An artificial intelligence-based transformer winding deformation monitoring method and system

Abstract

The present invention discloses a method and system for monitoring transformer winding deformation based on artificial intelligence, belonging to the technical field of deformation monitoring. The method includes: data acquisition, signal data processing, constructing a transformer winding deformation monitoring model, and real-time deformation monitoring. This solution selects IMF components based on contribution degree for wavelet transform, designs a guiding value and an update strategy to search for the threshold, finds the optimal threshold for each wavelet coefficient, introduces an adjustment coefficient to design a threshold processing function, and performs threshold processing on each wavelet coefficient based on the optimal threshold; screens and deletes the majority class nodes based on the reference node, generates new nodes for the minority class nodes based on the auxiliary node and the balance factor and calculates the embedding, sets the weight of the edges of the new nodes according to the correlation degree, and performs multi-layer update on the node embedding based on the node embedding matrix and the new adjacency matrix, improving the quality of the transformer winding signal and enhancing the timeliness and reliability of the winding deformation monitoring.

Description

Technical Field

[0001] The present invention belongs to the technical field of deformation monitoring, and specifically refers to a method and system for monitoring transformer winding deformation based on artificial intelligence. Background Art

[0002] The method for monitoring transformer winding deformation utilizes advanced artificial intelligence technology to analyze and process the signal data of the transformer winding in real time, achieve precise monitoring of the deformation state of the transformer winding, ensure the safe and stable operation of the transformer, and reduce the fault risk. However, in the existing methods for monitoring transformer winding deformation, the operation of the transformer is affected by various electromagnetic signals, there is a large amount of noise in the collected signal data of the transformer winding, and the winding signal itself is relatively complex, containing multiple components, resulting in difficult accurate analysis and inability to timely and accurately detect winding deformation; in the existing methods for monitoring transformer winding deformation, the deformation conditions of the transformer winding are complex and diverse, the data has complex non-linear relationships, and there are imbalances in various data types, making it difficult to comprehensively and accurately analyze the characteristics of the signal data of the transformer winding, resulting in misjudgment of transformer winding deformation and inability to timely detect hidden deformation conditions. Summary of the Invention

[0003] In view of the above situation, to overcome the defects of the prior art, the present invention provides a method and system for monitoring transformer winding deformation based on artificial intelligence. In the existing transformer winding deformation monitoring methods, the operation of the transformer is affected by various electromagnetic signals, a large amount of noise exists in the collected transformer winding signal data, and the winding signal itself is relatively complex, containing multiple components, resulting in difficult accurate analysis and inability to timely and accurately detect winding deformation. In this solution, the transformer winding signal data is decomposed, and the IMF components are selected for wavelet transform based on the contribution degree, and the low contribution degree IMF components with a high noise content are accurately selected; a guiding value and an update strategy are designed to search for the threshold, the optimal threshold for each wavelet coefficient is found, an adjustment coefficient is introduced to design a threshold processing function, and each wavelet coefficient is threshold processed based on the optimal threshold to find the optimal threshold to effectively remove the noise in the wavelet coefficient, better adapt to the noise removal requirements in different situations, improve the quality of the transformer winding signal, and enhance the reliability of the winding deformation monitoring; in view of the problems that in the existing transformer winding deformation monitoring methods, the deformation conditions of the transformer winding are complex and diverse, the data has complex non-linear relationships, and the data types are unbalanced, it is difficult to comprehensively and accurately analyze the characteristics of the transformer winding signal data, resulting in misjudgment of the transformer winding deformation and inability to timely detect hidden deformation conditions. In this solution, the embedding is calculated according to the node attribute data, the majority class nodes are screened and deleted based on the reference node to avoid missing important deformations due to the masking of the majority class; new nodes are generated for the minority class nodes based on the auxiliary node and the balance factor and the embedding is calculated to enhance the expression ability of the minority class deformation; the weights of the edges of the new nodes are set according to the correlation degree to more reasonably reflect the actual connection between the winding deformation data; the node embedding is updated layer by layer based on the node embedding matrix and the new adjacency matrix to ensure that hidden deformation conditions can be timely detected and improve the timeliness and reliability of the monitoring.

[0004] The technical solution adopted by the present invention is as follows: A method for monitoring transformer winding deformation based on artificial intelligence provided by the present invention includes the following steps:

[0005] Step S1: Data acquisition;

[0006] Step S2: Signal data processing;

[0007] Step S3: Construct a transformer winding deformation monitoring model;

[0008] Step S4: Real-time deformation monitoring.

[0009] Further, in step S1, the data acquisition is to acquire historical transformer winding signal data; the historical transformer winding signal data includes current signal data, voltage signal data, vibration signal data, temperature signal data, acoustic wave signal data, time stamps, and winding states; the winding state is used as a data label.

[0010] Further, in step S2, the signal data processing specifically includes the following steps:

[0011] Step S21: Decomposition. Empirically decompose the collected transformer winding signal data using the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) method to obtain N IMF components and a residual component;

[0012] Step S22: Calculate the contribution degree. Calculate the contribution degree of the IMF components based on the energy of each component, and preset a contribution threshold. The formula used is as follows:

[0013] ;

[0014] ;

[0015] In the formula, E i is the energy of the i-th component, c i is the i-th component, is the energy of the j-th IMF component, is the contribution degree of the j-th IMF component;

[0016] Step S23: Component processing. Perform wavelet transform on the IMF components with a contribution degree less than the contribution threshold. The wavelet basis is db6 and the number of wavelet layers is 5 to obtain wavelet coefficients;

[0017] Step S24: Search for the optimal threshold. Apply a threshold function to the wavelet coefficients for denoising, and search for the threshold to find the optimal threshold for each wavelet coefficient, including the following:

[0018] Step S241: Initial position. Based on the threshold of each wavelet coefficient, establish a search space, randomly initialize the individual positions within the search space, use the individual positions as representatives of the threshold, and use the signal-to-noise ratio of the wavelet coefficients after threshold processing as the fitness value corresponding to the individual positions;

[0019] Step S242: Design the guiding value. Calculate the guiding value based on the best fitness value at each iteration. The formula used is as follows:

[0020] ;

[0021] In the formula, D t is the guiding value at the t-th iteration, and are the best fitness values at the (t - 1)-th and (t - 2)-th iterations respectively, is rounding down, θ t is the dynamic adjustment benchmark, , exp(·) is the exponential function;

[0022] Step S243: Design an update strategy. Sort the individual positions in descending order according to the fitness value, select the top 10% of the individual positions as the elite individual positions, and randomly generate a random number within the range of (0, 1) for each individual position. Compare the size of the random number and the guiding value, select the update strategy for the individual position, and complete the update of the individual position. The formula used is as follows:

[0023] ;

[0024] In the formula, and are the positions of the b-th individual at the (t + 1)-th and t-th iterations respectively, β1 and β2 are the first guiding factor and the second guiding factor respectively, γ is the dynamic factor, is the Euclidean distance between the global best position and the position of the b-th individual at the t-th iteration, is the global best position at the t-th iteration, is the Euclidean distance between a random elite individual position and the position of the b-th individual at the t-th iteration, is the random elite individual position at the t-th iteration, is the random number of the b-th individual position at the t-th iteration, is the Euclidean distance between a random individual position and the position of the b-th individual at the t-th iteration, is the random individual position at the t-th iteration, is the Hadamard product, o is a constant, I is a d-dimensional uniformly random vector, and d is the dimension of the individual position;

[0025] Step S244: Determine the best threshold. Preset the fitness threshold and the maximum number of iterations, update the fitness value of the individual position. When the best fitness value is higher than the fitness threshold, use the threshold represented by the global best position as the best threshold; otherwise, if the maximum number of iterations is reached, set the number of iterations to 0 and return to Step S241; otherwise, increment the number of iterations by 1 and return to Step S242 to continue the iteration;

[0026] Step S25: Denoising. Introduce an adjustment coefficient to design a threshold processing function, perform threshold processing on each wavelet coefficient based on the best threshold, and remove the noise. The formula used is as follows:

[0027] ;

[0028] In the formula, x a and are the original wavelet coefficient of the a-th and the wavelet coefficient after threshold processing respectively, δ and μ are the first adjustment coefficient and the second adjustment coefficient respectively, Q ais the optimal threshold of the ath wavelet coefficient;

[0029] Step S26: Feature extraction. Perform inverse wavelet transform on the denoised wavelet coefficients to obtain the processed IMF components. Combine and superimpose the processed IMF components, the unprocessed IMF components, and the residual components to obtain the denoised transformer winding signal data. Extract time-domain features and frequency-domain features from the denoised transformer winding signal data, and construct a monitoring data set in combination with data labels; and divide the feature data in the monitoring data set into majority-class data and minority-class data.

[0030] Further, in step S3, the construction of the transformer winding deformation monitoring model specifically includes the following steps:

[0031] Step S31: Initialize the graph structure. Generate a node for each feature data in the monitoring data set, and use each feature data as the attribute data of the corresponding node to obtain a node set and a node attribute matrix; preset a connection threshold. If the Euclidean distance between the attribute data of two nodes is less than the connection threshold, there is an edge between the two nodes, and the weight of the edge is 1; otherwise, there is no edge between the corresponding two nodes, and the weight of the edge is 0; construct an adjacency matrix G based on the weight of the edge; if the feature data corresponding to the node is majority-class data, the node is a majority-class node, otherwise it is a minority-class node.

[0032] Step S32: Calculate the embedding of the node, and the formula used is as follows:

[0033] ;

[0034] In the formula, is the embedding of the vth node, W 1 is the first weight matrix, σ(·) is the activation function, CONCAT(·) is the concatenation function, M is the node attribute matrix, is the attribute data of the vth node, is the vth column of the adjacency matrix G;

[0035] Step S33: Node processing, including the following steps:

[0036] Step S331: Majority-class node processing. Calculate the Euclidean distance between the attribute data of each majority-class node and the attribute data of the remaining nodes, and select the first k nodes with the smallest Euclidean distance as reference nodes; if among the k reference nodes of the majority-class node the number of nodes belonging to the same data label as is greater than or equal to , then delete the majority-class node from the node set and delete the corresponding edge; where, is the nth majority-class node;

[0037] Step S332: Minority class node processing. Find the node with the closest Euclidean distance to each minority class node and belonging to the same data label, and use it as an auxiliary node. Generate a new node for each minority class node based on the auxiliary node, add the new node to the node set, and obtain the node embedding matrix F. The formula used is as follows:

[0038] ;

[0039] In the formula, is the embedding of the new node generated by the u-th minority class node, is the embedding of the u-th minority class node, is the embedding of the auxiliary node corresponding to the u-th minority class node, and α is the balance factor;

[0040] Step S34: Edge processing. Calculate the correlation degree between each new node B r and the remaining nodes B p . Preset a correlation threshold. If the correlation degree is greater than the correlation threshold, there is an edge between the new node B r and the remaining nodes B p , and the weight of the edge is 1; otherwise, there is no edge between the two nodes, and the weight of the edge is 0; and obtain the new adjacency matrix . The formula used is as follows:

[0041] ;

[0042] In the formula, is the correlation degree between B r and B p . B r and B p are the r-th new node and the p-th node in the node set respectively, and are the embeddings of B r and B p respectively, and W 2 is the second weight matrix;

[0043] Step S35: Embedding update. Based on the node embedding matrix F and the new adjacency matrix , perform a Z1-layer update on the embedding of the nodes. The formula used is as follows:

[0044] ;

[0045] In the formula, and are the embeddings of the p-th node in the node set at the (z + 1)-th layer and the z-th layer respectively, is the new adjacency matrix the p-th column of, W z+1 is the weight matrix of the (z + 1)-th layer;

[0046] Step S36: Monitor. Based on the softmax function, normalize the embeddings of the nodes in the Z1-th layer to obtain the probability distribution corresponding to each data label, and select the data label with the maximum probability as the monitoring output of the transformer winding deformation monitoring model for the nodes.

[0047] Furthermore, in step S4, the real-time deformation monitoring is to collect real-time transformer winding signal data; the real-time transformer winding signal data includes current signal data, voltage signal data, vibration signal data, temperature signal data, acoustic wave signal data and time stamps. After signal data processing, it is input into the transformer winding deformation monitoring model for analysis, and based on the output data label, the real-time state of the transformer winding is obtained.

[0048] A transformer winding deformation monitoring system based on artificial intelligence provided by the present invention includes a data acquisition module, a signal data processing module, a module for constructing a transformer winding deformation monitoring model, and a real-time deformation monitoring module;

[0049] The data acquisition module collects historical transformer winding signal data and sends the data to the signal data processing module;

[0050] The signal data processing module decomposes the transformer winding signal data, selects IMF components for wavelet transform based on the contribution degree to obtain wavelet coefficients, designs a guiding value and an update strategy to search for the threshold, finds the optimal threshold for each wavelet coefficient, introduces an adjustment coefficient to design a threshold processing function, performs threshold processing on each wavelet coefficient based on the optimal threshold, extracts features and constructs a monitoring data set, and sends the data to the module for constructing a transformer winding deformation monitoring model;

[0051] The module for constructing a transformer winding deformation monitoring model calculates embeddings according to node attribute data, filters and deletes majority class nodes based on reference nodes, generates new nodes for minority class nodes based on auxiliary nodes and a balance factor and calculates embeddings, sets the weights of the edges of the new nodes according to the correlation degree, performs multi-layer updates on the node embeddings based on the node embedding matrix and the new adjacency matrix, outputs data labels, and sends the data to the real-time deformation monitoring module;

[0052] The real-time deformation monitoring module collects real-time transformer winding signal data, performs signal data processing, and then inputs it into the transformer winding deformation monitoring model for analysis to obtain the real-time state of the transformer winding.

[0053] The beneficial effects achieved by the present invention using the above solution are as follows:

[0054] (1)In view of the problems existing in the existing transformer winding deformation monitoring methods, such as the operation of the transformer being interfered by various electromagnetic signals, a large amount of noise existing in the collected transformer winding signal data, and the winding signal itself being relatively complex, containing various components, resulting in difficult accurate analysis and inability to timely and accurately detect winding deformation, this solution decomposes the transformer winding signal data, which can effectively separate the IMF components and residual components with different characteristics; selects the IMF components based on the contribution degree for wavelet transform to obtain wavelet coefficients, and accurately selects the low contribution degree IMF components with a high noise content; designs a guiding value and an update strategy to search for the threshold, finds the optimal threshold for each wavelet coefficient, introduces an adjustment coefficient to design a threshold processing function, performs threshold processing on each wavelet coefficient based on the optimal threshold, extracts features and constructs a monitoring data set, finds the optimal threshold to effectively remove the noise in the wavelet coefficients, improves the purity of the data, reduces noise interference, and at the same time makes the threshold processing more flexible and accurate, better adapting to the noise removal requirements in different situations, improving the quality of the transformer winding signal, and enhancing the reliability of the winding deformation monitoring.

[0055] (2)In view of the problems existing in the existing transformer winding deformation monitoring methods, such as the complex and diverse winding deformation situations of the transformer, the data having complex non-linear relationships, and the imbalance of each data type, it is difficult to comprehensively and accurately analyze the characteristics of the transformer winding signal data, resulting in misjudgment of the transformer winding deformation and inability to timely detect hidden deformation situations. This solution calculates the embedding based on the node attribute data, converts the complex winding deformation characteristics into processable embedding vectors, which can better capture various deformation characteristics; screens and deletes the majority class nodes based on the reference node, more accurately identifies the relationships between nodes, and avoids missing important deformations due to being covered by the majority class; generates new nodes for the minority class nodes based on the auxiliary node and the balance factor and calculates the embedding, enhancing the expression ability of the minority class deformations, enabling the model to effectively monitor less common but key deformation situations, and improving the sensitivity and comprehensiveness of the monitoring; sets the weights of the new node edges according to the correlation degree, which can more reasonably reflect the actual connections between the winding deformation data, and improves the understanding and monitoring effect of complex deformation relationships; performs multi-layer updates on the node embedding based on the node embedding matrix and the new adjacency matrix, outputs data labels, continuously improves the analysis ability and accuracy of the winding deformation, ensures the timely detection of hidden deformation situations, and improves the timeliness and reliability of the monitoring. Description of the Drawings

[0056] Figure 1 It is a schematic flow chart of a transformer winding deformation monitoring method based on artificial intelligence provided by the present invention;

[0057] Figure 2 It is a schematic diagram of a transformer winding deformation monitoring system based on artificial intelligence provided by the present invention;

[0058] Figure 3 is a schematic flow diagram of step S2;

[0059] Figure 4 is a schematic flow diagram of step S3.

[0060] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. Detailed implementation manners

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0062] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation to the present invention.

[0063] Example 1. Refer to Figure 1 , a method for monitoring transformer winding deformation based on artificial intelligence provided by the present invention, the method includes the following steps:

[0064] Step S1: Data acquisition, acquiring historical transformer winding signal data;

[0065] Step S2: Signal data processing, decomposing the transformer winding signal data, selecting IMF components for wavelet transform based on contribution degree to obtain wavelet coefficients, designing a guiding value and an update strategy to search for thresholds, finding the optimal threshold for each wavelet coefficient, introducing an adjustment coefficient to design a threshold processing function, performing threshold processing on each wavelet coefficient based on the optimal threshold, extracting features and constructing a monitoring data set;

[0066] Step S3: Constructing a transformer winding deformation monitoring model, calculating embeddings according to node attribute data, screening and deleting majority class nodes based on reference nodes, generating new nodes for minority class nodes based on auxiliary nodes and a balance factor and calculating embeddings, setting the weights of the edges of the new nodes according to the correlation degree, performing multi-layer updates on the node embeddings based on the node embedding matrix and the new adjacency matrix, and outputting data labels;

[0067] Step S4: Real-time deformation monitoring. Collect real-time transformer winding signal data. After signal data processing, input it into the transformer winding deformation monitoring model for analysis to obtain the real-time state of the transformer winding.

[0068] Example 2, refer to Figure 1 , this example is based on the above example. In step S1, the historical transformer winding signal data includes current signal data, voltage signal data, vibration signal data, temperature signal data, acoustic wave signal data, time stamp, and winding state; the winding state includes normal, slight deformation, moderate deformation, and severe deformation, and the winding state is used as the data label.

[0069] Example 3, refer to Figure 1 and Figure 3 , this example is based on the above example. In step S2, the signal data processing specifically includes the following steps:

[0070] Step S21: Decomposition. Use the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) method to perform empirical mode decomposition on the collected transformer winding signal data to obtain N IMF components and a residual component; analyze the different frequency components and characteristics of the signal more finely.

[0071] Step S22: Calculate the contribution degree. Calculate the contribution degree of the IMF components based on the energy of each component, and preset a contribution threshold. The formula used is as follows:

[0072] ;

[0073] ;

[0074] In the formula, E i is the energy of the i-th component, c i is the i-th component, is the energy of the j-th IMF component, is the contribution degree of the j-th IMF component;

[0075] Step S23: Component processing. Perform wavelet transform on the IMF components with a contribution degree less than the contribution threshold. The wavelet basis is db6 and the number of wavelet layers is 5 to obtain wavelet coefficients.

[0076] Step S24: Search for the optimal threshold. Apply a threshold function to the wavelet coefficients for denoising, search for the threshold, and find the optimal threshold for each wavelet coefficient to improve the denoising effect and adaptability; it includes the following content:

[0077] Step S241: Initial position. Based on the thresholds of each wavelet coefficient, a search space is established. The individual positions are randomly initialized within the search space, and the individual positions are used as representatives of the thresholds. The signal-to-noise ratio of the wavelet coefficients after threshold processing is used as the fitness value corresponding to the individual positions.

[0078] Step S242: Design the guiding value. Based on the best fitness value at each iteration, the guiding value is calculated using the following formula:

[0079] ;

[0080] In the formula, D t is the guiding value at the t-th iteration, and are the best fitness values at the (t - 1)-th and (t - 2)-th iterations respectively. The best fitness value is the maximum fitness value. is the floor function, θ t is the dynamic adjustment benchmark, , exp(·) is the exponential function;

[0081] Step S243: Design the update strategy. The individual positions are sorted in descending order according to the fitness values. The top 10% of the individual positions are selected as the elite individual positions, and a random number within the range (0, 1) is generated for each individual position. The random number is compared with the guiding value to select the update strategy for the individual positions, and the update of the individual positions is completed using the following formula:

[0082] ;

[0083] In the formula, and are the b-th individual positions at the (t + 1)-th and t-th iterations respectively. β1 and β2 are the first guiding factor and the second guiding factor within the range (0.5, 1] respectively, and γ is the dynamic factor within the range (0.8, 1]. is the Euclidean distance between the global best position and the b-th individual position at the t-th iteration, is the global best position at the t-th iteration. The global best position is the individual position with the best fitness value. is the Euclidean distance between a random elite individual position and the b-th individual position at the t-th iteration, is the random elite individual position at the t-th iteration, is the random number of the b-th individual position at the t-th iteration, is the Euclidean distance between a random individual position and the b-th individual position at the t-th iteration, is the random individual position at the t-th iteration, is the Hadamard product, o is a constant within the range of (0, 1), and I is a d-dimensional uniform random vector within the range, where d is the dimension of the individual position;

[0084] Step S244: Optimal threshold determination. Predetermine the fitness threshold and the maximum number of iterations, update the fitness value of the individual position. When the best fitness value is higher than the fitness threshold, use the threshold represented by the global best position as the optimal threshold; otherwise, if the maximum number of iterations is reached, set the number of iterations to 0 and return to Step S241; otherwise, increment the number of iterations by 1 and return to Step S242 to continue the iteration.

[0085] Step S25: Denoising. Introduce an adjustment coefficient to design a threshold processing function, perform threshold processing on each wavelet coefficient based on the optimal threshold to remove noise. The formula used is as follows:

[0086] ;

[0087] In the formula, x a and are the a-th original wavelet coefficient and the wavelet coefficient after threshold processing respectively, δ and μ are the first adjustment coefficient and the second adjustment coefficient respectively, and Q a is the optimal threshold of the a-th wavelet coefficient;

[0088] Step S26: Feature extraction. Perform inverse wavelet transform on the denoised wavelet coefficients to obtain the processed IMF components. Combine and superimpose the processed IMF components, the unprocessed IMF components, and the residual components to obtain the denoised transformer winding signal data. Extract time-domain features and frequency-domain features from the denoised transformer winding signal data, and construct a monitoring dataset in combination with the data labels; and divide the feature data in the monitoring dataset into majority-class data and minority-class data.

[0089] By performing the above operations, in the existing transformer winding deformation monitoring methods, the operation of the transformer is affected by various electromagnetic signal interferences, there is a lot of noise in the collected transformer winding signal data, and the winding signal itself is relatively complex, containing various components, resulting in difficult accurate analysis and inability to timely and accurately detect winding deformation problems. This solution decomposes the transformer winding signal data, which can effectively separate IMF components and residual components with different characteristics; selects IMF components for wavelet transform based on contribution degree to obtain wavelet coefficients, and accurately selects low-contribution IMF components with high noise content; designs a guiding value and an update strategy to search for thresholds, finds the optimal threshold for each wavelet coefficient, introduces an adjustment coefficient to design a threshold processing function, performs threshold processing on each wavelet coefficient based on the optimal threshold, extracts features and constructs a monitoring data set, finds the optimal threshold to effectively remove the noise in the wavelet coefficients, improves the purity of the data, reduces noise interference, makes the threshold processing more flexible and accurate, better adapts to the noise removal requirements in different situations, improves the quality of the transformer winding signal, and enhances the reliability of the winding deformation monitoring.

[0090] Example 4, refer to Figure 1 and Figure 4 , based on the above example, in step S3, constructing a transformer winding deformation monitoring model specifically includes the following steps:

[0091] Step S31: Initialize the graph structure, generate a node for each feature data in the monitoring data set, and use each feature data as the attribute data of the corresponding node to obtain a node set and a node attribute matrix; preset a connection threshold. If the Euclidean distance between the attribute data of two nodes is less than the connection threshold, there is an edge between the two nodes, and the weight of the edge is 1; otherwise, there is no edge between the corresponding two nodes, and the weight of the edge is 0; construct an adjacency matrix G based on the weight of the edge; if the feature data corresponding to the node is majority class data, the node is a majority class node, otherwise it is a minority class node.

[0092] Step S32: Calculate the embedding of the node, calculate the node embedding using the node attributes and adjacency relationships, which can more comprehensively capture the node features; the formula used is as follows:

[0093] ;

[0094] In the formula, is the embedding of the v-th node, W 1 is the first weight matrix, σ(·) is the ReLU activation function, CONCAT(·) is the concatenation function, M is the node attribute matrix, is the attribute data of the v-th node, is the v-th column of the adjacency matrix G.

[0095] Step S33: Node processing, including the following steps:

[0096] Step S331: Majority class node processing. Calculate the Euclidean distance between each majority class node and the attribute data of the remaining nodes, and select the top k nodes with the smallest Euclidean distance as reference nodes; if among the k reference nodes of the majority class node the number of nodes belonging to the same data label as is greater than or equal to , then delete the majority class node from the node set and delete the corresponding edge; where is the nth majority class node; this avoids the excessive influence of majority class nodes on deformation monitoring and improves the efficiency and accuracy of monitoring;

[0097] Step S332: Minority class node processing. Find the node with the closest Euclidean distance to each minority class node and belonging to the same data label, use it as an auxiliary node, generate a new node for each minority class node based on the auxiliary node, add the new node to the node set, and obtain the node embedding matrix F, which enhances the attention to minority class information and improves the comprehensiveness of monitoring; the formula used is as follows:

[0098] ;

[0099] In the formula, is the embedding of the new node generated by the u-th minority class node, is the embedding of the u-th minority class node, is the embedding of the auxiliary node corresponding to the u-th minority class node, and α is a balance factor in the range of [0, 1];

[0100] Step S34: Edge processing. Calculate the correlation degree between each new node B r and the remaining nodes B p . Preset a correlation threshold. If the correlation degree is greater than the correlation threshold, there is an edge between the new node B r and the remaining nodes B p , and the weight of the edge is 1; otherwise, there is no edge between the two nodes and the weight of the edge is 0; and obtain the new adjacency matrix , which realizes the fine regulation of the relationship between nodes and improves the flexibility and adaptability of monitoring; the formula used is as follows:

[0101] ;

[0102] In the formula, is the correlation degree between B r and B p , B r and Bp They are the r-th new node and the p-th node in the node set respectively, and They are the embeddings of B r and B p respectively, and W 2 is the second weight matrix;

[0103] Step S35: Embedding update. Based on the node embedding matrix F and the new adjacency matrix , perform a Z1-layer update on the node embeddings, which improves the timeliness and accuracy of monitoring; the formula used is as follows:

[0104] ;

[0105] In the formula, and are the embeddings of the p-th node in the node set at the (z + 1)-th layer and the z-th layer respectively, is the p-th column of the new adjacency matrix , and W z+1 is the weight matrix at the (z + 1)-th layer;

[0106] Step S36: Monitoring. Based on the softmax function, normalize the embeddings of the nodes at the Z1 layer to obtain the probability distribution corresponding to each data label, and select the data label with the highest probability as the monitoring output of the transformer winding deformation monitoring model for the nodes.

[0107] By performing the above operations, aiming at the problems existing in the existing transformer winding deformation monitoring methods, such as the complex and diverse deformation situations of the transformer winding, the complex non-linear relationships of the data, and the imbalance of various data types, it is difficult to comprehensively and accurately analyze the characteristics of the transformer winding signal data, resulting in misjudgment of the transformer winding deformation and the inability to timely detect hidden deformation situations. In this solution, the embeddings are calculated according to the node attribute data, and the complex winding deformation characteristics are transformed into processable embedding vectors, which can better capture various deformation characteristics; based on the reference nodes, the majority-class nodes are screened and deleted, and the relationships between the nodes are more accurately identified, avoiding the omission of important deformations due to the masking of the majority class; based on the auxiliary nodes and the balance factor, new nodes are generated for the minority-class nodes and the embeddings are calculated, enhancing the expression ability of the minority-class deformations, enabling the model to effectively monitor less common but critical deformation situations, and improving the sensitivity and comprehensiveness of the monitoring; the weights of the new node edges are set according to the correlation degree, which can more reasonably reflect the actual connection between the winding deformation data, improving the understanding and monitoring effect of the complex deformation relationship; based on the node embedding matrix and the new adjacency matrix, the node embeddings are updated in multiple layers, and the data labels are output, continuously improving the analysis ability and accuracy of the winding deformation, ensuring the timely detection of hidden deformation situations, and improving the timeliness and reliability of the monitoring.

[0108] Embodiment 5, refer to Figure 1 , based on the above embodiment, in step S4, real-time deformation monitoring is to collect real-time transformer winding signal data; the real-time transformer winding signal data includes current signal data, voltage signal data, vibration signal data, temperature signal data, acoustic wave signal data and time stamps. After signal data processing, it is input into the transformer winding deformation monitoring model for analysis. Based on the output data label, the real-time state of the transformer winding is obtained, and the real-time deformation monitoring of the transformer winding is completed.

[0109] Embodiment 6, refer to Figure 2 , based on the above embodiment, an artificial intelligence-based transformer winding deformation monitoring system provided by the present invention includes a data acquisition module, a signal data processing module, a module for constructing a transformer winding deformation monitoring model, and a real-time deformation monitoring module;

[0110] The data acquisition module collects historical transformer winding signal data and sends the data to the signal data processing module;

[0111] The signal data processing module decomposes the transformer winding signal data, selects IMF components for wavelet transform based on contribution degree to obtain wavelet coefficients, designs a guiding value and an update strategy to search for thresholds, finds the optimal threshold for each wavelet coefficient, introduces an adjustment coefficient to design a threshold processing function, performs threshold processing on each wavelet coefficient based on the optimal threshold, extracts features and constructs a monitoring data set, and sends the data to the module for constructing a transformer winding deformation monitoring model;

[0112] The module for constructing a transformer winding deformation monitoring model calculates embeddings according to node attribute data, filters and deletes majority class nodes based on reference nodes, generates new nodes for minority class nodes based on auxiliary nodes and balance factors and calculates embeddings, sets the weights of the edges of the new nodes according to the correlation degree, performs multi-layer updates on the node embeddings based on the node embedding matrix and the new adjacency matrix, outputs data labels, and sends the data to the real-time deformation monitoring module;

[0113] The real-time deformation monitoring module collects real-time transformer winding signal data, performs signal data processing, and then inputs it into the transformer winding deformation monitoring model for analysis to obtain the real-time state of the transformer winding.

[0114] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.

[0115] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention.

[0116] The above description of the present invention and its embodiments is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In general, if those of ordinary skill in the art are inspired by it and design similar structural forms and embodiments without creative efforts without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A transformer winding deformation monitoring method based on artificial intelligence, characterized in that: The method comprises the following steps: Step S1: data collection, collecting historical transformer winding signal data; Step S2: signal data processing; Step S3: constructing a transformer winding deformation monitoring model; Step S4: real-time deformation monitoring, collecting real-time transformer winding signal data, processing the signal data, and inputting it into the transformer winding deformation monitoring model for analysis to obtain the real-time status of the transformer winding; Step S3 includes step S34: edge processing, calculating each new node B r With the rest of the Node B p The correlation between them is pre-set with a correlation threshold. If the correlation is greater than the correlation threshold, the new node B r With the rest of the Node B p There is an edge between the two nodes, and the weight of the edge is 1; otherwise, there is no edge between the two nodes, and the weight of the edge is 0; and a new adjacency matrix is ​​obtained , the formula used is as follows: ； In the formula, It is B r and B p The correlation between r and B p are the rth new node and the pth node in the node set, respectively. and B r and B p The embedding of W 2 is the second weight matrix, σ(•) is the activation function; In step S2, the signal data processing specifically includes the following steps: Step S21: Decomposition: using an improved adaptive noise complete set empirical mode decomposition method ICEEMDAN to perform empirical mode decomposition on the collected transformer winding signal data to obtain N IMF components and residual components; Step S22: Calculate the contribution. Calculate the contribution of the IMF component based on the energy of each component. Preset the contribution threshold. The formula used is as follows: ； ； In the formula, E i is the energy of the ith component, c i is the i-th component, is the energy of the jth IMF component, is the contribution of the jth IMF component; Step S23: Component processing: performing wavelet transform on the IMF components whose contribution is less than the contribution threshold, using the wavelet basis of db6 and the wavelet layer number of 5 to obtain wavelet coefficients; Step S24: searching for the best threshold, applying the threshold function to the wavelet coefficients for denoising, searching for the threshold, and finding the best threshold for each wavelet coefficient; Step S25: Denoising, introducing the adjustment coefficient to design the threshold processing function, performing threshold processing on each wavelet coefficient based on the optimal threshold to remove noise, the formula used is as follows: ； In the formula, x a and are the ath original wavelet coefficient and the wavelet coefficient after threshold processing, δ and μ are the first adjustment coefficient and the second adjustment coefficient, respectively. a is the optimal threshold of the ath wavelet coefficient; Step S26: feature extraction, performing inverse wavelet transform on the denoised wavelet coefficients to obtain processed IMF components, combining and superimposing the processed IMF components, unprocessed IMF components and residual components to obtain denoised transformer winding signal data, extracting time domain features and frequency domain features from the denoised transformer winding signal data, and constructing a monitoring data set in combination with data labels; and dividing the feature data in the monitoring data set into majority class data and minority class data; In step S24, the searching for the optimal threshold value specifically includes the following steps: Step S241: Initial position, establish a search space based on the threshold of each wavelet coefficient, randomly initialize individual positions in the search space, use the individual positions as representatives of the threshold, and use the signal-to-noise ratio of the wavelet coefficients after threshold processing as the fitness value of the corresponding individual position; Step S242: Design the guide value, and calculate the guide value based on the best fitness value in each iteration. The formula used is as follows: ； Where D t is the bootstrap value at the tth iteration, and are the best fitness values ​​at the t-1th and t-2th iterations, respectively. is rounded down, θ t is a dynamically adjusted benchmark. , exp(·) is the exponential function; Step S243: Design an update strategy, sort the individual positions in descending order according to the size of the fitness value, select the top 10% of the individual positions as the elite individual positions, and randomly generate a random number in the range of (0, 1) for each individual position, compare the size of the random number and the guide value, select the update strategy of the individual position, and complete the update of the individual position. The formula used is as follows: ； In the formula, and are the positions of the bth individual at the t+1th and tth iterations, β1 and β2 are the first and second guidance factors, γ is the dynamic factor, is the Euclidean distance between the global best position and the bth individual position at the tth iteration, is the global optimal position at the tth iteration, is the Euclidean distance between a random elite individual position and the bth individual position at the tth iteration, is the position of the random elite individual at the tth iteration, is the random number of the bth individual position at the tth iteration, is the Euclidean distance between a random individual position and the bth individual position at the tth iteration, is the random individual position at the tth iteration, is the Hadamard product, o is a constant, I is a d-dimensional uniform random vector, and d is the dimension of the individual position; Step S244: The optimal threshold is determined, the fitness threshold and the maximum number of iterations are pre-set, and the fitness value of the individual position is updated. When the optimal fitness value is higher than the fitness threshold, the threshold represented by the global optimal position is used as the optimal threshold; otherwise, if the maximum number of iterations is reached, the number of iterations is set to 0 and returns to step S241; otherwise, the number of iterations is increased by 1 and returns to step S242 to continue iterating.

2. The transformer winding deformation monitoring method based on artificial intelligence according to claim 1 is characterized in that: In step S3, the construction of the transformer winding deformation monitoring model specifically includes the following steps: Step S31: Initialize the graph structure, generate a node for each feature data in the monitoring data set, and use each feature data as the attribute data of the corresponding node to obtain a node set and a node attribute matrix; pre-set a connection threshold, if the Euclidean distance between the attribute data of two nodes is less than the connection threshold, there is an edge between the two nodes, and the weight of the edge is 1; otherwise, there is no edge between the corresponding two nodes, and the weight of the edge is 0; construct an adjacency matrix G based on the weight of the edge; if the feature data corresponding to the node is majority class data, the node is a majority class node, otherwise it is a minority class node; Step S32: Calculate the embedding of the node, using the following formula: ； In the formula, is the embedding of the vth node, W 1 is the first weight matrix, σ(·) is the activation function, CONCAT(·) is the connection function, M is the node attribute matrix, is the attribute data of the vth node, is the vth column of the adjacency matrix G; Step S33: node processing; Step S34: edge processing; Step S35: Embedding update based on the node embedding matrix F and the new adjacency matrix , perform Z1 layer update on the node embedding, the formula used is as follows: ； In the formula, and are the embeddings of the p-th node in the node set at the z+1th layer and the z-th layer, respectively. is the new adjacency matrix The pth column, W z+1 is the weight matrix of the z+1th layer; Step S36: Monitoring, normalizing the embedding of the Z1th layer nodes based on the softmax function to obtain the probability distribution corresponding to each data label, and selecting the data label with the largest probability as the monitoring output of the transformer winding deformation monitoring model for the node.

3. The transformer winding deformation monitoring method based on artificial intelligence according to claim 2 is characterized in that: In step S33, the node processing specifically includes the following steps: Step S331: majority class node processing, calculate the Euclidean distance between each majority class node and the attribute data of the remaining nodes, and select the first k nodes with the smallest Euclidean distance as reference nodes; if Among the k reference nodes, The number of nodes belonging to the same data label is greater than or equal to , then the majority class nodes Delete from the node set and delete the corresponding edge; where, is the nth majority class node; Step S332: Minority node processing: find the node with the shortest Euclidean distance to each minority node and belonging to the same data label, use it as an auxiliary node, generate a new node for each minority node based on the auxiliary node, add the new node to the node set, and obtain the node embedding matrix F. The formula used is as follows: ； In the formula, is the embedding of the new node generated by the u-th minority class node, is the embedding of the u-th minority class node, is the embedding of the auxiliary node corresponding to the u-th minority class node, and α is the balancing factor.

4. The transformer winding deformation monitoring method based on artificial intelligence according to claim 1 is characterized in that: In step S1, the data collection is to collect historical transformer winding signal data; the historical transformer winding signal data includes current signal data, voltage signal data, vibration signal data, temperature signal data, sound wave signal data, timestamp and winding status; the winding status is used as a data label.

5. The transformer winding deformation monitoring method based on artificial intelligence according to claim 1 is characterized in that: In step S4, the real-time deformation monitoring is to collect real-time transformer winding signal data; the real-time transformer winding signal data includes current signal data, voltage signal data, vibration signal data, temperature signal data, sound wave signal data and timestamp, and after signal data processing, it is input into the transformer winding deformation monitoring model for analysis, and the real-time status of the transformer winding is obtained based on the output data label.

6. A transformer winding deformation monitoring system based on artificial intelligence, used to implement a transformer winding deformation monitoring method based on artificial intelligence as claimed in any one of claims 1 to 5, characterized in that: It includes a data acquisition module, a signal data processing module, a transformer winding deformation monitoring model building module and a real-time deformation monitoring module; The data acquisition module collects historical transformer winding signal data and sends the data to the signal data processing module; The signal data processing module decomposes the transformer winding signal data, selects the IMF component based on the contribution to perform wavelet transform, obtains the wavelet coefficient, designs the guide value and the update strategy to search the threshold, finds the optimal threshold of each wavelet coefficient, introduces the adjustment coefficient to design the threshold processing function, performs threshold processing on each wavelet coefficient based on the optimal threshold, extracts features and constructs a monitoring data set, and sends the data to the module for constructing the transformer winding deformation monitoring model; The transformer winding deformation monitoring model building module calculates embedding according to node attribute data, screens and deletes majority class nodes based on reference nodes, generates new nodes for minority class nodes based on auxiliary nodes and balance factors and calculates embedding, sets weights of new node edges according to association, performs multi-layer update of node embedding based on node embedding matrix and new adjacency matrix, outputs data labels, and sends data to the real-time deformation monitoring module; The real-time deformation monitoring module collects real-time transformer winding signal data, processes the signal data, and then inputs the data into the transformer winding deformation monitoring model for analysis to obtain the real-time status of the transformer winding.

Images