warning of transformer faults, analysis of propagation paths, and prediction of future states, solve the problem that traditional knowledge representation cannot perform dynamic analysis using discrete parameters, and provide intelligent and accurate knowledge support for operation and maintenance decisions. Brief Description of the Drawings

[0022] Figure 1 : Schematic diagram of the process for power transformer status evaluation and reliable data cleaning; Figure 2 : Schematic diagram of the principle of the power transformer status evaluation and reliable data cleaning system. Detailed Embodiments

[0023] To further understand the content of the present invention, the following provides a detailed description of the present invention in combination with the accompanying drawings and specific embodiments. It should be understood that the embodiments are only for explaining the present invention and not for limiting it.

[0024] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Embodiment 1 Refer to Figure 1 , in the embodiments of the present invention, a method and system for power transformer status evaluation and reliable data cleaning are provided to effectively improve the accuracy and utilization rate of transformer data. The specific embodiments include the following steps.

[0026] S1: Obtain the original data of the power transformer, where the original data includes basic information, historical test data, and real-time operation data; as Figure 1 shown, the basic information includes the status quantities in the links of the power transformer's factory information, handover test, outage test, live detection, on-line monitoring, operation, etc.; among them, the transformer's factory information specifically includes nameplate parameters (rated capacity / voltage / connection group), type test report, and structure drawings. The historical test data includes handover test records (partial discharge quantity / insulation resistance) and annual outage test data (dielectric loss factor / winding deformation). The real-time operation data includes oil chromatography (H2 / CH4 / C2H6 / C2H4 / C2H2), core grounding current, top oil temperature, infrared thermal imaging map, ultrasonic partial discharge signal, and vibration spectrum data. The above basic information, historical test data, and real-time operation data form a data set.

[0027] S2: Data cleaning, detecting abnormal data based on methods such as physical rule filtering, outlier detection, and time series pattern analysis, and filling missing values and correcting error values based on artificial intelligence deep learning methods. Specifically: First, use the triple ratio method to perform boundary constraints on the gas concentration in the real-time operation data. The measured value of the gas concentration should conform to certain physical boundary constraints to avoid incorrect data caused by sensor failures or data transmission problems. Any data that does not meet the above boundary conditions will be marked as abnormal, and the abnormal data that does not meet the boundary conditions will be further processed and eliminated. Taking oil chromatogram data as an example, traverse the gas content data in the oil chromatogram data, and mark the measured values that exceed the preset threshold range as initial abnormal data. For the 、 and gas content data, the boundary constraints are as follows:

[0028]

[0029]

[0030] Among them, reflects the gas generation characteristics in the initial stage of the fault, is sensitive to high-temperature overheating faults, and can assist in judging the fault energy level; is highly sensitive to discharge faults such as partial discharge and spark discharge, and is a key indicator for identifying discharge faults; reflects the development degree of overheating faults, and different ratio ranges correspond to subdivided types such as low-temperature, medium-temperature, and high-temperature overheating.

[0031] By performing boundary constraints on the gas concentration in the real-time operation data, preliminary screening of the abnormal data in the real-time operation data is carried out.

[0032] Furthermore, establish the correlation equation of each data in the real-time operation data and perform constraints on the data. Based on the basic information of the power transformer, such as temperature and humidity data, construct a physical model, calculate the theoretical value range of each data point according to this physical model, and compare the actual measured value with the theoretical value range. If the actual measured value deviates from the range, the data point is regarded as abnormal, and the abnormal data is eliminated to further preliminarily screen the real-time operation data.

[0033] Furthermore, after the preliminary screening of abnormal data through gas concentration boundary constraints and theoretical value range constraints, and after the preliminary physical rule cleaning, the distribution in the real-time operation data is statistically analyzed, discrete points in the real-time operation data are detected and processed. Specifically, the Isolation Forest algorithm is used for outlier detection. The Isolation Forest evaluates the abnormality degree of data points by constructing multiple decision trees, thereby effectively detecting outlier points that are significantly different from other data points. The gas balance constraint is added, that is, the gas concentration should follow the physical laws of gas diffusion and chemical reactions. Within a certain range, the increase in gas concentration should not exceed a certain rate. If the detected change in gas concentration exceeds the expected rate, it is determined as an outlier point. Combining the Isolation Forest algorithm with the frequency-domain energy distribution constraint, the energy distribution of the real-time operation data should not be concentrated in unconventional frequency bands. If there is an abnormal energy distribution in the real-time operation data, it is identified as an outlier point.

[0034] Furthermore, since the real-time operation data of power transformers often exhibit time-series characteristics, according to the characteristics and correlations between different data, the Dynamic Time Warping algorithm is used to capture the morphological similarity between different data curves. The Dynamic Time Warping algorithm accumulates the point-by-point distances of two curves to find an optimal matching path, thereby evaluating the overall similarity of the two curves. If the similarity is high, it indicates that the changes in the real-time operation data are normal; if the similarity is low, there may be abnormal situations. In this embodiment, the similarity threshold is set to 0.5. If the calculated similarity is lower than this threshold, the time period is marked as an abnormal time-series segment, and the abnormal time-series data is repaired or removed to avoid interfering with the subsequent analysis results. Taking the load rate and oil temperature as two typical time-series data indicators as an example. The load rate represents the working load of the transformer, and the oil temperature is one of the key indicators of the health status of the transformer. Especially during overload or long-term operation, the oil temperature usually rises. The relationship between the two is very close. The change in the load rate usually affects the change in the oil temperature, but this relationship is not completely linear. Using the Dynamic Time Warping algorithm to capture the morphological similarity between the load rate curve and the oil temperature curve, the two curves may have different time-series lengths or asynchronous time axes. The Dynamic Time Warping algorithm accumulates the point-by-point distances of two curves to find an optimal matching path, thereby evaluating the overall similarity of the two curves. If the similarity is high, it indicates that the changes in the load rate and the oil temperature are normal; if the similarity is low, there may be abnormal situations.

[0035] Furthermore, data repair processing is performed on the detected abnormal data. Abnormal data includes missing data, outliers, and incorrect data. For missing data and outliers, the outliers are removed and conditional generative adversarial networks are used for data filling; for incorrect data, LSTM model transfer learning is used to correct the incorrect values. Specifically, for missing data and outliers, after removing the outliers, generative adversarial networks can be used to fill in the missing values. In specific implementation, the LSTM model adopts conditional generative adversarial networks, taking the real-time operation data after initially removing abnormal data as the input, and generating filling data that conforms to physical laws (such as gas diffusion kinetics). According to historical test data, the LSTM model can generate reasonable missing values by learning the gas generation patterns in the historical test data, filling the corresponding missing positions without introducing additional noise. For example, an LSTM model is trained on a 300MW generator to learn the normal vibration mode; then, the LSTM model is transferred to a 500MW generator, and the LSTM model is fine-tuned by receiving a small amount of 500MW data. The LSTM model can quickly adapt to the data characteristics of the 500MW generator, achieving the purpose of correcting incorrect data caused by sensor failures or data loss.

[0036] S3: Data fusion. Based on the multi-modal data spatio-temporal alignment method, a time scale synchronization mechanism and heterogeneous data feature extraction technology for real-time operation data are established based on the data set of basic information, historical test data, and real-time operation data, and a cross-modal attention mechanism that fuses frequency domain dielectric response and infrared thermogram is used.

[0037] Specifically, in multi-modal data fusion, the real-time operation data comes from different sensors and acquisition devices, so it is necessary to solve their spatio-temporal alignment problems. The equipment condition state information of the power transformer is extracted according to historical test data. Based on the equipment condition information, by establishing a synchronization mechanism based on the equipment condition state, the consistency of timestamps of different data sources is ensured, and the changes in the state of the power transformer can be accurately reflected. For example, for the working state of a transformer or generator, a condition state model is established, considering factors such as the load, temperature, and operation time of the transformer or generator, for dynamically adjusting the time scale synchronization between different data sources. Through the equipment condition state information, a time series alignment algorithm is used to synchronize different data in the real-time operation data of the transformer or generator. The specific implementation of the data synchronization of the transformer is as follows: First, the timestamps of the data collected by different sensors are standardized, and then the time series of different data are translated and adjusted through the equipment condition state to ensure that the comparison data is within the same time window, reflecting the real equipment operation state. The time series alignment algorithms adopted in this embodiment include the dynamic time warping algorithm, the mutual information method, and the time window method, realizing data fusion on different time scales while ensuring data accuracy.

[0038] Furthermore, feature extraction needs to be performed on heterogeneous data. Heterogeneous data refers to data from different types of sensors, such as oil chromatogram data, vibration signals, temperature data, atlas data, and voiceprint data, which have different physical meanings, data distributions, and data structures. Appropriate feature extraction techniques need to be designed to extract key information from different data sources and achieve feature fusion. For image data in real-time operation data, using image segmentation techniques, the thermal image is divided into different regions, and features such as the temperature mean and standard deviation of each region are extracted. Object detection or image classification techniques are used to identify abnormal hot spots (such as hot spots) in the thermal image and extract their shape, area, intensity, and other features. For voiceprint data, in the time domain, features such as the mean, standard deviation, maximum value, and minimum value of the voiceprint signal can be calculated to help describe the distribution and fluctuation characteristics of the sound signal. For the voiceprint signal, FFT is used to extract the frequency components with high energy proportion in the spectrum to analyze the frequency characteristics generated by the operation of the equipment. For data other than image data and voiceprint data in the operation data, a method combining the Fourier transform algorithm and the wavelet transform algorithm is used to extract features such as frequency and amplitude; dimensionality reduction methods such as principal component analysis (PCA) and t-SNE are adopted to reduce redundant features and improve the calculation efficiency. On the basis of ensuring data integrity, representative features are extracted to reduce noise interference. Through the above feature extraction of heterogeneous data, data features of different types are obtained.

[0039] Furthermore, according to different data types, feature extraction methods are formulated respectively. After extracting data features of different types, cross-modal feature fusion of the extracted data features of different types is required. The specific fusion methods are as follows: (2) Weighted fusion: According to different types of data features, different weights are assigned to each data feature, and weighted fusion is performed according to the importance of each data feature for equipment status evaluation; (2) Feature-level fusion: Different data features are concatenated together to construct a comprehensive feature vector; (3) Attention mechanism: According to the comprehensive feature vector, a cross-modal attention mechanism is used to dynamically adjust the importance of different data features in the final fusion so that the subsequent constructed model can automatically focus on the most valuable features.

[0040] S4: Construct an initial model for fault propagation reasoning. Based on the fault propagation reasoning method and the dynamic ontology update mechanism, update and iterate the initial model for fault propagation reasoning so that it can better adapt to the changes in the operating state of the power transformer and provide more accurate knowledge support for subsequent fault diagnosis and health assessment.

[0041] Based on the spatio-temporal alignment and feature extraction of multi-modal data, in this embodiment, a causal relationship model of the equipment working state and fault type is established through a graph-structured method, based on the data features after cross-modal feature fusion, according to the basic information of the power transformer and historical test data. Through the causal relationship model of the equipment working state and fault type, the monitoring data of each component and parameter of the equipment is analyzed to identify how the fault propagates in the equipment after it occurs. Through data flow, information transmission and other channels, the "propagation chain" of the equipment fault is found to realize the identification of the fault propagation path and form a fault propagation knowledge graph.

[0042] The knowledge graph provides a comprehensive and structured knowledge framework for subsequent fault diagnosis, health assessment and predictive maintenance. The nodes in the knowledge graph usually represent entities such as equipment, states, fault types, etc. For example, a power transformer can be a node, the working state of the power transformer (such as "normal" or "overheated") can be another node, and the fault type (such as "insulation fault", "short circuit") can also be a node. The connection between nodes is called an edge, which represents the relationship between nodes. The relationship between "load" and "oil temperature" can be connected by an edge, indicating their relevance; the relationship between the "oil chromatogram data" node and the "fault type" node indicates the impact of oil chromatogram data on fault diagnosis.

[0043] Further, in this embodiment, a rule-based inference engine is used in combination with a Bayesian network-based inference method to form an inference engine. By analyzing the operation data and fault information of each part of the equipment, the occurrence probability, influence range and possible fault modes of the equipment fault are dynamically inferred. For example, when the vibration signal appears abnormal, combined with the oil chromatogram data and temperature data, it can be inferred that the transformer may have faults such as overheating or internal short circuit. Through fault propagation reasoning, it is further analyzed that this fault may affect other parts of the transformer or adjacent equipment. According to the fault propagation knowledge graph and the inference engine, an initial model of fault propagation reasoning is constructed.

[0044] The above steps have established a preliminary initial model of fault propagation reasoning. In this embodiment, on this basis, a dynamic ontology update mechanism is introduced to dynamically iterate this initial model of fault propagation reasoning. The specific implementation method is as follows: First, a 3D model of the power transformer is constructed according to the basic information and historical test data of the power transformer to realize the mapping of entities and attributes. The real-time operation data is preprocessed by the aforementioned data cleaning. Subsequently, the initial model of fault propagation reasoning is used to compare the real-time operation data with the established normal mode in the 3D model of the power transformer, and the deviation or error index of each parameter is calculated through statistical indicators and machine learning algorithms. When these errors exceed the preset threshold, the dynamic update process is triggered.

[0045] In the update process, first, a deep learning model is used to perform pattern recognition on newly collected abnormal data, extracting possible new fault features and patterns. The deep learning model in this embodiment can be a convolutional neural network or a long short-term memory network trained with historical data. Then, a semantic matching algorithm is used to compare the newly recognized patterns with the fault classifications in the existing ontology. The semantic matching algorithm in this embodiment is a similarity calculation algorithm based on word vectors or graph embeddings. For fault patterns with low matching degrees or completely new ones, the system automatically generates new entities and their associated relationships and dynamically adjusts the weights of relevant edges. At the same time, for existing entities and relationships, their attribute values and associated weights are recalibrated according to real-time data feedback to form a fault propagation inference model, ensuring that the fault propagation inference model can accurately reflect the current device state.

[0046] The entire update process adopts an incremental update strategy, avoiding the high computational overhead of full graph reconstruction, enabling the knowledge graph to respond quickly in an environment with a large amount of data and frequent changes in device states, and being synchronized to the fault diagnosis and health assessment modules in real time to support device state prediction and fault warning based on the latest knowledge. By continuously optimizing the fault propagation inference and ontology update algorithms, this mechanism significantly improves the intelligence and automation levels of the power transformer condition monitoring system, providing a solid knowledge foundation for equipment maintenance and preventive maintenance.

[0047] Embodiment 2 This embodiment proposes a power transformer condition assessment and trustworthy data cleaning system, which can implement the steps of the power transformer condition assessment and trustworthy data cleaning method described in Embodiment 1. The schematic diagram of the system principle is as Figure 2 shown in the figure, and mainly includes a data acquisition module, a data cleaning module, a data fusion module, a model construction module, and a device condition assessment module; The data acquisition module is used to acquire the original data of the power transformer to form a dataset of basic information, historical test data, and real-time operation data; The data cleaning module is used to detect abnormal data in the original data of the power transformer based on physical rule filtering, outlier detection, and time series pattern analysis, and correct the abnormal data based on artificial intelligence deep learning methods; The data fusion module is used to perform spatio-temporal alignment of multi-modal data on the corrected data, extract different types of data features according to the data types, and perform cross-modal feature fusion on the extracted different types of data features; The model construction module is used to construct an initial fault propagation inference model based on the data features after cross-modal feature fusion, the basic information of the power transformer, and the historical test data; The device status evaluation module is used to dynamically iterate the initial model of fault propagation reasoning through a dynamic ontology update mechanism to form a fault propagation reasoning model for power transformer status evaluation.

[0048] In addition, it should be understood that although this specification is described according to the embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art. The above content is only to illustrate the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention fall within the protection scope of the claims of the present invention.

Claims

1. A method for power transformer condition assessment and trustworthy data cleaning, characterized in that, Specifically, it includes the following steps: Obtain the original data of the power transformer to form a dataset of basic information, historical test data, and real-time operation data; Perform abnormal data detection on the real-time operation data of the power transformer based on physical rule filtering, outlier detection, and time series pattern analysis, and perform data correction on the abnormal data based on the artificial intelligence deep learning method; Perform spatio-temporal alignment of multi-modal data on the corrected data, extract different types of data features according to the data type, and perform cross-modal feature fusion on the extracted different types of data features; Construct an initial model of fault propagation reasoning based on the data features after cross-modal feature fusion, the basic information of the power transformer, and the historical test data; Through the dynamic ontology update mechanism, dynamically iterate the initial model of fault propagation reasoning to form a fault propagation reasoning model for power transformer status assessment.

2. A method for power transformer status evaluation and reliable data cleaning according to claim 1, characterized in that The specific method for performing abnormal data detection on the real-time operation data of the power transformer is as follows: Use the Isolation Forest algorithm for outlier detection to detect outliers in the real-time operation data of the power transformer; then adopt the Dynamic Time Warping algorithm to test the morphological similarity of different data curves in the real-time operation data of the power transformer, and mark the abnormal data according to the outliers and morphological similarity.

3. A method for power transformer status evaluation and reliable data cleaning according to claim 2, characterized in that, Abnormal data includes missing data, outliers, and incorrect data. For missing data and outliers, eliminate the outliers and use the Conditional Generative Adversarial Network for data filling; for incorrect data, use the LSTM model transfer learning for error value correction.

4. A method for power transformer status evaluation and reliable data cleaning according to claim 1, characterized in that The specific method for performing spatio-temporal alignment of multi-modal data on the corrected data is as follows: For the heterogeneous data in the corrected data, establish a synchronization mechanism based on the equipment working condition status to reflect the equipment working condition status information, and use the time series alignment algorithm to translate and adjust the heterogeneous data according to the working condition status information.

5. A method for power transformer condition assessment and trustworthy data cleaning according to claim 4, characterized in that, Adopt the Fourier transform algorithm and the wavelet transform algorithm to extract features from the data after spatio-temporal alignment of multi-modal data, and adopt the principal component analysis algorithm and the t-SNE dimensionality reduction algorithm to reduce the redundant features in the data to complete the extraction of different types of data features.

6. A method for power transformer status evaluation and reliable data cleaning according to claim 5, characterized in that, According to different types of data features, assign different weights to each data feature. According to different weights, splice different data features together to construct a comprehensive feature vector. According to the comprehensive feature vector, use the cross-modal attention mechanism to perform cross-modal feature fusion on different types of data features.

7. A method for power transformer condition assessment and reliable data cleaning according to claim 6, characterized in that, Based on the data features after cross-modal feature fusion, establish a causal relationship model between the equipment working state and the fault type according to the basic information of the power transformer and the historical test data, analyze each component and parameter of the equipment, identify how the fault propagates in the equipment after the fault occurs, and find the "propagation chain" of the equipment fault through data flow, information transmission, etc., to realize the identification of the fault propagation path and form a fault propagation knowledge graph.

8. A method for power transformer status assessment and reliable data cleaning according to claim 7, characterized in that Use a rule-based inference engine combined with a Bayesian network-based inference method to form an inference engine, and construct an initial model of fault propagation reasoning according to the fault propagation knowledge graph and the inference engine.

9. A method for power transformer status assessment and reliable data cleaning according to claim 8, characterized in that, The initial model of fault propagation reasoning is dynamically iterated using a dynamic ontology update mechanism, specifically: the initial model of fault propagation reasoning is used to compare real-time data with the established normal mode in the ontology, and deviation or error indicators of each parameter are calculated through statistical indicators and machine learning algorithms. When these deviation or error indicator data exceed a preset threshold, a dynamic update process is triggered. A deep learning model is used to perform pattern recognition on the deviation or error indicator data exceeding the threshold, extract possible new fault features and patterns, and use a semantic matching algorithm to compare the possible new fault features and patterns with the fault classification in the existing ontology. For fault patterns with low matching degree or completely new ones, new entities and their association relationships are generated, and the weights of relevant edges are dynamically adjusted; for existing entities and relationships, their attribute values and association weights are recalibrated according to real-time data feedback to ensure that the data in the initial model of fault propagation reasoning can accurately reflect the current device state.

10. A power transformer condition assessment and reliable data cleaning system, based on the power transformer condition assessment and reliable data cleaning method according to any one of claims 1 to 9, characterized in that It includes a data acquisition module, a data cleaning module, a data fusion module, a model construction module, and a state evaluation module; The data acquisition module is used to acquire the original data of the power transformer to form a data set of basic information, historical test data, and real-time operation data; The data cleaning module is used to detect abnormal data in the original data of the power transformer based on physical rule filtering, outlier detection, and time series pattern analysis, and correct the abnormal data based on artificial intelligence deep learning methods; The data fusion module is used to perform spatio-temporal alignment of multi-modal data on the corrected data, extract different types of data features according to the data type, and perform cross-modal feature fusion on the extracted different types of data features; The model construction module is used to construct an initial model of fault propagation reasoning according to the data features after cross-modal feature fusion, the basic information of the power transformer, and historical test data; The state evaluation module is used to dynamically iterate the initial model of fault propagation reasoning through a dynamic ontology update mechanism to form a fault propagation reasoning model for power transformer state evaluation.

Images