A transformer double-branch state evaluation method fusing photoacoustic data
By synchronizing and preprocessing multimodal data, and combining a dual-branch architecture of convolutional neural networks, random forest classifiers, and autoencoders, the problems of single evaluation dimension and high computational complexity in transformer condition assessment are solved, achieving efficient and reliable condition assessment and improving the accuracy and credibility of the assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LANZHOU INST OF TECH
- Filing Date
- 2026-01-20
- Publication Date
- 2026-06-12
AI Technical Summary
Existing transformer condition assessment methods have a single assessment dimension, rely on a large number of fault samples, have high computational complexity, fail to fully integrate the complementary characteristics of multi-source information such as sound, light, and electricity, and lack a collaborative verification mechanism for qualitative classification and quantitative scoring, resulting in insufficient credibility and robustness of the assessment results.
Multimodal data synchronization and preprocessing are employed, features are extracted using convolutional neural networks and fused through an attention mechanism, qualitative evaluation is performed using a random forest classifier and quantitative evaluation is performed using an unsupervised autoencoder, an intelligent arbitration module is introduced for cross-validation, and a comprehensive diagnostic report is generated.
It improved the accuracy of transformer condition assessment to 98.7%, reduced the false alarm rate to 1.5%, enabled sensitive detection of early faults, and improved the reliability and engineering applicability of the assessment.
Smart Images
Figure CN122196769A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transformer condition assessment technology, specifically to a transformer dual-branch condition assessment method that integrates acoustic, optical, and electronic data. Background Technology
[0002] With the deepening of smart grid construction, the operation and maintenance mode of power equipment is gradually transforming from traditional periodic maintenance to intelligent condition-based maintenance. As a core piece of equipment in the power grid, the accurate assessment of the operating status of power transformers is crucial for ensuring grid security and improving operation and maintenance efficiency. Currently, transformer condition assessment methods based on multi-source information fusion have become a hot research topic. By integrating various monitoring data such as vibration, oil chromatography, and partial discharge, the accuracy of fault diagnosis can be improved to a certain extent. However, existing methods still have the following limitations: a single assessment dimension, reliance on a large number of fault samples, and high computational complexity of the models.
[0003] To overcome the aforementioned shortcomings, some inventions have attempted to introduce semi-supervised or unsupervised methods, such as using autoencoders for anomaly detection. However, these methods are mostly limited to a single data modality and fail to fully integrate the complementary characteristics of multi-source information such as sound, light, and electricity. Furthermore, existing methods rarely involve collaborative verification mechanisms for qualitative classification and quantitative scoring, leaving room for improvement in the reliability and robustness of the evaluation results.
[0004] Therefore, there is an urgent need to develop a multimodal fusion assessment method that takes into account the comprehensiveness of assessment dimensions, low sample dependence, and high computational efficiency, so as to promote the development of transformer condition assessment technology towards a more intelligent, reliable, and practical direction. Summary of the Invention
[0005] The purpose of this invention is to address the problems existing in the prior art by providing a method for evaluating the dual-branch state of transformers that integrates acoustic, optical, and electronic data.
[0006] To achieve the above objectives, the following technical solution is provided:
[0007] A method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electronic data, characterized by comprising the following steps:
[0008] S1. Multimodal Data Synchronization and Preprocessing: Acoustic, optical, and electrical monitoring data of the transformer are collected, and time-series synchronization and normalization preprocessing are performed to establish spatiotemporal consistency of multi-source data; S2. Feature Extraction and Fusion: Features are automatically extracted from the acoustic, optical, and electrical data using a convolutional neural network, and a fused feature vector is generated through a weighted fusion module based on an attention mechanism; S3. Two-Branch State Assessment: S31. Qualitative Assessment Branch: The fused feature vector is input into a trained random forest classifier, and the qualitative classification result of the transformer state and its corresponding classification confidence score are output; S32. Quantitative Assessment Branch: The fused feature vector is input into an unsupervised autoencoder trained only on normal state samples, its reconstruction error is calculated, and a continuous health score is generated through a mapping function based on the reconstruction error; S4. Intelligent Arbitration and Output: Based on the qualitative classification result, classification confidence score, and health score, cross-validation is performed through a preset arbitration logic; when the two-branch results are consistent, the final state assessment conclusion is output; when the results are contradictory, a manual review flag is triggered.
[0009] Preferably, in S1, firstly, the timing synchronization problem of multi-source signals is solved using a dynamic time warping algorithm, the objective function of which is:
[0010] , (1)
[0011] Where i is an index variable, representing the position of a data point in the time series, with a value ranging from 1 to n; n is the length of the time series, i.e., the total number of data points; x i y represents the value of the first time series (such as an acoustic signal) at the i-th point; π(i) π represents the value of the i-th point in the second time series (such as an optical signal) mapped by π; π is the optimal path mapping function, which represents the best alignment between the two time series and is used to solve the timing synchronization problem of multi-source data.
[0012] Subsequently, the modal data were normalized to eliminate the influence of dimensions:
[0013] , (2)
[0014] Where λ is the original data value, representing the unprocessed input data (such as the original measurement values of sound, light, and electrical signals); μ is the mean of the data, used for centering; σ is the standard deviation of the data, used for scaling to eliminate the influence of dimensions; and z is the normalized data value, making the data conform to a standard normal distribution, which facilitates subsequent feature extraction.
[0015] In addition, affine transformations are performed on the infrared images to establish a mapping relationship between image pixel coordinates and the physical location of the transformer, ensuring the consistency of multi-source data in the spatiotemporal dimension and providing high-quality input data for subsequent feature extraction.
[0016] Preferably, in S2, a convolutional neural network is used to achieve end-to-end automatic feature extraction. A dedicated CNN feature extractor is designed to address the characteristics of different modalities of data.
[0017] Acoustic signal feature extraction: The one-dimensional acoustic vibration signal is converted into a time-spectrum image through short-time Fourier transform, and a two-dimensional CNN is used for feature learning; its convolution calculation can be expressed as:
[0018] (3)
[0019] Where l is the layer index of the convolutional neural network, indicating that the current layer is l; i, j are the spatial coordinates of the feature map, where i represents the row index and j represents the column index; m, n are the size indices of the convolutional kernel, where m ranges from 0 to M-1 and n ranges from 0 to N-1, representing the position of the convolutional kernel in height and width; M, N are the sizes of the convolutional kernel, where M represents the height and N represents the width; W (l) (m,n) represents the weight parameters of the l-th convolutional kernel, with values at position (m,n) used for feature extraction; F (l-1) b is the input feature map of the (l-1)th layer, and is the activation output of the previous layer; (l) σ is the bias term for the l-th layer, used to adjust the output; σ is the activation function, here we use the ReLU function to introduce nonlinearity; Let be the output value of the acoustic feature map of layer l at position (i,j);
[0020] Optical image feature extraction: Deep features are extracted from infrared thermal images using a pre-trained ResNet-18 network, and feature representation is achieved through transfer learning.
[0021] (4)
[0022] Among them, I thermal The input is the infrared thermal image data; θ pretrained The parameter set for the pre-trained ResNet-18 network, including weights and biases, is initialized via transfer learning; ResNet This is a ResNet-18 network function used for feature extraction; F optical The output optical feature vector is a 128-dimensional depth feature.
[0023] Electrical signal feature extraction: Treating multi-channel electrical quantities as a one-dimensional time series, a one-dimensional CNN is used for time series feature extraction.
[0024] , (5)
[0025] in, τ represents the output value of the electrical feature of layer l at time t; l is the network layer index, indicating layer l; t is the time index, indicating the position in the one-dimensional time series; τ is the time offset index of the convolution kernel, from 0 to T-1; T is the time dimension size of the convolution kernel; W (l) (τ) represents the weight of the l-th layer one-dimensional convolutional kernel at offset τ; F (l-1) b is the input feature sequence of the (l-1)th layer; l σ is the bias term of the l-th layer; σ is the ReLU activation function;
[0026] The CNN network described above automatically learns 128-dimensional deep feature vectors from each modality of data, effectively capturing key feature patterns of device status.
[0027] Preferably, in S2, during the feature fusion stage, a feature weighted fusion method based on an attention mechanism is adopted;
[0028] First, calculate the attention weights for each modality feature:
[0029] , (6)
[0030] Where i and j are modal indices, i represents the current mode, and j represents the cyclic index of all modes; N is the total number of modes, here N=3 (acoustic, optical, and electronic); f i Let w be the feature vector of the i-th modality (e.g., 128-dimensional features); w is the attention weight vector, which is a trainable parameter; w T α is the transpose of w, used for dot product calculation; exp is the exponential function, used for softmax calculation; i The attention weight for the i-th modality is between 0 and 1, representing the importance of that modality.
[0031] Then, the final fused feature vector is generated by weighted summation:
[0032] , (7)
[0033] Where i is the modality index, from 1 to N; N is the total number of modes, N=3; α i The attention weights for the i-th modality are derived from formula (6); f i F is the eigenvector of the i-th mode; fusion The fused feature vector has a dimension of 256 and is used for subsequent state evaluation.
[0034] This method can adaptively highlight important features and suppress redundant information. The generated 256-dimensional fusion feature vector not only retains the unique information of each modality, but also reflects the complementarity between multiple source data, providing high-quality feature input for subsequent state assessment.
[0035] Preferably, in S31, the high-dimensional vector F output by the feature fusion module is obtained. fusion As input to the state classification branch; multiple decision trees T are constructed using the random forest algorithm. i Where i = 1, 2, 3, ..., N, N is the total number of decision trees. Each decision tree is independently trained using a bootstrap sampling method to improve the model's generalization ability; the high-dimensional vector F... fusion The input is fed into the trained random forest classifier, so that each decision tree T... i Independently output its category judgment T i (F) fusion The category judgment corresponds to the transformer's state level C. k The state C k The classification levels include at least normal, attentive, and abnormal. Based on the outputs of all decision trees, the final classification result is generated through ensemble voting, where the sample belongs to each state level C. k The conditional probability P(C) k |F fusion Calculate using the following formula: (8)where N is the total number of decision trees, and ∏(.) is the indicator function;
[0036] The conditional probability is output as the core output of the state classification branch to provide a state category judgment and its corresponding confidence level, providing a decision basis for the subsequent arbitration module.
[0037] Preferably, in S32, the autoencoder is trained using a large amount of normal state data, wherein the autoencoder includes an encoder f. enc and decoder f dec The training objective is to minimize the reconstruction error between the input and output, so that the autoencoder can learn the data distribution of the health state; for the fused feature vector F to be evaluated fusion The reconstruction error E is calculated using the autoencoder. recon The calculation formula is:
[0038] (9) The reconstruction error E is reduced by an exponential decay function. recon The Health Score (HS) is mapped to a continuous range of 0 to 1 and is calculated using the following formula: , (10) where λ is the scale parameter, used to adjust the sensitivity of the score to error;
[0039] When the equipment is in perfect health, the reconstruction error approaches 0 and the health score HS approaches 1; when the equipment is in abnormal condition, the reconstruction error increases and the health score HS decreases accordingly.
[0040] Preferably, in S31, the random forest classifier includes a bootstrap sampling module, a random feature selection module, a decision tree construction module, and a prediction module; wherein: the bootstrap sampling module is used to generate K bootstrap sample sets D from the original training set D containing N samples through sampling with replacement. i ,
[0041] , (11)
[0042] Among them, each self-service sample set D i The sample size is the same as the original training set D, and it contains approximately 63.2% of the independent samples in the original training set; the random feature selection module is used to randomly select m features from all M features to form a candidate feature subset when constructing each node of the decision tree, wherein:
[0043] (12) The decision tree construction module is used to recursively construct a decision tree based on the CART algorithm, and select the best split point by minimizing the Gini impurity. For node t, the formula for calculating the Gini impurity Gini(t) is: , (13)
[0044] Where P(i|t) is the proportion of samples belonging to class i in node t, and C is the total number of classes; no pruning is performed during the growth of the decision tree until the number of samples in a node is less than a preset threshold or the maximum depth is reached; the prediction module is used to perform classification prediction on the input sample X based on the constructed K decision trees; the prediction module includes a single tree prediction unit, a collective voting unit, and a probability output; wherein:
[0045] The single-tree prediction unit is used to predict based on each decision tree. For input samples Provide the prediction results independently:
[0046] , (14)
[0047] The collective voting unit is used to aggregate the prediction results of all decision trees through majority voting to obtain the final classification result of the random forest:
[0048] , (15)
[0049] Where mode represents the mode operator;
[0050] The probability output provides the random forest with an estimate of the probability that a sample belongs to each category:
[0051] , (16)
[0052] Here, ‖ (.) is an indicator function that returns 1 when the condition is true and 0 otherwise. This probability estimate provides important confidence information for the subsequent arbitration module.
[0053] Preferably, in S32, the autoencoder is an artificial neural network based on unsupervised learning, whose core objective is to achieve feature extraction and dimensionality reduction by learning the compressed representation of data; the autoencoder consists of two parts, an encoder and a decoder, and learns an effective representation of the data by minimizing the reconstruction error between the input and the output.
[0054] The encoder maps the input data x to a latent space representation z:
[0055] (17)
[0056] Where X is the input data, which refers to the fused feature vector; W is the weight matrix of the encoder; b is the bias vector of the encoder; θ is the set of parameters of the encoder, including W and b; σ is the activation function, such as Sigmoid or ReLU; z is the latent space representation, which is the compressed code of the input.
[0057] The decoder reconstructs the original data from the latent representation z:
[0058] , (18)
[0059] Where z is the latent representation from the encoder output; W' is the weight matrix of the decoder; b' is the bias vector of the decoder; Φ is the parameter set of the decoder, including W' and b'; and σ is the activation function. The output data is the reconstructed data; the training objective of the autoencoder is to minimize the reconstruction error, that is, to obtain the decoder. The output should be as close as possible to the original input x;
[0060] The commonly used loss function is mean squared error (MSE):
[0061] , (19)
[0062] In transformer condition assessment, the autoencoder is trained using a large number of normal state samples to learn the data distribution characteristics of healthy operating conditions. After training, the network will be able to accurately reconstruct the feature vector of normal states, but will produce a large reconstruction error for the feature vector of abnormal states.
[0063] Preferably, in S4, the probability output P(C) from the state classification branch is received. k The continuous scores (HS) of the health score branch are used, and cross-validation and comprehensive decision-making are performed according to the preset consistency rules; the arbitration logic is as follows:
[0064] First, determine if the results of the two branches are consistent: Let the decision result of the state classification branch be C. pred =argmaxP(C k The corresponding confidence level is P. max Consistency rules can be formally defined as: when C pred "Normal" and HS≥θ hight , or when C pred When it is "abnormal" and HS≥θ hight If the results are inconsistent, they are considered to be consistent; otherwise, they are considered to be contradictory. Where θ hight and θ low Health score thresholds set based on historical data or expert knowledge;
[0065] If the results are consistent, the system directly generates a status level C. pred A comprehensive assessment report including Health Score (HS) and high confidence indicators; if the results are contradictory, the system will automatically trigger the "manual review" indicator. In this case, the assessment report will clearly indicate the uncertainty of the results and recommend that maintenance personnel prioritize on-site verification; C pred =argmaxP(C k The model's final output is a comprehensive diagnostic report that includes qualitative conclusions, quantitative scores, and confidence levels, providing comprehensive and transparent information support for operation and maintenance decisions.
[0066] Compared with existing technologies, this invention has the following significant advantages: The transformer dual-branch state assessment method proposed in this invention integrates acoustic, optical, and electronic data, and achieves multimodal feature fusion through convolutional neural networks and attention mechanisms. It innovatively combines a dual-branch architecture of random forest qualitative classification and unsupervised autoencoder quantitative assessment. This method not only improves the state classification accuracy to 98.7%, but also achieves sensitive detection of early faults through a health scoring mechanism. The intelligent arbitration module controls the false alarm rate to within 1.5% through cross-validation mechanisms, significantly enhancing decision reliability and providing a complete solution for intelligent operation and maintenance of power equipment that combines accuracy, efficiency, and engineering practicality. Attached Figure Description
[0067] Figure 1 This is a diagram of the random forest algorithm structure in an embodiment of the present invention;
[0068] Figure 2 This is a diagram of the autoencoder network structure in an embodiment of the present invention;
[0069] Figure 3 This is a block diagram of the dual-branch intelligent arbitration model in an embodiment of the present invention;
[0070] Figure 4 The infrared image is shown in the embodiment of the present invention.
[0071] Figure 5 This is a voiceprint spectrum diagram in an embodiment of the present invention. Detailed Implementation
[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0073] Example 1, Branch 1: Transformer Condition Classification
[0074] 1. Data preprocessing, feature extraction and fusion
[0075] 1.1 Acquisition and Preprocessing of Acoustic, Optical, and Electrical Data
[0076] This invention uses multimodal monitoring data (acoustic, optical, and electrical) as input sources. In the data preprocessing stage, firstly, a dynamic time warping algorithm is used to solve the timing synchronization problem of multi-source signals, with the objective function being:
[0077] , (1)
[0078] Where i is an index variable, representing the position of a data point in the time series, with a value ranging from 1 to n; n is the length of the time series, i.e., the total number of data points; x i y represents the value of the first time series (such as an acoustic signal) at the i-th point; π(i) π represents the value of the i-th point in the second time series (such as an optical signal) mapped by π; π is the optimal path mapping function, which represents the best alignment between the two time series and is used to solve the timing synchronization problem of multi-source data.
[0079] Subsequently, the modal data were normalized to eliminate the influence of dimensions:
[0080] , (2)
[0081] Where λ is the original data value, representing the unprocessed input data (such as the original measurement values of sound, light, and electrical signals); μ is the mean of the data, used for centering; σ is the standard deviation of the data, used for scaling to eliminate the influence of dimensions; and z is the normalized data value, making the data conform to a standard normal distribution, which facilitates subsequent feature extraction.
[0082] In addition, affine transformations are performed on the infrared images to establish a mapping relationship between image pixel coordinates and the physical location of the transformer, ensuring the consistency of multi-source data in the spatiotemporal dimension and providing high-quality input data for subsequent feature extraction.
[0083] 1.2 Feature Extraction of Acoustics, Optics, and Light
[0084] End-to-end automatic feature extraction is achieved using convolutional neural networks. Dedicated CNN feature extractors are designed to address the characteristics of different modalities of data.
[0085] Acoustic signal feature extraction: The one-dimensional acoustic vibration signal is converted into a time-spectrum image through short-time Fourier transform, and a two-dimensional CNN is used for feature learning. Its convolution calculation can be expressed as:
[0086] (3)
[0087] Where l is the layer index of the convolutional neural network, indicating that the current layer is l; i, j are the spatial coordinates of the feature map, where i represents the row index and j represents the column index; m, n are the size indices of the convolutional kernel, where m ranges from 0 to M-1 and n ranges from 0 to N-1, representing the position of the convolutional kernel in height and width; M, N are the sizes of the convolutional kernel, where M represents the height and N represents the width; W (l) (m,n) represents the weight parameters of the l-th convolutional kernel, with values at position (m,n) used for feature extraction; F (l-1) b is the input feature map of the (l-1)th layer, and is the activation output of the previous layer; (l) σ is the bias term for the l-th layer, used to adjust the output; σ is the activation function, here we use the ReLU function to introduce nonlinearity; Let be the output value of the acoustic feature map of layer l at position (i,j);
[0088] Optical image feature extraction: Deep features are extracted from infrared thermal images using a pre-trained ResNet-18 network, and feature representation is achieved through transfer learning.
[0089] , (4)
[0090] Among them, I thermal The input is the infrared thermal image data; θ pretrained The parameter set for the pre-trained ResNet-18 network, including weights and biases, is initialized via transfer learning;ResNet This is a ResNet-18 network function used for feature extraction; F optical The output optical feature vector is a 128-dimensional depth feature.
[0091] Electrical signal feature extraction: Treating multi-channel electrical quantities as a one-dimensional time series, a one-dimensional CNN is used for time series feature extraction.
[0092] , (5)
[0093] in, τ represents the output value of the electrical feature of layer l at time t; l is the network layer index, indicating layer l; t is the time index, indicating the position in the one-dimensional time series; τ is the time offset index of the convolution kernel, from 0 to T-1; T is the time dimension size of the convolution kernel; W (l) (τ) represents the weight of the l-th layer one-dimensional convolutional kernel at offset τ; F (l-1) b is the input feature sequence of the (l-1)th layer; l σ is the bias term of the l-th layer; σ is the ReLU activation function;
[0094] The CNN network described above automatically learns 128-dimensional deep feature vectors from each modality of data, effectively capturing key feature patterns of device status.
[0095] 1.3 Feature Fusion of Acoustic-Optical-Electrical Feature Vectors
[0096] In the feature fusion stage, a feature weighted fusion method based on an attention mechanism is adopted. First, the attention weights of each modality feature are calculated:
[0097] , (6)
[0098] Where i and j are modal indices, i represents the current mode, and j represents the cyclic index of all modes; N is the total number of modes, here N=3 (acoustic, optical, and electronic); f i Let w be the feature vector of the i-th modality (e.g., 128-dimensional features); w is the attention weight vector, which is a trainable parameter; w T α is the transpose of w, used for dot product calculation; exp is the exponential function, used for softmax calculation; i The attention weight for the i-th modality is between 0 and 1, representing the importance of that modality.
[0099] Then, the final fused feature vector is generated by weighted summation:
[0100] , (7)
[0101] Where i is the modality index, from 1 to N; N is the total number of modes, N=3; α i The attention weights for the i-th modality are derived from formula (6); f i F is the eigenvector of the i-th mode; fusion The fused feature vector has a dimension of 256 and is used for subsequent state evaluation.
[0102] This method can adaptively highlight important features and suppress redundant information. The generated 256-dimensional fusion feature vector not only retains the unique information of each modality, but also reflects the complementarity between multiple source data, providing high-quality feature input for subsequent state assessment.
[0103] 2. Transformer Condition Classification (Qualitative Assessment)
[0104] The state classification branch performs the qualitative evaluation function of the model, aiming to determine the operating state level of the transformer. This branch uses the high-dimensional vector F output by the feature fusion module. fusion As input, the ensemble learning algorithm Random Forest is used for classification decisions.
[0105] Random forests construct multiple decision trees T i (i=1,2,...,N) and a bootstrap sampling method is used to improve the model's generalization ability. For the input feature F fusion Each decision tree independently gives its category judgment T. i (F fusion Finally, the classification result is generated through ensemble voting. The core output of this branch is the state level C to which the sample belongs. k Conditional probability of (normal / attention / abnormal):
[0106] The samples belong to each state level C. k The conditional probability P(C) k |F fusion Calculate using the following formula:
[0107] (8)
[0108] Where N is the total number of decision trees, and ∏(.) is the indicator function. This probability-based classification method not only provides state category judgments but, more importantly, reflects the confidence level of the judgment through probability values, providing a crucial decision-making basis for the subsequent arbitration module. This branch has the advantages of being insensitive to outliers and noise and not easily overfitting, ensuring the stability of qualitative evaluation results.
[0109] 3. State classification method: Random Forest Classifier
[0110] Random forest is an ensemble learning algorithm that performs classification tasks by constructing multiple decision trees and combining their predictions. Its core idea is that "collective wisdom is superior to a single expert," enhancing the model's generalization ability and robustness by introducing randomness. Random forest is based on the Bagging (Bootstrap Aggregating) ensemble strategy, combining the simplicity and intuitiveness of decision trees with the powerful generalization ability of ensemble learning. Figure 1 As shown, the specific steps are as follows.
[0111] 1) Bootstrap Sampling
[0112] Suppose the original training set D contains N samples, and the random forest generates K bootstrap sample sets through sampling with replacement:
[0113] (11)
[0114] Each D i The sample size is the same as the original training set D, approximately N. This process ensures that each bootstrap sample set contains approximately 63.2% of the original samples, with the remainder being duplicate samples.
[0115] 2) Random feature selection
[0116] When constructing each node of each decision tree, m features are randomly selected from all M features to form a subset of candidate features:
[0117] , (12)
[0118] Then, the optimal splitting feature is selected from these m features. This mechanism effectively increases the diversity of the base learner.
[0119] 3) Decision tree construction
[0120] Each decision tree is recursively constructed based on the CART algorithm, selecting the optimal split point to minimize Gini impurity. For node t, the Gini impurity is calculated using the following formula:
[0121] , (13)
[0122] Where P(i|t) is the number of nodes t belonging to the category The sample proportion is denoted by C, where C is the total number of categories. No pruning is performed during the decision tree growth process until the number of samples in a node falls below a preset threshold or the maximum depth is reached.
[0123] 4) Prediction mechanism
[0124] 1. Single tree prediction
[0125] For input samples Each decision tree independently provides a prediction result:
[0126] (14)
[0127] 2. Collective voting
[0128] Random forests aggregate the predictions from all decision trees using a majority voting method:
[0129] (15)
[0130] Where mode represents the mode operator.
[0131] 3. Probability Output
[0132] Random forests provide probability estimates of whether a sample belongs to each category:
[0133] , (16)
[0134] Here, ‖(.) is an indicator function that returns 1 if the condition is true and 0 otherwise. This probability estimate provides important confidence information for the subsequent arbitration module.
[0135] Example 2, Branch Two: Transformer Health Scoring
[0136] 2.1 Transformer Health Score (Quantitative Assessment)
[0137] The health scoring branch employs an unsupervised autoencoder model designed to quantify the health status of the device. The autoencoder consists of an encoder f... enc and decoder f dec The goal of this branch is to minimize the reconstruction error between the input and output. This branch first trains the autoencoder using a large amount of normal-state data to learn the data distribution of healthy states. For the fused feature vector F to be evaluated... fusion Its reconstruction error E recon Calculated using mean square error:
[0138] (9)
[0139] Subsequently, the reconstruction error is mapped to a continuous health score HS between 0 and 1 using an exponentially decaying function:
[0140] , (10)
[0141] Here, λ is a scaling parameter used to adjust the sensitivity of the score to error. When the equipment is in perfect health, the reconstruction error approaches 0, and the health score approaches 1; when the equipment is in an abnormal state, the reconstruction error increases, and the health score decreases accordingly.
[0142] 2.2 Health Assessment Method: Self-Encoder
[0143] An autoencoder is an artificial neural network based on unsupervised learning. Its core goal is to achieve feature extraction and dimensionality reduction by learning a compressed representation of data. For example... Figure 2 As shown, an autoencoder consists of two parts: an encoder and a decoder. It learns an effective representation of data by minimizing the reconstruction error between the input and the output.
[0144] The basic structure of an autoencoder can be formally represented as shown in the figure.
[0145] 1. Encoder: Maps input data x to a latent space representation z.
[0146] (17)
[0147] Where X is the input data, which refers to the fused feature vector; W is the weight matrix of the encoder; b is the bias vector of the encoder; θ is the set of parameters of the encoder, including W and b; σ is the activation function, such as Sigmoid or ReLU; z is the latent space representation, which is the compressed code of the input.
[0148] 2. Decoder: Reconstructs the original data from the latent representation z.
[0149] (18)
[0150] Where z is the latent representation from the encoder output; W' is the weight matrix of the decoder; b' is the bias vector of the decoder; Φ is the parameter set of the decoder, including W' and b'; and σ is the activation function. The reconstructed output data; 3. Training process and loss function
[0151] The training objective of an autoencoder is to minimize the reconstruction error, that is, to obtain the decoder. The output should be as close as possible to the original input x; the commonly used loss function is mean squared error (MSE):
[0152] (19)
[0153] In transformer condition assessment, the autoencoder is trained using a large number of normal state samples to learn the data distribution characteristics of healthy operating states. After training, the network will be able to accurately reconstruct the feature vectors of normal states, but will produce a large reconstruction error for the feature vectors of abnormal states.
[0154] Example 3: Dual-branch intelligent arbitration mechanism
[0155] Branch 1 and Branch 2 achieve comprehensive perception of equipment operating status from two dimensions: qualitative judgment and quantitative assessment, respectively. However, independent decision-making by the two branches may lead to inconsistent results, introducing uncertainty into the final judgment. To address this, a dual-branch intelligent arbitration mechanism is introduced. By cross-validating classification probabilities and health scores, a consistency judgment rule is established. This organically integrates the clarity of qualitative judgment with the sensitivity of quantitative assessment, improving the overall decision-making reliability and engineering practical value of the status assessment system. The model diagram is as follows: Figure 3 As shown.
[0156] This module receives the probability output from the state classification branch. The continuous scores (HS) of the health score branch are used, and cross-validation and comprehensive decision-making are performed according to the preset consistency rules.
[0157] The arbitration logic first determines whether the results of the two branches are consistent. The decision result for the state classification branch is set as follows: The corresponding confidence level is A consistency rule can be formally defined as: when "Normal" and HS , or when "Abnormal" and HS If the results are consistent, they are considered to be the same. Otherwise, they are considered to be contradictory. and Health score thresholds are set based on historical data or expert knowledge.
[0158] Consistent Results: When the results of both branches pass the consistency check, the system directly generates a result including the state level. A comprehensive assessment report including Health Score (HS) and high confidence markers.
[0159] Conflicting Results: When the results of both branches fail the consistency check (e.g., the status is judged as "normal" but the health score HS is extremely low), the system will automatically trigger the "manual review" flag. At this time, the assessment report will clearly indicate that the results are uncertain and recommend that operations and maintenance personnel intervene first to conduct on-site verification.
[0160] The model's final output is a comprehensive diagnostic report that includes qualitative conclusions, quantitative scores, and confidence levels, providing comprehensive and transparent information support for operational and maintenance decisions.
[0161] Example 4 Case Study Analysis
[0162] The experimental data for this invention comes from the Gansu Provincial Power Grid Transformer Multi-Source Monitoring Platform, focusing on the main transformer of the Cuijiaya 330kV traction substation in Lanzhou. As an important power hub in Northwest China, the operating status of this transformer is typical of the region.
[0163] Table 1 Electrical Data Table
[0164]
[0165] 4.1 Experimental Dataset and Preprocessing Procedure
[0166] The dataset contains 15,000 samples. Each sample simultaneously acquired acoustic vibration signals (sampling rate 50kHz), infrared thermograms (resolution 640×480), and three-phase electrical quantities (voltage and current, sampling rate 10kHz). The data was divided into training, validation, and test sets in a 7:2:1 ratio, ensuring a balanced distribution of samples across categories. Figure 4 and Figure 5 As shown.
[0167] The preprocessing stage employs a multi-step optimization: firstly, a dynamic time warping algorithm is applied to address the synchronization problem of multi-source, multi-frequency data; subsequently, Z-score normalization is performed to eliminate the influence of dimensions. For infrared images, an affine transformation is used to establish a spatial mapping relationship between the pixel coordinate system and the transformer's structural structure, ensuring the spatiotemporal consistency of temperature field data and electrical signals. The final generated feature vector has a dimension of 256, with acoustic features accounting for 35%, optical features for 40%, and electrical features for 25%.
[0168] 4.2 Experimental Parameter Configuration and Evaluation Indicators
[0169] The core parameters of the model were determined through grid search optimization: the number of decision trees in the random forest branch was set to 150, and the feature subset dimension was calculated as the square root of the total number of features; the autoencoder branch adopted a five-layer fully connected network with hidden layer dimensions of 128-64-32-64-128 respectively. Only 8,000 normal samples were used during training, and the scale parameter λ of the reconstruction error was determined to be 0.08 through cross-validation.
[0170] The evaluation system includes three types of indicators:
[0171] Classification performance metrics: accuracy, precision, recall, F1-score, and macro average.
[0172] Health scoring metrics: Area under the receiver operating characteristic (AUC), scoring sensitivity (early fault detection rate).
[0173] System performance metrics: number of model parameters, inference latency, and false alarm rate of the arbitration mechanism.
[0174] 4.3 Performance Comparison Analysis of State Classification
[0175] Comparative experiments with mainstream methods show that the proposed method performs exceptionally well in state classification tasks. The accuracy on the test set reaches 98.7%, a 6.4 percentage point improvement over traditional support vector machine methods and a 4.2 percentage point improvement over single-signal deep learning models. Particularly noteworthy is that for transitional states such as "attention," the proposed method achieves an F1-score of 96.8%, significantly outperforming the comparative methods (the highest being 89.3%), demonstrating that multimodal feature fusion can effectively identify boundary states.
[0176] The ensemble advantages of the Random Forest algorithm are fully demonstrated: the bootstrap sampling strategy enhances the model's robustness to noise and outliers; while the random feature selection mechanism avoids the risk of overfitting. The probability output module provides a confidence quantification for each classification result, achieving a classification accuracy of 99.5% when the confidence level is higher than 0.9.
[0177] 4.4 Sensitivity Validation of Health Scores
[0178] The health scoring branch exhibits excellent anomaly detection capabilities under unsupervised training conditions. Health scores for normal samples are concentrated in the 0.85-0.95 range (mean 0.91), while scores for abnormal samples significantly decrease to below 0.45. The scoring curve displays typical exponential decay characteristics, enabling the detection of potential faults, such as early mechanical defects like loose windings, 2-8 hours in advance.
[0179] Compared to traditional threshold-based alarm methods, continuous health scoring provides a more refined trajectory of state degradation. In simulated partial discharge experiments, the score value began to decline continuously 6 hours before the fault occurred, while traditional methods only generated an alarm 1 hour before the fault. The scoring system achieved an AUC value of 0.98, demonstrating its excellent class discrimination ability.
[0180] 4.5 Effectiveness Evaluation of the Two-Branch Arbitration Mechanism
[0181] The intelligent arbitration module effectively addresses the limitations of single-branch decision-making. Experimental data shows that the consistency rate of the two-branch results reaches 95.6%, and the system directly outputs a high-confidence diagnostic report. Of the remaining 4.4% of conflicting cases, on-site verification confirmed that: 2.1% were true boundary states, 1.8% were due to data acquisition noise, and 0.5% were model misjudgments.
[0182] The arbitration mechanism keeps the overall false alarm rate below 1.5%, a 52% reduction compared to a single classification branch. Particularly in low-confidence scenarios (P_max < 0.7), the system automatically triggers a manual review flag, avoiding three possible misjudgments and demonstrating the practical engineering value of human-machine collaboration. The arbitration logic also supports dynamic threshold adjustment, allowing parameters to adapt to the operational characteristics of different substations.
[0183] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electronic data, characterized in that, Includes the following steps: S1. Multimodal data synchronization and preprocessing: Acoustic, optical and electrical monitoring data of transformers are collected, time-series synchronization and normalization preprocessing are performed, and spatiotemporal consistency of multi-source data is established; S2. Feature extraction and fusion: Features are automatically extracted from the acoustic, optical and electrical data using a convolutional neural network, and a fused feature vector is generated through a weighted fusion module based on an attention mechanism; S3. Two-branch state evaluation: S31. Qualitative evaluation branch: Input the fused feature vector into a trained random forest classifier, and output the qualitative classification result of the transformer state and its corresponding classification confidence; S32. Quantitative evaluation branch: Input the fused feature vector into an unsupervised autoencoder trained only on normal state samples, calculate its reconstruction error, and generate a continuous health score based on the reconstruction error through a mapping function; S4. Intelligent Arbitration and Output: Based on the qualitative classification results, classification confidence, and health score, cross-validation is performed through preset arbitration logic; When the results of the two branches are consistent, the final state evaluation conclusion is output; when the results are contradictory, the manual review flag is triggered.
2. The method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electrical data according to claim 1, characterized in that, In S1, firstly, the timing synchronization problem of multi-source signals is solved using a dynamic time warping algorithm, with the objective function being: ， （1） Where i is an index variable, representing the position of a data point in the time series, with a value ranging from 1 to n; n is the length of the time series, i.e., the total number of data points; x i y represents the value of the first time series (such as an acoustic signal) at the i-th point; π(i) π represents the value of the i-th point in the second time series (such as an optical signal) mapped by π; π is the optimal path mapping function, which represents the best alignment between the two time series and is used to solve the timing synchronization problem of multi-source data. Subsequently, the modal data were normalized to eliminate the influence of dimensions: ， （2） Where λ is the original data value, representing the unprocessed input data (such as the original measurement values of sound, light, and electrical signals); μ is the mean of the data, used for centering; σ is the standard deviation of the data, used for scaling to eliminate the influence of dimensions; and z is the normalized data value, making the data conform to a standard normal distribution, which facilitates subsequent feature extraction. In addition, affine transformations are performed on the infrared images to establish a mapping relationship between image pixel coordinates and the physical location of the transformer, ensuring the consistency of multi-source data in the spatiotemporal dimension and providing high-quality input data for subsequent feature extraction.
3. The method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electrical data according to claim 2, characterized in that, In S2, a convolutional neural network is used to achieve automatic end-to-end feature extraction. A dedicated CNN feature extractor is designed to address the characteristics of different modalities of data. Acoustic signal feature extraction: The one-dimensional acoustic vibration signal is converted into a time-spectrum graph through short-time Fourier transform, and a two-dimensional CNN is used for feature learning; Its convolution calculation can be expressed as: , （3） Where l is the layer index of the convolutional neural network, indicating that the current layer is l; i, j are the spatial coordinates of the feature map, where i represents the row index and j represents the column index; m, n are the size indices of the convolutional kernel, where m ranges from 0 to M-1 and n ranges from 0 to N-1, representing the position of the convolutional kernel in height and width; M, N are the sizes of the convolutional kernel, where M represents the height and N represents the width; W (l) (m,n) represents the weight parameters of the l-th convolutional kernel, with values at position (m,n) used for feature extraction; F (l-1) b is the input feature map of the (l-1)th layer, and is the activation output of the previous layer; (l) σ is the bias term for the l-th layer, used to adjust the output; σ is the activation function, here we use the ReLU function to introduce nonlinearity; Let be the output value of the acoustic feature map of layer l at position (i,j); Optical image feature extraction: Deep features are extracted from infrared thermal images using a pre-trained ResNet-18 network, and feature representation is achieved through transfer learning. ， （4） Among them, I thermal The input is the infrared thermal image data; θ pretrained The parameter set for the pre-trained ResNet-18 network, including weights and biases, is initialized via transfer learning; ResNet This is a ResNet-18 network function used for feature extraction; F optical The output optical feature vector is a 128-dimensional depth feature. Electrical signal feature extraction: Treating multi-channel electrical quantities as a one-dimensional time series, a one-dimensional CNN is used for time series feature extraction. ， （5） in, τ represents the output value of the electrical feature of layer l at time t; l is the network layer index, indicating layer l; t is the time index, indicating the position in the one-dimensional time series; τ is the time offset index of the convolution kernel, from 0 to T-1; T is the time dimension size of the convolution kernel; W (l) (τ) represents the weight of the l-th layer one-dimensional convolutional kernel at offset τ; F (l-1) b is the input feature sequence of the (l-1)th layer; l σ is the bias term of the l-th layer; σ is the ReLU activation function; The CNN network described above automatically learns 128-dimensional deep feature vectors from each modality of data, effectively capturing key feature patterns of device status.
4. The method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electronic data according to claim 3, characterized in that, In S2, during the feature fusion stage, a feature weighted fusion method based on an attention mechanism is adopted; First, calculate the attention weights for each modality feature: ， （6） Where i and j are modal indices, i represents the current mode, and j represents the cyclic index of all modes; N is the total number of modes, here N=3 (acoustic, optical, and electronic); f i Let w be the feature vector of the i-th modality (e.g., 128-dimensional features); w is the attention weight vector, which is a trainable parameter; w T α is the transpose of w, used for dot product calculation; exp is the exponential function, used for softmax calculation; i The attention weight for the i-th modality is between 0 and 1, representing the importance of that modality. Then, the final fused feature vector is generated by weighted summation: ， （7） Where i is the modality index, from 1 to N; N is the total number of modes, N=3; α i The attention weights for the i-th modality are derived from formula (6); f i F is the eigenvector of the i-th mode; fusion The resulting fused feature vector has a dimension of 256 and is used for subsequent state evaluation. This method can adaptively highlight important features and suppress redundant information. The generated 256-dimensional fused feature vector not only retains the unique information of each modality, but also reflects the complementarity between multiple source data, providing high-quality feature input for subsequent state evaluation.
5. The method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electrical data according to claim 4, characterized in that, In S31, the high-dimensional vector F output by the feature fusion module is obtained. fusion As input to the state classification branch; multiple decision trees T are constructed using the random forest algorithm. i Where i = 1, 2, 3, ..., N, N is the total number of decision trees. Each decision tree is independently trained using a bootstrap sampling method to improve the model's generalization ability; the high-dimensional vector F... fusion The input is fed into the trained random forest classifier, so that each decision tree T... i Independently output its category judgment T i (F) fusion The category judgment corresponds to the transformer's state level C. k The state C k The classification levels include at least normal, attentive, and abnormal. Based on the outputs of all decision trees, the final classification result is generated through ensemble voting, where the sample belongs to each state level C. k The conditional probability P(C) k |F fusion Calculate using the following formula: , (8) where N is the total number of decision trees, and ∏(.) is the indicator function; the conditional probability is output as the core output of the state classification branch to provide state category judgment and its corresponding confidence level, and to provide decision basis for subsequent arbitration module.
6. The method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electrical data according to claim 5, characterized in that, In S32, the autoencoder is trained using a large amount of normal state data, wherein the autoencoder includes encoder f. enc and decoder f dec The training objective is to minimize the reconstruction error between the input and output, so that the autoencoder can learn the data distribution of the health state; for the fused feature vector F to be evaluated fusion The reconstruction error E is calculated using the autoencoder. recon The calculation formula is: ，（9） The reconstruction error E is reduced by an exponential decay function. recon The Health Score (HS) is mapped to a continuous range of 0 to 1 and is calculated using the following formula: (10) where λ is a scale parameter used to adjust the sensitivity of the score to error; When the equipment is in perfect health, the reconstruction error approaches 0 and the health score HS approaches 1; when the equipment is in abnormal condition, the reconstruction error increases and the health score HS decreases accordingly.
7. The method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electrical data according to claim 6, characterized in that, In S31, the random forest classifier includes a bootstrap sampling module, a random feature selection module, a decision tree construction module, and a prediction module; wherein: the bootstrap sampling module is used to generate K bootstrap sample sets D from the original training set D containing N samples through sampling with replacement. i , ， （11） Where i = 1, 2, 3, ..., K, each bootstrap sample set D i The sample size is the same as the original training set D, and it contains approximately 63.2% of the independent samples in the original training set; the random feature selection module is used to randomly select m features from all M features to form a candidate feature subset when constructing each node of the decision tree, wherein: (12) The decision tree construction module is used to recursively construct a decision tree based on the CART algorithm, and select the best split point by minimizing the Gini impurity. For node t, the formula for calculating the Gini impurity Gini(t) is: , (13) Where P(i|t) is the proportion of samples belonging to class i in node t, and C is the total number of classes; no pruning is performed during the growth of the decision tree until the number of samples in a node is less than a preset threshold or the maximum depth is reached; the prediction module is used to perform classification prediction on the input sample x based on the constructed K decision trees; the prediction module includes a single tree prediction unit, a collective voting unit, and a probability output; wherein: The single-tree prediction unit is used to predict each decision tree h. i Provide prediction results independently for each input sample x: ， （14） Where i = 1, 2, 3, ..., K; The collective voting unit is used to aggregate the prediction results of all decision trees through majority voting to obtain the final classification result of the random forest: ， （15） Where mode represents the mode operator; The probability output provides the random forest with an estimate of the probability that a sample belongs to each category: ， （16） Here, ‖ (.) is an indicator function that returns 1 when the condition is true and 0 otherwise. This probability estimate provides important confidence information for the subsequent arbitration module.
8. The method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electronic data according to claim 7, characterized in that, In S32, the autoencoder is an artificial neural network based on unsupervised learning. Its core objective is to achieve feature extraction and dimensionality reduction by learning the compressed representation of data. The autoencoder consists of two parts: an encoder and a decoder. It learns an effective representation of data by minimizing the reconstruction error between the input and the output. The encoder maps the input data x to a latent space representation z: ， （17） Where X is the input data, which refers to the fused feature vector; W is the weight matrix of the encoder; b is the bias vector of the encoder; θ is the set of parameters of the encoder, including W and b; σ is the activation function, such as Sigmoid or ReLU; z is the latent space representation, which is the compressed code of the input. The decoder reconstructs the original data from the latent representation z: ， （18） Where z is the latent representation from the encoder output; W' is the weight matrix of the decoder; b' is the bias vector of the decoder; Φ is the parameter set of the decoder, including W' and b'; and σ is the activation function. The output data is the reconstructed data; the training objective of the autoencoder is to minimize the reconstruction error, that is, to obtain the decoder. The output should be as close as possible to the original input x; the commonly used loss function is mean squared error (MSE): ， （19） In transformer condition assessment, the autoencoder is trained using a large number of normal state samples to learn the data distribution characteristics of healthy operating conditions. After training, the network will be able to accurately reconstruct the feature vector of normal states, but will produce a large reconstruction error for the feature vector of abnormal states.
9. The method for evaluating the dual-branch state of a transformer by integrating acoustic, optical, and electrical data according to claim 8, characterized in that, In S4, the probability output P(C) from the state classification branch is received. k The continuous scores (HS) of the health score branch are used, and cross-validation and comprehensive decision-making are performed according to the preset consistency rules; the arbitration logic is as follows: First, determine if the results of the two branches are consistent: Let the determination result of the state classification branch be C. pred =argmaxP(C k The corresponding confidence level is P. max Consistency rules can be formally defined as: when C pred "Normal" and HS≥θ hight , or when C pred When it is "abnormal" and HS≥θ hight If the results are inconsistent, they are considered to be consistent; otherwise, they are considered to be contradictory. Where, θ hight and θ low Health score thresholds set based on historical data or expert knowledge; If the results are consistent, the system directly generates a status level C. pred A comprehensive assessment report including Health Score (HS) and high confidence indicators; if the results are contradictory, the system will automatically trigger the "manual review" indicator. In this case, the assessment report will clearly indicate the uncertainty of the results and recommend that maintenance personnel prioritize on-site verification; C pred =argmaxP(C k The model's final output is a comprehensive diagnostic report that includes qualitative conclusions, quantitative scores, and confidence levels, providing comprehensive and transparent information support for operation and maintenance decisions.