Transformer state evaluation method based on improved deep learning algorithm

Abstract

The application relates to the field of state evaluation, and discloses a transformer state evaluation method based on an improved deep learning algorithm, which is used for intelligent state evaluation and health degree analysis of a transformer. The method comprises the following steps: firstly, collecting multi-source monitoring time sequence data to generate characteristic data containing a trend entropy index; inputting the deep learning network to perform dimension reduction; introducing an information entropy maximization principle to optimize state code vector distribution during training; forming a state clustering area by using historical data; determining the state by calculating the distance between a real-time vector and each area and outputting a health degree score. In addition, the method can also identify a new abnormal state, generate an expert analysis report, and expand a knowledge base. The application effectively constrains network training, accurately clusters and matches states, and has the abilities of new abnormal state identification and knowledge expansion, so that the accuracy, reliability and adaptability of transformer state evaluation can be improved, and the safe and efficient operation of a power grid can be ensured.

Description

Technical Field

[0001] This invention relates to the field of state assessment, and more particularly to a transformer state assessment method based on an improved deep learning algorithm. Background Technology

[0002] In power systems, transformers are critical equipment, and their operational stability and reliability directly affect the safe and efficient operation of the entire power grid. With the continuous expansion of the power grid and its increasing intelligence, higher demands are placed on the accuracy and real-time performance of transformer condition assessments. Traditional transformer condition assessment methods mainly rely on periodic manual inspections and offline tests. These methods are not only time-consuming and labor-intensive, but also fail to reflect the actual operating status of the transformer in real time, making it difficult to detect potential faults in a timely manner.

[0003] In recent years, with the rapid development of IoT, big data, and AI technologies, transformer condition assessment methods based on multi-source monitoring data have gradually become a research hotspot. By collecting multi-source monitoring time-series data such as oil chromatography data, electrical load data, and temperature data during transformer operation, and using advanced algorithms such as deep learning for data fusion and analysis, real-time monitoring and accurate assessment of transformer condition can be achieved. However, existing deep learning-based transformer condition assessment methods still have some shortcomings, limiting their effectiveness and widespread application in practice.

[0004] Existing methods, when processing multi-source monitoring data, often only focus on the static features or simple time-series features of the data, while ignoring dynamic change characteristics such as trend entropy and key information such as fault chemical characteristics. This results in incomplete feature extraction and affects the accuracy of condition assessment.

[0005] During the training process of deep learning networks, existing methods typically rely solely on supervised learning signals, lacking effective constraints on the probability distribution characteristics of state encoding vectors. This results in the encoded vectors generated by the trained network potentially failing to accurately reflect the actual operating state of the transformer, thus reducing the reliability of state assessment.

[0006] Existing methods often rely solely on simple spatial partitioning of historical state encoding vectors when forming state clustering regions, neglecting the geometric distribution characteristics between vectors and the consistency of state labels. This results in low reliability of state clustering regions, which in turn affects the accuracy of real-time state matching.

[0007] Therefore, we propose a transformer state evaluation method based on an improved deep learning algorithm to solve the above problems. Summary of the Invention

[0008] This invention provides a transformer condition assessment method based on an improved deep learning algorithm, which is used for intelligent condition assessment and health analysis of transformers.

[0009] The first aspect of the present invention provides a transformer condition assessment method based on an improved deep learning algorithm, the transformer condition assessment method based on the improved deep learning algorithm includes: collecting multi-source monitoring time-series data during transformer operation, calculating and generating feature data; The feature data is input into a deep learning network for fusion and dimensionality reduction, and the output is a state encoding vector. During the training process of the deep learning network, the probability distribution characteristics of the state encoding vector are optimized by adjusting the network parameters to complete the training of the network. The trained deep learning network is used to process historical monitoring data of known states to obtain the corresponding historical state encoding vector set, and spatial division is performed to form state clustering regions corresponding to various historical operating states. The real-time monitoring data of the transformer is input into the trained deep learning network to obtain a real-time state encoding vector. The distance between the vector and each state clustering region is calculated. The current operating state is determined based on the minimum distance, and a continuously quantified health score is output.

[0010] Optionally, in a first implementation of the first aspect of the present invention, the method includes: Construct a training sample set containing historical monitoring data of transformers and their corresponding status labels; Based on the state encoding vector and multiple preset state categories, the similarity relationship between the state encoding vector of each training sample and each state category is calculated. Based on the similarity relationship, determine the probability distribution of each training sample belonging to each state category; Calculate the information entropy of the probability distribution, which serves as the entropy value characterizing the uncertainty of the sample state; The parameters of the deep learning network are adjusted with the goal of optimizing the entropy value, so that the state encoding vector generated by the network after training has a distribution characteristic constrained by the information entropy theory.

[0011] Optionally, in a second implementation of the first aspect of the present invention, the samples in the training sample set are grouped based on the state labels, and the distribution density of the state encoding vector of the samples in each group in space is calculated respectively. Based on the distribution density, a theoretical entropy target value reflecting the typical degree of uncertainty is set for each state category; The information entropy of the probability distribution of each training sample is compared with the theoretical entropy target value of its state category to generate an entropy deviation. With minimizing the entropy deviation as one of the objectives, and combining the supervision signal generated by the similarity relationship, a comprehensive optimization objective is formed; The network parameters are adjusted according to the comprehensive optimization objective, so that the uncertainty of the state encoding vector generated during the training process converges to a level that matches the inherent characteristics of each category.

[0012] Optionally, in a third implementation of the first aspect of the present invention, the method includes: The historical monitoring data of the known state is input into the trained deep learning network to obtain the corresponding historical state encoding vector, thus forming a historical state encoding vector set; Based on the status labels of the historical monitoring data, the historical status encoded vector set is grouped to obtain a vector group set corresponding to different historical operating states. For each vector group set, calculate the spatial distribution center of its internal vectors to generate a state cluster center point representing the historical running state. Taking each state cluster center as the core, and combining the distribution range of its corresponding vector group set, the corresponding state clustering region is delineated in the state coding vector space; The spatial location information of each state clustering region and its corresponding state clustering center point is stored to form a state clustering region.

[0013] Optionally, in a fourth implementation of the first aspect of the present invention, for each state clustering region, the geometric distribution characteristics of its internal historical state encoding vector are calculated to obtain a geometric characteristic description of the compactness and consistency of the region. Based on the geometric characteristic description and the consistency of the state labels of the corresponding historical monitoring data, the credibility of each state clustering region is evaluated, and a region credibility evaluation result is generated. Based on the regional credibility assessment results, historical state encoding vectors within state clustering regions that are below a preset credibility threshold are filtered out, and vectors that deviate too far from the state cluster center point are removed to form a high-credibility vector subset. The state cluster centers are recalculated using the high-confidence vector subset to generate optimized state cluster centers; Based on the distribution range of the optimized state cluster centers and their corresponding high-confidence vector subsets, the corresponding state clustering regions are redefined to form optimized state clustering regions.

[0014] Optionally, in a fifth implementation of the first aspect of the present invention, the method includes: The real-time monitoring data is input into the trained deep learning network to obtain the corresponding real-time state encoding vector; Calculate the spatial distance between the real-time state encoding vector and each of the state cluster centers to form a distance set; The minimum distance value is selected from the distance set, and the state clustering region corresponding to the minimum distance value is determined as the current running state; Based on the minimum distance value, a continuous health score reflecting the health status of the transformer is calculated through a preset mapping relationship; The output includes the current operating status and the continuous health score, resulting in a final status assessment.

[0015] Optionally, in a sixth implementation of the first aspect of the present invention, the method further includes identifying and expanding knowledge of novel abnormal states: When the minimum distance between the real-time state encoding vector and all the state clustering regions exceeds a preset threshold, it is determined that the current transformer is in a new abnormal state, and a new abnormal state label is generated. The real-time monitoring data carrying the novel abnormal state marker and its corresponding real-time state encoding vector are stored to form a novel abnormal state dataset. Based on the novel abnormal state dataset, an expert analysis report containing the causes of the fault and handling suggestions is generated by combining expert diagnostic analysis. Based on the expert analysis report, corresponding state clustering regions are added to the state encoding vector space to form an updated state clustering pattern.

[0016] Optionally, in the seventh implementation of the first aspect of the present invention, the real-time monitoring data in the novel abnormal state dataset is submitted to an expert system for diagnosis to obtain an expert diagnosis conclusion that includes the fault type and severity level. Based on the expert diagnosis, a status description vector is generated, which includes the fault type code, severity level value, and recommended processing priority. The state description vector is associated and bound with the corresponding real-time state encoding vector in the novel abnormal state dataset to form a state vector pair with expert labels. Collect all the state vector pairs with expert labels to construct an expert knowledge set; Based on the aforementioned expert knowledge set, an expert analysis report containing the causes of the fault and suggested solutions is generated.

[0017] Beneficial effects: Feature data is screened and reconstructed based on information entropy, a subset of features with stable information is selected, and further processed by mutual information analysis and principal component analysis to remove redundant information, extract key features, improve the quality and effectiveness of feature data, and enhance the model's ability to identify transformer states. Based on the characteristics of state distribution, the entropy target is guided. The theoretical entropy target value is set according to the typical uncertainty of different state categories. The network parameters are adjusted with minimizing the entropy deviation as one of the objectives, so that the uncertainty of the generated state encoding vector converges to a level that matches the inherent characteristics of each category, thereby further improving the accuracy and reliability of the model. This improves the reliability of state clustering regions, making real-time state matching more accurate and enabling more precise determination of the transformer's current operating status. Not only can it determine the current operating status, but it can also calculate a continuous health score based on the distance between the real-time state encoding vector and the state clustering region through a preset mapping relationship. This provides more detailed and intuitive quantitative indicators for transformer health management, facilitating maintenance personnel to promptly grasp the trend of transformer status changes. By combining novel abnormal state data with expert diagnostic analysis to generate expert analysis reports, and adding corresponding state clustering regions to the state coding vector space, the system's knowledge is dynamically expanded. This enables the assessment system to continuously adapt to changes in transformer operating conditions, and with the passage of time and the accumulation of data, the accuracy and comprehensiveness of the assessment are continuously improved. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of an embodiment of the transformer state assessment method based on an improved deep learning algorithm in this invention. Detailed Implementation

[0019] This invention provides a transformer condition assessment method based on an improved deep learning algorithm for intelligent transformer condition assessment and health analysis. The terms "first," "second," "third," "fourth," etc. (if applicable) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the transformer state assessment method based on an improved deep learning algorithm in this invention includes: 101. Collect multi-source monitoring time-series data during transformer operation, calculate and generate feature data including trend entropy indicators.

[0021] It is understood that the executing entity of this invention can be a transformer state assessment device based on an improved deep learning algorithm, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0022] Specifically, transformer oil chromatographic data, electrical load data, and temperature data are collected to form raw time-series data for multi-source monitoring. The various data types in the raw time-series data are normalized to eliminate physical dimensions, resulting in normalized time-series data. Based on the normalized time-series data, the information entropy of the value distribution of each key parameter sequence is calculated according to a preset time window, yielding a trend entropy index reflecting the disorder of parameter fluctuations. Based on the normalized time-series data, the concentration ratios between preset gas components are calculated to generate fault-indicating characteristic ratio data. Finally, the trend entropy index and characteristic ratio data are integrated to construct a multi-dimensional feature set containing time-series variation characteristics and fault chemical characteristics, serving as feature data.

[0023] It should be noted that, taking a 110kV oil-immersed power transformer as an example, firstly, during a continuous 7-day monitoring period, multi-source monitoring time-series data of the transformer were collected at fixed times each day. The raw data included: oil chromatographic data (concentrations of five characteristic gases: hydrogen, methane, ethane, ethylene, and acetylene, in μL / L). The data for the first day were H2:50, CH4:20, C2H6:10, C2H4:5, C2H2:2; the data for the seventh day were H2:55, CH4:22, C2H6:12, C2H4:6, C2H2:3, with reasonable fluctuations in data throughout the period. Electrical load data was the daily average load rate (percentage), gradually increasing from 80% on the first day to 85% on the seventh day. Temperature data was the daily highest top-layer oil temperature (in °C), gradually increasing from 65 °C to 68 °C.

[0024] The min-max normalization method was used to independently scale each parameter sequence to the [0,1] interval. For the hydrogen concentration sequence, the minimum value was 50 μL / L and the maximum value was 55 μL / L. Therefore, the normalized value for the first day was (50-50) / (55-50)=0, and the normalized value for the seventh day was (55-50) / (55-50)=1. The load rate sequence was scaled with a minimum of 80% and a maximum of 85%, and the oil temperature sequence was scaled with a minimum of 65℃ and a maximum of 68℃ to obtain the normalized time-series data for all parameters.

[0025] A 3-day time window was set, and the information entropy of the value distribution was calculated for each key parameter sequence (normalized hydrogen concentration sequence). Taking the first three days of the hydrogen sequence (normalized values ​​0, 0.2, 0.4) as an example, its value range was divided into 5 equal intervals, and the frequency distribution of values ​​falling into each interval was statistically analyzed. The frequency distribution was [0.33, 0.33, 0.33, 0, 0]. Based on this, the information entropy was calculated to be approximately 1.1 (a higher entropy value indicates stronger fluctuation and disorder). The trend entropy of each parameter sequence was calculated sequentially to form a trend entropy index.

[0026] The key gas ratios were calculated to diagnose arc discharge (acetylene / ethylene ratio, C2H2 / C2H4, 2 / 5=0.4 on the first day and 3 / 6=0.5 on the seventh day) and low-temperature overheating (methane / hydrogen ratio, CH4 / H2, 20 / 50=0.4 on the first day and 22 / 55=0.4 on the seventh day), and the ratio sequence was obtained.

[0027] The five gas trend entropies, load trend entropy, and oil temperature trend entropy, along with the two gas ratios, are combined to form eight features, forming preliminary feature data.

[0028] 102. Input the feature data into a deep learning network for fusion and dimensionality reduction, and output a state encoding vector that reflects the overall operating status of the transformer.

[0029] Specifically, the feature data is separated into static feature data reflecting static characteristics and time-series feature data reflecting temporal change characteristics; a high-dimensional static feature vector is obtained by performing a nonlinear transformation on the static feature data through a first neural network path; a time-series pattern is extracted from the time-series feature data through a second neural network path to obtain a time-series feature representation vector; the high-dimensional static feature vector and the time-series feature representation vector are concatenated in the feature dimension to generate a fused feature vector; and the fused feature vector is reduced in dimension and encoded through a third neural network path to output a state encoding vector. Furthermore, it also includes feature selection and reconstruction based on information entropy of the feature data: calculating the information entropy value of each feature dimension sequence in the feature data to obtain the information entropy sequence of each feature; comparing the information entropy sequence with a preset entropy threshold range to select features whose entropy values ​​are within the threshold range, forming an information-stable feature subset; based on the information-stable feature subset, calculating the mutual information between different feature dimensions to generate a mutual information matrix representing the correlation between features; according to the mutual information matrix, grouping and aggregating the features in the information-stable feature subset to form multiple feature aggregation groups with strong internal correlation; performing principal component analysis on each feature aggregation group, extracting its principal components as new feature dimensions, and combining all principal components into reconstructed feature data for subsequent separation and processing.

[0030] It should be noted that, following the 8-dimensional feature data obtained in step 101, the implementation process of step 102 will now be illustrated with a specific example. However, before that, the features need to be filtered and reconstructed based on information entropy.

[0031] An 8-dimensional feature data matrix for 7 days was obtained from step 101, and the information entropy of each feature dimension sequence was calculated. Assuming that after calculation, the entropy values ​​are: Feature 1 (hydrogen trend entropy) 1.3, Feature 2 (methane trend entropy) 0.7, Feature 3 (ethane trend entropy) 1.5, Feature 4 (ethylene trend entropy) 1.0, Feature 5 (acetylene trend entropy) 0.5, Feature 6 (load trend entropy) 1.2, Feature 7 (oil temperature trend entropy) 0.9, and Feature 8 (C2H2 / C2H4 ratio) 1.1. The preset entropy threshold range is [0.8, 1.5]. After filtering, Features 1, 3, 4, 6, and 8 are retained to form a stable feature subset. Next, the mutual information among these five retained features is calculated. It is found that Feature 1 (hydrogen trend entropy) and Feature 6 (load trend entropy) are highly correlated, Feature 3 and Feature 4 (ethane and ethylene trend entropy) are highly correlated, while Feature 8 (ratio) is relatively independent. Based on this, the first four features are divided into two groups, and feature 8 forms its own group. Principal component analysis is performed on each group. One principal component (F1) is extracted from the first group (features 1 and 6), one principal component (F2) is extracted from the second group (features 3 and 4), and the third group (feature 8) is retained (F3). Finally, the reconstructed feature data contains three new dimensions: F1, F2, and F3, which are the actual features input into the deep learning network.

[0032] The network separates this reconstructed feature (F1, F2, F3). F3 (gas ratio feature) is regarded as static feature data reflecting static chemical properties; F1 and F2 (derived from the fusion of multiple trend entropies) are regarded as time-series feature data reflecting time-series change patterns.

[0033] The static feature data F3 is processed through the first neural network path (a three-layer fully connected network). Assuming F3 is a scalar value of 0.45, after nonlinear transformation by the network, it is expanded into an 8-dimensional high-dimensional static feature vector [0.1, 0.3, -0.2, 0.4, 0.0, 0.5, 0.2, -0.1].

[0034] The temporal feature data F1 and F2 (a 7-day × 2-dimensional sequence) are processed through a second neural network path (a Long Short-Term Memory network LSTM). This network extracts the temporal patterns within the 7 days and finally outputs a 5-dimensional temporal feature representation vector [0.4, 0.8, -0.3, 0.6, 0.1].

[0035] The 8-dimensional high-dimensional static feature vector is concatenated with the 5-dimensional temporal feature representation vector to generate a 13-dimensional fused feature vector [0.1,0.3,-0.2,0.4,0.0,0.5,0.2,-0.1,0.4,0.8,-0.3,0.6,0.1].

[0036] The 13-dimensional fused feature vector is input into the third neural network path (a two-layer fully connected network) for dimensionality reduction and encoding, outputting a low-dimensional, comprehensive 4-dimensional state encoding vector [0.7, -0.2, 0.5, 0.3].

[0037] 103. In the training process of deep learning networks, the principle of maximizing information entropy is introduced to construct a regularization objective. By adjusting the network parameters, the probability distribution characteristics of the state encoding vector are optimized to complete the training of the network.

[0038] Specifically, a training sample set containing historical monitoring data of transformers and their corresponding state labels is constructed; based on the state encoding vector and multiple preset state categories, the similarity relationship between the state encoding vector of each training sample and each state category is calculated; according to the similarity relationship, the probability distribution of each training sample belonging to each state category is determined; the information entropy of the probability distribution is calculated as the entropy value characterizing the uncertainty of the sample's state; the parameters of the deep learning network are adjusted with the goal of optimizing the entropy value, so that the state encoding vector generated by the network after training has a distribution characteristic constrained by the information entropy theory. Furthermore, it also includes entropy target guidance based on state distribution characteristics: the samples in the training sample set are grouped based on state labels, and the distribution density of the state encoding vector of the samples in each group in space is calculated; according to the distribution density, a theoretical entropy target value reflecting its typical uncertainty is set for each state category; the information entropy of the probability distribution of each training sample is compared with the theoretical entropy target value of its state category to generate an entropy deviation; with minimizing the entropy deviation as one of the objectives, combined with the supervision signal generated by the similarity relationship, a comprehensive optimization objective is formed; the network parameters are adjusted according to the comprehensive optimization objective so that the uncertainty of the state encoding vector generated by the training process converges to a level that matches the inherent characteristics of each category.

[0039] It should be noted that the core of step 103 is to introduce information entropy as a regularization objective during network training to optimize the probability distribution characteristics of the state encoding vector. This will be illustrated below with specific data examples.

[0040] Currently, there is a training sample set containing 1000 sets of historical data, each set corresponding to a known state label, such as "normal state," "low-energy discharge," or "high-temperature overheating." The training data should cover as many known typical state patterns as possible.

[0041] A set of sample data (features of a transformer over 7 consecutive days) is input into the deep learning network to be trained (its structure is as described in step 102). The network finally outputs a 4-dimensional state encoding vector, S=[0.7,-0.2,0.5,0.3]. Three state categories ("normal", "low-energy discharge", and "high-temperature overheating") are preset, each with a prototype point in the vector space, C1=[1.0,0.0,0.8,0.2], C2=[-0.5,0.5,0.0,0.9], C3=[0.2,1.0,-0.3,0.4].

[0042] The cosine similarity between vector S and each category prototype point is calculated, assuming similarities of 0.85 (with normal), 0.30 (with low-energy discharge), and 0.45 (with high-temperature overheating). These three similarity values ​​are then transformed into a probability distribution using the Softmax function, resulting in a ternary probability vector P = [0.55, 0.20, 0.25]. This indicates that the network considers the sample to belong to "normal," "low-energy discharge," and "high-temperature overheating" with probabilities of 55%, 20%, and 25%, respectively.

[0043] Calculate the information entropy of this probability distribution P. For this distribution P = [0.55, 0.20, 0.25], the calculated information entropy value is approximately 1.05. This entropy value indicates that the network has some uncertainty in its judgment of the state of this sample.

[0044] A key objective of network training is to "guide the process with entropy objectives based on state distribution characteristics." Specifically: Based on the true state labels of the samples, all training samples are grouped. The spatial distribution density of the sample state encoding vectors within each group is calculated. Assuming that the encoding distribution of the "normal" group samples is highly compact and its "typical uncertainty" is low, a low theoretical entropy target value T is set for it. normal =0.3. The "low-energy discharge" group samples may exhibit diverse characteristics and be widely distributed; therefore, a higher theoretical entropy target T is set. discharge =1.2. The "High Temperature Overheating" group is set to T. overheat =0.8. For potential unknown states that have not appeared in the training data, a higher general entropy target (1.5) can be preset to enhance the model's generalization ability.

[0045] For the sample with the true label "normal" above, its calculated entropy value of 1.05 is much higher than the theoretical target of 0.3 for the "normal" group, resulting in an entropy bias of 0.75. The overall optimization goal of network training is to minimize this entropy bias while reducing the distance between the sample encoding and its true class prototype (the traditional supervised learning objective).

[0046] Using the backpropagation algorithm, all parameters in the network are adjusted based on this comprehensive optimization objective that combines classification accuracy and entropy bias. After iterative training with a large number of samples, the network gradually learns to generate state encoding vectors such that for a "normal" sample, its encoding not only approaches the prototype point of the "normal" class, but its corresponding probability distribution also becomes very deterministic ([0.9, 0.05, 0.05], with very low entropy), thus matching the inherent characteristic of low uncertainty in the "normal" state.

[0047] The uncertainty (entropy) of the probability distribution of the state encoding vector generated by the trained network will be effectively constrained, and it can more reasonably reflect the inherent characteristics of different operating state categories.

[0048] 104. Use the trained deep learning network to process historical monitoring data of known states, obtain the corresponding historical state encoding vector set, and spatially divide the vector set to form state clustering regions corresponding to various historical operating states.

[0049] Specifically, historical monitoring data of known states are input into the trained deep learning network to obtain corresponding historical state encoding vectors, forming a historical state encoding vector set. Based on the state labels of the historical monitoring data, the historical state encoding vector set is grouped to obtain vector group sets corresponding to different historical operating states. For each vector group set, the spatial distribution center of its internal vectors is calculated to generate state cluster center points representing that historical operating state. With each state cluster center point as the core and combined with the distribution range of its corresponding vector group set, a corresponding state clustering region is delineated in the state encoding vector space. The spatial location information of each state clustering region and its corresponding state cluster center point is stored to form a state clustering region for subsequent state matching. Furthermore, the method includes credibility assessment and optimization of state clustering regions: For each state clustering region, the geometric distribution characteristics of its internal historical state encoding vectors are calculated to obtain a geometric characteristic description of the region's compactness and consistency; based on the geometric characteristic description and the consistency of the state labels in the corresponding historical monitoring data, the credibility of each state clustering region is assessed, generating a region credibility assessment result; based on the region credibility assessment result, historical state encoding vectors in state clustering regions below a preset credibility threshold are filtered out, removing vectors that deviate too far from the state cluster center point to form a high-credibility vector subset; the state cluster center point is recalculated using the high-credibility vector subset to generate optimized state cluster center points; based on the distribution range of the optimized state cluster center points and their corresponding high-credibility vector subsets, the corresponding state clustering regions are redefined to form optimized state clustering regions for subsequent state matching.

[0050] It should be noted that we have 1000 sets of labeled historical monitoring data, with state labels including "normal," "low-energy discharge," and "high-temperature overheating." These 1000 sets of data are fed one by one into a pre-trained deep learning network. The network outputs a 4-dimensional state encoding vector for each set of data. For example, a "normal" sample outputs [0.85, 0.12, 0.78, 0.18], and a "low-energy discharge" sample outputs [-0.48, 0.52, 0.05, 0.88]. All vectors constitute the historical state encoding vector set.

[0051] After grouping, the "normal" state group has 600 vectors, the "low-energy discharge" group has 300 vectors, and the "high-temperature overheating" group has 100 vectors.

[0052] Calculate the state cluster centroids for each group. Taking the "Normal" group as an example, calculate the average of its 600 vectors across all dimensions. Assume the calculated centroid coordinates are [0.90, 0.10, 0.80, 0.20]. Similarly, the centroids for the "Low Energy Discharge" group are approximately [-0.50, 0.50, 0.00, 0.90], and the centroids for the "High Temperature Overheating" group are approximately [0.20, 0.95, -0.30, 0.40].

[0053] The state clustering region is defined with each center point as the core. The definition method is typically based on the distribution range of the vectors in that group. The Euclidean distances from the 600 vectors of the "normal" group to its center point [0.90, 0.10, 0.80, 0.20] are calculated, yielding an average distance of 0.2 and a maximum distance of 0.5. We can define the "normal" state clustering region as a spherical space centered at this center point with a radius of 0.3 (a value slightly larger than the average distance). Similarly, regions are defined for the other two states.

[0054] To further ensure quality, the clustered regions underwent credibility assessment and optimization. An examination of the "low-energy discharge" region revealed 20 vectors that deviated excessively from the center point [-0.50, 0.50, 0.00, 0.90] (distance greater than 0.6), and the original data labels of 5 of these vectors had questionable records. Therefore, the initial credibility of this region was assessed as low. Optimization was performed: these 20 excessively deviated vectors were removed, and the center point was recalculated using the remaining 280 high-credibility vectors, resulting in an optimized center point [-0.52, 0.51, 0.02, 0.89]. Based on the distribution of these 280 vectors, the average distance (0.18) and maximum distance (0.4) were recalculated, and the region radius was reset to 0.25, thus forming a more compact and reliable "low-energy discharge" state clustered region.

[0055] The spatial information of these three optimized state clustering regions (defined by their center point coordinates and region radius) is stored to form a baseline map for subsequent real-time state matching.

[0056] 105. Input the real-time monitoring data of the transformer into the trained deep learning network to obtain the real-time state encoding vector, calculate the distance between the vector and each state clustering region, determine the current operating state based on the minimum distance, and output a continuously quantified health score.

[0057] Specifically, real-time monitoring data is input into a trained deep learning network to obtain the corresponding real-time state encoding vector; the spatial distance between the real-time state encoding vector and the center point of each state cluster is calculated to form a distance set; the minimum distance value is selected from the distance set, and the state cluster region corresponding to the minimum distance value is determined as the current operating state; based on the minimum distance value, a continuous health score reflecting the health level of the transformer is calculated through a preset mapping relationship; and the final state evaluation result containing the current operating state and the continuous health score is output.

[0058] It should be noted that, taking the real-time state assessment of a 110kV transformer as an example, it is assumed that a deep learning network has been trained based on historical data, and three state clustering regions have been constructed, with their center points and radii as shown in Table 1 below: Table 1 Real-time monitoring data of the transformer, including oil chromatography, load, and temperature, are collected over a given period. After preprocessing, feature construction, and screening / reconstruction consistent with the training phase, the data is input into the trained deep learning network. The network outputs a 4-dimensional real-time state encoding vector: V real =[0.85,0.15,0.75,0.25].

[0059] Calculate V real Euclidean distances between the center points in the table above: Distance to the center point of "normal state": ; Distance to the center point of "low-energy discharge": ; Distance to the center point of "high temperature overheating": ; Comparing the three distances, the smallest distance is 0.10, corresponding to the "normal state" cluster region, and this distance of 0.10 is less than the radius of the "normal state" region of 0.25. Therefore, the current operating state of the transformer is determined to be "normal".

[0060] The health score is calculated using a preset mapping relationship. The scoring rule is set as a piecewise function: when the distance from the real-time vector to the center point of the matched state is ≤ the radius of the state region, the health score = 100 - (distance / radius) × 20; when the distance > the radius, the health score = 80 - (distance - radius) / radius × 40. In this example, the distance 0.10 ≤ radius 0.25, therefore: Health score = 100 - (0.10 / 0.25) × 20 = 100 - 8 = 92.

[0061] The output is shown in Table 2 below: Table 2 The results indicate that the transformer is currently operating normally and in good health, consistent with the actual situation where the characteristic gas content in the oil chromatography data showed no significant abnormalities and the load and temperature remained stable. The entire process achieved automated evaluation from real-time data to status determination and quantitative scoring.

[0062] 106. It also includes the identification and knowledge expansion of novel abnormal states: when the minimum distance between the real-time state coding vector and all state clustering regions exceeds a preset threshold, the current transformer is determined to be in a novel abnormal state, and a novel abnormal state marker is generated; the real-time monitoring data carrying the novel abnormal state marker and its corresponding real-time state coding vector are stored to form a novel abnormal state dataset; based on the novel abnormal state dataset, combined with expert diagnostic analysis, an expert analysis report containing fault causes and handling suggestions is generated; according to the expert analysis report, corresponding state clustering regions are added to the state coding vector space to form an updated state clustering pattern; in subsequent state assessments, the updated state clustering pattern is used for state matching and health score calculation.

[0063] Furthermore, real-time monitoring data from the novel abnormal state dataset is submitted to an expert system for diagnosis to obtain expert diagnostic conclusions containing fault type and severity level. Based on the expert diagnostic conclusions, a structured state description vector is generated, which includes fault type code, severity level value, and suggested handling priority. The state description vector is associated and bound with the corresponding real-time state code vector in the novel abnormal state dataset to form state vector pairs with expert labels. All state vector pairs with expert labels are collected to construct an expert knowledge set for expanding the system knowledge base. Based on the expert knowledge set, an expert analysis report containing fault causes and handling suggestions is generated.

[0064] It should be noted that in the evaluation scenario of step 105, if the distance between the real-time state encoding vector of a transformer and all known state clustering regions is significantly larger, the novel abnormal state identification and knowledge expansion mechanism in step 106 will be triggered. This will be explained below with specific data.

[0065] The real-time monitoring data of a certain transformer, after being processed by the network, yields a state coding vector of V. new =[0.60,0.70,-0.50,0.10]. Calculate its Euclidean distance to the center points of the three existing state cluster regions. The results are shown in Table 3 below (after correction): Table 3 The preset distance threshold is 0.5 (usually set based on the statistical characteristics of historical clustering regions). Real-time vector V new The minimum distance to all regions is 0.59. Since the minimum distance of 0.59 is greater than the threshold of 0.5, the system determines that the transformer is in a "new type of abnormal state" and generates a flag.

[0066] Then store the real-time monitoring data and V. new Vector. The example real-time data characteristics are: acetylene (C2H2) content increased to 15 μL / L (significantly higher than normal), hydrogen (H2) reached 200 μL / L, but the carbon monoxide to carbon dioxide ratio was also abnormal, and load and temperature data also fluctuated. These data are related to V new The vectors together form a new abnormal state record, which is stored in the dataset to be analyzed.

[0067] This record was subsequently submitted to an expert system or diagnosed by a professional engineer. Expert analysis concluded that the combination of gas characteristics (significantly increased acetylene and hydrogen, accompanied by specific changes in carbon oxides) indicated a complex fault of "medium-temperature overheating with partial discharge," rated as "moderate" severity and "high" priority. Based on this, a structured state description vector was generated: [Fault Type Code: 205, Severity Level: 6, Priority: 8].

[0068] This state description vector and V new The system associates and binds these anomalies, forming state vector pairs with expert labels, and stores them in a structured expert knowledge set. When a certain number of these novel anomalies accumulate (5 cases), the system expands its knowledge: in the state coding vector space, based on the state coding vectors corresponding to these novel anomaly data, its distribution center is calculated [0.58, 0.72, -0.48, 0.12], and an initial region radius (0.15) is set, thus adding a new "intermediate-temperature overheating with discharge" state clustering region. This region information is updated in the system's state clustering pattern library. Subsequently, when similar anomalies reappear, their real-time vectors will be more accurately identified and matched due to their proximity to this new region, and diagnosis will be performed by referring to the bound expert knowledge, achieving self-evolution of the evaluation system.

[0069] The present invention also provides a transformer condition assessment device based on an improved deep learning algorithm. The transformer condition assessment device based on the improved deep learning algorithm includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the transformer condition assessment method based on the improved deep learning algorithm in the above embodiments.

[0070] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the transformer state assessment method based on the improved deep learning algorithm.

[0071] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0072] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A transformer state assessment method based on an improved deep learning algorithm, characterized in that, include: Collect multi-source monitoring time-series data during transformer operation, calculate and generate characteristic data; The feature data is input into a deep learning network for fusion and dimensionality reduction, and the output is a state encoding vector. During the training process of the deep learning network, the probability distribution characteristics of the state encoding vector are optimized by adjusting the network parameters to complete the training of the network. The trained deep learning network is used to process historical monitoring data of known states to obtain the corresponding historical state encoding vector set, and spatial division is performed to form state clustering regions corresponding to various historical operating states. The real-time monitoring data of the transformer is input into the trained deep learning network to obtain a real-time state encoding vector. The distance between the vector and each state clustering region is calculated. The current operating state is determined based on the minimum distance, and a continuously quantified health score is output.

2. The transformer state assessment method based on an improved deep learning algorithm according to claim 1, characterized in that, include: Construct a training sample set containing historical monitoring data of transformers and their corresponding status labels; Based on the state encoding vector and multiple preset state categories, the similarity relationship between the state encoding vector of each training sample and each state category is calculated. Based on the similarity relationship, determine the probability distribution of each training sample belonging to each state category; Calculate the information entropy of the probability distribution, which serves as the entropy value characterizing the uncertainty of the sample state; The parameters of the deep learning network are adjusted with the goal of optimizing the entropy value, so that the state encoding vector generated by the network after training has a distribution characteristic constrained by the information entropy theory.

3. The transformer state assessment method based on an improved deep learning algorithm according to claim 2, characterized in that, Based on the state labels, the samples in the training sample set are grouped, and the distribution density of the state encoding vector of the samples in each group in space is calculated. Based on the distribution density, a theoretical entropy target value reflecting the typical degree of uncertainty is set for each state category; The information entropy of the probability distribution of each training sample is compared with the theoretical entropy target value of its state category to generate an entropy deviation. With minimizing the entropy deviation as one of the objectives, and combining the supervision signal generated by the similarity relationship, a comprehensive optimization objective is formed; The network parameters are adjusted according to the comprehensive optimization objective, so that the uncertainty of the state encoding vector generated during the training process converges to a level that matches the inherent characteristics of each category.

4. The transformer state assessment method based on an improved deep learning algorithm according to claim 1, characterized in that, include: The historical monitoring data of the known state is input into the trained deep learning network to obtain the corresponding historical state encoding vector, thus forming a historical state encoding vector set; Based on the status labels of the historical monitoring data, the historical status encoded vector set is grouped to obtain a vector group set corresponding to different historical operating states. For each vector group set, calculate the spatial distribution center of its internal vectors to generate a state cluster center point representing the historical running state. Taking each state cluster center as the core, and combining the distribution range of its corresponding vector group set, the corresponding state clustering region is delineated in the state coding vector space; The spatial location information of each state clustering region and its corresponding state clustering center point is stored to form a state clustering region.

5. The transformer state assessment method based on an improved deep learning algorithm according to claim 4, characterized in that, For each state clustering region, the geometric distribution characteristics of its internal historical state encoding vectors are calculated to obtain a geometric characteristic description of the compactness and consistency of the region. Based on the geometric characteristic description and the consistency of the state labels of the corresponding historical monitoring data, the credibility of each state clustering region is evaluated, and a region credibility evaluation result is generated. Based on the regional credibility assessment results, historical state encoding vectors within state clustering regions that are below a preset credibility threshold are filtered out, and vectors that deviate too far from the state cluster center point are removed to form a high-credibility vector subset. The state cluster centers are recalculated using the high-confidence vector subset to generate optimized state cluster centers; Based on the distribution range of the optimized state cluster centers and their corresponding high-confidence vector subsets, the corresponding state clustering regions are redefined to form optimized state clustering regions.

6. The transformer state assessment method based on an improved deep learning algorithm according to claim 1, characterized in that, include: The real-time monitoring data is input into the trained deep learning network to obtain the corresponding real-time state encoding vector; Calculate the spatial distance between the real-time state encoding vector and each of the state cluster centers to form a distance set; The minimum distance value is selected from the distance set, and the state clustering region corresponding to the minimum distance value is determined as the current running state; Based on the minimum distance value, a continuous health score reflecting the health status of the transformer is calculated through a preset mapping relationship; The output includes the current operating status and the continuous health score, resulting in a final status assessment.

7. The transformer state assessment method based on an improved deep learning algorithm according to claim 1, characterized in that, It also includes the identification and knowledge expansion of new abnormal states: When the minimum distance between the real-time state encoding vector and all the state clustering regions exceeds a preset threshold, it is determined that the current transformer is in a new abnormal state, and a new abnormal state label is generated. The real-time monitoring data carrying the novel abnormal state marker and its corresponding real-time state encoding vector are stored to form a novel abnormal state dataset. Based on the novel abnormal state dataset, an expert analysis report containing the causes of the fault and handling suggestions is generated by combining expert diagnostic analysis. Based on the expert analysis report, corresponding state clustering regions are added to the state encoding vector space to form an updated state clustering pattern.

8. The transformer state assessment method based on the improved deep learning algorithm according to claim 7, characterized in that, The real-time monitoring data in the new abnormal state dataset is submitted to the expert system for diagnosis to obtain expert diagnostic conclusions that include fault type and severity level. Based on the expert diagnosis, a status description vector is generated, which includes the fault type code, severity level value, and recommended processing priority. The state description vector is associated and bound with the corresponding real-time state encoding vector in the novel abnormal state dataset to form a state vector pair with expert labels. Collect all the state vector pairs with expert labels to construct an expert knowledge set; Based on the aforementioned expert knowledge set, an expert analysis report containing the causes of the fault and suggested solutions is generated.

Images