Transformer operation abnormality monitoring method based on sound feature recognition

Abstract

The application provides a transformer operation abnormality monitoring method based on sound feature recognition, which comprises the following steps: 1, constructing a data set based on sound data and converting sound signals to image data by using a Gram angle field; 2, training the image data obtained in step 1 by using a MobileNetV3 deep learning model; and 3, performing real-time sound monitoring by using the trained MobileNetV3 model, and detecting and diagnosing the abnormal state of the transformer. Through this innovative processing and analysis method, the speed and efficiency of noise detection can be effectively improved while ensuring accuracy. In addition, this method can also accurately identify various different types of environmental noise, further improving the wide applicability of environmental noise monitoring.

Description

Technical Field

[0001] The present invention relates to the fields of sound processing and neural networks, and in particular to a method for monitoring abnormal operation of a transformer based on sound feature recognition. Background Art

[0002] With the rapid development of my country's economy, the power grid system has continued to evolve towards ultra-high voltage (UHV), long-distance transmission, and large capacity. The global interconnection of UHV systems has posed even more stringent challenges to grid operation. As core equipment in the power grid, the reliable operation of power transformers is crucial for ensuring the stability of the entire grid. During long-term operation, transformers are inevitably subject to risks such as short-circuit current shocks and accidents. If latent faults are not promptly addressed, they can cause sudden transformer failures, leading to widespread power outages.

[0003] Currently, power transformer condition diagnosis methods are primarily categorized into two main categories: those based on electrical quantities and those based on non-electrical quantities. The former, including frequency response analysis, low voltage impulse, short-circuit impedance, and transfer function methods, are relatively mature in transformer condition diagnosis. However, these methods typically require the transformer to be taken out of service, making them unsuitable for online monitoring and difficult to detect in the latent stages of a fault. Consequently, they are only suitable for diagnosis before shipment, after fault repair, or during regular maintenance.

[0004] In contrast, transformer condition diagnosis methods based on non-electrical quantities offer greater flexibility for online monitoring, allowing signal data to be monitored without affecting transformer operation. However, this technology currently faces challenges such as an inadequate theoretical foundation, overly simplistic data types, and incomplete condition and fault data, which severely impact its effectiveness in actual field applications.

[0005] Against this backdrop, a transformer anomaly monitoring method based on sound signature recognition offers a new development direction for non-electrical diagnostic technology. This method analyzes the sound signatures generated during transformer operation to identify potential anomalies. This approach aims to address the limitations of existing technologies, which often lack the ability to meet online monitoring requirements and detect problems during the incubation period. This approach, in turn, helps further improve grid stability and transformer reliability. Summary of the Invention

[0006] In order to solve the above problems, the present invention aims to provide an environmental noise detection method based on Gram's angle field processing and combined with the MobileNetV3 neural network model, which has higher detection accuracy and lower computational complexity.

[0007] To achieve the above objectives, the present invention adopts a technical solution: a method for detecting environmental noise, comprising the following steps:

[0008] Step 1: Construct a dataset based on sound data and convert the sound signal into image data using Gram angular field processing;

[0009] Step 2: Use the MobileNetV3 deep learning model to train the image data obtained in step 1;

[0010] Step 3: Use the trained MobileNetV3 model to perform real-time sound monitoring and detect and diagnose abnormal transformer conditions.

[0011] In a possible embodiment, step 1 includes the following steps:

[0012] Step 11: Collect the sound dataset and extract the sound information and sampling frequency from each audio;

[0013] Step 12: Calculate the number of sliding window and sliding step samples by sampling time and step size;

[0014] Step 13: Based on step 12, the audio signal is converted into a Gram image through the Gram angle field;

[0015] Step 14: Construct a dataset. All Gram images are labeled with noise-free and noise-containing labels, and are divided into training and validation sets in proportion.

[0016] In a possible embodiment, step 13 includes the following steps:

[0017] Step 131: performing normalization processing on the intercepted portion of the sampling signal;

[0018] Step 132: Correlate the vibration signal value and the corresponding sampling point sequence value through angle and radius, and map them to polar coordinates using the arccos function;

[0019] Step 133: Encode the sound signal features into a Gram matrix graph.

[0020] In a possible embodiment, step 133 includes:

[0021] The data obtained in step 132 is used as a unit vector group, and the Gram matrix is ​​composed of the inner product of each pair of unit vectors from left to right and from top to bottom in the unit vector group.

[0022] In a possible embodiment, step S2 includes:

[0023] Step 21: Perform random inversion and normalization preprocessing on the training set and validation set of Gram images respectively;

[0024] Step 22: Build the MobileNetV3 network model structure;

[0025] Step 23: Input the preprocessed dataset into the mobilenetv3 network model to obtain the network weights.

[0026] In a possible embodiment, the MobileNetV3 network model in step 22 includes:

[0027] an input module configured to receive image data;

[0028] a convolution module, connected to the input module and configured to perform convolution calculation on the received image data;

[0029] The Depthwise Separable Convolution module is connected to the convolution module. The separable convolution module includes depthwise convolution and pointwise convolution to reduce computation and memory consumption.

[0030] The Squeeze-and-Excitation module is connected to the separable convolution module and configured to learn the weight of each channel through global average pooling and two fully connected layers to adaptively adjust the importance of feature channels;

[0031] an average pooling module, connected to the squeeze-excitation module and configured to reduce the size of the feature map at the end of the feature extraction stage;

[0032] The fully connected layer module is connected to the average pooling module and configured to map the extracted features to the final number of categories or labels.

[0033] In a possible embodiment, step 23 includes the following steps:

[0034] Step 231: Input the training set into the MobileNetV3 network model in batches, and clear the gradient values ​​of all weights in the optimizer.

[0035] Step 232: Calculate the gap between the actual output of the network model and the target label, and calculate the network loss using the cross entropy loss function. The network loss calculation formula is as follows:

[0036]

[0037] Wherein: loss represents a loss function; x represents the probability of the actual output of the network corresponding to each class, and class represents the probability of the real class of the image;

[0038] Step 233: the loss function is back-propagated, the gradient of all weights that need to be calculated is calculated, and accumulation is performed, and the gradient accumulation calculation formula is as follows:

[0039]

[0040] Wherein: x.grad is the gradient of variable x;

[0041] Step 234: the value of variable x is updated using the Adam optimizer;

[0042] Step 235: the MobileNetV3 network model of this round of training is verified, the verification set is input for prediction, and the accuracy of the prediction result is calculated, and after multiple rounds of training, the model weight with the best accuracy is saved.

[0043] In a possible embodiment, the step 3 comprises:

[0044] Step 31: the microphone of the computer collects the sound of the surrounding environment in real time according to the sampling rate and sampling time of the audio data set, and saves it in a designated folder;

[0045] Step 32: at the same time, the sound collected and saved in the last round is input into the trained MobileNetV3 prediction network model for real-time noise detection.

[0046] The beneficial effects of the present application are:

[0047] 1. Through the gram image, the frequency spectrum information of the audio signal can be more comprehensively extracted and utilized, and the accuracy of the transformer noise detection is enhanced.

[0048] 2. By converting the linearly related time series signal into an orthogonal signal set, the interference and overlap between signals can be reduced, and the accuracy and reliability of noise processing are improved.

[0049] 3. By converting the audio signal into a two-dimensional vibration signal matrix, more sound features, geometric properties and internal data structures can be extracted from high-dimensional image data, thereby improving the effect of noise recognition.

[0050] 4. The MobileNetV3 network model adopted in the present application reduces the parameter quantity and computational complexity of the model, so that the present application has lower requirements for computing resources and is suitable for application in resource-limited devices and embedded systems.

[0051] 5、In addition to the detection of environmental noise, the present application is also applicable to other sound classification scenarios, showing wide applicability. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 For the transformer operation abnormality monitoring method based on sound feature recognition in the embodiment of the present application, a whole flow chart is shown.

[0053] Figure 2 For the network structure diagram of the mobilenetv3 neural network model in the embodiment of the present application.

[0054] Figure 3 For the Bneck basic unit structure diagram in the embodiment of the present application.

[0055] Figure 4 For the gram angle field processing data flow chart in the embodiment of the present application, wherein Figure 4 (a) represents the extracted sound signal; Figure 4 (b) represents the normalized sound information; Figure 4 (c) is a diagram of the sampling point mapped to the polar coordinate; Figure 4 (d) represents the gram matrix diagram. DETAILED DESCRIPTION

[0056] In order for those skilled in the art to better understand the technical solutions of the present application, the present application will be described in detail below in combination with the drawings, and the description in this part is only exemplary and explanatory, and should not have any limiting effect on the protection scope of the present application.

[0057] The present application proposes a transformer operation abnormality monitoring method combining Gram angle field processed original audio with an image classification neural network, which provides full spectrum information of the audio signal, is very important for the case of extracting global spectrum features from the audio in the image classification task, like having uniform frequency resolution, can provide more detailed frequency resolution. The Gram image can better capture the harmonic structure and frequency changes of the audio signal, and can provide more abundant information for the case of analyzing the harmonic features and frequency patterns of the audio signal. Through Gram angle field processing, linearly related time series signals are converted into a set of orthogonal signals, and the orthogonal signals have the characteristics of being independent of each other, which can reduce the mutual interference and overlap between signals, and improve the accuracy and reliability of signal processing. The high-order signal description capability of the image converts the time domain and frequency domain signals into a two-dimensional vibration signal matrix, and extracts sound features from the texture information in the converted signal matrix. Compared with the original sound signal, the high-order description in the high-dimensional image data can show more geometric properties and internal data structures, and also has the speed of accelerating the convolution network feature fusion, and improves the operation precision. The present application adopts the MobileNetV3 neural network model, reduces the parameter quantity and computational complexity of the model, has smaller model volume and lower calculation resource requirement, so that the present application is suitable for resource-limited mobile devices and embedded systems. The present application can also be used in sound classification, for example, as described below.

[0058] Reference Figure 1 A transformer operation abnormality monitoring method based on sound feature recognition, comprising the following steps:

[0059] S1: Construct a sound feature based data set, and perform Gram angle field processing on the original sound data, specifically including:

[0060] S11: Collect the sound of the transformer during operation, and extract the sound information and sampling frequency in the audio.

[0061] S12: Set appropriate sampling window time and step, and calculate the corresponding sample number.

[0062] S13: Convert the audio signal into a Gram image that can be presented on an image through the Gram angle field method, specifically including:

[0063] S131: Normalize the audio signal segment obtained by sampling to ensure that the signal is evaluated under the same standard.

[0064] S132: Combine the value of the vibration signal and the sampling point sequence value, and map it to the polar coordinate system through the arccos function.

[0065] S133: In order to encode the sound signal feature into the Gram matrix diagram, the data obtained in S132 is taken as a unit vector group, which corresponds to the order structure of the sampling points from left to right and from top to bottom, and the Gram matrix is the inner product of each pair of unit vectors.

[0066] S14: Constructing a data set, constructing a data set according to the noise-free and noisy labels of all Gram images, and dividing them into training sets and validation sets in proportion.

[0067] S2: Training the model by using the MobileNetV3 neural network;

[0068] S21: Randomly reversing and standardizing the Gram image training set and validation set for preprocessing;

[0069] S22: Constructing the MobileNetV3 network model structure, which specifically includes:

[0070] An input module configured to receive image data;

[0071] A convolution module connected to the input module and configured to perform convolution calculation on the received image data;

[0072] A separate convolution module connected to the convolution module, which includes depth convolution and point-by-point convolution, for reducing the amount of calculation and memory consumption;

[0073] A compressed excitation module connected to the separate convolution module and configured to learn the weight of each channel through global average pooling and two fully connected layers, and to adaptively adjust the importance of the feature channel;

[0074] An average pooling module connected to the compressed excitation module and configured to reduce the size of the feature map at the end of the feature extraction stage;

[0075] A fully connected layer module connected to the average pooling module and configured to map the extracted features to the final class or label number.

[0076] S23: Inputting the preprocessed data set into the MobileNetV3 network model to obtain the network weight.

[0077] Step 231: Batch inputting the training set into the MobileNetV3 network model to perform gradient value zero operation on all weights inside the optimizer;

[0078] Step 232: Calculating the gap between the actual output of the network model and the target label, and calculating the network loss through the cross-entropy loss function, the network loss calculation formula is as follows:

[0079]

[0080] wherein: x represents the actual output of the network corresponding to the probability of each category, x[class] is the probability of the real category of the image; ∑_j exp(x[j]) calculates the sum of the exponentials of all probabilities in the input vector.

[0081] Step 233: Backpropagation is performed on the loss function to calculate the gradients of all weights that need to be calculated, and the gradients are accumulated, and the gradient accumulation calculation formula is as follows:

[0082]

[0083] Step 233: Update the x value using the Adam optimizer;

[0084] Step 234: Verify the MobileNetV3 network model of this round of training, input the verification set for prediction, and calculate the accuracy of the prediction result. After multiple rounds of training, the model weight with the best accuracy is saved.

[0085] S3: Real-time noise monitoring of the environment around the transformer using the trained network model.

[0086] Step 31: The computer's microphone collects the sound around the transformer in real time according to the sampling rate and sampling time of the audio data set, and saves it in a designated folder;

[0087] Step 32: Input the newly collected audio data into the trained MobileNetV3 model to detect whether there is noise.

[0088] In use, as shown in Figure 1 , the noise detection method includes the following specific steps:

[0089] A. Collect noise-free and noisy audio in the environment around the transformer under different noise levels (such as 5dB, 10dB, and 15dB), and generate 640x640 pixel Gram images through Gram angle field processing.

[0090] B. Construct a mobilenetv3 neural network model, divide the audio Gram images obtained in step (A) into a training set and a data set, and input them into the mobilenetv3 neural network model for iterative training until the model test accuracy reaches a saturated state, and save the network model weight file with the highest accuracy.

[0091] C. Record audio through a microphone in paInt16 sampling format at a sampling rate of 8000Hz, and output 640 sample points in real time to generate 640x640 resolution images in the Gram angle field.

[0092] D, input the image obtained in step (C) into the neural network model obtained in step (B) to output a determination result and a probability of whether noise exists in the audio. A mobilenetv3 lightweight neural network is collected and constructed, which has the characteristics of few parameters, small calculation amount and short inference time. It is more suitable for scenarios with limited storage space and power consumption, such as mobile embedded devices and other edge computing devices. The method specifically comprises:

[0093] a, the input 640x640 gram image is convolved by a 3x3 convolution kernel and 11 basic units bneck, and the network output is calculated by a 1x1 convolution, and then enters the Avg-pool layer through the activation function H-Swish, once 1x1 convolution kernel convolution and H-Swish activation, and the output result after once 1x1 convolution kernel convolution.

[0094] b, the Bneck unit is the basic module of the network, which mainly realizes channel separable convolution + SE channel attention mechanism + residual connection, and the structural diagram is as shown in Figure 2 Channel separable convolution is divided into two processes, depth convolution (convolution kernel channel number is equal to input feature map channel number, only changes the size of feature map) and pointwise convolution (through the size of 1x1 convolution kernel, only changes the channel number of feature map). The input feature map is raised by pointwise convolution, then depth convolution is performed, and finally pointwise convolution is performed to reduce the dimension, forming an inverted residual structure, and the activation function generally uses H-Swish and RelU, and NBN represents that BN operation is not used, as shown in Figure 3 .

[0095] c, the SE channel attention mechanism is to process the feature map of each channel in the Bneck unit through depth convolution into a 1-dimensional vector through average pooling, and then multiply the original channel feature map parameters after passing through two fully connected layers to form a feature map equal in size and channel number to the original feature map.

[0096] In this embodiment, the audio data of the transformer under different operating conditions is collected, including normal operating sound and various abnormal noises such as electrical fault noise, mechanical fault noise, device abnormal heat dissipation noise and external interference noise, etc. After unifying the audio format, the audio data is processed by the gram angle field. The specific operation process includes:

[0097] 1) In the field, the audio is recorded in paInt16 sampling format and 8000Hz sampling rate, and 0.08s audio is extracted by window sliding for feature extraction, 0.08s audio is 640 sampling points, and the obtained sound signal is as shown in Figure 4 (a).

[0098] 2) The intercepted sampling points are normalized, and the active sound model is as shown in Figure 4 (b). The specific calculation formula is as follows:

[0099]

[0100] Wherein: is the normalized data point; X i is a certain data determined characteristic value; max(X) is the maximum value of the characteristic in the data set; min(X) is the minimum value of the characteristic in the data set;

[0101] The value and the corresponding sampling point order value are associated with the angle and the radius, and the arccos function is used for calculation. The specific calculation formula is as follows:

[0102]

[0103] is the azimuth angle, r is the radial distance; ti represents the discrete step length, and N represents a scaling factor.

[0104] With the expression of the sampling points one by one, the corresponding value will produce bending across the circle step by step. The normalized vibration signal value and the value mapped into polar coordinates are in one-to-one correspondence, as shown in Figure 4 (c).

[0105] (3) In the Gram matrix, the order structure of the corresponding sampling points from the top left to the bottom right, and the time characteristics of the time series signal are expressed by the image, as shown in Figure 4 (d). The signal is regarded as a unit vector group, and the inner product of two vectors reflects their direct similarity. Therefore, the Gram matrix can be regarded as the inner product definition matrix of each pair of vectors, and the specific matrix is as follows:

[0106]

[0107] The embodiment of the present application comprises the following steps: ① constructing a sound data set: initially collecting sound data generated by the transformer under normal operation and different abnormal conditions, covering electrical fault noise, mechanical fault noise, device heat dissipation abnormal noise and external interference noise, etc. Extract the sound information and sampling frequency of each collected audio. ② sound data preprocessing: using the Gram angle field conversion technology to process the original audio data, and converting the audio signal into a Gram image. This step includes the normalization processing of the sampling points, and the sampling points are mapped to the polar coordinate system through the arccos function, thereby forming a Gram matrix image that can encode the characteristics of the sound signal. ③ Constructing the final data set, classifying the Gram image according to the sound state (noise or no noise) it represents, and dividing it into training set and validation set according to the proportion. ④ Model training and verification: using MobileNetV3 neural network to train the model. For the Gram image of the training set and the validation set, data enhancement and standardization preprocessing are performed. The MobileNetV3 network model is constructed and optimized, which includes multiple core modules: convolution module, separated convolution module, Squeeze-and-Excitation module, average pooling module and fully connected layer module. ⑤ Model training, including batch input, loss calculation, back propagation and weight update, etc., until the model reaches the best accuracy on the validation set. ⑥ Real-time noise monitoring: using the trained network model for real-time noise detection. The computer microphone is used to collect the sound of the surrounding environment in real time, and the collected audio is converted into a Gram image input into the model to determine whether the current transformer has an abnormal operation. Through this method, the sound characteristics of the transformer can be effectively utilized for accurate operation state monitoring, which is of great significance for preventing transformer faults in the power system in advance and ensuring the stable operation of the equipment. This method not only can greatly reduce the risk of sudden power failure caused by transformer abnormalities, but also can effectively improve the operation efficiency and safety level of the power grid.

[0108] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or equipment.

[0109] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea. The above description is only the preferred implementation manner of the present application, and it should be noted that, due to the limited expression of the text, there are objectively infinite specific structures, and for ordinary skilled persons in the technical field, some improvements, refinements or changes can be made without departing from the principles of the present application, and the above technical features can also be combined in an appropriate manner; the improvements, refinements, changes or combinations, or the direct application of the inventive concept and technical solution to other occasions without improvement, shall be regarded as the protection scope of the present application.

Claims

1. A transformer operation abnormality monitoring method based on sound feature recognition, characterized in that: The following steps are included: Step 1: Construct a dataset based on sound data and convert the sound signal into image data using Gram angular field processing; Step 2: Using the MobileNetV3 deep learning model to train the image data obtained in step 1; step 2 includes: Step 21: Perform random inversion and normalization preprocessing on the training set and validation set of Gram images respectively; Step 22: Build the MobileNetV3 network model structure; the mobilenetv3 network model in step 22 includes: an input module configured to receive image data; a convolution module, connected to the input module and configured to perform convolution calculation on the received image data; A separation convolution module is connected to the convolution module. The separation convolution module includes depthwise convolution and pointwise convolution to reduce computational complexity and memory consumption. The compression-excitation module is connected to the separation convolution module and configured to learn the weight of each channel through global average pooling and two fully connected layers to adaptively adjust the importance of feature channels; an average pooling module, connected to the squeeze-excitation module and configured to reduce the size of the feature map at the end of the feature extraction stage; A fully connected layer module, connected to the average pooling module and configured to map the extracted features to the final number of categories or labels; Step 23: Input the preprocessed data set into the MobileNetV3 network model to obtain network weights; Step 23 includes the following steps: Step 231: Input the training set into the MobileNetV3 network model in batches, and clear the gradient values ​​of all weights in the optimizer. Step 232: Calculate the gap between the actual output of the network model and the target label, and calculate the network loss using the cross entropy loss function. The network loss calculation formula is as follows: ， in: loss Represents the loss function; x represents the probability of the actual output of the network corresponding to each category, and class represents the probability that this image is the true category; Step 233: Backpropagate the loss function, calculate the gradients of all weights that need to be calculated, and accumulate them. The gradient accumulation calculation formula is as follows: ， in: x.grad is a variable x gradient; Step 234: Use the Adam optimizer to update the value of the variable x; Step 235: Validate the MobileNetV3 network model trained in this round, input the validation set for prediction, and calculate the accuracy of the prediction results. After multiple rounds of training, save the model weights with the best accuracy. Step 3: Use the trained MobileNetV3 model to perform real-time sound monitoring and detect and diagnose abnormal transformer conditions.

2. The transformer operation abnormality monitoring method according to claim 1, characterized in that: The step 1 comprises the following steps: Step 11: Collect the sound dataset and extract the sound information and sampling frequency from each audio; Step 12: Calculate the number of sliding window and sliding step samples by sampling time and step size; Step 13: Based on step 12, the audio signal is converted into a Gram image through the Gram angle field; Step 14: Construct a dataset. All Gram images are labeled with noise-free and noise-containing labels, and are divided into training and validation sets in proportion.

3. The transformer operation abnormality monitoring method according to claim 2, characterized in that: The step 13 comprises the following steps: Step 131: performing normalization processing on the intercepted portion of the sampling signal; Step 132: Correlate the vibration signal value and the corresponding sampling point sequence value through angle and radius, and map them to polar coordinates using the arccos function; Step 133: Encode the sound signal features into a Gram matrix graph.

4. The method for monitoring abnormal operation of a transformer according to claim 3, characterized in that: The step 133 includes: The data obtained in step 132 is used as a unit vector group, and the Gram matrix is ​​composed of the inner product of each pair of unit vectors from left to right and from top to bottom in the unit vector group.

5. The transformer operation abnormality monitoring method according to claim 1, characterized in that: The step 3 includes: Step 31: Based on the audio sampling rate and sampling time of the dataset, collect the sound of the transformer's surrounding environment in real time and save it in a specified folder; Step 32: At the same time, the sound collected and saved in the previous round is input into the trained MobileNetV3 prediction network model for real-time noise detection.

Images