Mutual inductor fault diagnosis method based on multi-sensor data fusion and transfer learning

Abstract

The invention provides a mutual inductor fault diagnosis method based on multi-sensor data fusion and transfer learning, and the method comprises the following steps: S1, installing a plurality of high-precision sensors on the secondary side of a voltage transformer to be subjected to fault diagnosis based on the voltage transformer to be subjected to fault diagnosis; s2, on the basis of the step S1, performing data acquisition and processing on data acquired by various high-precision sensors through a DHDAS dynamic signal acquisition instrument; and S3, based on the step S2, carrying out fault diagnosis on the voltage transformer through a PC (Personal Computer) carrying acquisition analysis software on the DHDAS dynamic signal acquisition instrument. According to the method, multi-mode signals such as voltage, current and temperature are synchronously collected, time-frequency domain features are extracted through continuous wavelet transform, the dynamic characteristics of interlayer short circuit faults can be more comprehensively represented, meanwhile, by combining transfer learning and a deep neural network, the intelligent diagnosis model can remarkably improve the classification performance under the condition that sample data is small, and the classification efficiency is improved. And the limitation of a single-mode signal diagnosis method is effectively overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault diagnosis, and in particular to a mutual inductor fault diagnosis method based on multi-sensor data fusion and transfer learning. Background Art

[0002] As the core equipment for measurement and protection in the power system, the operating status of the voltage transformer is crucial to the safety and stability of the power grid. In actual operation, the voltage transformer winding may have interlayer short circuit faults due to insulation aging, external interference or the complexity of the operating environment. According to the degree of fault, interlayer short circuit faults are usually divided into minor, medium and severe types. Among them, minor faults are easy to be ignored because the signal changes are not obvious. However, minor faults that are not discovered in time may gradually deteriorate over time and eventually evolve into serious faults that threaten the normal operation of the power grid.

[0003] At present, traditional fault diagnosis methods on the market usually rely on manual inspection and analysis based on physical models. However, manual inspection has obvious limitations. It is not only inefficient but also has limited ability to identify hidden faults. Physical model-based diagnosis is not able to describe complex electromagnetic characteristics, and shows obvious lack of adaptability when dealing with multi-source interference and complex working conditions. In addition, these methods generally lack the ability to dynamically track the evolution of faults and cannot meet the needs of modern power grids for rapid fault diagnosis. At the same time, with the rapid development of machine learning technology, data-driven fault diagnosis methods have gradually become a research hotspot. Such methods can effectively improve the accuracy and efficiency of diagnosis by mining the deep features of fault signals. However, most of the current research mainly relies on fault signals collected by single sensors as input sources, and fails to fully consider the complementarity of multi-source data in characterizing fault characteristics. When single sensor signals are affected by noise and power grid interference, it is easy to cause unstable feature extraction, especially when detecting minor faults, it often cannot provide sufficient diagnostic capabilities.

[0004] Therefore, based on the above problems, the present invention provides a transformer fault diagnosis method based on multi-sensor data fusion and transfer learning. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a transformer fault diagnosis method based on multi-sensor data fusion and transfer learning, which solves the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0007] The transformer fault diagnosis method based on multi-sensor data fusion and transfer learning includes the following steps:

[0008] S1: Based on the voltage transformer to be diagnosed for fault, multiple high-precision sensors are installed on the secondary side of the voltage transformer to be diagnosed for fault;

[0009] S2: Based on step S1, the data collected by various high-precision sensors are collected and processed by the DHDAS dynamic signal acquisition instrument;

[0010] S3: Based on step S2, the voltage transformer is fault diagnosed by using a PC equipped with acquisition and analysis software on a DHDAS dynamic signal acquisition instrument.

[0011] Furthermore, the multiple high-precision sensors in step S1 include voltage sensors, current sensors and wireless temperature sensors.

[0012] Furthermore, based on step S1, effective features are further extracted from the collected multimodal signals such as voltage, current and temperature in step S2, and continuous wavelet transform (CWT) is used for time-frequency domain analysis. The core idea of ​​CWT is to analyze the signal at multiple scales (frequencies) by convolving the signal with a set of mother wavelets. Given a signal and a mother wavelet function, the continuous wavelet transform formula is:

[0013]

[0014] Where f(t) is the input signal, is the mother wavelet function. Usually, a waveform with time-frequency localization characteristics is selected. Morlet wavelet is selected as the mother wavelet. a is the scale factor, which controls the expansion and contraction of the wavelet and represents the frequency resolution. b is the translation factor, which controls the time position of the wavelet and determines the analysis area of ​​the signal on the time axis. ψ * (t) is the complex conjugate of the mother wavelet, which ensures that the transformation process has good time-frequency localization characteristics;

[0015] The Morlet wavelet is a complex signal formed by multiplying a sine wave with a Gaussian window function, usually expressed as:

[0016]

[0017] Where σ is the standard deviation of the Gaussian window, which controls the width of the wavelet in the time domain. ω0 is the center frequency, which controls the frequency of the wavelet in the frequency domain. i is the imaginary unit, which ensures that the wavelet has time-frequency localization characteristics in the complex domain.

[0018] Furthermore, the collected multimodal signals are subjected to wavelet transformation respectively, which specifically includes the following steps:

[0019] A: Wavelet transformation of voltage signal. The collected voltage signal is V(t). The continuous wavelet transform is performed by Morlet wavelet to obtain the time-frequency diagram CWT of the voltage signal. v (a, b), its continuous wavelet transform is defined as:

[0020]

[0021] B: Wavelet transform of current signal. The collected current signal is I(t). The continuous wavelet transform is performed through Morlet wavelet to obtain the time-frequency diagram CWT of the current signal. I (a, b), its continuous wavelet transform is defined as:

[0022]

[0023] C: Wavelet transformation of temperature signal. The collected temperature signal is T(t). Through the Morlet wavelet, continuous wavelet transform is performed to obtain the time-frequency diagram CWT of the temperature signal. T (a, b), its continuous wavelet transform is defined as:

[0024]

[0025] Furthermore, the collected multimodal signals such as voltage, current and temperature are further subjected to multimodal feature fusion, the importance weight of each modal signal is calculated by the entropy weight method, and their time-frequency features are weightedly fused to generate a comprehensive time-frequency feature graph.

[0026] Furthermore, the entropy weight method for calculating the importance weight of each modal signal includes the following steps:

[0027] S11. The entropy weight method calculates the weights as follows:

[0028] a: Time-frequency feature extraction: First determine the wavelet basis function, scale factor, and center frequency of the wavelet, and then bring them into the above formula to calculate the time-frequency characteristics of the three signals:

[0029] Time-frequency diagram of voltage signal: CWT v (a, b)

[0030] Time-frequency diagram of current signal: CWT I (a, b)

[0031] Time-frequency diagram of temperature signal: CWT T (a, b)

[0032] b: Discretization of time-frequency graph: After discretizing the time-frequency graph according to time and frequency, construct the signal feature matrix:

[0033]

[0034] Among them, x ij It represents the time-frequency amplitude at the i-th time point and the j-th frequency point, where n is the number of time points and m is the number of frequency points.

[0035] c: Normalization: Normalize the feature matrix to eliminate the influence of different signal dimensions on the eigenvalues. The normalization formula is:

[0036]

[0037] After normalization, the value range of each column is [0, 1], which is used for the subsequent calculation of information entropy.

[0038] d: Information entropy calculation: Calculate the entropy value of each column based on the normalized result:

[0039]

[0040] in, is the normalization coefficient, H j Represents the information entropy of the signal at frequency point j.

[0041] S12. Weight allocation:

[0042] The entropy value reflects the uncertainty of the signal. The smaller the entropy value, the more important the signal feature is to the overall diagnosis. The weight is calculated based on the entropy value:

[0043]

[0044] Combine the weights of the three signals:

[0045] W=[w V ,w I , w T ]

[0046] Among them, w V is the weight of the voltage signal, w I is the weight of the current signal, w T is the weight of the temperature signal, and the weight W represents the relative contribution of the three signals to the overall diagnosis, ensuring that the importance of each signal is fully considered during fusion.

[0047] S13. Feature fusion:

[0048] Based on the weight W obtained by the entropy weight method, the time-frequency diagram CWT of the three signals is v (a, b), CWT I (a, b) and CWT T (a, b) perform weighted fusion to generate a comprehensive time-frequency feature map:

[0049] CWTfusion (a,b)=w V CWT V (a,b)+w I CWT I (a,b)+w T CWT T (a,b)

[0050] Among them, CWT fusion (a, b) are comprehensive time-frequency characteristic diagrams, which combine the characteristics of voltage, current and temperature signals; w V 、w I 、w T are the weights of the three signals respectively.

[0051] Furthermore, based on steps S1 and S2, a fault diagnosis model is further constructed in step S3, including the following steps:

[0052] a1: Transfer learning model selection: Introduce transfer learning technology, initialize the GoogLeNet network by loading the model weights pre-trained on a large image classification dataset (such as ImageNet), so that it has good feature extraction capabilities. The migration of the pre-trained model can be expressed as:

[0053]

[0054] in: represents the weights transferred from the pre-trained model; represents the randomly initialized weights, is the set of layers shared with the target task; l is the layer index in the network.

[0055] b1: Model training: The details are as follows:

[0056] b11: Comprehensive feature input: According to the above method, the collected signal is first subjected to continuous wavelet transform to generate a spectrum diagram, and then the importance weight of each modal signal is calculated by the entropy weight method, and its time-frequency features are weighted fused to generate a comprehensive time-frequency feature diagram CWT. fusion (a, b), let the comprehensive time-frequency feature map be F, and use it as the input of GoogLeNet. The input layer processes the feature map through normalization operation to unify the dimensions of different physical quantities. The normalization formula is as follows:

[0057] F input =Normalize(F)

[0058] Among them, Normalize is a normalization operation used to unify the dimensions of different physical quantities.

[0059] The Normalize operation can be expressed as:

[0060]

[0061] Make sure all eigenvalues ​​are normalized to the range [0, 1] to facilitate model processing.

[0062] b12: Data enhancement and expansion: Data enhancement technology can be used to artificially transform the spatial, brightness, contrast and other features of the samples so that the model can learn more diverse features and improve its performance on unseen data. The specific enhancement methods are as follows:

[0063] b121: Random cropping and flipping: Random cropping refers to randomly selecting an area from the original image and cropping it. By changing the spatial distribution of the image, the model can learn features from different perspectives and enhance its adaptability to different spatial positions. The mathematical expression is:

[0064] F′=T crop，flip (F;θ)

[0065] Among them, T crop，flip is the cropping and flipping operation function; θ is the random parameter of the augmentation operation.

[0066] b122: Noise addition: Use Gaussian noise model, assume that the noise has a mean of 0 and a variance of σ 2 The mathematical representation of Gaussian noise is:

[0067]

[0068] Where F(x, y) is the pixel value of the original feature map at position (x, y); Indicates that the mean is 0 and the variance is σ 2 Gaussian noise; F′(x, y) is the feature map after adding noise; the intensity of the noise is controlled by σ, and a larger σ value will increase the impact of noise and simulate a more severe sensor environment.

[0069] b123: Random grayscale adjustment: Grayscale adjustment is used to simulate different lighting conditions and contrast changes. The specific grayscale adjustment operation can be expressed as:

[0070] F′(x,y)=F(x,y)·α+β

[0071] Among them, α is a randomly generated contrast factor; β is a randomly generated brightness offset; F(x, y) is the pixel value of the original image, and F′(x, y) is the adjusted image pixel value.

[0072] b124: Mathematical framework of data enhancement: All enhancement operations can be uniformly represented by the transformation function T, specifically:

[0073] F′=T(F;θ)

[0074] Where T is the data enhancement transformation function, including the above-mentioned cropping, flipping, noise addition and grayscale adjustment; θ is the random parameter of the enhancement operation, which may include cropping size, noise intensity σ, contrast factor α, etc.

[0075] c1: Model training and optimization, the details are as follows:

[0076] c11: Network structure adjustment: The final dropout layer (pool5-drop_7x7_s1) in the GoogLeNet network is replaced with a dropout layer with a probability of 0.6 to enhance the regularization effect and reduce the risk of overfitting. The last learnable layer (loss3-classifier) ​​of GoogLeNet is redesigned. This layer is used to map the features extracted by the network to class probabilities. To adapt to the voltage transformer fault classification task, it is replaced with a new fully connected layer. The number of filters is set to 4 to correspond to the 4 states of the voltage transformer operation.

[0077] c12: Loss function: The cross entropy loss function is used as the model optimization target, and its formula is:

[0078]

[0079] Among them, N is the total number of samples; C is the total number of categories; y ij and are the true label and predicted probability respectively.

[0080] c13: Iteration strategy: Set the initial learning rate η0 and introduce a learning rate decay strategy to dynamically adjust the learning rate to ensure that the model can converge more stably in the later stages of training. The decay formula is as follows:

[0081]

[0082] Among them, λt is the decay factor, and t is the current iteration number.

[0083] d1: Cross validation: K-fold cross validation is used, the steps are as follows:

[0084] d11: Divide the data set into K subsets;

[0085] d12: Select K-1 subsets each time as training sets, and the remaining 1 subset as validation set;

[0086] d13: Repeat K times and calculate the average performance index.

[0087] The cross-validation model performance indicators include accuracy, precision, recall, etc.:

[0088]

[0089] Among them, TP, TN, FP and FN are the number of true positive, true negative, false positive and false negative samples respectively.

[0090] The present invention provides a transformer fault diagnosis method based on multi-sensor data fusion and transfer learning. Compared with the prior art, it has the following beneficial effects:

[0091] 1. Multimodal data fusion: This patent overcomes the problem of single modal signals being affected by noise and interference by integrating multiple sensor signals such as voltage, current and temperature. The entropy weight method is used to perform weighted fusion on the feature graphs of each modal signal, which can give full play to the complementarity of different signals and improve the accuracy of fault diagnosis. Especially in minor fault detection, the fusion of multi-source information significantly enhances the system's sensitivity to small changes and effectively improves the stability and robustness of the diagnosis results.

[0092] 2. Accurate feature extraction in the time-frequency domain: By using CWT to perform time-frequency domain analysis on voltage, current and temperature signals, the non-stationary characteristics of the signals can be accurately captured and converted into feature maps with dual time-frequency information. This technology effectively enhances the time-frequency characteristic expression of fault signals, enabling the model to better identify different types of faults in complex power grid environments, especially for fault modes such as interlayer short circuits.

[0093] 3. Transfer learning technology solves the problem of data scarcity: In order to solve the problem of scarce voltage transformer fault data, this patent uses transfer learning technology to use pre-trained models and fine-tune them, greatly reducing the dependence on a large amount of labeled data. In the case of insufficient sample data, transfer learning helps the model quickly adapt to different fault modes, significantly improving the classification accuracy and generalization ability of the model, especially in new fault modes or unseen working conditions, and can maintain a high diagnostic performance;

[0094] 4. Optimize fault classification performance: This patent combines the GoogLeNet deep neural network for feature learning and fault classification, making full use of its deep convolution structure and multi-scale feature extraction capabilities to efficiently identify complex fault modes. The deep learning model performs well in fault classification accuracy and real-time performance, and can monitor the status of voltage transformers in real time during power system operation, provide fast and accurate fault diagnosis results, and significantly improve the reliability and safety of the power system;

[0095] In summary, the method of the present invention can more comprehensively characterize the dynamic characteristics of interlayer short-circuit faults by synchronously collecting multimodal signals such as voltage, current, and temperature, and extracting time-frequency domain features using continuous wavelet transform (CWT). At the same time, combined with transfer learning and deep neural networks (such as GoogLeNet), the intelligent diagnosis model can significantly improve the classification performance when there is less sample data, and effectively overcome the limitations of single-modal signal diagnosis methods. The diagnosis strategy based on multi-source information fusion can not only significantly improve the sensitivity to minor faults, but also greatly improve the robustness and real-time performance of the diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0097] Figure 1 Shows the overall step diagram of the present invention;

[0098] Figure 2 The present invention shows a voltage transformer fault diagnosis flow chart;

[0099] Figure 3 The fault data acquisition platform of the present invention is shown; DETAILED DESCRIPTION

[0100] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0101] Example

[0102] In order to solve the technical problems in the background technology, the following transformer fault diagnosis method based on multi-sensor data fusion and transfer learning is given:

[0103] Combination Figure 1-3 As shown, the transformer fault diagnosis method based on multi-sensor data fusion and transfer learning provided by the present invention includes the following steps:

[0104] S1: Based on the voltage transformer to be diagnosed for fault, multiple high-precision sensors are installed on the secondary side of the voltage transformer to be diagnosed for fault;

[0105] S2: Based on step S1, the data collected by various high-precision sensors are collected and processed by the DHDAS dynamic signal acquisition instrument;

[0106] S3: Based on step S2, the voltage transformer is fault diagnosed by using a PC equipped with acquisition and analysis software on a DHDAS dynamic signal acquisition instrument.

[0107] The multiple high-precision sensors in step S1 include a voltage sensor, a current sensor, and a wireless temperature sensor.

[0108] Specifically, in order to accurately obtain the dynamic response information of the voltage transformer, this solution installs a variety of high-precision sensors (voltage sensor, current sensor and temperature sensor) on the secondary side of the voltage transformer. Through reasonable sensor configuration and high-precision data acquisition equipment, the electrical parameters and thermal stress of the voltage transformer can be accurately and synchronously monitored to ensure efficient data collection and real-time processing. The following is the specific configuration and working principle of each sensor:

[0109] The voltage sensor is installed on the secondary side of the voltage transformer and is mainly used to monitor voltage fluctuations and mutations in real time. The selected Weidmueller UTI-12 / 0.5 model voltage sensor has an accuracy of ±1%. It uses voltage conversion technology to convert voltage signals into standard output signals to ensure accurate capture of voltage fluctuations and instantaneous changes. The sensor has strong dynamic response capabilities and is suitable for real-time monitoring of voltage mutations in power systems. The output signal is an analog signal of 4-20mA or 0-10V, which is convenient for docking with data acquisition systems.

[0110] The current sensor is installed on the secondary side of the voltage transformer and is responsible for monitoring the current changes and reflecting the working status of the electrical system. The LEM LF 205-S model current sensor is used, which has an accuracy of ±0.5%. The sensor performs current detection based on the Hall effect principle and can capture the changes in the current waveform in real time. It is particularly suitable for monitoring current changes when a fault occurs. Its output signal is a standard current signal of 0-20mA or 4-20mA to ensure compatibility with the data acquisition system.

[0111] The temperature sensor is installed on the surface of the voltage transformer housing to monitor the temperature change of the equipment and provide thermal stress information before and after the failure. The selected GSensor-T60 wireless temperature sensor has a high accuracy of ±0.3℃ and can effectively capture the temperature change of the voltage transformer. The working status and potential fault hazards of the equipment are analyzed through the temperature change trend, and the data is transmitted to the monitoring system in real time through wireless signals to realize remote temperature monitoring. Its wireless transmission function makes the sensor installation more flexible and suitable for use in complex working environments.

[0112] To ensure synchronous acquisition of signals from voltage, current and temperature sensors, a DHDAS dynamic signal acquisition instrument is used. The device supports a sampling frequency of up to 1kHz, has excellent dynamic response capabilities, and can accurately acquire real-time signals from each sensor to ensure that the accuracy of time alignment reaches the millisecond level. The DHDAS dynamic signal acquisition instrument has built-in advanced anti-interference technology to effectively reduce the impact of external noise on data acquisition, and provides strong support for high-precision acquisition and synchronization of voltage, current and temperature signals, thereby ensuring the reliability of data analysis and the accuracy of diagnostic models.

[0113] In order to extract effective features from the collected multimodal signals such as voltage, current and temperature, this embodiment uses continuous wavelet transform (CWT) for time-frequency domain analysis. CWT can provide detailed information of the signal in both the time domain and the frequency domain, and is particularly suitable for the processing of non-stationary signals. By converting the signal into a time-frequency diagram, CWT can effectively enhance the expression ability of the signal features and capture the time-varying characteristics of the signal in different frequency ranges, thereby providing rich information for subsequent feature extraction and pattern recognition. Based on step S1, in step S2, the collected multimodal signals such as voltage, current and temperature are further extracted for effective features, and continuous wavelet transform (CWT) is used for time-frequency domain analysis. The core idea of ​​CWT is to analyze the signal at multiple scales (frequencies) by convolving the signal with a set of mother wavelets. Given a signal and a mother wavelet function, its continuous wavelet transform formula is:

[0114]

[0115] Where f(t) is the input signal, is the mother wavelet function. Usually, a waveform with time-frequency localization characteristics is selected. Morlet wavelet is selected as the mother wavelet. a is the scale factor, which controls the expansion and contraction of the wavelet and represents the frequency resolution. b is the translation factor, which controls the time position of the wavelet and determines the analysis area of ​​the signal on the time axis. ψ * (t) is the complex conjugate of the mother wavelet, which ensures that the transformation process has good time-frequency localization characteristics;

[0116] The Morlet wavelet is a complex signal formed by multiplying a sine wave with a Gaussian window function, usually expressed as:

[0117]

[0118] Among them, σ is the standard deviation of the Gaussian window, which controls the width of the wavelet in the time domain. ω0 is the center frequency, which controls the frequency of the wavelet in the frequency domain. i is an imaginary unit, which ensures that the wavelet has time-frequency localization characteristics in the complex domain. The time-frequency localization characteristics of the Morlet wavelet enable it to effectively capture the changes of the signal in the time domain and frequency domain, which is particularly applicable to the voltage, current and temperature signals collected from the voltage transformer.

[0119] The collected multimodal signals are subjected to wavelet transformation respectively, which specifically includes the following steps:

[0120] A: Wavelet transformation of voltage signal. The collected voltage signal is V(t). The continuous wavelet transform is performed by Morlet wavelet to obtain the time-frequency diagram CWT of the voltage signal. v (a, b), its continuous wavelet transform is defined as:

[0121]

[0122] B: Wavelet transform of current signal. The collected current signal is I(t). The continuous wavelet transform is performed through Morlet wavelet to obtain the time-frequency diagram CWT of the current signal. I (a, b), its continuous wavelet transform is defined as:

[0123]

[0124] C: Wavelet transformation of temperature signal. The collected temperature signal is T(t). Through the Morlet wavelet, continuous wavelet transform is performed to obtain the time-frequency diagram CWT of the temperature signal. T (a, b), its continuous wavelet transform is defined as:

[0125]

[0126] in, is the mother wavelet function, usually a waveform with time-frequency localization characteristics is selected. The present invention selects Morlet wavelet as the mother wavelet. a is the scale factor, which controls the expansion and contraction of the wavelet and represents the frequency resolution. b is the translation factor, which controls the time position of the wavelet and determines the analysis area of ​​the signal on the time axis. * (t) is the complex conjugate of the mother wavelet, which ensures that the transformation process has good time-frequency localization characteristics.

[0127] By performing continuous wavelet transform on the voltage V(t), current I(t) and temperature T(t) signals respectively, we can get their respective time-frequency diagrams CWT v (a, b), CWT I (a, b), CWT T(a, b). These time-frequency diagrams have high resolution in the time and frequency domains, and can fully characterize the time-frequency characteristics of different modal signals. These features provide basic data for subsequent feature extraction and multimodal signal fusion, while enhancing the signal's expressiveness, capturing the characteristic differences of signals under different fault modes, and improving the accuracy and robustness of diagnosis.

[0128] The collected multimodal signals such as voltage, current and temperature are further fused for multimodal features, the importance weight of each modal signal is calculated by the entropy weight method, and their time-frequency features are weighted fused to generate a comprehensive time-frequency feature map.

[0129] The entropy weight method calculates the importance weight of each modal signal and includes the following steps:

[0130] S11. The entropy weight method calculates the weights as follows:

[0131] a: Time-frequency feature extraction: First determine the wavelet basis function, scale factor, and center frequency of the wavelet, and then bring them into the above formula to calculate the time-frequency characteristics of the three signals:

[0132] Time-frequency diagram of voltage signal: CWT v (a, b)

[0133] Time-frequency diagram of current signal: CWT I (a, b)

[0134] Time-frequency diagram of temperature signal: CWT T (a, b)

[0135] b: Discretization of time-frequency graph: After discretizing the time-frequency graph according to time and frequency, construct the signal feature matrix:

[0136]

[0137] Among them, x ij It represents the time-frequency amplitude at the i-th time point and the j-th frequency point, where n is the number of time points and m is the number of frequency points.

[0138] c: Normalization: Normalize the feature matrix to eliminate the influence of different signal dimensions on the eigenvalues. The normalization formula is:

[0139]

[0140] After normalization, the value range of each column is [0, 1], which is used for the subsequent calculation of information entropy.

[0141] d: Information entropy calculation: Calculate the entropy value of each column based on the normalized result:

[0142]

[0143] in, is the normalization coefficient, H j Represents the information entropy of the signal at frequency point j.

[0144] S12. Weight allocation:

[0145] The entropy value reflects the uncertainty of the signal. The smaller the entropy value, the more important the signal feature is to the overall diagnosis. The weight is calculated based on the entropy value:

[0146]

[0147] Combine the weights of the three signals:

[0148] W=[w V ,w I , w T ]

[0149] Among them, w V is the weight of the voltage signal, w I is the weight of the current signal, w T is the weight of the temperature signal, and the weight W represents the relative contribution of the three signals to the overall diagnosis, ensuring that the importance of each signal is fully considered during fusion.

[0150] S13. Feature fusion:

[0151] Based on the weight W obtained by the entropy weight method, the time-frequency diagram CWT of the three signals is v (a, b), CWT I (a, b) and CWT T (a, b) perform weighted fusion to generate a comprehensive time-frequency feature map:

[0152] CWT fusion (a,b)=w V CWT V (a,b)+w I CWT I (a,b)+w T CWT T (a,b)

[0153] Among them, CWT fusion (a, b) are comprehensive time-frequency characteristic diagrams, which combine the characteristics of voltage, current and temperature signals; w V 、w I 、w T are the weights of the three signals respectively.

[0154] Multimodal signal fusion calculates the weights of voltage, current and temperature signals through the entropy weight method, and combines the weighted feature fusion method to generate a comprehensive time-frequency feature map. This map makes full use of the characteristics of different signals, makes up for the shortcomings of single signal features, and provides reliable feature input for subsequent fault mode recognition, thereby significantly improving the accuracy and robustness of fault diagnosis.

[0155] Based on steps S1 and S2, a fault diagnosis model is further constructed in step S3, including the following steps:

[0156] a1: Transfer learning model selection: In order to improve the accuracy and efficiency of voltage transformer fault diagnosis and reduce the dependence of training deep learning models on large-scale data sets, this paper selects the GoogLeNet deep learning model as the basic framework for diagnosis. GoogLeNet is a convolutional neural network based on the Inception module. It is characterized by extracting features of different granularities through multi-scale convolution kernels and has high feature expression capabilities. Due to the limited labeled data in the field of voltage transformer fault diagnosis, training deep neural networks directly from scratch may lead to underfitting or overfitting. For this reason, this paper introduces transfer learning technology. By loading the model weights pre-trained on large image classification datasets (such as ImageNet), the GoogLeNet network is initialized to have good feature extraction capabilities.

[0157] The transfer learning technology is introduced to initialize the GoogLeNet network by loading the model weights pre-trained on a large image classification dataset (such as ImageNet), so that it has good feature extraction capabilities. The migration of the pre-trained model can be expressed as:

[0158]

[0159] in: represents the weights transferred from the pre-trained model; represents the randomly initialized weights, is the set of layers shared with the target task; l is the layer index in the network. Through transfer learning, the model has a higher optimization starting point on the target task and converges faster during training.

[0160] b1: Model training: The details are as follows:

[0161] b11: Comprehensive feature input: According to the above method, the collected signal is first subjected to continuous wavelet transform to generate a spectrum diagram, and then the importance weight of each modal signal is calculated by the entropy weight method, and its time-frequency features are weighted fused to generate a comprehensive time-frequency feature diagram CWT. fusion(a, b), let the comprehensive time-frequency feature map be F, and use it as the input of GoogLeNet. The input layer processes the feature map through normalization operation to unify the dimensions of different physical quantities. The normalization formula is as follows:

[0162] F input =Normalize(F)

[0163] Among them, Normalize is a normalization operation used to unify the dimensions of different physical quantities.

[0164] The Normalize operation can be expressed as:

[0165]

[0166] Make sure all eigenvalues ​​are normalized to the range [0, 1] to facilitate model processing.

[0167] b12: Data enhancement and expansion: Data enhancement technology can be used to artificially transform the spatial, brightness, contrast and other features of the samples so that the model can learn more diverse features and improve its performance on unseen data. The specific enhancement methods are as follows:

[0168] b121: Random cropping and flipping: Random cropping refers to randomly selecting an area from the original image and cropping it. By changing the spatial distribution of the image, the model can learn features from different perspectives and enhance its adaptability to different spatial positions. The mathematical expression is:

[0169] F′=T crop，flip (F;θ)

[0170] Among them, T crop，flip is the cropping and flipping operation function; θ is the random parameter of the augmentation operation.

[0171] b122: Noise addition: Use Gaussian noise model, assume that the noise has a mean of 0 and a variance of σ 2 The mathematical representation of Gaussian noise is:

[0172]

[0173] Where F(x, y) is the pixel value of the original feature map at position (x, y); Indicates that the mean is 0 and the variance is σ 2Gaussian noise; F′(x, y) is the feature map after adding noise; the intensity of the noise is controlled by σ, and a larger σ value will increase the impact of noise and simulate a more severe sensor environment. Noise addition simulates the measurement error or external interference of the sensor. Common noise types include Gaussian noise and salt and pepper noise. By adding noise, the error of the sensor in actual operation can be simulated and the robustness of the model to noise data can be improved.

[0174] b123: Random grayscale adjustment: Grayscale adjustment is used to simulate different lighting conditions and contrast changes. The specific grayscale adjustment operation can be expressed as:

[0175] F′(x,y)=F(x,y)·α+β

[0176] Among them, α is a randomly generated contrast factor; β is a randomly generated brightness offset; F(x, y) is the pixel value of the original image, and F′(x, y) is the adjusted image pixel value.

[0177] Grayscale adjustment is used to simulate different lighting conditions and contrast changes. In practical applications, the brightness and contrast of the sensor data may change due to changes in ambient lighting conditions or device aging. Random grayscale adjustment can increase the diversity of the data and enable the model to better cope with such changes.

[0178] b124: Mathematical framework of data enhancement: All enhancement operations can be uniformly represented by the transformation function T, specifically:

[0179] F′=T(F;θ)

[0180] Where T is the data enhancement transformation function, including the above-mentioned cropping, flipping, noise addition and grayscale adjustment; θ is the random parameter of the enhancement operation, which may include cropping size, noise intensity σ, contrast factor α, etc.

[0181] Through these data augmentation operations, the original training data is expanded into diverse samples, which enhances the diversity of the training set, thereby improving the robustness of the model and reducing the risk of overfitting.

[0182] c1: Model training and optimization, the details are as follows:

[0183] c11: Network structure adjustment: The final dropout layer (pool5-drop_7x7_s1) in the GoogLeNet network is replaced with a dropout layer with a probability of 0.6 to enhance the regularization effect and reduce the risk of overfitting. The last learnable layer (loss3-classifier) ​​of GoogLeNet is redesigned. This layer is used to map the features extracted by the network to class probabilities. To adapt to the voltage transformer fault classification task, it is replaced with a new fully connected layer. The number of filters is set to 4 to correspond to the 4 states of the voltage transformer operation.

[0184] c12: Loss function: The cross entropy loss function is used as the model optimization target, and its formula is:

[0185]

[0186] Among them, N is the total number of samples; C is the total number of categories; y ij and are the true label and predicted probability respectively.

[0187] c13: Iteration strategy: Set the initial learning rate η0 and introduce a learning rate decay strategy to dynamically adjust the learning rate to ensure that the model can converge more stably in the later stages of training. The decay formula is as follows:

[0188]

[0189] Among them, λt is the decay factor, and t is the current iteration number.

[0190] d1: Cross validation: K-fold cross validation is used, the steps are as follows:

[0191] d11: Divide the data set into K subsets;

[0192] d12: Select K-1 subsets each time as training sets, and the remaining 1 subset as validation set;

[0193] d13: Repeat K times and calculate the average performance index.

[0194] The cross-validation model performance indicators include accuracy, precision, recall, etc.:

[0195]

[0196] Among them, TP, TN, FP and FN are the number of true positive, true negative, false positive and false negative samples respectively; the GoogLeNet model is trained and evaluated by the above method to generate the optimal model suitable for voltage transformer fault diagnosis and use it for subsequent online diagnosis of voltage transformer faults.

[0197] The relevant voltage, current and temperature data are collected by the above method, and the collected signals are analyzed in the time-frequency domain by continuous wavelet transform to generate a time-frequency feature diagram. Then, the obtained time-frequency diagram is weighted fused by the entropy weight method to generate a comprehensive time-frequency feature diagram (including multimodal information of voltage, current and temperature). Finally, the fused time-frequency feature diagram is input into the trained fault diagnosis model (GoogLeNet) for fault classification (normal state, minor fault, moderate fault and severe fault). The fault diagnosis results can be presented through the human-machine interface (HMI), providing real-time voltage transformer fault diagnosis function.

[0198] In summary, the present invention proposes a transformer fault diagnosis method based on multi-sensor data fusion and transfer learning. This method synchronously collects multi-modal signals such as voltage, current, and temperature, and uses continuous wavelet transform (CWT) to extract time-frequency domain features. It can more comprehensively characterize the dynamic characteristics of interlayer short-circuit faults. At the same time, combined with transfer learning and deep neural networks (such as GoogLeNet), this intelligent diagnosis model can significantly improve the classification performance when there are fewer sample data, and effectively overcome the limitations of single-modal signal diagnosis methods. The diagnosis strategy based on multi-source information fusion can not only significantly improve the sensitivity to minor faults, but also greatly improve the robustness and real-time performance of the diagnosis results.

[0199] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0200] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A transformer fault diagnosis method based on multi-sensor data fusion and transfer learning, characterized in that: The steps include: S1: Based on the voltage transformer to be diagnosed for fault, multiple high-precision sensors are installed on the secondary side of the voltage transformer to be diagnosed for fault; S2: Based on step S1, the data collected by various high-precision sensors are collected and processed by the DHDAS dynamic signal acquisition instrument; S3: Based on step S2, the voltage transformer is fault diagnosed by using a PC equipped with acquisition and analysis software on the DHDAS dynamic signal acquisition instrument.

2. The transformer fault diagnosis method based on multi-sensor data fusion and transfer learning according to claim 1 is characterized in that: The multiple high-precision sensors in step S1 include a voltage sensor, a current sensor, and a wireless temperature sensor.

3. The transformer fault diagnosis method based on multi-sensor data fusion and transfer learning according to claim 1 is characterized in that: Based on step S1, in step S2, the collected multimodal signals such as voltage, current and temperature are further extracted for effective features, and continuous wavelet transform (CWT) is used for time-frequency domain analysis. The core idea of ​​CWT is to analyze the signal at multiple scales (frequencies) by convolving the signal with a set of mother wavelets. Given a signal and a mother wavelet function, the continuous wavelet transform formula is: Where f(t) is the input signal, is the mother wavelet function. Usually, a waveform with time-frequency localization characteristics is selected. Morlet wavelet is selected as the mother wavelet. a is the scale factor, which controls the expansion and contraction of the wavelet and represents the frequency resolution. b is the translation factor, which controls the time position of the wavelet and determines the analysis area of ​​the signal on the time axis. ψ * (t) is the complex conjugate of the mother wavelet, which ensures that the transformation process has good time-frequency localization characteristics; The Morlet wavelet is a complex signal formed by multiplying a sine wave with a Gaussian window function, usually expressed as: Where σ is the standard deviation of the Gaussian window, which controls the width of the wavelet in the time domain. ω0 is the center frequency, which controls the frequency of the wavelet in the frequency domain. i is the imaginary unit, which ensures that the wavelet has time-frequency localization characteristics in the complex domain.

4. The transformer fault diagnosis method based on multi-sensor data fusion and transfer learning according to claim 3 is characterized in that: The collected multimodal signals are subjected to wavelet transformation respectively, which specifically includes the following steps: A: Wavelet transformation of voltage signal. The collected voltage signal is V(t). The continuous wavelet transform is performed by Morlet wavelet to obtain the time-frequency diagram CWT of the voltage signal. v (a, b), its continuous wavelet transform is defined as: B: Wavelet transform of current signal. The collected current signal is I(t). The continuous wavelet transform is performed through Morlet wavelet to obtain the time-frequency diagram CWT of the current signal. I (a, b), its continuous wavelet transform is defined as: C: Wavelet transformation of temperature signal. The collected temperature signal is T(t). Through the Morlet wavelet, continuous wavelet transform is performed to obtain the time-frequency diagram CWT of the temperature signal. T (a, b), its continuous wavelet transform is defined as:

5. The transformer fault diagnosis method based on multi-sensor data fusion and transfer learning according to claim 3 is characterized in that: The collected multimodal signals such as voltage, current and temperature are further fused for multimodal features, the importance weight of each modal signal is calculated by the entropy weight method, and their time-frequency features are weighted fused to generate a comprehensive time-frequency feature map.

6. The transformer fault diagnosis method based on multi-sensor data fusion and transfer learning according to claim 5 is characterized in that: The entropy weight method calculates the importance weight of each modal signal and includes the following steps: S11. The entropy weight method calculates the weights as follows: a: Time-frequency feature extraction: First determine the wavelet basis function, scale factor, and center frequency of the wavelet, and then bring them into the above formula to calculate the time-frequency characteristics of the three signals: Time-frequency diagram of voltage signal: CWT v (a, b) Time-frequency diagram of current signal: CWT I (a, b) Time-frequency diagram of temperature signal: CWT T (a, b) b: Discretization of time-frequency graph: After discretizing the time-frequency graph according to time and frequency, construct the signal feature matrix: Among them, x ij It represents the time-frequency amplitude at the i-th time point and the j-th frequency point, where n is the number of time points and m is the number of frequency points. c: Normalization: Normalize the feature matrix to eliminate the influence of different signal dimensions on the eigenvalues. The normalization formula is: After normalization, the value range of each column is [0, 1], which is used for the subsequent calculation of information entropy. d: Information entropy calculation: Calculate the entropy value of each column based on the normalized result: in, is the normalization coefficient, H j Represents the information entropy of the signal at frequency point j. S12. Weight allocation: The entropy value reflects the uncertainty of the signal. The smaller the entropy value, the more important the signal feature is to the overall diagnosis. The weight is calculated based on the entropy value: Combine the weights of the three signals: In=[in V ,In I ,In T ] Among them, w V is the weight of the voltage signal, w I is the weight of the current signal, w T is the weight of the temperature signal, and the weight W represents the relative contribution of the three signals to the overall diagnosis, ensuring that the importance of each signal is fully considered during fusion. S13. Feature fusion: Based on the weight W obtained by the entropy weight method, the time-frequency diagram CWT of the three signals is v (a, b), CWT I (a, b) and CWT T (a, b) perform weighted fusion to generate a comprehensive time-frequency feature map: CWT fusion (a,b)=w V ·CWT V (a,b)+w I ·CWT I (a,b)+w T ·CWT T (a,b) Among them, CWT fusion (a, b) are comprehensive time-frequency characteristic diagrams, which combine the characteristics of voltage, current and temperature signals; w V 、w I 、w T are the weights of the three signals respectively.

7. The transformer fault diagnosis method based on multi-sensor data fusion and transfer learning according to claim 5 is characterized in that: Based on steps S1 and S2, a fault diagnosis model is further constructed in step S3, including the following steps: a1: Transfer learning model selection: Introduce transfer learning technology, initialize the GoogLeNet network by loading the model weights pre-trained on a large image classification dataset (such as ImageNet), so that it has good feature extraction capabilities. The migration of the pre-trained model can be expressed as: in: represents the weights transferred from the pre-trained model; represents the randomly initialized weights, is the set of layers shared with the target task; l is the layer index in the network. b1: Model training: The details are as follows: b11: Comprehensive feature input: According to the above method, the collected signal is first subjected to continuous wavelet transform to generate a spectrum diagram, and then the importance weight of each modal signal is calculated by the entropy weight method, and its time-frequency features are weighted fused to generate a comprehensive time-frequency feature diagram CWT. fusion (a, b), let the comprehensive time-frequency feature map be F, and use it as the input of GoogLeNet. The input layer processes the feature map through normalization operation to unify the dimensions of different physical quantities. The normalization formula is as follows: F input =Normalize(F) Among them, Normalize is a normalization operation used to unify the dimensions of different physical quantities. The Normalize operation can be expressed as: Make sure all eigenvalues ​​are normalized to the range [0, 1] to facilitate model processing. b12: Data enhancement and expansion: Data enhancement technology can be used to artificially transform the spatial, brightness, contrast and other features of the samples so that the model can learn more diverse features and improve its performance on unseen data. The specific enhancement methods are as follows: b121: Random cropping and flipping: Random cropping refers to randomly selecting an area from the original image and cropping it. By changing the spatial distribution of the image, the model can learn features from different perspectives and enhance its adaptability to different spatial positions. The mathematical expression is: F′=T crop，flip (F;θ) Among them, T crop，flip is the cropping and flipping operation function; θ is the random parameter of the augmentation operation. b122: Noise addition: Use Gaussian noise model, assume that the noise has a mean of 0 and a variance of σ 2 The mathematical representation of Gaussian noise is: Where F(x, y) is the pixel value of the original feature map at position (x, y); Indicates that the mean is 0 and the variance is σ 2 Gaussian noise; F′(x, y) is the feature map after adding noise; the intensity of the noise is controlled by σ, and a larger σ value will increase the impact of noise and simulate a more severe sensor environment. b123: Random grayscale adjustment: Grayscale adjustment is used to simulate different lighting conditions and contrast changes. The specific grayscale adjustment operation can be expressed as: F′(x, y)=F(x, y)·α+β Among them, α is a randomly generated contrast factor; β is a randomly generated brightness offset; F(x, y) is the pixel value of the original image, and F′(x, y) is the adjusted image pixel value. b124: Mathematical framework of data enhancement: All enhancement operations can be uniformly represented by the transformation function T, specifically: F′=T(F;θ) Where T is the data enhancement transformation function, including the above-mentioned cropping, flipping, noise addition and grayscale adjustment; θ is the random parameter of the enhancement operation, which may include cropping size, noise intensity σ, contrast factor α, etc. c1: Model training and optimization, the details are as follows: c11: Network structure adjustment: The final dropout layer (poo15-drop_7x7_s1) in the GoogLeNet network is replaced with a dropout layer with a probability of 0.6 to enhance the regularization effect and reduce the risk of overfitting. The last learnable layer (loss3-classifier) ​​of GoogLeNet is redesigned. This layer is used to map the features extracted by the network to class probabilities. To adapt to the voltage transformer fault classification task, it is replaced with a new fully connected layer. The number of filters is set to 4 to correspond to the 4 states of the voltage transformer operation. c12: Loss function: The cross entropy loss function is used as the model optimization target, and its formula is: Among them, N is the total number of samples; C is the total number of categories; y ij and are the true label and predicted probability respectively. c13: Iteration strategy: Set the initial learning rate η0 and introduce a learning rate decay strategy to dynamically adjust the learning rate to ensure that the model can converge more stably in the later stages of training. The decay formula is as follows: Among them, λt is the decay factor, and t is the current iteration number. d1: Cross validation: K-fold cross validation is used, the steps are as follows: d11: Divide the data set into K subsets; d12: Select K-1 subsets each time as training sets, and the remaining 1 subset as validation set; d13: Repeat K times and calculate the average performance index. The cross-validation model performance indicators include accuracy, precision, recall, etc.: Among them, TP, TN, FP and FN are the number of true positive, true negative, false positive and false negative samples respectively.

Images