Transformer state prediction method based on multivariate sensing and feature extraction

Abstract

A transformer state prediction method based on multivariate sensing and feature extraction is disclosed. In this method, vibration signals are collected from the surface of the transformer tank under normal and fault conditions. The vibration signals are preprocessed and visualized as grayscale feature images. A portion of the grayscale feature images are stored as a fault feature image library, and the remaining grayscale feature images are used as a model dataset. Based on the model dataset, an algorithm combining convolutional neural networks and long short-term memory networks is used to extract features. The original grayscale feature images are encoded and decoded using a 3D convolutional neural network. The feature images of the vibration signals under fault conditions are used as a benchmark. The corresponding decoder outputs the feature images from all time steps, and the spatial similarity between the decoder and the fault feature image library is calculated. The similarity index (IoU) is used to classify the features. When the similarity is greater than a predetermined threshold, the two are considered to belong to the same feature state, and the label of the feature state in the fault feature image library is output.

Description

Technical Field

[0001] This invention belongs to the field of transformer monitoring technology, specifically a transformer state prediction method based on multivariate sensing and feature extraction. Background Technology

[0002] Power transformers are crucial power equipment in power grids, enabling long-distance transmission and distribution of electricity. Their stability is vital to grid security. The faults and operating status of power transformers are affected by multiple factors, including natural environmental factors, human factors, and defects in the transformer itself and its associated equipment. Transformer vibration is caused by the vibration of the transformer body (core, windings, etc.) and the cooling system. Core vibration is primarily caused by the magnetostriction of silicon steel sheets. Winding vibration is caused by the leakage magnetic field generated by the load current flowing through the coils, resulting in dynamic electromagnetic forces between windings, coils, and turns. Long-term vibration effects lead to mechanical defects such as deformation of the transformer's internal structure. These changes are irreversible, and the vibration pattern changes accordingly. Therefore, transformer vibration signals contain rich information about mechanical faults. Current methods for identifying the mechanical condition of transformer windings include dissolved gas analysis and fault sound analysis. While these electrical quantity-based methods are relatively mature, they require the transformer to be taken out of service, failing to meet the needs of online monitoring and making it difficult to quantitatively analyze fault symptoms. Fault sound analysis methods include ultrasonic detection, noise detection, and vibration detection. These condition diagnosis methods based on non-electrical quantities are more flexible and can monitor signal data without affecting transformer operation. However, this technology currently lacks theoretical research, has limited data types, and incomplete condition fault data, making practical field applications quite difficult.

[0003] The information disclosed in the background section is only intended to enhance the understanding of the background of the present invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a transformer state prediction method based on multivariate sensing and feature extraction, which can monitor signal data without affecting the operation of the transformer and has high monitoring accuracy.

[0005] The objective of this invention is achieved through the following technical solution: a transformer state prediction method based on multivariate sensing and feature extraction, comprising,

[0006] Vibration signals were collected from the surface of the transformer tank under both normal and fault conditions.

[0007] The vibration signal is preprocessed and visualized as a grayscale feature image. A portion of the grayscale feature images are stored as a fault feature image library, and the remaining grayscale feature images are used as a model dataset. In the preprocessing, the vibration signal is read, the feature values ​​are calculated, and the result is saved as an feature value matrix. Z-score normalization is performed based on the feature value matrix, as shown in formula (1):

[0008]

[0009] Where T′ is the feature value, μ is the sample mean, σ is the sample standard deviation, and T is the feature value after standardization, which follows a normal distribution with a mean of 0 and a standard deviation of 1. Then, the gray-scale feature image M is constructed using the ratio of the extreme values ​​of each channel element and the ratio of the absolute value of the adjacent difference to the extreme value, formula (2):

[0010]

[0011] Where T is the matrix form of the standardized eigenvalues, with a matrix size of L*1, where L is the number of eigenvalues ​​contained in an image, and i, j, a, b range from 0 to L. i Let |T| be the value of the i-th row of the sequence. max T is the absolute value of the maximum value. i -T j |T is the difference between adjacent elements in the sequence. a -T b | max The maximum value of the difference between any two elements in the sequence is the difference between the maximum and minimum values. The resulting grayscale image matrix M has a size of L*L and an eigenvalue range of [0, 1].

[0012] Based on the model dataset, the algorithm combines convolutional neural networks and long short-term memory networks to extract features, and uses 3D convolutional neural networks to encode and decode the original grayscale feature images. In the encoder part, attention modules and dynamic semantic vector sequences are used to replace fixed semantic vectors as input information encoding, and a subset of vectors is automatically selected during the decoder decoding process. The feature image output by the decoder is quantitatively compared with the grayscale feature image using peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).

[0013] Using the feature image of the vibration signal under fault conditions as a benchmark, the feature image output by the matching decoder is selected from all time steps, and its spatial similarity with the fault feature library is calculated. The similarity IoU index is used to classify the features. When it is greater than a predetermined threshold, it is considered that the two belong to the same feature state, and the label of the feature state in the fault feature library is output.

[0014] In the transformer state prediction method based on multivariate perception and feature extraction, the load current of the transformer vibration and noise online detection system is adjusted under normal and fault conditions, starting with a load current of 4A and increasing to 8A in step size of 0.8A. Vibration signals are collected from the surface of the tank under both normal and fault conditions to serve as samples for state prediction.

[0015] In the transformer state prediction method based on multivariate perception and feature extraction, the fault states include core and winding clamp loosening faults, winding warping faults, winding misalignment faults, winding bulging faults, and partial discharge faults.

[0016] In the transformer state prediction method based on multivariate perception and feature extraction, a winding loosening fault is set for the transformer. A torque wrench is used to loosen the screw fixing the winding by three stops, namely 4 N·m, 8 N·m and 12 N·m, and the time-domain vibration signal is obtained as the vibration signal under the winding clamp loosening fault.

[0017] In the transformer state prediction method based on multivariate perception and feature extraction, a winding warping fault is set for the transformer. By applying pressure to the tap during the winding process, the winding is deformed, and the time-domain vibration signal is obtained as the vibration signal under the winding warping fault.

[0018] In the transformer state prediction method based on multivariate sensing and feature extraction, a winding misalignment fault is set for the transformer, and the time-domain vibration signal under rated current is obtained as the vibration signal under the winding misalignment fault.

[0019] In the transformer state prediction method based on multivariate perception and feature extraction, a winding bulge fault is set for the transformer. A medium is added between the high-voltage winding and the low-voltage winding to make the high-voltage winding bulge outward and the low-voltage winding bulge inward. At the same time, the high-voltage winding generates axial displacement, and the time-domain vibration signal is obtained as the vibration signal under the winding bulge fault.

[0020] In the transformer state prediction method based on multivariate perception and feature extraction, a partial discharge fault is set for the transformer. Metal spikes are set on the high-voltage output side of the dry-type transformer to simulate a partial discharge fault in the dry-type transformer. The dry-type transformer is then stepped up, and the time-domain vibration signal between the metal spikes and the metal clamps from partial discharge to breakdown is measured in real time as the vibration signal under the partial discharge fault.

[0021] In the transformer state prediction method based on multivariate perception and feature extraction, the feature values ​​include fundamental frequency amplitude, fundamental frequency weight, spectral energy, and spectral complexity.

[0022] In the transformer state prediction method based on multivariate perception and feature extraction, the formula (3) for classifying features using the similarity IoU index is:

[0023]

[0024] Where intersection is the set of elements in two images, union is the set of elements in two images, and IoU is the similarity.

[0025] Compared with existing technologies, this invention has the following advantages: The transformer state prediction method based on multivariate perception and feature extraction described in this invention reads multiple sets of different feature values ​​from the dataset and performs vectorization processing to generate multidimensional data. This data is visualized as a grayscale feature image and fed into a convolutional neural network to extract features. By combining the convolutional neural network with a long short-term memory network, the hidden states of the recurrent network are extended to three dimensions, and all matrix multiplication operations of the traditional LSTM are replaced with convolutions. This allows spatial features to flow between nodes of the recurrent network in the form of three-dimensional tensors. This multivariate perception gives ConvLSTM a strong predictive ability. The predicted feature images are then compared with a fault feature image library for classification, and the model outputs an expectation of the future state of the transformer. The fault prediction process in this invention is convenient and accurate. On the one hand, it uses a feature extraction method—multi-set feature value imaging—to comprehensively depict the state of transformer equipment from an image perspective, making up for the limitation of past methods that only analyzed a single feature value. Moreover, the images output by the model are highly intuitive. On the other hand, it processes image features through a combination of two neural networks, thereby enabling the analysis of a large number and variety of feature values ​​that reflect the state of transformer equipment, providing a reference for transformer fault prediction and in-depth data mining of main power grid equipment. Attached Figure Description

[0026] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0027] In the attached diagram:

[0028] Figure 1 This is a schematic diagram of a transformer vibration signal acquisition system based on a transformer state prediction method using multivariate sensing and feature extraction according to an embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram of the grayscale feature image of the features to be extracted in a transformer state prediction method based on multivariate perception and feature extraction according to an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of an Encoder-Decoder model based on the Attention mechanism for a transformer state prediction method based on multivariate perception and feature extraction according to an embodiment of the present invention.

[0031] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0032] Specific embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0033] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0034] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0035] To better understand, such as Figures 1 to 3 As shown, the transformer state prediction method based on multivariate sensing and feature extraction includes:

[0036] Vibration signals were collected from the surface of the transformer tank under both normal and fault conditions.

[0037] The vibration signal is preprocessed and visualized as a grayscale feature image. A portion of the grayscale feature images are stored as a fault feature image library, and the remaining grayscale feature images are used as a model dataset. In the preprocessing, the vibration signal is read, the feature values ​​are calculated, and the result is saved as an feature value matrix. Z-score normalization is performed based on the feature value matrix, as shown in formula (1):

[0038]

[0039] Where T′ is the feature value, μ is the sample mean, σ is the sample standard deviation, and T is the feature value after standardization, which follows a normal distribution with a mean of 0 and a standard deviation of 1. Then, the gray-scale feature image M is constructed using the ratio of the extreme values ​​of each channel element and the ratio of the absolute value of the adjacent difference to the extreme value, formula (2):

[0040]

[0041] Where T is the matrix form of the standardized eigenvalues, i.e., the complete feature sequence, with a matrix size of L*1, where L is the number of eigenvalues ​​contained in an image, and i, j, a, b range from 0 to L. i Let |T| be the value of the i-th row of the sequence. max T is the absolute value of the maximum value. i -T j |T is the difference between adjacent elements in the sequence. a -T b | max The maximum value of the difference between any two elements in the sequence is the difference between the maximum and minimum values. The resulting grayscale image matrix M has a size of L*L and an eigenvalue range of [0, 1].

[0042] Based on the model dataset, the algorithm combines convolutional neural networks and long short-term memory networks to extract features, and uses 3D convolutional neural networks to encode and decode the original grayscale feature images. In the encoder part, attention modules and dynamic semantic vector sequences are used to replace fixed semantic vectors as input information encoding, and a subset of vectors is automatically selected during the decoder decoding process. The feature image output by the decoder is quantitatively compared with the grayscale feature image using peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).

[0043] Using the feature image of the vibration signal under fault conditions as a benchmark, the feature image output by the matching decoder is selected from all time steps, and its spatial similarity with the fault feature library is calculated. The similarity IoU index is used to classify the features. When it is greater than a predetermined threshold, it is considered that the two belong to the same feature state, and the label of the feature state in the fault feature library is output.

[0044] By combining Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, ConvLSTM not only retains the advantages of Full-Connection LSTMs (FC-LSTMs), possessing the temporal modeling capabilities of LSTMs, but also, due to its inherent convolutional structure, can characterize local features, making it highly suitable for predicting spatiotemporal data. By introducing the ConvLSTM model into the encoding prediction structure, a transformer state prediction model was established. This algorithm has no fixed requirements on the amount of data, high learning freedom, good robustness, fast training speed, and high accuracy.

[0045] In a preferred embodiment of the transformer state prediction method based on multivariate perception and feature extraction, the load current of the transformer vibration and noise online detection system is adjusted under normal and fault conditions, starting with a load current of 4A and increasing to 8A in step size of 0.8A. Vibration signals are collected from the surface of the tank under both normal and fault conditions to serve as samples for state prediction.

[0046] In a preferred embodiment of the transformer state prediction method based on multivariate perception and feature extraction, the fault states include core and winding clamp loosening faults, winding warping faults, winding misalignment faults, winding bulging faults, and partial discharge faults.

[0047] In a preferred embodiment of the transformer state prediction method based on multivariate perception and feature extraction, a winding loosening fault is set for the transformer. A torque wrench is used to loosen the screw fixing the winding by three stops, namely 4 N·m, 8 N·m and 12 N·m, and the time-domain vibration signal is obtained as the vibration signal under the winding clamp loosening fault.

[0048] In a preferred embodiment of the transformer state prediction method based on multivariate perception and feature extraction, a winding warping fault is set for the transformer. By applying pressure to the tap during the winding process, the winding is deformed, and the time-domain vibration signal is obtained as the vibration signal under the winding warping fault.

[0049] In a preferred embodiment of the transformer state prediction method based on multivariate sensing and feature extraction, a winding misalignment fault is set for the transformer, and the time-domain vibration signal under rated current is obtained as the vibration signal under the winding misalignment fault.

[0050] In a preferred embodiment of the transformer state prediction method based on multivariate perception and feature extraction, a winding bulge fault is set for the transformer. A medium is added between the high-voltage winding and the low-voltage winding to make the high-voltage winding bulge outward and the low-voltage winding bulge inward. At the same time, the high-voltage winding generates axial displacement, and the time-domain vibration signal is obtained as the vibration signal under the winding bulge fault.

[0051] In a preferred embodiment of the transformer state prediction method based on multivariate perception and feature extraction, a partial discharge fault is set up on the transformer. Metal spikes are set on the high-voltage output side of the dry-type transformer to simulate a partial discharge fault in the dry-type transformer. The dry-type transformer is then stepped up, and the time-domain vibration signal between the metal spikes and the metal clamps from partial discharge to breakdown is measured in real time as the vibration signal under the partial discharge fault.

[0052] In a preferred embodiment of the transformer state prediction method based on multivariate perception and feature extraction, the feature values ​​include fundamental frequency amplitude, fundamental frequency weight, spectral energy, and spectral complexity.

[0053] In a preferred embodiment of the transformer state prediction method based on multivariate perception and feature extraction, the formula (3) for classifying features using the similarity IoU index is:

[0054]

[0055] Where intersection is the set of elements in two images, union is the set of elements in two images, and IoU is the similarity.

[0056] In one embodiment, a transformer vibration and noise online detection system is used to collect signals from the surface of the transformer tank under both normal and fault conditions. The collected typical fault and normal vibration signals are preprocessed to obtain multiple sets of feature value data, which are then visualized as grayscale feature texture images. A portion of these images is stored in a typical fault identification library, while the majority is used as the model dataset. The ratio of data used for training, validation, and testing is 7:2:1. After the model has been trained, validated, and tested using the dataset from the first 16 days, the feature images predicted by the model are compared with the image library to output a prediction of the transformer's condition on the 17th day. This prediction allows power engineers to take appropriate maintenance measures in advance.

[0057] A complete feature sequence contains 500 sequence feature images. In every 17 feature sequence images, the first 16 images are used as the input feature sequence length, and the last image is the output sequence. The stride is set to 8. All training data is labeled, so the sequence feature images used for the first input sequence in model training are {x0, x1, ..., x...}. 15 The corresponding label for the output feature image sequence is {x}. 16}, the second input sequence feature image {x9, x 10 , ..., x 24 The corresponding label for the output feature image sequence is {x}. 25}, and so on.

[0058] Signal preprocessing involves first reading the original vibration signal and then calculating and saving its eigenvalues. Among these, the eigenvalues ​​selected are those that best characterize the mechanical state of the transformer, such as fundamental frequency amplitude, fundamental frequency weight, spectral energy, and spectral complexity.

[0059] The eigenvalue matrix is ​​standardized by Z-Score. Compared with the min-max normalization method, this method can not only remove the dimensions, but also treat all dimensions of variables equally. Because each dimension follows a normal distribution with a mean of 0 and a variance of 1, the data of each dimension plays the same role when calculating the difference. This avoids the huge impact of different dimensions on the distance calculation. The specific formula is as follows (1):

[0060]

[0061] Where T′ is the original data, μ is the sample mean, σ is the sample standard deviation, and T is the standardized feature value, which follows a normal distribution with a mean of 0 and a standard deviation of 1. Then, the feature image M is constructed using the ratio of the extreme values ​​of each channel element, the absolute value of the adjacent difference, and the ratio of the extreme values, as shown in (2):

[0062]

[0063] Here, T represents the complete feature sequence, and the matrix size is L*1, where L is the number of feature values ​​contained in an image. The element values ​​of the resulting image matrix M are all in the range of [0, 1], thus ensuring that the image retains as much transformer feature information as possible, while also accelerating the convergence speed of the neural network and avoiding overfitting.

[0064] In one embodiment, the image size per frame is adjusted to 128*128 pixels. The prediction method steps include:

[0065] 1) Using an online transformer vibration and noise detection system, signals are collected from the surface of the transformer tank under both normal and fault conditions. The acquired signals can reflect the internal mechanical state of the transformer. The signal acquisition device includes... Figure 1 As shown, the online transformer vibration and noise detection system includes a vibration acceleration sensor and an acoustic sensor, which are connected to the monitoring device and transmit data to the monitoring terminal via remote data transmission.

[0066] 2) Adjust the load current under normal and winding fault conditions respectively, starting from 4A and in 0.8A increments, and adjust it to 8A. Under these conditions, collect vibration signals on the surface of the oil tank and use them as samples for condition prediction.

[0067] 3) Set up faults such as loose windings, warping, misalignment, and bulging on the transformer, and collect vibration signals on the surface of the oil tank under these conditions respectively;

[0068] 4) Metal spikes were installed on the high-voltage output side of the transformer to simulate a partial discharge fault. The transformer was stepped up, and the vibration signal between the metal spikes and metal clamps from partial discharge to breakdown was measured in real time.

[0069] 5) Preprocess the collected typical transformer fault and normal vibration signals, and then visualize them as grayscale feature images, such as... Figure 2 As shown.

[0070] 6) Features are extracted using an algorithm combining convolutional neural networks and long short-term memory networks. The original grayscale feature image is then encoded and decoded using a combination of multivariate perceptron-based convolutional neural networks and a special long short-term memory network. The Encoder-Decoder model diagram is shown below. Figure 3 As shown.

[0071] 7) Quantitatively compare the decoded feature image with the actual feature image, and use the evaluation metrics on this dataset to quantitatively evaluate the performance of this model training: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM).

[0072] 8) Using typical fault feature images derived from a large amount of raw data as a benchmark, select matching decoded feature images from all time steps and calculate their spatial similarity with the fault feature library. Use the similarity IoU index to classify the features. The calculation formula is as shown in (3). The calculation result can indicate which state the sample belongs to.

[0073]

[0074] Where intersection is the set of elements in two images, and union is the set of elements in two images. A threshold is set, and when the value is greater than the threshold, the two images are considered to belong to the same feature state. The model will output a reference judgment on the transformer state on day 17 at each time step: normal operation, loose core and winding clamps, winding misalignment, winding warping, winding bulging, and partial discharge.

[0075] In one embodiment, an online transformer vibration and noise detection system is used to collect raw vibration signals from the surface of the transformer tank under normal and typical fault conditions. The preprocessed signals are converted into grayscale feature images. A ConvLSTM algorithm based on multivariate perception is used to extract texture features from the images. The mean squared error loss value and structural similarity are used to evaluate the judgment results. A similarity index is then used to classify the feature images. The model outputs the classification results, thereby achieving the effect of predicting the internal mechanical state of the transformer and thus achieving the purpose of fault prediction. The specific implementation steps are as follows:

[0076] S1: Under normal internal mechanical conditions of the transformer, adjust the load current in steps of 0.8A to obtain the time-domain vibration signal of the transformer at load currents of 4A, 4.8A, 5.6A, 6.4A, 7.2A and 8A respectively. The horizontal distance between the microphone and the surface of the transformer tank is 1m.

[0077] S2: Set a winding loosening fault for the transformer, and use a torque wrench to loosen the screw fixing the winding in three positions, namely 4N·m, 8N·m and 12N·m, and obtain the time-domain vibration signal corresponding to the state.

[0078] S3: Set a winding warping fault for the transformer. Apply pressure to the tap during the winding process to deform the winding and obtain the time-domain vibration signal corresponding to the state.

[0079] S4: Set a winding misalignment fault for the transformer and obtain the time-domain vibration signal of this state under rated current;

[0080] S5: To set up a winding bulge fault in the transformer, a medium is added between the high-voltage and low-voltage windings to make the high-voltage winding bulge outward and the low-voltage winding dent inward. At the same time, the high-voltage winding is made to generate a certain axial displacement, and the time-domain vibration signal corresponding to the state is obtained.

[0081] S6: A partial discharge fault was simulated on the transformer. Metal spikes were placed on the high-voltage output side of the dry-type transformer to simulate a partial discharge fault. The transformer was stepped up, and the time-domain vibration signal between the metal spikes and metal clamps from partial discharge to breakdown was measured in real time.

[0082] S7: The collected typical transformer fault and normal vibration signals were preprocessed using MATLAB to obtain multiple sets of feature value data, which were then visualized as grayscale feature images. The specific visualization method is referred to (1) and (2). The matrix dimension of the complete feature sequence image is 48*48. Since the time interval of the original time domain data is half an hour, the number of feature values ​​L contained in one image is 48 to ensure that the generated image contains features within one day. The resolution of the grayscale feature map is changed to 128*128. A portion is stored as a typical fault identification library for feature comparison with the subsequent model prediction results. Most of the remaining feature images are stored in the training, validation, and testing folders as model datasets, which are used to generate file paths for training, validation, and testing models.

[0083] S8: Feature extraction is performed using an algorithm combining convolutional neural networks and long short-term memory networks, and a 3D convolutional neural network is used to encode and decode the original grayscale feature image. In the encoder part, an attention module is employed, where a dynamic semantic vector sequence replaces the fixed semantic vector as input information for encoding, and a subset of vectors is automatically selected during the decoder's decoding process. This approach assigns different weight coefficients based on the varying contributions of the input sequence data to the output prediction sequence; this is the main idea behind the Attention mechanism.

[0084] S9: The decoded feature image is quantitatively compared with the actual feature image using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). Similar to MSE, PSNR evaluates the pixel difference between two frames. Higher PSNR and SSIM indicate better prediction performance of the model and more accurate results compared with the actual feature image.

[0085] S10: Using typical fault feature images derived from a large amount of raw data as a benchmark, matching decoded feature images are selected from all time steps, and their spatial similarity with the fault feature image library is calculated. The similarity IoU index is used to classify the features, and the calculation result can indicate which state the sample belongs to. The threshold is set to 0.5. When it is greater than this threshold, it is considered that the two belong to the same feature state, and the model will output the label of the state in the image library, that is, the reference judgment of the transformer state on day 17 at each time step: normal operation state, core and winding clamp loosening fault, winding misalignment fault, winding warping fault, winding bulging fault, and partial discharge fault.

[0086] In one embodiment, the ConvLSTM network combines a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) network. The main structure of the CNN includes an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The CNN part of the model uses a two-dimensional convolution method. The input image is an RGB image; therefore, in the first convolution, there are 3 in channels and 32 out channels, a kernel size of 4, a stride of 2, and padding of 1 (the number of layers that pad the image with zeros during convolution). The output data size is 64*64*32. The out channels of the first convolution are the same as the in channels of the second convolution, which has 64 out channels. All other parameters remain unchanged, so the output size of the second convolution is 32*32*64. This process continues, undergoing 7 two-dimensional convolutions, resulting in a final output data size of 1*1*1024. Since the output data has a height and width of 1, no additional pooling is needed for feature dimensionality reduction. The activation function used throughout the process is the Leaky ReLU function, which addresses the "death" problem of neurons. Unlike the ReLU activation function, Leaky ReLU treats the less-than-zero parts of the input as negative and has a small gradient. Data processed by the convolutional neural network is input into an LSTM network. The forget gate, input gate, and output gate in the LSTM network are iteratively trained with a large amount of data to adjust their parameters, enabling them to learn the temporal fitting relationships between data extracted from the convolutional neural network. This allows for effective dynamic modeling of the predicted time series input and output data. Finally, the trained data is fitted by a ConvLSTM network and output as a predicted image through neurons in the fully connected layers. Since SEQ_SIZE is 16, the network iterates 16 times before outputting the preset result. The entire prediction process begins with input data for model training to determine the model parameters. The training and validation process involves initializing the model structure and optimizer. If a GPU is available, CUDA is used to accelerate model loading. The MSELoss loss function is chosen, and the training dataset is read into the input and label variables one by one. Training is performed epoch by epoch based on the set EPOCH (number of training iterations). Input data is fed into the model for forward propagation, error calculation, gradient zeroing, backpropagation of error, and finally parameter updates. Validation is performed every five training iterations. The validation process involves reading the validation dataset into the input and label variables one by one, following the same procedure as training. Finally, the model is saved. During training and validation, epochs and loss values ​​are printed line by line, allowing real-time observation of how the loss value changes with each epoch. The loss value is allowed to fluctuate within a small range; a generally decreasing value indicates that the current training and validation are effective.Typically, the initial number of iterations (EPOCH) is set to 300, and the learning rate is set to 0.1%. After training and validation, open the loss curve. When the curve stabilizes within the given number of iterations, if the loss values ​​of the training and test sets are very similar, then the current parameters are appropriate and no adjustment is needed. If the performance of the training set is significantly better than that of the test set, this is called overfitting, and methods such as data augmentation and increasing the dataset can be considered to suppress it. Conversely, underfitting occurs, and increasing the network depth and the number of units in the hidden layers can be considered. As for other parameters that are continuously accumulated, updated, and reset during training, they will be saved with the model and do not require manual adjustment.

[0087] This prediction method employs a multi-sensor-based ConvLSTM algorithm to extract features from grayscale images, evaluates the prediction results using mean squared error loss and structural similarity, and evaluates the classification results using IoU. The online transformer vibration and noise detection system acquires raw vibration signals, which are preprocessed into grayscale feature images. A multi-sensor-based ConvLSTM model is used to extract texture features from the images. The mean squared error loss and structural similarity are used to evaluate the judgment results, and a similarity index is used to classify the feature images. The model outputs the classification results, thus achieving the effect of predicting the internal mechanical state of the transformer. This method has no fixed requirements for the amount of data, high training freedom, good robustness, high classification accuracy of feature states, and fast state judgment speed. Furthermore, the prediction algorithm employs other prediction algorithms such as DDPAE, DFN, CDNA, FRNN, VPN Baseline, PredRNN, PredRNN+, MIM, BP-Net, and E3D-LSTM.

[0088] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A transformer state prediction method based on multivariate sensing and feature extraction, characterized in that, It includes, Vibration signals were collected from the surface of the transformer tank under both normal and fault conditions. The vibration signal is preprocessed and visualized as a grayscale feature image. A portion of the grayscale feature images are stored as a fault feature image library, and the remaining grayscale feature images are used as a model dataset. In the preprocessing, the vibration signal is read, the feature values ​​are calculated and saved as a feature value matrix, and Z-Score standardization is performed based on the feature value matrix, as shown in formula (1): （1） in, It is an eigenvalue. It is the sample standard deviation. The features are standardized and follow a normal distribution with a mean of 0 and a standard deviation of 1. A grayscale feature image is then constructed using the ratio of the maximum and minimum values ​​of each channel element, and the absolute value and ratio of adjacent differences. , formula (2): （2） in, This is a matrix of standardized eigenvalues, with a size of L*1, where L is the number of eigenvalues ​​contained in an image. , , , The range is from 0 to L. For the sequence number The value of the row, The absolute value of the maximum value. It represents the difference between adjacent elements in the sequence. The grayscale image matrix is ​​obtained by finding the maximum difference between any two elements in the sequence, that is, the difference between the maximum and minimum values. The size is L*L, and the eigenvalue range is [0, 1]. Based on the model dataset, the algorithm combines convolutional neural networks and long short-term memory networks to extract features, and uses 3D convolutional neural networks to encode and decode grayscale feature images. In the encoder part, attention modules and dynamic semantic vector sequences are used to replace fixed semantic vectors as input information encoding, and a subset of vectors is automatically selected during the decoder decoding process. The feature image output by the decoder is quantitatively compared with the grayscale feature image using peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Using the feature image of the vibration signal under fault conditions as a benchmark, the feature image output by the matching decoder is selected from all time steps, and its spatial similarity with the fault feature library is calculated. The similarity IoU index is used to classify the features. When it is greater than a predetermined threshold, it is considered that the two belong to the same feature state, and the label of the feature state in the fault feature library is output.

2. The transformer state prediction method based on multivariate sensing and feature extraction according to claim 1, characterized in that, Under normal and fault conditions, the load current of the transformer vibration and noise online detection system is adjusted. Starting with a load current of 4A, the load current is adjusted to 8A in increments of 0.8A. Vibration signals are collected from the surface of the tank under both normal and fault conditions to serve as samples for condition prediction.

3. The transformer state prediction method based on multivariate sensing and feature extraction according to claim 1, characterized in that, Fault conditions include loose core and winding clamps, winding warping, winding misalignment, winding bulging, and partial discharge.

4. The transformer state prediction method based on multivariate sensing and feature extraction according to claim 3, characterized in that, To simulate a winding loosening fault in the transformer, a torque wrench was used to loosen the screw fixing the winding at three settings: 4 N•m, 8 N•m, and 12 N•m. The time-domain vibration signal was then obtained as the vibration signal under the winding clamp loosening fault.

5. The transformer state prediction method based on multivariate sensing and feature extraction according to claim 3, characterized in that, A winding warping fault is set up for the transformer. Pressure is applied to the tap during the winding process to cause the winding to deform. The time-domain vibration signal is then obtained as the vibration signal under the winding warping fault.

6. The transformer state prediction method based on multivariate sensing and feature extraction according to claim 3, characterized in that, A winding misalignment fault is set up on the transformer, and the time-domain vibration signal under rated current is obtained as the vibration signal under the winding misalignment fault.

7. The transformer state prediction method based on multivariate sensing and feature extraction according to claim 3, characterized in that, A winding bulge fault is simulated in the transformer. A medium is added between the high-voltage winding and the low-voltage winding, causing the high-voltage winding to bulge outward and the low-voltage winding to dent inward. At the same time, the high-voltage winding generates axial displacement. The time-domain vibration signal is obtained as the vibration signal under the winding bulge fault.

8. The transformer state prediction method based on multivariate sensing and feature extraction according to claim 3, characterized in that, A partial discharge fault was simulated on the transformer. Metal spikes were installed on the high-voltage output side of the dry-type transformer to simulate a partial discharge fault. The transformer was then stepped up, and the time-domain vibration signal between the metal spikes and metal clamps from partial discharge to breakdown was measured in real time as the vibration signal under the partial discharge fault.

9. The transformer state prediction method based on multivariate sensing and feature extraction according to claim 1, characterized in that, The eigenvalues ​​include fundamental frequency amplitude, fundamental frequency weight, spectral energy, and spectral complexity.

10. The transformer state prediction method based on multivariate sensing and feature extraction according to claim 1, characterized in that, Formula (3) for classifying features using the similarity IoU index is: （3） Where intersection is the set of elements in two images, union is the set of elements in two images, and IoU is the similarity.

Images