Transformer life prediction method and system based on machine learning and Internet of Things

Abstract

The invention provides a transformer life prediction method and system based on machine learning and the Internet of Things, and relates to the technical field of transformer life prediction, and the method comprises the steps: collecting and preprocessing the multi-modal operation data of a transformer, constructing a multi-modal data fusion model based on a graph neural network, and carrying out the prediction of the life of the transformer; and mapping the dimensionality-reduced feature data into a graph structure, inputting transformer multi-modal operation data collected in real time into an optimized fuzzy inference system to obtain a health state evaluation result and a final residual life prediction result of the transformer, and obtaining a prediction result of the residual life of the transformer on the basis of transformer maintenance suggestions and historical maintenance records. Constructing a transformer digital twinborn model for maintenance strategy optimization, inputting the health state evaluation result, the residual life prediction result and the transformer maintenance suggestion into the transformer digital twinborn model, performing fault simulation and predictive maintenance analysis, and based on the fault simulation and predictive maintenance analysis result, obtaining a maintenance strategy of the transformer. And generating an intelligent maintenance scheme in combination with the expert knowledge base.

Description

Technical Field

[0001] The present invention relates to transformer life prediction technology, and in particular to a transformer life prediction method and system based on machine learning and the Internet of Things. Background Art

[0002] As a key equipment in the power system, the reliability and life of the transformer directly affect the stability and economy of the power supply. With the development of smart grids and the widespread application of Internet of Things technology, higher requirements are placed on the health status assessment and remaining life prediction of transformers. Traditional transformer life prediction methods mainly rely on empirical models and regular inspections, and have problems such as incomplete data acquisition, limited processing capabilities, low accuracy of prediction models, and lack of real-time performance. At the same time, traditional methods also face challenges in intelligent maintenance decision-making, data security, historical data utilization, multi-dimensional analysis, and adaptability. Existing machine learning-based methods still have limitations such as insufficient data security, the effect of multimodal data fusion needs to be improved, and the interpretability of prediction models is poor.

[0003] In view of this, it is urgent to develop a new transformer life prediction method that can make full use of multimodal data, improve prediction accuracy and reliability, and achieve an organic combination with intelligent maintenance decision-making. This method should be able to effectively handle the long-term dependencies of time series data, effectively model uncertainties, and consider the dynamic changes of the transformer operating environment. By addressing these challenges, the new method will provide more powerful support for the full life cycle management of transformers and promote the development of power systems in a smarter and more reliable direction. Summary of the invention

[0004] The embodiments of the present invention provide a transformer life prediction method and system based on machine learning and the Internet of Things, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] A transformer life prediction method based on machine learning and the Internet of Things is provided, including:

[0007] Collect transformer multimodal operation data and preprocess it, encrypt and transmit and store the preprocessed multimodal operation data to form a historical data set, apply adaptive wavelet transform to the historical data set, extract multi-scale time-frequency features, and obtain time-frequency feature data. At the same time, use a non-negative matrix decomposition algorithm to reduce the dimension and select features of the time-frequency feature data to obtain feature data after dimension reduction, build a multimodal data fusion model based on a graph neural network, map the feature data after dimension reduction to a graph structure, obtain graph representation data, and apply a dynamic time warping algorithm to perform time series alignment on the graph representation data to obtain aligned graph representation data;

[0008] Using a quantum support vector machine to perform nonlinear classification on the aligned graph representation data to obtain a transformer health status classification result, based on the transformer health status classification result, constructing a fuzzy inference system to generate an initial prediction result of the transformer remaining life, using a multi-objective optimization algorithm to optimize the parameters of the fuzzy inference system to obtain an optimized fuzzy inference system, inputting the real-time collected transformer multimodal operation data into the optimized fuzzy inference system to obtain the transformer health status assessment result and the final remaining life prediction result, and generating transformer maintenance recommendations based on the health status assessment result and the final remaining life prediction result in combination with predefined maintenance rules;

[0009] Based on the transformer maintenance recommendations and historical maintenance records, a transformer digital twin model for maintenance strategy optimization is constructed, the health status assessment results, the remaining life prediction results and the transformer maintenance recommendations are input into the transformer digital twin model, fault simulation and predictive maintenance analysis are performed, and based on the fault simulation and predictive maintenance analysis results, combined with an expert knowledge base, an intelligent maintenance plan is generated.

[0010] In an optional embodiment,

[0011] A multimodal data fusion model based on a graph neural network is constructed, the feature data after dimension reduction is mapped to a graph structure to obtain graph representation data, and a dynamic time warping algorithm is applied to perform time series alignment on the graph representation data. The aligned graph representation data includes:

[0012] Collecting transformer multimodal operation data, the multimodal operation data including electrical parameter data, acoustic data, thermal imaging data and oil-gas data, performing dimension reduction preprocessing on the multimodal operation data, using a non-negative matrix decomposition algorithm to retain the non-negative characteristics of the data, and obtaining feature data after dimension reduction;

[0013] The feature data after dimensionality reduction is mapped to nodes in a graph structure, each node represents multimodal data at a time point, directed edges are added to adjacent time point nodes based on timestamps, edge weights are calculated using an exponential decay function, temporal relationships are constructed, an adaptive k-nearest neighbor algorithm is used to establish edge connections between different time points of the same modality, the k value is dynamically adjusted according to data density, intra-modal relationships are constructed, a multi-head attention mechanism is designed to calculate the correlation strength between different modalities as the weight of the inter-modal edges, and inter-modal relationships are constructed, and a multi-level heterogeneous graph structure is formed based on the constructed temporal relationships, intra-modal relationships, and inter-modal relationships;

[0014] Construct a graph convolution layer containing spatial and temporal convolutions, process multi-level heterogeneous graph structures, capture local spatial-temporal correlations, design an adaptive graph attention layer, process the output of the graph convolution layer, dynamically adjust the importance of different nodes and edges, introduce residual connections and layer normalization, and design a skip connection structure to fuse feature representations at different levels to obtain fused graph representation data;

[0015] Based on the principal component analysis method, the target operation cycle graph representation is selected as the reference sequence from the fused graph representation data. A dynamic time warping distance metric based on graph structure similarity is designed. Considering the node characteristics and topological structure, a dynamic time warping algorithm is adopted to calculate the alignment path between the fused graph representation data and the reference sequence. The fused graph representation data is aligned using an interpolation algorithm to obtain the aligned graph sequence.

[0016] In an optional embodiment,

[0017] The aligned graph representation data is nonlinearly classified using a quantum support vector machine to obtain a transformer health status classification result. Based on the transformer health status classification result, a fuzzy inference system is constructed to generate an initial prediction result of the transformer remaining life. The parameters of the fuzzy inference system are optimized using a multi-objective optimization algorithm to obtain an optimized fuzzy inference system. The real-time collected transformer multimodal operation data is input into the optimized fuzzy inference system to obtain the transformer health status assessment result and the final remaining life prediction result, including:

[0018] Mapping the aligned graph representation data to the quantum state space through a parameterized quantum circuit to obtain a quantum state representation, defining a quantum kernel function in the quantum state space, calculating the similarity between different quantum state representations, constructing a quantum support vector machine optimization problem, applying a quantum phase estimation algorithm to solve the optimization problem, obtaining an optimal Lagrangian multiplier, and constructing a quantum decision function based on the optimal Lagrangian multiplier;

[0019] Using the quantum decision function to classify the health status of the new graph representation data, obtaining the health status classification result, converting the health status classification result, transformer operation time, and load rate into a fuzzy set with fuzzy membership, and constructing a fuzzy rule base containing multiple fuzzy rules based on historical data and expert knowledge;

[0020] The fuzzy set is inferred by the Mamdani reasoning method, including calculating the rule triggering strength, truncating the rule conclusion and aggregating all rule outputs to obtain the fuzzy reasoning result, and the fuzzy reasoning result is defuzzified by the adaptive centroid method to obtain the preliminary prediction value of the remaining life of the transformer;

[0021] An improved α cutoff method is applied to calculate the confidence interval of the preliminary prediction value, and the health status classification result, the preliminary prediction value and its confidence interval are input into a pre-trained deep neural network. The deep neural network combines long short-term memory units and an attention mechanism to perform time series feature extraction and multi-scale fusion. The output of the deep neural network is combined with a multi-objective optimization algorithm to optimize the preliminary prediction value to obtain the final transformer health status assessment result and remaining life prediction value.

[0022] In an optional embodiment,

[0023] The health status classification result, the preliminary prediction value and its confidence interval are input into a pre-trained deep neural network. The deep neural network combines long short-term memory units and attention mechanisms to perform time series feature extraction and multi-scale fusion. The output of the deep neural network is combined with a multi-objective optimization algorithm to optimize the preliminary prediction value, and the final transformer health status assessment result and remaining life prediction value are obtained, including:

[0024] Receive the transformer health status classification results, preliminary prediction values ​​and confidence intervals, preprocess the received data, apply principal component analysis to the preprocessed data to reduce the dimension, obtain the reduced feature data, and input it into a pre-trained deep neural network, which contains two parallel processing branches: time series feature flow and state feature flow;

[0025] In the time series feature stream, time series data is processed through a multi-layer stacked bidirectional long short-term memory network, and the adaptive time window technology is applied to optimize the time series features by dynamically adjusting the input sequence length. At the same time, in the state feature stream, a multi-scale residual convolutional network is used to process the state data to obtain the optimized state features.

[0026] The optimized time series features and state features are input into the feature pyramid network to generate a comprehensive feature representation, which is then input into the multi-task learning module to achieve knowledge transfer between tasks through the task correlation matrix, and to simultaneously predict the health status and remaining lifespan, thus obtaining a preliminary health status assessment result and a remaining lifespan prediction value.

[0027] Construct a multi-objective optimization problem including prediction accuracy, model complexity and prediction uncertainty, solve the multi-objective optimization problem by a multi-objective optimization algorithm, obtain a set of optimal solutions, fine-tune the deep neural network parameters based on the obtained optimal solutions, and obtain multiple deep neural network models with different parameter configurations;

[0028] Each deep neural network model is used to predict the health status and remaining life of the transformer respectively, and multiple groups of prediction results are obtained. The multiple groups of prediction results are fused using an integration strategy to obtain the final transformer health status assessment results and remaining life prediction results.

[0029] In an optional embodiment,

[0030] The health status assessment result, the remaining life prediction result and the transformer maintenance suggestion are input into the transformer digital twin model to perform fault simulation and predictive maintenance analysis. Based on the fault simulation and predictive maintenance analysis results and combined with the expert knowledge base, an intelligent maintenance plan is generated, including:

[0031] Construct a transformer digital twin model, which includes a physical layer, a data layer, an analysis layer, and a decision layer. Collect real-time monitoring data, historical operation data, environmental data, and manual inspection data of the transformer, align and fuse the collected data through tensor decomposition and transfer learning methods to obtain fused data, input the fused data into the data layer of the digital twin model, and use a recurrent neural network to process the time series data to obtain time series features;

[0032] In the analysis layer, the convolutional neural network and the graph neural network are integrated to extract the spatial-temporal features of the time series features, and the fuzzy reasoning system and evidence theory are combined to output the health status scores and credibility of each component of the transformer. According to the health status scores and credibility, an adaptive Monte Carlo simulation method is used in the digital twin model to perform fault simulation and evolution path prediction to obtain fault simulation results and evolution path prediction results;

[0033] Based on the fault simulation results and the evolution path prediction results, a dynamic Bayesian network model is constructed to calculate the failure probability and mutual influence relationship of each component of the transformer and generate a risk assessment matrix;

[0034] Using the risk assessment matrix, combined with a preset maintenance rule base, a heuristic algorithm is used to generate multiple candidate maintenance plans, and the candidate maintenance plans are sorted according to a cost-benefit analysis to form an optimized maintenance plan set;

[0035] A multimodal knowledge graph is constructed, and maintenance knowledge related to the health status score is extracted through graph embedding and relational reasoning. An intelligent maintenance plan is generated by combining the optimized maintenance plan set and the maintenance knowledge.

[0036] In an optional embodiment,

[0037] Constructing a multimodal knowledge graph, extracting maintenance knowledge related to the health status score through graph embedding and relational reasoning, combining the optimized maintenance plan set with the maintenance knowledge, and generating an intelligent maintenance plan including:

[0038] The extracted named entities, relations, equipment parts, defects, timing patterns, and abnormal features are integrated into the attribute graph model to form a multimodal knowledge graph containing equipment entities, fault types, symptom characteristics, maintenance operations, expert rules, and historical cases;

[0039] Introducing an uncertainty representation method based on evidence theory, assigning basic probabilities to the relationships in the multimodal knowledge graph, adopting an improved transfer distance model to embed the multimodal knowledge graph, constructing an objective function of the transfer distance model based on the graph embedding result, and obtaining a low-dimensional vector representation of entities and relationships in the multimodal knowledge graph by optimizing the objective function;

[0040] A subgraph related to the health status score is extracted from the multimodal knowledge graph using a bidirectional attention network, and the extracted subgraph is encoded using a graph attention network, a reasoning path is generated in the encoded subgraph using a policy gradient reinforcement learning algorithm, and the generated multiple reasoning paths are weightedly fused to obtain maintenance knowledge related to the health status score;

[0041] The maintenance plan generation problem is modeled as a multi-objective Markov decision process. Based on the obtained maintenance knowledge, the multi-objective Markov decision process problem is solved, the strategy network and multiple state-action value functions are learned, and based on the learned strategy, candidate maintenance plans are generated. The generated candidate maintenance plans are screened and sorted, and finally an intelligent maintenance plan is obtained.

[0042] In an optional embodiment,

[0043] The calculation formula for constructing the objective function of the transfer distance model based on the graph embedding results is as follows:

[0044]

[0045] Among them, J represents the objective function, E represents the edge set of the graph, H represents the low-dimensional representation matrix of the node, and α ij represents the attention weight of the relationship between nodes, h i represents the low-dimensional representation of node vi, h j Represents node v j The low-dimensional representation of λ1 represents the weight coefficient of the transfer distance, D ij (·) represents the transfer distance function, λ2 represents the weight coefficient of the regularization term, and A i Represents node v i The feature matrix of .

[0046] According to a second aspect of the embodiments of the present invention,

[0047] A transformer life prediction system based on machine learning and the Internet of Things is provided, including:

[0048] The first unit is used to collect and preprocess the multimodal operation data of the transformer, encrypt, transmit and store the preprocessed multimodal operation data to form a historical data set, apply adaptive wavelet transform to the historical data set, extract multi-scale time-frequency features, and obtain time-frequency feature data, and use a non-negative matrix decomposition algorithm to reduce the dimension and select features of the time-frequency feature data to obtain feature data after dimension reduction, build a multimodal data fusion model based on a graph neural network, map the feature data after dimension reduction to a graph structure to obtain graph representation data, and apply a dynamic time warping algorithm to perform time series alignment on the graph representation data to obtain aligned graph representation data;

[0049] The second unit is used to use a quantum support vector machine to perform nonlinear classification on the aligned graph representation data to obtain a transformer health status classification result, build a fuzzy reasoning system based on the transformer health status classification result, generate an initial prediction result of the transformer remaining life, optimize the parameters of the fuzzy reasoning system using a multi-objective optimization algorithm to obtain an optimized fuzzy reasoning system, input the real-time collected transformer multimodal operation data into the optimized fuzzy reasoning system to obtain the transformer health status assessment result and the final remaining life prediction result, and generate a transformer maintenance recommendation based on the health status assessment result and the final remaining life prediction result in combination with predefined maintenance rules;

[0050] The third unit is used to build a transformer digital twin model for maintenance strategy optimization based on the transformer maintenance suggestions and historical maintenance records, input the health status assessment results, the remaining life prediction results and the transformer maintenance suggestions into the transformer digital twin model, perform fault simulation and predictive maintenance analysis, and generate an intelligent maintenance plan based on the fault simulation and predictive maintenance analysis results in combination with an expert knowledge base.

[0051] According to a third aspect of the embodiments of the present invention,

[0052] An electronic device is provided, comprising:

[0053] processor;

[0054] a memory for storing processor-executable instructions;

[0055] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0056] A fourth aspect of the embodiments of the present invention is:

[0057] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0058] In this embodiment, data security can be ensured by collecting and preprocessing the multimodal operation data of the transformer, and then encrypting and transmitting and storing the historical data set after the data is formed. The multi-scale time-frequency features extracted by the adaptive wavelet transform are combined with non-negative matrix decomposition to effectively reduce the dimension and select key features, build a multimodal data fusion model based on a graph neural network, map the features to the graph structure and perform time series alignment, thereby enhancing the availability of data. The aligned graph representation data is nonlinearly classified using a quantum support vector machine. After obtaining the transformer health status classification result, a fuzzy reasoning system is constructed to generate an initial remaining life prediction, and then the parameters of the system are optimized through a multi-objective optimization algorithm to more accurately evaluate the health status and remaining life of the transformer. Combined with maintenance recommendations and historical records, the constructed digital twin model can perform fault simulation and predictive maintenance analysis, and finally generate an intelligent maintenance plan based on an expert knowledge base to improve the maintenance efficiency and reliability of the transformer. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic diagram of a flow chart of a transformer life prediction method based on machine learning and the Internet of Things according to an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the structure of a transformer life prediction system based on machine learning and the Internet of Things according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1 FIG. 1 is a flow chart of a transformer life prediction method based on machine learning and the Internet of Things according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] S101. Collect transformer multimodal operation data and preprocess it, encrypt and transmit and store the preprocessed multimodal operation data to form a historical data set, apply adaptive wavelet transform to the historical data set, extract multi-scale time-frequency features, obtain time-frequency feature data, and use non-negative matrix decomposition algorithm to reduce dimension and select features of the time-frequency feature data to obtain feature data after dimension reduction, build a multimodal data fusion model based on graph neural network, map the feature data after dimension reduction to the graph structure, obtain graph representation data, apply dynamic time warping algorithm to time series alignment of the graph representation data, and obtain aligned graph representation data;

[0065] First, the transformer's multimodal operating data is collected, including parameters such as voltage, current, temperature, vibration, and dissolved gas in oil. This data is collected in real time using a sensor network, and is preprocessed by edge computing devices, such as denoising and outlier detection. The preprocessed data is encrypted using the AES encryption algorithm, and then transmitted to the cloud server for storage via the MQTT protocol to form a historical data set.

[0066] Adaptive wavelet transform is applied to the stored historical data set for multi-scale time-frequency analysis. Specifically, Morlet wavelet is used as the mother wavelet to decompose different frequency components and extract time-frequency features of multiple scales. For example, the voltage data is decomposed by 5 layers of wavelet to obtain detail coefficients of 5 scales and 1 approximate coefficient. These coefficients constitute the time-frequency feature data.

[0067] Next, the non-negative matrix factorization (NMF) algorithm is used to reduce the dimension and select features of the time-frequency feature data. NMF decomposes the original high-dimensional matrix into the product of two low-dimensional non-negative matrices, one of which represents the basis vectors and the other represents the coefficients. By setting an appropriate dimensionality reduction ratio (such as 50%), the reduced-dimensional feature data can be obtained, which effectively reduces the computational complexity of subsequent processing.

[0068] Then, a multimodal data fusion model based on graph neural network is constructed. The feature data after dimensionality reduction is mapped to a graph structure, where nodes represent different sensors and edges represent the correlation between sensors. The graph structure data is processed using a graph convolutional network (GCN), the topological relationship between nodes is learned, and the graph representation data is obtained.

[0069] The dynamic time warping (DTW) algorithm is applied to the graph representation data for time series alignment. DTW solves the time offset problem between different sensor data by calculating the optimal alignment path between two time series through dynamic programming. For example, the voltage data is used as the reference sequence to align other sensor data to obtain the aligned graph representation data.

[0070] In an optional embodiment,

[0071] A multimodal data fusion model based on a graph neural network is constructed, the feature data after dimension reduction is mapped to a graph structure to obtain graph representation data, and a dynamic time warping algorithm is applied to perform time series alignment on the graph representation data. The aligned graph representation data includes:

[0072] Collecting transformer multimodal operation data, the multimodal operation data including electrical parameter data, acoustic data, thermal imaging data and oil-gas data, performing dimension reduction preprocessing on the multimodal operation data, using a non-negative matrix decomposition algorithm to retain the non-negative characteristics of the data, and obtaining feature data after dimension reduction;

[0073] The feature data after dimensionality reduction is mapped to nodes in a graph structure, each node represents multimodal data at a time point, directed edges are added to adjacent time point nodes based on timestamps, edge weights are calculated using an exponential decay function, temporal relationships are constructed, an adaptive k-nearest neighbor algorithm is used to establish edge connections between different time points of the same modality, the k value is dynamically adjusted according to data density, intra-modal relationships are constructed, a multi-head attention mechanism is designed to calculate the correlation strength between different modalities as the weight of the inter-modal edges, and inter-modal relationships are constructed, and a multi-level heterogeneous graph structure is formed based on the constructed temporal relationships, intra-modal relationships, and inter-modal relationships;

[0074] Construct a graph convolution layer containing spatial and temporal convolutions, process multi-level heterogeneous graph structures, capture local spatial-temporal correlations, design an adaptive graph attention layer, process the output of the graph convolution layer, dynamically adjust the importance of different nodes and edges, introduce residual connections and layer normalization, and design a skip connection structure to fuse feature representations at different levels to obtain fused graph representation data;

[0075] Based on the principal component analysis method, the target operation cycle graph representation is selected as the reference sequence from the fused graph representation data. A dynamic time warping distance metric based on graph structure similarity is designed. Considering the node characteristics and topological structure, a dynamic time warping algorithm is adopted to calculate the alignment path between the fused graph representation data and the reference sequence. The fused graph representation data is aligned using an interpolation algorithm to obtain the aligned graph sequence.

[0076] This embodiment provides a transformer multimodal data fusion method based on graph neural network. The method first collects the transformer multimodal operation data, including electrical parameter data, acoustic data, thermal imaging data and oil gas data. The collected multimodal data is preprocessed by dimensionality reduction, and the non-negative matrix decomposition algorithm is used to retain the non-negative characteristics of the data to obtain the feature data after dimensionality reduction.

[0077] The feature data after dimensionality reduction is mapped to the graph structure to construct a multi-level heterogeneous graph. Specifically, the multimodal data at each time point is represented as a node in the graph, and directed edges are added to the nodes of adjacent time points based on the timestamp. The edge weights are calculated using the exponential decay function to construct the temporal relationship. The adaptive k-nearest neighbor algorithm is used to establish edge connections between different time points of the same modality, and the k value is dynamically adjusted according to the data density to construct the intra-modal relationship. A multi-head attention mechanism is designed to calculate the correlation strength between different modalities as the weight of the inter-modal edge to construct the inter-modal relationship. Based on the constructed temporal relationship, intra-modal relationship and inter-modal relationship, a multi-level heterogeneous graph structure is formed.

[0078] Next, we construct a graph convolution layer containing spatial and temporal convolutions to process multi-level heterogeneous graph structures and capture local spatial-temporal correlations. We design an adaptive graph attention layer to process the output of the graph convolution layer and dynamically adjust the importance of different nodes and edges. We introduce residual connections and layer normalization, and design a skip connection structure to fuse feature representations at different levels to obtain fused graph representation data.

[0079] Then, the target operation cycle graph representation is selected as the reference sequence from the fused graph representation data based on the principal component analysis method. A dynamic time warping distance metric based on graph structure similarity is designed, taking into account node features and topological structure. The dynamic time warping algorithm is used to calculate the alignment path between the fused graph representation data and the reference sequence, and the interpolation algorithm is used to align the fused graph representation data to obtain the aligned graph sequence.

[0080] The specific implementation process is as follows:

[0081] First, the multimodal operation data of the transformer is collected. The electrical parameter data includes voltage, current, active power, reactive power, etc., with a sampling frequency of 1Hz. The acoustic data uses a microphone array to collect the transformer operation noise, with a sampling frequency of 44.1kHz. The thermal imaging data uses an infrared thermal imager to collect the surface temperature distribution image of the transformer, with a resolution of 640x480 and a sampling frequency of 0.1Hz. The gas data in the oil includes the dissolved gas content of hydrogen, methane, ethane, ethylene, acetylene, etc., with a sampling frequency of 0.01Hz.

[0082] The collected multimodal data was preprocessed by dimensionality reduction. For the electrical parameter data and gas in oil data, the non-negative matrix factorization algorithm was used to reduce the dimension of the original 100-dimensional features to 20 dimensions. For the acoustic data, the Mel-frequency cepstral coefficient (MFCC) features were extracted and reduced to 20 dimensions using non-negative matrix factorization. For the thermal imaging data, the statistical features and texture features of the thermal image were extracted and reduced to 20 dimensions using non-negative matrix factorization. Finally, 80-dimensional feature data after dimensionality reduction was obtained.

[0083] Map the reduced feature data into the graph structure. Use the adaptive k-nearest neighbor algorithm to establish edge connections between different time points of the same modality. The initial k value is set to 5, and the k value is dynamically adjusted according to the local data density. Calculate the Pearson correlation coefficient between different modal data, design a 4-head attention mechanism, use the correlation coefficient as the attention score, and calculate the weight of the edge between modalities.

[0084] A neural network model with three layers of graph convolution is constructed. Each layer of graph convolution contains spatial convolution and temporal convolution. The spatial convolution is approximated by Chebyshev polynomials, and the temporal convolution is approximated by one-dimensional convolution. An adaptive graph attention layer is added after each layer of graph convolution to dynamically adjust the importance of nodes and edges. Residual connections and layer normalization are introduced, and a jump connection structure is designed. After three layers of graph convolution processing, the fused 64-dimensional graph representation data is obtained.

[0085] Based on the principal component analysis method, the 24-hour operation cycle graph corresponding to the principal component with the largest variance contribution rate is selected as the reference sequence. A graph structure similarity metric based on the Euclidean distance of node features and the Jaccard similarity of topological structure is designed. The dynamic time warping algorithm is used to calculate the alignment path between the fused graph representation data and the reference sequence. The fused graph representation data is aligned using the cubic spline interpolation algorithm to obtain the aligned graph sequence.

[0086] Through the above steps, the graph representation and time series alignment of transformer multimodal data are realized, which provides a basis for subsequent state assessment and fault diagnosis. This method makes full use of the graph structure to express the time series relationship, intra-modal relationship and inter-modal relationship of multimodal data, effectively integrates multi-source heterogeneous data, and improves data utilization efficiency.

[0087] S102. Use a quantum support vector machine to perform nonlinear classification on the aligned graph representation data to obtain a transformer health status classification result; based on the transformer health status classification result, construct a fuzzy inference system to generate an initial prediction result of the transformer remaining life; use a multi-objective optimization algorithm to optimize the parameters of the fuzzy inference system to obtain an optimized fuzzy inference system; input the real-time collected transformer multimodal operation data into the optimized fuzzy inference system to obtain the transformer health status assessment result and the final remaining life prediction result; generate transformer maintenance recommendations based on the health status assessment result and the final remaining life prediction result in combination with predefined maintenance rules;

[0088] The aligned graph representation data is nonlinearly classified using a quantum support vector machine (QSVM). QSVM uses the parallelism of quantum computing to process high-dimensional data faster. The data is mapped to a high-dimensional feature space through a kernel function (such as the RBF kernel), and then the optimal separation hyperplane is found to obtain the classification results of the transformer health status, such as "normal", "mildly degraded", "moderately degraded", "severely degraded", etc.

[0089] Based on the classification results of the transformer health status, a fuzzy inference system is constructed. Input variables (such as load rate, temperature, etc.) and output variables (remaining life) are defined, and a fuzzy rule base is designed. For example, "if the load rate is high and the temperature is high, the remaining life is short." Fuzzy inference is performed using the Mamdani reasoning method to obtain the initial prediction results of the transformer's remaining life.

[0090] The parameters of the fuzzy inference system are optimized using a multi-objective optimization algorithm (such as NSGA-II). The optimization objectives include prediction accuracy and model complexity. By adjusting the membership function parameters and rule weights, an optimal solution that balances performance and complexity is obtained to form an optimized fuzzy inference system.

[0091] The real-time multi-modal operation data of the transformer is input into the optimized fuzzy inference system to obtain the health status assessment results and the final remaining life prediction results of the transformer. For example, the health status may be "moderately deteriorated" and the remaining life prediction is 5.3 years.

[0092] Based on the health status assessment results and the remaining life prediction results, combined with predefined maintenance rules, transformer maintenance recommendations are generated. The maintenance rules can be decision trees based on expert experience, such as "if the health status is moderately degraded and the remaining life is less than 6 years, a comprehensive overhaul is recommended."

[0093] In an optional embodiment,

[0094] The aligned graph representation data is nonlinearly classified using a quantum support vector machine to obtain a transformer health status classification result. Based on the transformer health status classification result, a fuzzy inference system is constructed to generate an initial prediction result of the transformer remaining life. The parameters of the fuzzy inference system are optimized using a multi-objective optimization algorithm to obtain an optimized fuzzy inference system. The real-time collected transformer multimodal operation data is input into the optimized fuzzy inference system to obtain the transformer health status assessment result and the final remaining life prediction result, including:

[0095] Mapping the aligned graph representation data to the quantum state space through a parameterized quantum circuit to obtain a quantum state representation, defining a quantum kernel function in the quantum state space, calculating the similarity between different quantum state representations, constructing a quantum support vector machine optimization problem, applying a quantum phase estimation algorithm to solve the optimization problem, obtaining an optimal Lagrangian multiplier, and constructing a quantum decision function based on the optimal Lagrangian multiplier;

[0096] Using the quantum decision function to classify the health status of the new graph representation data, obtaining the health status classification result, converting the health status classification result, transformer operation time, and load rate into a fuzzy set with fuzzy membership, and constructing a fuzzy rule base containing multiple fuzzy rules based on historical data and expert knowledge;

[0097] The fuzzy set is inferred by the Mamdani reasoning method, including calculating the rule triggering strength, truncating the rule conclusion and aggregating all rule outputs to obtain the fuzzy reasoning result, and the fuzzy reasoning result is defuzzified by the adaptive centroid method to obtain the preliminary prediction value of the remaining life of the transformer;

[0098] An improved α cutoff method is applied to calculate the confidence interval of the preliminary prediction value, and the health status classification result, the preliminary prediction value and its confidence interval are input into a pre-trained deep neural network. The deep neural network combines long short-term memory units and an attention mechanism to perform time series feature extraction and multi-scale fusion. The output of the deep neural network is combined with a multi-objective optimization algorithm to optimize the preliminary prediction value to obtain the final transformer health status assessment result and remaining life prediction value.

[0099] In this embodiment, the aligned graph representation data is first mapped to the quantum state space through a parameterized quantum circuit. Specifically, a parameterized quantum circuit can be constructed using quantum gates such as Hadamard gates and rotation gates to encode classical data into quantum states. For example, for an 8-dimensional classical data vector, a parameterized quantum circuit with a depth of 4 can be constructed using 3 quantum bits, and the vector can be mapped into an 8-dimensional quantum state by adjusting the parameters of the quantum gate.

[0100] Next, a quantum kernel function is defined in the quantum state space to calculate the similarity between different quantum state representations. Commonly used quantum kernel functions include linear kernels, polynomial kernels, and Gaussian kernels. Based on the defined quantum kernel function, an optimization problem for a quantum support vector machine is constructed. The goal of this optimization problem is to maximize the classification interval while minimizing the classification error. To solve this optimization problem, a quantum phase estimation algorithm is applied. First, a quantum operator is constructed whose eigenvalue is related to the optimization objective function. Then, a quantum phase estimation circuit is applied to the initial quantum state containing the optimization variable information, and the auxiliary quantum bit is measured to obtain the phase estimation result. By repeating this process multiple times and taking the average, the optimal Lagrangian multiplier can be obtained, and the quantum decision function is constructed using the obtained optimal Lagrangian multiplier.

[0101] The constructed quantum decision function is used to classify the health status of the new graph-represented data. For example, a new set of data of a transformer is classified, and the health status classification result is "sub-healthy". The classification result is combined with the current operation time of the transformer for 5 years and the average load rate of 80%, and converted into a fuzzy set with fuzzy membership. At the same time, a fuzzy rule base containing multiple fuzzy rules is constructed based on historical data and expert knowledge. The rule example is "IF health status = sub-healthy AND operation time = medium AND load rate = high THEN remaining life = medium".

[0102] The Mamdani reasoning method is used to reason about the above fuzzy sets. First, the trigger strength of each rule is calculated, that is, the membership value of the input variable is substituted into the antecedent part of the rule, and the minimum value is taken as the trigger strength of the rule. Then, the conclusion of the rule is truncated according to the trigger strength, that is, the membership function of the trigger strength and the conclusion part is minimized. Finally, the truncated outputs of all rules are aggregated, and the maximum value is taken to obtain the comprehensive fuzzy reasoning result.

[0103] The fuzzy reasoning results are defuzzified to obtain the initial prediction results of the transformer's remaining life. The adaptive centroid method is used here to adaptively adjust the centroid calculation method according to the shape of the fuzzy reasoning results. For example, for fuzzy sets with approximately normal distribution, the weighted average method can be used; for fuzzy sets with skewed distribution, the maximum membership mean method can be used. Assume that the initial prediction result is 8.5 years of remaining life.

[0104] The improved α cut method is used to calculate the confidence interval of the preliminary prediction value. First, determine the confidence level α, such as 95%. Then find all elements with a membership greater than or equal to 1-α in the fuzzy inference results. These elements constitute the α cut. The lower and upper bounds of the α cut are the lower and upper bounds of the confidence interval. For example, the 95% confidence interval obtained is [7.2 years, 9.8 years].

[0105] The health status classification result "sub-health", the preliminary prediction value of 8.5 years and its confidence interval [7.2 years, 9.8 years] are input into the pre-trained deep neural network. The network combines long short-term memory (LSTM) units and attention mechanisms to effectively extract time series features and perform multi-scale fusion. The LSTM unit captures long-term dependencies through a gating mechanism, and the attention mechanism adaptively assigns weights to different time steps. The output of the network includes the revised health status assessment results and the remaining life expectancy prediction value.

[0106] Finally, the output of the deep neural network is used to optimize the preliminary prediction value in combination with a multi-objective optimization algorithm. The optimization objectives include minimizing the prediction error and maximizing the prediction confidence. The improved non-dominated sorting genetic algorithm (NSGA-III) is used for multi-objective optimization. New candidate solutions are generated through operations such as crossover and mutation, and the optimal solution is selected based on the non-dominated sorting and reference point method. The final health status assessment result of the transformer is "mild sub-health", the remaining life prediction value is 8.2 years, and the confidence interval is [7.5 years, 8.9 years].

[0107] Through the above steps, a method for transformer health status assessment and remaining life prediction based on quantum support vector machine and fuzzy reasoning is realized. This method comprehensively utilizes technologies such as quantum computing, fuzzy logic and deep learning, and can effectively process transformer multimodal data and improve the accuracy and reliability of assessment and prediction.

[0108] In an optional embodiment,

[0109] The health status classification result, the preliminary prediction value and its confidence interval are input into a pre-trained deep neural network. The deep neural network combines long short-term memory units and attention mechanisms to perform time series feature extraction and multi-scale fusion. The output of the deep neural network is combined with a multi-objective optimization algorithm to optimize the preliminary prediction value, and the final transformer health status assessment result and remaining life prediction value are obtained, including:

[0110] Receive the transformer health status classification results, preliminary prediction values ​​and confidence intervals, preprocess the received data, apply principal component analysis to the preprocessed data to reduce the dimension, obtain the reduced feature data, and input it into a pre-trained deep neural network, which contains two parallel processing branches: time series feature flow and state feature flow;

[0111] In the time series feature stream, time series data is processed through a multi-layer stacked bidirectional long short-term memory network, and the adaptive time window technology is applied to optimize the time series features by dynamically adjusting the input sequence length. At the same time, in the state feature stream, a multi-scale residual convolutional network is used to process the state data to obtain the optimized state features.

[0112] The optimized time series features and state features are input into the feature pyramid network to generate a comprehensive feature representation, which is then input into the multi-task learning module to achieve knowledge transfer between tasks through the task correlation matrix, and to simultaneously predict the health status and remaining lifespan, thus obtaining a preliminary health status assessment result and a remaining lifespan prediction value.

[0113] Construct a multi-objective optimization problem including prediction accuracy, model complexity and prediction uncertainty, solve the multi-objective optimization problem by a multi-objective optimization algorithm, obtain a set of optimal solutions, fine-tune the deep neural network parameters based on the obtained optimal solutions, and obtain multiple deep neural network models with different parameter configurations;

[0114] Each deep neural network model is used to predict the health status and remaining life of the transformer respectively, and multiple groups of prediction results are obtained. The multiple groups of prediction results are fused using an integration strategy to obtain the final transformer health status assessment results and remaining life prediction results.

[0115] In a specific implementation, the transformer health status classification result, preliminary prediction value and confidence interval are used as input data, and a series of processing steps are performed to obtain the final transformer health status assessment result and remaining life prediction value.

[0116] First, the input data is preprocessed. Preprocessing includes data cleaning, normalization, and standardization. For example, the health status classification results can be encoded as numerical variables; the preliminary prediction values ​​can be normalized to be distributed between 0 and 1; the confidence intervals can be converted into standard deviations. The preprocessed data is more regular, which is conducive to subsequent analysis.

[0117] Next, principal component analysis (PCA) is applied to the preprocessed data for dimensionality reduction. PCA can extract the main features in the data and remove redundant information. In the specific operation, the covariance matrix of the data is first calculated, and then the eigenvalues ​​and eigenvectors are solved, and the eigenvectors corresponding to the first k largest eigenvalues ​​are selected as the principal components. Through PCA, the original high-dimensional data can be reduced to a lower dimension while retaining the main information. For example, the original 100-dimensional data can be reduced to 20 dimensions, retaining 95% of the information.

[0118] Then, the reduced feature data is input into a pre-trained deep neural network. The network contains two parallel processing branches: the time series feature stream and the state feature stream. The time series feature stream uses a multi-layer stacked bidirectional long short-term memory (BiLSTM) network to process time series data. The BiLSTM network can consider both past and future information at the same time to better capture time series dependencies. For example, a 3-layer BiLSTM can be used, each containing 128 hidden units. At the same time, an adaptive time window technology is applied to dynamically adjust the input sequence length. In specific implementation, the initial window size can be set to 50, and then dynamically adjusted according to the prediction error, with the window size ranging from 30-100.

[0119] The state feature flow uses a multi-scale residual convolutional network to process state data. The multi-scale residual convolutional network contains multiple parallel convolution branches, each branch uses a convolution kernel of different sizes, and can extract features of different scales. For example, 3 parallel branches can be used, and the convolution kernel sizes are 3x3, 5x5, and 7x7 respectively. Residual connections can alleviate the gradient vanishing problem of deep networks and improve model performance. Through this structure, optimized state features can be obtained.

[0120] The optimized time series features and state features are input into the Feature Pyramid Network (FPN). FPN can generate multi-scale feature representations, which is beneficial for subsequent prediction tasks. FPN contains top-down paths and lateral connections, which can fuse features at different levels. For example, a 4-layer FPN can be constructed, and the size of the feature map of each layer is halved successively. Through FPN, a comprehensive feature representation can be obtained.

[0121] Next, the generated comprehensive feature representation is input into the multi-task learning module. This module realizes knowledge transfer between tasks through the task relevance matrix and predicts the health status and remaining life span at the same time. In specific implementation, a shared feature extraction layer can be designed, and then the health status prediction branch and the remaining life prediction branch can be connected respectively. The task relevance matrix is ​​used to adjust the weights between different tasks and promote knowledge sharing. Through multi-task learning, preliminary health status assessment results and remaining life prediction values ​​can be obtained.

[0122] In order to further optimize the prediction results, a multi-objective optimization problem including prediction accuracy, model complexity and prediction uncertainty is constructed. Prediction accuracy can be measured by root mean square error (RMSE), model complexity can be expressed by the number of parameters, and prediction uncertainty can be expressed by the variance of the prediction results. A multi-objective optimization algorithm (such as NSGA-II) is used to solve the problem and obtain a set of optimal solutions. Each optimal solution corresponds to a set of network parameter configurations.

[0123] Based on the optimal solution obtained, the parameters of the deep neural network are fine-tuned to obtain multiple models with different parameter configurations. For example, 10 different model configurations can be obtained. These models have their own advantages in different objectives and can model problems from different angles.

[0124] These different models are used to predict the health status and remaining life of the transformer respectively, and multiple sets of prediction results are obtained. For example, for a certain transformer sample, 10 models may give 10 different sets of prediction results. Then, an integration strategy is used to fuse multiple sets of prediction results to obtain the final prediction result. The integration strategy can use simple averaging, weighted averaging or stacking. For example, different weights can be assigned to each model based on its performance on the validation set, and weighted averaging can be performed.

[0125] Through the above steps, the health status assessment results and remaining life prediction values ​​of the transformer are finally obtained. This method makes comprehensive use of deep learning, multi-task learning, multi-objective optimization and other technologies, which can effectively improve the accuracy and robustness of the prediction. At the same time, by integrating the results of multiple models, the limitations of a single model can be reduced and more reliable prediction results can be obtained.

[0126] S103. Based on the transformer maintenance recommendations and historical maintenance records, a transformer digital twin model for maintenance strategy optimization is constructed, the health status assessment results, the remaining life prediction results and the transformer maintenance recommendations are input into the transformer digital twin model, fault simulation and predictive maintenance analysis are performed, and based on the fault simulation and predictive maintenance analysis results, combined with the expert knowledge base, an intelligent maintenance plan is generated.

[0127] Based on transformer maintenance recommendations and historical maintenance records, a transformer digital twin model is built for maintenance strategy optimization. The digital twin model includes physical models (such as thermal models, electrical models) and data-driven models (such as life prediction models). Use tools such as Unity 3D to build a three-dimensional visualization model of the transformer to achieve real-time data mapping and status display.

[0128] The health status assessment results, remaining life prediction results and transformer maintenance recommendations are input into the transformer digital twin model for fault simulation and predictive maintenance analysis. For example, the degradation process of the transformer under different maintenance strategies can be simulated to evaluate the cost-effectiveness of various maintenance plans.

[0129] Based on the results of fault simulation and predictive maintenance analysis, combined with the expert knowledge base, an intelligent maintenance plan is generated. The expert knowledge base contains information such as historical fault cases and maintenance experience. Using technologies such as case-based reasoning (CBR), similar cases are retrieved from the knowledge base, and personalized maintenance plans are formulated based on the current situation, including detailed information such as maintenance items, required resources, and estimated duration.

[0130] Finally, the generated intelligent maintenance plan is presented to the operation and maintenance personnel through a visual interface, and relevant instructions are pushed through the mobile terminal App to guide the on-site maintenance work. At the same time, the maintenance results are fed back to the digital twin model to continuously optimize the prediction model and maintenance strategy.

[0131] In an optional embodiment,

[0132] The health status assessment result, the remaining life prediction result and the transformer maintenance suggestion are input into the transformer digital twin model to perform fault simulation and predictive maintenance analysis. Based on the fault simulation and predictive maintenance analysis results and combined with the expert knowledge base, an intelligent maintenance plan is generated, including:

[0133] Construct a transformer digital twin model, which includes a physical layer, a data layer, an analysis layer, and a decision layer. Collect real-time monitoring data, historical operation data, environmental data, and manual inspection data of the transformer, align and fuse the collected data through tensor decomposition and transfer learning methods to obtain fused data, input the fused data into the data layer of the digital twin model, and use a recurrent neural network to process the time series data to obtain time series features;

[0134] In the analysis layer, the convolutional neural network and the graph neural network are integrated to extract the spatial-temporal features of the time series features, and the fuzzy reasoning system and evidence theory are combined to output the health status scores and credibility of each component of the transformer. According to the health status scores and credibility, an adaptive Monte Carlo simulation method is used in the digital twin model to perform fault simulation and evolution path prediction to obtain fault simulation results and evolution path prediction results;

[0135] Based on the fault simulation results and the evolution path prediction results, a dynamic Bayesian network model is constructed to calculate the failure probability and mutual influence relationship of each component of the transformer and generate a risk assessment matrix;

[0136] Using the risk assessment matrix, combined with a preset maintenance rule base, a heuristic algorithm is used to generate multiple candidate maintenance plans, and the candidate maintenance plans are sorted according to a cost-benefit analysis to form an optimized maintenance plan set;

[0137] A multimodal knowledge graph is constructed, and maintenance knowledge related to the health status score is extracted through graph embedding and relational reasoning. An intelligent maintenance plan is generated by combining the optimized maintenance plan set and the maintenance knowledge.

[0138] In a transformer health assessment and intelligent maintenance method, in order to generate an intelligent maintenance plan, it is necessary to input the health status assessment results, remaining life prediction results and transformer maintenance suggestions into the transformer digital twin model for fault simulation and predictive maintenance analysis. The specific implementation method is as follows:

[0139] First, a transformer digital twin model is constructed. The model consists of four layers: physical layer, data layer, analysis layer, and decision layer. The physical layer corresponds to the actual transformer and its components, the data layer is responsible for data collection and processing, the analysis layer performs data analysis and fault diagnosis, and the decision layer is used to generate maintenance decisions.

[0140] In the data collection stage, various sensors are used to collect real-time monitoring data of the transformer, such as oil temperature, load current, gas content, etc.; historical operation data of the transformer is extracted from the historical database; environmental data such as ambient temperature and humidity are obtained from the weather station; and data recorded by manual inspections are collected at the same time. For example, the oil temperature data of a 330kV transformer is 45°C, the load current is 500A, the H2 content of dissolved gas in the top oil is 15μL / L, the CO content is 150μL / L, the ambient temperature is 30°C, and the relative humidity is 60%.

[0141] Next, tensor decomposition and transfer learning methods are used to align and fuse the collected multi-source heterogeneous data. Specifically, data from different sources are aligned by timestamp, a third-order tensor is constructed, and common features are extracted through Tucker decomposition. Then, the features of different data sources are mapped to the same feature space using transfer learning methods. For example, continuous numerical data such as oil temperature and current and discrete data such as gas content are uniformly mapped to values ​​in the range of 0-1. The fused data contains comprehensive information on all aspects of the transformer.

[0142] The fused data is input into the data layer of the digital twin model. In the data layer, recurrent neural networks such as long short-term memory networks (LSTM) are used to process time series data and extract time series features. For example, for oil temperature data, LSTM can learn time series features such as the rising / falling trend and periodic changes of temperature.

[0143] In the analysis layer, convolutional neural network (CNN) and graph neural network (GNN) are integrated to extract the spatial-temporal features of time series features. CNN is used to extract local features, such as the distribution characteristics of oil temperature at different locations; GNN is used to capture the correlation between the various components of the transformer. Combining fuzzy reasoning system and evidence theory, multi-dimensional features are integrated to output the health status score and credibility of each component of the transformer. For example, the health status score of a transformer insulation system is 85 points (out of 100 points) and the credibility is 0.9.

[0144] Based on the health status score and credibility, an adaptive Monte Carlo simulation method is used in the digital twin model to simulate faults and predict evolution paths. Specifically, a large number of possible fault evolution scenarios are generated based on the current state, the fault development process is simulated, and possible future fault paths are predicted. For example, the simulation results show that the probability of severe deterioration of the transformer insulation system within 3 years is 5%, and within 10 years is 30%.

[0145] Based on the fault simulation results and evolution path prediction results, a dynamic Bayesian network model is constructed to calculate the failure probability and mutual influence relationship of each component of the transformer, and generate a risk assessment matrix. This matrix reflects the failure risk level of each component and its impact on the overall system reliability. For example, the failure risk level of the insulation system is medium, and the impact weight on the overall reliability of the transformer is 0.3.

[0146] Using the risk assessment matrix, combined with the pre-established maintenance rule base, heuristic algorithms such as genetic algorithms are used to generate multiple candidate maintenance plans. The maintenance rule base contains rules such as maintenance measures and maintenance cycles corresponding to various fault types. The algorithm generates different maintenance combination plans based on factors such as risk level and impact. Then, the candidate plans are sorted according to the cost-benefit analysis to form an optimized maintenance plan set. For example, multiple candidate plans such as replacing insulating oil, local filtration, and comprehensive overhaul are generated. After sorting, local filtration is selected as the optimal plan.

[0147] Finally, a multimodal knowledge graph is constructed, which contains knowledge on equipment parameters, historical faults, maintenance experience, etc. Maintenance knowledge related to the current health status score is extracted from the knowledge graph through graph embedding and relational reasoning methods. This knowledge is combined with the optimized maintenance plan set to generate the final intelligent maintenance plan. For example, combined with historical experience, it is recommended to test the moisture content of insulation paper while performing local filtration.

[0148] Through the above steps, the whole process from data collection, feature extraction, status assessment to maintenance decision-making is intelligentized, providing strong support for the scientific maintenance and management of transformers. This method makes full use of advanced technologies such as digital twins and artificial intelligence to improve the accuracy and efficiency of transformer health assessment and maintenance decision-making.

[0149] In an optional embodiment,

[0150] Constructing a multimodal knowledge graph, extracting maintenance knowledge related to the health status score through graph embedding and relational reasoning, combining the optimized maintenance plan set with the maintenance knowledge, and generating an intelligent maintenance plan including:

[0151] The extracted named entities, relations, equipment parts, defects, timing patterns, and abnormal features are integrated into the attribute graph model to form a multimodal knowledge graph containing equipment entities, fault types, symptom characteristics, maintenance operations, expert rules, and historical cases;

[0152] Introducing an uncertainty representation method based on evidence theory, assigning basic probabilities to the relationships in the multimodal knowledge graph, adopting an improved transfer distance model to embed the multimodal knowledge graph, constructing an objective function of the transfer distance model based on the graph embedding result, and obtaining a low-dimensional vector representation of entities and relationships in the multimodal knowledge graph by optimizing the objective function;

[0153] A subgraph related to the health status score is extracted from the multimodal knowledge graph using a bidirectional attention network, and the extracted subgraph is encoded using a graph attention network, a reasoning path is generated in the encoded subgraph using a policy gradient reinforcement learning algorithm, and the generated multiple reasoning paths are weightedly fused to obtain maintenance knowledge related to the health status score;

[0154] The maintenance plan generation problem is modeled as a multi-objective Markov decision process. Based on the obtained maintenance knowledge, the multi-objective Markov decision process problem is solved, the strategy network and multiple state-action value functions are learned, and based on the learned strategy, candidate maintenance plans are generated. The generated candidate maintenance plans are screened and sorted, and finally an intelligent maintenance plan is obtained.

[0155] In a specific implementation, the process of constructing a multimodal knowledge graph and extracting maintenance knowledge through graph embedding and relational reasoning, and then generating an intelligent maintenance plan in combination with an optimized maintenance plan set is as follows:

[0156] First, named entities, relationships, equipment components, defects, timing patterns, and abnormal features are extracted from multi-source heterogeneous data and integrated into the attribute graph model. Specifically, named entity recognition and relationship extraction techniques can be used to identify entities such as equipment names, fault types, symptom features, and the relationships between them from text data; signal processing and pattern recognition methods are used to extract timing patterns and abnormal features from sensor data; and image processing and computer vision techniques are used to identify equipment components and defects from image data. This information is stored in the attribute graph database in the form of nodes and edges to form a knowledge graph containing multiple modal information such as equipment entities, fault types, symptom features, maintenance operations, expert rules, and historical cases. For example, triple relationships such as "equipment A-exists-fault B", "fault B-manifests-symptom C", and "symptom C-needs-maintenance operation D" can be established.

[0157] Next, in order to deal with the uncertainty in the knowledge graph, a representation method based on evidence theory is introduced. Specifically, a basic probability value is assigned to each relationship in the knowledge graph to indicate the credibility of the relationship. For example, based on the statistical results of historical data, a basic probability of 0.8 can be assigned to the relationship "Fault B-Manifestation-Symptom C", indicating that there is an 80% chance that the relationship is established. For relationships that lack prior knowledge, a lower basic probability can be assigned.

[0158] Then, the improved transfer distance model is used to embed the multimodal knowledge graph. First, the embedding vectors of entities and relations, as well as the transfer function, are defined. For each triple in the knowledge graph, the distance between the head entity and the tail entity after the relationship transfer is calculated. An optimization objective function is constructed so that the distance of the correct triple is less than the distance of the incorrect triple. The objective function is iteratively optimized through optimization algorithms such as stochastic gradient descent, and finally a low-dimensional vector representation of entities and relations is obtained. For example, both entities and relations can be embedded in a 100-dimensional vector space.

[0159] Next, a bidirectional attention network is used to extract subgraphs related to health status scores from the multimodal knowledge graph. Specifically, the health status score is used as the query, and the attention score between each node in the knowledge graph and the query is calculated. Nodes with higher attention scores and their related edges are selected to form a subgraph related to the health status score. Then, a graph attention network is used to encode the extracted subgraph to obtain the contextual representation of the node.

[0160] After that, the policy gradient reinforcement learning algorithm is used to generate reasoning paths in the encoded subgraph. The path generation process is modeled as a Markov decision process, and the state space, action space, and reward function are defined. The policy network is used to select the next transfer action according to the current state until the termination condition is reached. The reward is calculated for the generated path and the policy network parameters are updated. Through multiple iterations, a strategy for efficient reasoning in the subgraph is learned. After generating multiple reasoning paths, they are weighted and fused according to the reliability of each path to obtain maintenance knowledge related to the health status score.

[0161] Finally, the maintenance plan generation problem is modeled as a multi-objective Markov decision process. Define the state space, action space, transition function and multiple objective functions (such as maintenance cost, downtime, etc.). Based on the maintenance knowledge obtained above, a multi-objective reinforcement learning algorithm is used to solve the problem. Specifically, a policy network and multiple state-action value functions are learned. The policy network selects actions based on the current state, and each value function estimates the long-term reward that can be obtained by taking a certain action in the current state. By alternately optimizing the policy network and the value function, the optimal strategy for weighing multiple objectives is learned. Based on the learned strategy, multiple candidate maintenance plans are generated. These candidate plans are screened and sorted, infeasible plans are removed, and the remaining plans are comprehensively scored according to the importance of multiple objectives. The plan with the highest score is selected as the final intelligent maintenance plan output.

[0162] In an optional embodiment,

[0163] The calculation formula for constructing the objective function of the transfer distance model based on the graph embedding results is as follows:

[0164]

[0165] Among them, J represents the objective function, E represents the edge set of the graph, H represents the low-dimensional representation matrix of the node, and α ij represents the attention weight of the relationship between nodes, h i represents the low-dimensional representation of node vi, h j Represents node v j The low-dimensional representation of λ1 represents the weight coefficient of the transfer distance, D ij (·) represents the transfer distance function, λ2 represents the weight coefficient of the regularization term, and A i Represents node v i The feature matrix of .

[0166] In this embodiment, by integrating the extracted named entities, relations, equipment components, defects, timing patterns and abnormal features into the multimodal knowledge graph, a comprehensive representation of equipment entities, fault types, symptom characteristics, maintenance operations, expert rules and historical cases can be achieved. The uncertainty representation method based on evidence theory is introduced to assign basic probabilities to the relations in the knowledge graph, and the graph embedding is combined with the improved transfer distance model. After optimizing the objective function, a low-dimensional vector representation can be obtained to improve the availability of information. The subgraph related to the health status score is extracted from the knowledge graph through a bidirectional attention network, and then encoded using the graph attention network, and the reasoning path is generated in combination with the policy gradient reinforcement learning algorithm, thereby achieving weighted fusion of maintenance knowledge and enhancing the accuracy of health status assessment. Finally, the maintenance plan generation problem is modeled as a multi-objective Markov decision process. With the help of maintenance knowledge, the process is solved. The learned policy network and state-action value function will support the generation of candidate maintenance plans, and finally obtain intelligent maintenance plans through screening and sorting, thereby improving the efficiency and intelligence level of equipment maintenance.

[0167] Figure 2 FIG. 1 is a schematic diagram of a transformer life prediction system based on machine learning and the Internet of Things according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0168] The first unit is used to collect and preprocess the multimodal operation data of the transformer, encrypt, transmit and store the preprocessed multimodal operation data to form a historical data set, apply adaptive wavelet transform to the historical data set, extract multi-scale time-frequency features, and obtain time-frequency feature data, and use a non-negative matrix decomposition algorithm to reduce the dimension and select features of the time-frequency feature data to obtain feature data after dimension reduction, build a multimodal data fusion model based on a graph neural network, map the feature data after dimension reduction to a graph structure to obtain graph representation data, and apply a dynamic time warping algorithm to perform time series alignment on the graph representation data to obtain aligned graph representation data;

[0169] The second unit is used to use a quantum support vector machine to perform nonlinear classification on the aligned graph representation data to obtain a transformer health status classification result, build a fuzzy reasoning system based on the transformer health status classification result, generate an initial prediction result of the transformer remaining life, optimize the parameters of the fuzzy reasoning system using a multi-objective optimization algorithm to obtain an optimized fuzzy reasoning system, input the real-time collected transformer multimodal operation data into the optimized fuzzy reasoning system to obtain the transformer health status assessment result and the final remaining life prediction result, and generate a transformer maintenance recommendation based on the health status assessment result and the final remaining life prediction result in combination with predefined maintenance rules;

[0170] The third unit is used to build a transformer digital twin model for maintenance strategy optimization based on the transformer maintenance suggestions and historical maintenance records, input the health status assessment results, the remaining life prediction results and the transformer maintenance suggestions into the transformer digital twin model, perform fault simulation and predictive maintenance analysis, and generate an intelligent maintenance plan based on the fault simulation and predictive maintenance analysis results in combination with an expert knowledge base.

[0171] According to a third aspect of the embodiments of the present invention,

[0172] An electronic device is provided, comprising:

[0173] processor;

[0174] a memory for storing processor-executable instructions;

[0175] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0176] A fourth aspect of the embodiments of the present invention is:

[0177] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0178] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A transformer life prediction method based on machine learning and the Internet of Things, characterized in that: include: Collect transformer multimodal operation data and preprocess it, encrypt and transmit and store the preprocessed multimodal operation data to form a historical data set, apply adaptive wavelet transform to the historical data set, extract multi-scale time-frequency features, and obtain time-frequency feature data. At the same time, use a non-negative matrix decomposition algorithm to reduce the dimension and select features of the time-frequency feature data to obtain feature data after dimension reduction, build a multimodal data fusion model based on a graph neural network, map the feature data after dimension reduction to a graph structure, obtain graph representation data, and apply a dynamic time warping algorithm to perform time series alignment on the graph representation data to obtain aligned graph representation data; Using a quantum support vector machine to perform nonlinear classification on the aligned graph representation data to obtain a transformer health status classification result, based on the transformer health status classification result, constructing a fuzzy inference system to generate an initial prediction result of the transformer remaining life, using a multi-objective optimization algorithm to optimize the parameters of the fuzzy inference system to obtain an optimized fuzzy inference system, inputting the real-time collected transformer multimodal operation data into the optimized fuzzy inference system to obtain the transformer health status assessment result and the final remaining life prediction result, and generating transformer maintenance recommendations based on the health status assessment result and the final remaining life prediction result in combination with predefined maintenance rules; Based on the transformer maintenance recommendations and historical maintenance records, a transformer digital twin model for maintenance strategy optimization is constructed, the health status assessment results, the remaining life prediction results and the transformer maintenance recommendations are input into the transformer digital twin model, fault simulation and predictive maintenance analysis are performed, and based on the fault simulation and predictive maintenance analysis results, combined with an expert knowledge base, an intelligent maintenance plan is generated.

2. The method according to claim 1, characterized in that: A multimodal data fusion model based on a graph neural network is constructed, the feature data after dimension reduction is mapped to a graph structure to obtain graph representation data, and a dynamic time warping algorithm is applied to perform time series alignment on the graph representation data. The aligned graph representation data includes: Collecting transformer multimodal operation data, the multimodal operation data including electrical parameter data, acoustic data, thermal imaging data and oil-gas data, performing dimension reduction preprocessing on the multimodal operation data, using a non-negative matrix decomposition algorithm to retain the non-negative characteristics of the data, and obtaining feature data after dimension reduction; The feature data after dimensionality reduction is mapped to nodes in a graph structure, each node represents multimodal data at a time point, directed edges are added to adjacent time point nodes based on timestamps, edge weights are calculated using an exponential decay function, temporal relationships are constructed, an adaptive k-nearest neighbor algorithm is used to establish edge connections between different time points of the same modality, the k value is dynamically adjusted according to data density, intra-modal relationships are constructed, a multi-head attention mechanism is designed to calculate the correlation strength between different modalities as the weight of the inter-modal edges, and inter-modal relationships are constructed, and a multi-level heterogeneous graph structure is formed based on the constructed temporal relationships, intra-modal relationships, and inter-modal relationships; Construct a graph convolution layer containing spatial and temporal convolutions, process multi-level heterogeneous graph structures, capture local spatial-temporal correlations, design an adaptive graph attention layer, process the output of the graph convolution layer, dynamically adjust the importance of different nodes and edges, introduce residual connections and layer normalization, and design a skip connection structure to fuse feature representations at different levels to obtain fused graph representation data; Based on the principal component analysis method, the target operation cycle graph representation is selected as the reference sequence from the fused graph representation data. A dynamic time warping distance metric based on graph structure similarity is designed. Considering the node characteristics and topological structure, a dynamic time warping algorithm is adopted to calculate the alignment path between the fused graph representation data and the reference sequence. The fused graph representation data is aligned using an interpolation algorithm to obtain the aligned graph sequence.

3. The method according to claim 1, characterized in that The aligned graph representation data is nonlinearly classified using a quantum support vector machine to obtain a transformer health status classification result. Based on the transformer health status classification result, a fuzzy inference system is constructed to generate an initial prediction result of the transformer remaining life. The parameters of the fuzzy inference system are optimized using a multi-objective optimization algorithm to obtain an optimized fuzzy inference system. The real-time collected transformer multimodal operation data is input into the optimized fuzzy inference system to obtain the transformer health status assessment result and the final remaining life prediction result, including: Mapping the aligned graph representation data to the quantum state space through a parameterized quantum circuit to obtain a quantum state representation, defining a quantum kernel function in the quantum state space, calculating the similarity between different quantum state representations, constructing a quantum support vector machine optimization problem, applying a quantum phase estimation algorithm to solve the optimization problem, obtaining an optimal Lagrangian multiplier, and constructing a quantum decision function based on the optimal Lagrangian multiplier; Using the quantum decision function to classify the health status of the new graph representation data, obtaining the health status classification result, converting the health status classification result, transformer operation time, and load rate into a fuzzy set with fuzzy membership, and constructing a fuzzy rule base containing multiple fuzzy rules based on historical data and expert knowledge; The fuzzy set is inferred by the Mamdani reasoning method, including calculating the rule triggering strength, truncating the rule conclusion and aggregating all rule outputs to obtain the fuzzy reasoning result, and the fuzzy reasoning result is defuzzified by the adaptive centroid method to obtain the preliminary prediction value of the remaining life of the transformer; An improved α cutoff method is applied to calculate the confidence interval of the preliminary prediction value, and the health status classification result, the preliminary prediction value and its confidence interval are input into a pre-trained deep neural network. The deep neural network combines long short-term memory units and an attention mechanism to perform time series feature extraction and multi-scale fusion. The output of the deep neural network is combined with a multi-objective optimization algorithm to optimize the preliminary prediction value to obtain the final transformer health status assessment result and remaining life prediction value.

4. The method according to claim 3, characterized in that The health status classification result, the preliminary prediction value and its confidence interval are input into a pre-trained deep neural network. The deep neural network combines long short-term memory units and attention mechanisms to perform time series feature extraction and multi-scale fusion. The output of the deep neural network is combined with a multi-objective optimization algorithm to optimize the preliminary prediction value, and the final transformer health status assessment result and remaining life prediction value are obtained, including: Receive the transformer health status classification results, preliminary prediction values ​​and confidence intervals, preprocess the received data, apply principal component analysis to the preprocessed data to reduce the dimension, obtain the reduced feature data, and input it into a pre-trained deep neural network, which contains two parallel processing branches: time series feature flow and state feature flow; In the time series feature stream, time series data is processed through a multi-layer stacked bidirectional long short-term memory network, and the adaptive time window technology is applied to optimize the time series features by dynamically adjusting the input sequence length. At the same time, in the state feature stream, a multi-scale residual convolutional network is used to process the state data to obtain the optimized state features. The optimized time series features and state features are input into the feature pyramid network to generate a comprehensive feature representation, which is then input into the multi-task learning module to achieve knowledge transfer between tasks through the task correlation matrix, and to simultaneously predict the health status and remaining lifespan, thus obtaining a preliminary health status assessment result and a remaining lifespan prediction value. Construct a multi-objective optimization problem including prediction accuracy, model complexity and prediction uncertainty, solve the multi-objective optimization problem by a multi-objective optimization algorithm, obtain a set of optimal solutions, fine-tune the deep neural network parameters based on the obtained optimal solutions, and obtain multiple deep neural network models with different parameter configurations; Each deep neural network model is used to predict the health status and remaining life of the transformer respectively, and multiple groups of prediction results are obtained. The multiple groups of prediction results are fused using an integration strategy to obtain the final transformer health status assessment results and remaining life prediction results.

5. The method according to claim 1, characterized in that The health status assessment result, the remaining life prediction result and the transformer maintenance suggestion are input into the transformer digital twin model to perform fault simulation and predictive maintenance analysis. Based on the fault simulation and predictive maintenance analysis results and combined with the expert knowledge base, an intelligent maintenance plan is generated, including: Construct a transformer digital twin model, which includes a physical layer, a data layer, an analysis layer, and a decision layer. Collect real-time monitoring data, historical operation data, environmental data, and manual inspection data of the transformer, align and fuse the collected data through tensor decomposition and transfer learning methods to obtain fused data, input the fused data into the data layer of the digital twin model, and use a recurrent neural network to process the time series data to obtain time series features; In the analysis layer, the convolutional neural network and the graph neural network are integrated to extract the spatial-temporal features of the time series features, and the fuzzy reasoning system and evidence theory are combined to output the health status scores and credibility of each component of the transformer. According to the health status scores and credibility, an adaptive Monte Carlo simulation method is used in the digital twin model to perform fault simulation and evolution path prediction to obtain fault simulation results and evolution path prediction results; Based on the fault simulation results and the evolution path prediction results, a dynamic Bayesian network model is constructed to calculate the failure probability and mutual influence relationship of each component of the transformer and generate a risk assessment matrix; Using the risk assessment matrix, combined with a preset maintenance rule base, a heuristic algorithm is used to generate multiple candidate maintenance plans, and the candidate maintenance plans are sorted according to a cost-benefit analysis to form an optimized maintenance plan set; A multimodal knowledge graph is constructed, and maintenance knowledge related to the health status score is extracted through graph embedding and relational reasoning. An intelligent maintenance plan is generated by combining the optimized maintenance plan set and the maintenance knowledge.

6. The method according to claim 5, characterized in that Constructing a multimodal knowledge graph, extracting maintenance knowledge related to the health status score through graph embedding and relational reasoning, combining the optimized maintenance plan set with the maintenance knowledge, and generating an intelligent maintenance plan including: The extracted named entities, relations, equipment parts, defects, timing patterns, and abnormal features are integrated into the attribute graph model to form a multimodal knowledge graph containing equipment entities, fault types, symptom characteristics, maintenance operations, expert rules, and historical cases; Introducing an uncertainty representation method based on evidence theory, assigning basic probabilities to the relationships in the multimodal knowledge graph, adopting an improved transfer distance model to embed the multimodal knowledge graph, constructing an objective function of the transfer distance model based on the graph embedding result, and obtaining a low-dimensional vector representation of entities and relationships in the multimodal knowledge graph by optimizing the objective function; A subgraph related to the health status score is extracted from the multimodal knowledge graph using a bidirectional attention network, and the extracted subgraph is encoded using a graph attention network, a reasoning path is generated in the encoded subgraph using a policy gradient reinforcement learning algorithm, and the generated multiple reasoning paths are weightedly fused to obtain maintenance knowledge related to the health status score; The maintenance plan generation problem is modeled as a multi-objective Markov decision process. Based on the obtained maintenance knowledge, the multi-objective Markov decision process problem is solved, the strategy network and multiple state-action value functions are learned, and based on the learned strategy, candidate maintenance plans are generated. The generated candidate maintenance plans are screened and sorted, and finally an intelligent maintenance plan is obtained.

7. The method according to claim 6, characterized in that The calculation formula for constructing the objective function of the transfer distance model based on the graph embedding results is as follows: Among them, J represents the objective function, E represents the edge set of the graph, H represents the low-dimensional representation matrix of the node, and α ij represents the attention weight of the relationship between nodes, h i represents the low-dimensional representation of node vi, h j Represents node v j The low-dimensional representation of λ1 represents the weight coefficient of the transfer distance, D ij (·) represents the transfer distance function, λ2 represents the weight coefficient of the regularization term, and A i Represents node v i The feature matrix of .

8. A transformer life prediction system based on machine learning and the Internet of Things, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect and preprocess the multimodal operation data of the transformer, encrypt, transmit and store the preprocessed multimodal operation data to form a historical data set, apply adaptive wavelet transform to the historical data set, extract multi-scale time-frequency features, and obtain time-frequency feature data, and use a non-negative matrix decomposition algorithm to reduce the dimension and select features of the time-frequency feature data to obtain feature data after dimension reduction, build a multimodal data fusion model based on a graph neural network, map the feature data after dimension reduction to a graph structure to obtain graph representation data, and apply a dynamic time warping algorithm to perform time series alignment on the graph representation data to obtain aligned graph representation data; The second unit is used to use a quantum support vector machine to perform nonlinear classification on the aligned graph representation data to obtain a transformer health status classification result, build a fuzzy reasoning system based on the transformer health status classification result, generate an initial prediction result of the transformer remaining life, optimize the parameters of the fuzzy reasoning system using a multi-objective optimization algorithm to obtain an optimized fuzzy reasoning system, input the real-time collected transformer multimodal operation data into the optimized fuzzy reasoning system to obtain the transformer health status assessment result and the final remaining life prediction result, and generate a transformer maintenance recommendation based on the health status assessment result and the final remaining life prediction result in combination with predefined maintenance rules; The third unit is used to build a transformer digital twin model for maintenance strategy optimization based on the transformer maintenance suggestions and historical maintenance records, input the health status assessment results, the remaining life prediction results and the transformer maintenance suggestions into the transformer digital twin model, perform fault simulation and predictive maintenance analysis, and generate an intelligent maintenance plan based on the fault simulation and predictive maintenance analysis results in combination with an expert knowledge base.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images