Transformer fault diagnosis method and system based on deep learning and sensor fusion

Abstract

The invention discloses a transformer fault diagnosis method and system based on deep learning and sensor fusion. The method comprises the steps that original signals generated by a multi-source sensor installed on a transformer are collected; calculating a noise parameter of the transformer, and preprocessing the original signal based on the noise parameter to generate a preprocessing feature tensor; inputting the preprocessed feature tensor into a deep learning model with multiple channels, and modeling the preprocessed feature tensor in each channel to obtain corresponding domain features; in a fusion layer of the deep learning model, adaptively adjusting a fusion weight based on a noise parameter, fusing domain features of the preprocessed feature tensor, and generating a fused feature tensor; and determining the fault type or health state of the transformer based on the fused feature tensor and the noise parameter. The method still has high precision and real-time performance under a low signal-to-noise ratio, has a generalization capability of cross-equipment and cross-working-condition, and remarkably improves the reliability and applicability of intelligent operation and maintenance and state evaluation of the transformer.

Description

Technical Field

[0001] This invention relates to the field of data acquisition and processing technology, and in particular to a method and system for transformer fault diagnosis based on deep learning and sensor fusion. Background Technology

[0002] As a key piece of equipment in the power system, the operating status of transformers directly affects the safety and stability of the power grid. However, existing transformer fault diagnosis methods still have many shortcomings.

[0003] Traditional methods often rely on single sensor signals or manual experience rules, resulting in limited diagnostic basis and difficulty in comprehensively reflecting the actual operating status of equipment. Secondly, in real-world operating environments, sensor signals are susceptible to noise interference, especially under low signal-to-noise ratio conditions, significantly reducing the accuracy and stability of existing diagnostic methods. Due to differences in equipment type and operating conditions, existing diagnostic methods lack generalization ability, typically requiring remodeling or retraining for different devices, making it difficult to meet the universal application requirements across various fields. Furthermore, existing methods also have shortcomings in real-time performance and computational efficiency, making it difficult to simultaneously meet the requirements of high-precision diagnosis and online applications.

[0004] Therefore, there is an urgent need for a new transformer fault diagnosis technology that can effectively suppress noise interference and improve the reliability of diagnostic results with the support of multi-source sensor data. It should also have strong cross-device adaptability and take into account both real-time performance and computational efficiency, thereby achieving accurate judgment of transformer fault type and health status. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a transformer fault diagnosis method and system based on deep learning and sensor fusion, in order to solve the problems of insufficient utilization of multi-source signals, poor diagnostic reliability in noisy environments, insufficient cross-device adaptability, and limited real-time performance and computational efficiency.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A transformer fault diagnosis method based on deep learning and sensor fusion includes the following steps: S1, Acquire raw signals generated by at least two types of multi-source sensors installed on the transformer to be monitored; S2, calculate the noise parameters of the transformer to be monitored, and preprocess the original signal based on the noise parameters to generate a preprocessed feature tensor; S3, input the preprocessed feature tensor into a deep learning model with multiple channels, and model the preprocessed feature tensor in each channel to obtain the corresponding time domain features, frequency domain features and sensor domain features; S4, in the fusion layer of the deep learning model, the fusion weights are adaptively adjusted based on the noise parameters to fuse the time domain features, frequency domain features and sensor domain features of the preprocessed feature tensor to generate a fused feature tensor; S5. Based on the fused feature tensor and noise parameters, determine the fault type or health status of the transformer to be monitored.

[0007] Further, in step S2, the noise parameters of the transformer to be monitored are calculated, specifically including: The pixel-level noise map is calculated based on the local fluctuation and instantaneous anomaly at each time and frequency point using the time-frequency tensor of the original signal, as shown in the following expression: ; In the above formula, This is a pixel-level noise map; The volatility in the frequency dimension; Anomalies in the time dimension; This is a compression function; , These are weight parameters; Indexed by frequency and time; The channel-level noise index is calculated based on the mean of the pixel-level noise map across all time frequency points, as shown in the following expression: ; In the above formula, The noise level is the channel level. , These represent the total number of frequency points and the total number of time points, respectively. This represents summation across all time-frequency points.

[0008] Furthermore, the variability in the frequency dimension is determined by the frequency points. The variance of the power series at all time points is obtained by calculating the following expression: ; In the above formula, This represents the time-frequency power value. Representing frequency point Average power at all time points; The anomaly in the time dimension is characterized by comparing the deviation of the power at that point from the mean of the time window, as expressed below: ; In the above formula, Based on time points Average power of the window centered on the center To prevent constants with a denominator of zero.

[0009] Further, in step S2, the original signal is preprocessed based on the noise parameters to generate a preprocessed feature tensor, specifically including: Based on the noise parameters, threshold adjustment is performed on the original signal to adjust the filtering coefficient, and nonlinear filtering is performed on the original signal to obtain the filtered signal; The filtered signal is subjected to time-frequency transformation with variable resolution, and time-frequency features are generated. Signal features are extracted at multiple time or frequency scales, and the signal features are weighted in combination with the noise parameters to generate the preprocessed feature tensor.

[0010] Furthermore, when performing threshold adjustment on the original signal based on the noise parameters, the threshold is specifically adjusted based on the pixel-level noise map and channel-level noise index in the noise parameters to perform shrinkage processing on the original signal, resulting in a denoised signal, as expressed below: ; In the above formula, For denoising signals; The original signal; The baseline threshold; This is a threshold adjustment factor; It is a positive part function; It is a symbolic function; The noise level is the channel level. This is a pixel-level noise map.

[0011] Further, in step S4, when adaptively adjusting the fusion weights based on the noise parameters, the weighting coefficients of the multi-domain features of the preprocessed feature tensor are specifically adjusted based on the channel-level noise index in the noise parameters, and the fusion features are generated, as shown in the following expression: ; In the above formula, These are multi-domain features, including time domain, frequency domain, and sensor domain features; For feature domain indexing, This indicates accumulation over all feature domains; These are weighting coefficients; These are the domain sensitivity coefficients determined during the model training phase. Features of fusion; This is the channel-level noise figure.

[0012] Further, step S5 includes: The noise-aware classifier network receives the fusion feature tensor output by the fusion layer and the channel-level noise index in the noise parameters. The classification threshold or normalization function parameters of the noise-sensing classifier network are adjusted based on the channel-level noise index. The fused feature tensor is input into the noise-aware classifier network and the corresponding fault category or health status label is output. Specifically, when the channel-level noise index is high, the classification threshold or the temperature coefficient of the normalization function is increased; when the channel-level noise index is low, the classification threshold or the temperature coefficient of the normalization function is decreased.

[0013] Furthermore, the deep learning model includes a time-domain modeling channel, a frequency-domain modeling channel, and a sensor-domain modeling channel. The time-domain modeling channel includes a multi-layer self-attention network, the frequency-domain modeling channel includes a combination structure of convolutional layers and self-attention layers, and the sensor-domain modeling channel includes a self-attention layer with residual connections. The attention weights of the multi-layer self-attention network, the convolutional kernel weights or the number of attention heads of the combination structure, and the proportional coefficient of the residual connections of the self-attention layer are adjusted based on the channel-level noise index in the noise parameters.

[0014] Furthermore, the multi-source sensor includes two or more of the following: current sensor, voltage sensor, temperature sensor, vibration sensor, acoustic emission sensor, or gas sensor, and each sensor is used to synchronously acquire the original signal under a unified clock.

[0015] A transformer fault diagnosis system based on deep learning and sensor fusion, wherein the system applies the above-mentioned method, and the system includes: The sensor acquisition module is used to acquire raw signals generated by at least two types of multi-source sensors installed on the transformer to be monitored. The preprocessing module is used to calculate the noise parameters of the transformer to be monitored, and to preprocess the original signal based on the noise parameters to generate a preprocessed feature tensor. The deep learning modeling module is used to input the preprocessed feature tensor into a multi-channel deep learning model, and to model the preprocessed feature tensor in each channel to obtain the corresponding time domain features, frequency domain features and sensor domain features. The feature fusion module is used to adaptively adjust the fusion weights based on the noise parameters in the fusion layer of the deep learning model, and fuse the time domain features, frequency domain features and sensor domain features of the preprocessed feature tensor to generate a fused feature tensor. The fault identification module is used to determine the fault type or health status of the transformer under monitoring based on the fused feature tensor and noise parameters.

[0016] Compared with the prior art, the advantages of the present invention are as follows: This invention utilizes collaborative data sensing from multiple sensors to comprehensively reflect the characteristics of transformers and improve the completeness of fault information acquisition. Preprocessing the original signal using noise parameters effectively suppresses noise interference and enhances robustness in low signal-to-noise ratio environments. Adaptively adjusting the fusion weights of features from different domains during the deep learning model fusion process achieves dynamic balance between features from different domains, ensuring the accuracy and completeness of feature extraction. This allows the model to maintain stable feature extraction capabilities under varying noise conditions, overcoming the diagnostic accuracy degradation caused by fixed fusion in traditional methods. By simultaneously utilizing fused features and noise information for fault determination, the accuracy of transformer fault identification is significantly improved, enabling precise differentiation of multiple types of faults and health statuses. Attached Figure Description

[0017] Figure 1 This is a flowchart of a transformer fault diagnosis method based on deep learning and sensor fusion, according to an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the structure of a transformer fault diagnosis system based on deep learning and sensor fusion according to an embodiment of the present invention. Detailed Implementation

[0019] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0020] Example 1 like Figure 1 As shown, the transformer fault diagnosis method based on deep learning and sensor fusion in this embodiment includes the following steps: S1, Acquire raw signals generated by at least two types of multi-source sensors installed on the transformer to be monitored; S2, calculate the noise parameters of the transformer to be monitored, and preprocess the original signal based on the noise parameters to generate a preprocessed feature tensor; S3, input the preprocessed feature tensor into a deep learning model with multiple channels, and model the preprocessed feature tensor in each channel to obtain the corresponding time domain features, frequency domain features and sensor domain features; S4, in the fusion layer of the deep learning model, the fusion weights are adaptively adjusted based on the noise parameters to fuse the time domain features, frequency domain features and sensor domain features of the preprocessed feature tensor to generate a fused feature tensor; S5. Based on the fused feature tensor and noise parameters, determine the fault type or health status of the transformer to be monitored.

[0021] In a specific application embodiment, step S1 is described in detail as follows: Various types of sensors can be installed at different locations on the transformer to comprehensively monitor its operating status. For example, current and voltage sensors can be installed near the windings or core to collect electrical signals; temperature sensors can be installed at the tank or coils to acquire local temperature rise data; vibration sensors can be placed on the casing or support to monitor mechanical vibration; and gas sensors can be installed in the transformer's oil and gas chamber to detect the composition of decomposition gases. Furthermore, all raw signals are acquired synchronously under a unified clock. Through the control of the unified clock, these sensors can acquire data on the same time reference, ensuring consistency of multi-source signals in the time dimension, thereby avoiding feature misalignment problems caused by sampling delays.

[0022] In this embodiment, the multi-source sensor includes two or more of the following: current sensor, voltage sensor, temperature sensor, vibration sensor, acoustic emission sensor, or gas sensor. Each sensor is used to synchronously acquire the original signal under a unified clock.

[0023] In one feasible embodiment, a combination of current and temperature sensors can be used to achieve multi-source signal acquisition. The current sensor is used to acquire the load current signal in real time, and the temperature sensor is used to sense changes in operating temperature. The combination of the two can effectively reflect the state of the transformer under different load and environmental conditions. In another embodiment, a combination of vibration and acoustic emission sensors can be used to simultaneously monitor abnormal vibration of the mechanical structure and partial discharge of the insulation material. By combining multiple sensors, the method of this embodiment can flexibly adapt to different monitoring needs and application scenarios.

[0024] In this embodiment, step S2, calculating the noise parameters (including pixel-level noise map and channel-level noise index) of the transformer to be monitored, specifically includes: The pixel-level noise map is calculated based on the local fluctuation and instantaneous anomaly at each time and frequency point using the time-frequency tensor of the original signal, as shown in the following expression: (1) In the above formula, This is a pixel-level noise map; The volatility in the frequency dimension; Anomalies in the time dimension; This is a compression function; , These are weight parameters; Indexed by frequency and time; The channel-level noise index is calculated based on the mean of the pixel-level noise map across all time frequency points, as shown in the following expression: (2) In the above formula, The noise level is the channel level. , These represent the total number of frequency points and the total number of time points, respectively. This represents summation across all time-frequency points.

[0025] In this embodiment, the variability in the frequency dimension is determined by frequency points. The variance of the power series at all time points is obtained by calculating the following expression: (3) In the above formula, This represents the time-frequency power value. Representing frequency point The average power at all time points; the larger the variance, the stronger the fluctuation at that frequency point. The anomaly in the time dimension is characterized by comparing the deviation of the power at that point from the mean of the time window, as expressed below: (4) In the above formula, Based on time points Average power of the window centered on the center To prevent constants with zero denominators; the greater the anomaly, the more significant the difference between the point and its neighborhood.

[0026] In this embodiment, step S2, which involves preprocessing the original signal based on the noise parameters to generate a preprocessed feature tensor, specifically includes: Based on the noise parameters, threshold adjustment is performed on the original signal to adjust the filtering coefficient, and nonlinear filtering is performed on the original signal to obtain the filtered signal; The filtered signal is subjected to time-frequency transformation with variable resolution, and time-frequency features are generated. Signal features are extracted at multiple time or frequency scales, and the signal features are weighted in combination with the noise parameters to generate the preprocessed feature tensor.

[0027] In this embodiment, when performing threshold adjustment on the original signal based on the noise parameters, the threshold is specifically adjusted based on the pixel-level noise map and channel-level noise index in the noise parameters to perform shrinkage processing on the original signal, resulting in a denoised signal, as expressed below: (5) In the above formula, For denoising signals; The original signal; The baseline threshold; This is a threshold adjustment factor; It is a positive part function; It is a symbolic function; The noise level is the channel level. This is a pixel-level noise map.

[0028] In a specific application embodiment, step S2 is described in detail as follows: The raw signal first enters the preprocessing module, where adaptive filtering for noise reduction, variable resolution time-frequency transformation, and multi-scale feature extraction are performed to generate a preprocessed feature tensor. Specifically, this module uses adaptive filtering to denoise the signal, ensuring that effective components are retained even in noisy environments. Subsequently, a time-frequency transformation method is used to map the signal from the time domain to the time-frequency domain, and a variable resolution strategy is employed based on the characteristics of different frequency components to obtain more refined spectral information. Furthermore, a multi-scale feature extraction method captures features at different time and frequency scales, enabling the final feature representation to comprehensively reflect the multidimensional characteristics of the signal. This module calculates noise parameters, where pixel-level noise maps reflect the noise situation at local frequency points, and channel-level noise indices are aggregated to reflect the overall noise level, providing a reference for subsequent modeling and feature fusion.

[0029] The preprocessing module includes: An adaptive filtering unit is used to adjust the filtering coefficients based on the noise parameters and perform nonlinear filtering on the original signal. The time-frequency analysis unit is used to perform variable-resolution time-frequency transformation on the filtered signal and generate time-frequency features; The multi-scale feature extraction unit is used to extract signal features at multiple time scales or frequency scales, and to perform weighted processing on the signal features in combination with the noise parameters to generate the preprocessed feature tensor.

[0030] In one embodiment, the adaptive filtering unit may employ a filter structure that dynamically adjusts based on noise levels. When the noise intensity is high, the filter automatically increases the filter coefficient to enhance suppression capability; when the noise is weak, the filter appropriately reduces the filter intensity to avoid over-smoothing, thereby ensuring that signal characteristics are not weakened.

[0031] In one embodiment, the time-frequency analysis unit can be based on wavelet transform or short-time Fourier transform framework, and can adapt to different frequency components by adjusting the time window or bandwidth, thereby obtaining high time resolution in the high-frequency part and high frequency resolution in the low-frequency part, so as to ensure the effectiveness and integrity of the features.

[0032] In one embodiment, the multi-scale feature extraction unit can extract features from different scales by constructing multi-layer convolutional or multi-layer wavelet decomposition structures, and dynamically adjust the feature weights based on noise parameters. This reduces the contribution of features at noisy scales and increases their weights at scales with high signal-to-noise ratios, thereby generating a more robust preprocessed feature tensor.

[0033] In a specific application embodiment, step S3 is described in detail as follows: The preprocessed feature tensor can be divided into three parts, corresponding to time series data, spectral data, and multi-sensor data, respectively. The time domain modeling channel receives serialized signal features and performs time-series correlation modeling; the frequency domain modeling channel receives spectral features and captures frequency patterns; and the sensor domain modeling channel receives features from multiple sensors and analyzes the correlation between different sensors. This multi-channel structure allows for the full extraction of features across different dimensions, avoiding information loss caused by single-domain modeling.

[0034] The multi-channel deep learning model includes: The time-domain modeling channel is used to receive and model the time-series noise-sensing features corresponding to the preprocessed feature tensor. The frequency domain modeling channel is used to receive and model the corresponding spectral noise sensing features in the preprocessed feature tensor. The sensor domain modeling channel is used to receive and model the multi-sensor noise sensing features corresponding to the preprocessed feature tensor. The outputs of the time-domain modeling channel, frequency-domain modeling channel, and sensor-domain modeling channel are used as multi-domain feature inputs to the fusion layer.

[0035] In one embodiment, the time-domain modeling channel can employ a deep neural network based on a self-attention mechanism to capture long-term dependencies and dynamic change patterns, ensuring that effective features can still be extracted under complex time-series conditions.

[0036] In one embodiment, the frequency domain modeling channel can adopt a structure combining convolutional layers and self-attention layers. The convolutional layers are used to extract local frequency features, and the self-attention layers are used to capture global frequency patterns. Furthermore, the weights of the convolutional kernels or the number of attention heads can be adjusted based on the channel-level noise index NNN, thereby enhancing the spectrum modeling capability under different noise levels.

[0037] In one embodiment, the sensor domain modeling channel can employ a self-attention layer with residual connections. The residual structure helps alleviate the gradient vanishing problem in deep networks, while the self-attention mechanism can dynamically model the interactions between different sensors. Based on the channel-level noise index, the scaling factor of the residual connections can be further adjusted, enabling multi-sensor modeling to adapt to different noise levels.

[0038] In one embodiment, the multi-domain features output from the three channels can be combined into a multi-domain feature vector by concatenation or weighted summation, providing input for subsequent feature fusion steps.

[0039] In this embodiment, the deep learning model includes a time-domain modeling channel, a frequency-domain modeling channel, and a sensor-domain modeling channel. The time-domain modeling channel includes a multi-layer self-attention network, the frequency-domain modeling channel includes a combination structure of convolutional layers and self-attention layers, and the sensor-domain modeling channel includes a self-attention layer with residual connections. The attention weights of the multi-layer self-attention network, the convolutional kernel weights or the number of attention heads of the combination structure, and the proportional coefficient of the residual connections of the self-attention layer are adjusted based on the channel-level noise index in the noise parameters.

[0040] In one embodiment, the temporal modeling channel includes a multi-layer self-attention network and is based on the channel-level noise index. Adjusting attention weights; the frequency domain modeling channel includes a combination structure of convolutional layers and self-attention layers, and is based on the channel-level noise index. Adjust the convolution kernel weights or the number of attention heads; the sensor domain modeling channel includes a self-attention layer with residual connections, and is based on the channel-level noise index. Adjusting the scaling factor of the residual connection; the outputs of the time-domain modeling channel, frequency-domain modeling channel, and sensor-domain modeling channel are combined into a multi-domain feature vector by feature concatenation or weighting before entering the fusion layer. .

[0041] This constraint further clarifies the implementation of each modeling channel and its adaptive relationship with noise parameters. In practice, the channel-level noise index can dynamically influence the weight settings of each network layer. For example, when the noise level increases, the attention mechanism of the time-domain channel automatically adjusts its weights to reduce the importance of time-series features affected by noise; the frequency-domain channel enhances features in key frequency bands by adjusting the convolution kernel weights; and the sensor-domain channel strengthens the contribution of a stable sensor by adjusting the residual ratio. In this way, the multi-channel network can dynamically optimize feature extraction performance according to noise conditions.

[0042] In this embodiment, in step S4, when adaptively adjusting the fusion weights based on the noise parameters, the weighting coefficients of the multi-domain features of the preprocessed feature tensor are specifically adjusted based on the channel-level noise index in the noise parameters, and the fusion features are generated, as shown in the following expression: (6) In the above formula, These are multi-domain features, including time domain, frequency domain, and sensor domain features; For feature domain indexing, This indicates accumulation over all feature domains; These are weighting coefficients; These are the domain sensitivity coefficients determined during the model training phase. Features of fusion; This is the channel-level noise figure.

[0043] In a specific application embodiment, step S4 is described in detail as follows: The fusion layer integrates features from different domains output by a multi-channel deep learning model. Unlike conventional static weighting, the fusion layer in this embodiment dynamically adjusts the weights using a channel-level noise index: when the signal in a certain domain is significantly affected by noise, the system automatically reduces the weight of that domain's features in the fusion result; when the signal-to-noise ratio of a certain domain is high, its weight contribution is increased. This results in a more robust fused feature tensor that can adapt to different noise conditions.

[0044] The noise parameters include pixel-level noise maps calculated from the original time-frequency tensor. and channel-level noise figure The data is simultaneously input to the preprocessing module and the fusion layer; in the preprocessing module, based on the pixel-level noise map... With channel-level noise figure Threshold adjustment is performed on the input signal; in the fusion layer, based on the channel-level noise index... The weighting coefficients of multi-domain features are dynamically adjusted.

[0045] In one embodiment, a pixel-level noise map can be obtained by calculating the local fluctuations and anomalies of the signal in the frequency and time dimensions. Subsequently, a channel-level noise index is formed by averaging across all time frequency points. The preprocessing module utilizes... and The threshold is jointly determined, and nonlinear contraction is applied to the input signal to reduce noise components; the fusion layer then utilizes... By adjusting the weighting coefficients of different feature domains, the proportion of each domain feature in the fusion result can be dynamically controlled.

[0046] In a specific application embodiment, the calculation and application of the noise parameters involved in steps S2-S4 are as follows: (1) Calculation of pixel-level noise map and channel-level noise figure First, the local fluctuation of the original time-frequency tensor at each time point and frequency point is calculated. and transient anomaly The calculation is then performed using formulas (1) and (2) to obtain the pixel-level noise map. Compared to existing technologies that rely solely on a single noise index, this approach simultaneously considers frequency domain stability and temporal anomalies, thereby more accurately characterizing the noise level at different time and frequency points. and In this embodiment, the parameters serve as a measure of the overall noise level, used for adaptive adjustment in subsequent feature processing and fusion stages. Unlike existing methods, this embodiment employs a "pixel-level—channel-level" dual-layer noise characterization mechanism, avoiding the problem that a single global indicator cannot reflect local noise differences.

[0047] By frequency points The variance of the power sequence at all time points is obtained by calculating the variance, and the specific calculation method is shown in formula (3). It can be characterized by comparing the deviation between the power at this point and the average value of the time window. The specific calculation method is shown in formula (4).

[0048] It should be noted that those skilled in the art can also use other statistical measures (such as spectral entropy, kurtosis, residual energy, etc.) to calculate volatility and anomaly, and this embodiment does not limit this.

[0049] (2) Adaptive denoising in the preprocessing stage In the preprocessing module, pixel-level noise maps are used. and channel-level noise figure Adjusting the threshold for the input signal Perform shrinkage processing to obtain a denoised signal. The specific calculation method is shown in formula (5).

[0050] It is understandable that the threshold depends not only on the overall noise figure. It also incorporates pixel-level noise maps. This allows for the application of a stronger threshold at high noise points and a weaker threshold at low noise points. Compared with traditional fixed threshold denoising, this embodiment can achieve adaptive nonlinear denoising and improve feature fidelity under low signal-to-noise ratio conditions.

[0051] (3) Noise-aware weighting in the fusion layer In the fusion layer, based on the channel-level noise index Adjusting multi-domain features The weighting coefficients are used to generate fusion features. The specific calculation method is shown in formula (6). This formula ensures that the weight of the noise-sensitive domain is automatically reduced when the overall noise level is high, while the weight balance is maintained when the noise level is low. Compared with existing fixed weighting or simple splicing fusion methods, this embodiment can realize dynamic weighted fusion of noise perception, thereby significantly improving the stability of diagnostic results in complex noise environments.

[0052] Through the above formula chain, this embodiment introduces a noise adaptive mechanism throughout the entire process of "noise parameter calculation - preprocessing denoising - feature fusion": Pixel-level noise map Provides noise characterization at a local time-frequency scale; Channel-level noise figure Provides a metric for the overall noise level; Noise-based adaptive threshold shrinkage is implemented in the preprocessing stage, and noise-based feature weighting is implemented in the fusion stage.

[0053] Compared with existing technologies, this embodiment not only maintains diagnostic accuracy in low signal-to-noise ratio environments, but also has stronger cross-domain generalization ability, significantly improving the engineering applicability of transformer fault diagnosis.

[0054] In this embodiment, step S5 includes: The noise-aware classifier network receives the fusion feature tensor output by the fusion layer and the channel-level noise index in the noise parameters. The classification threshold or normalization function parameters of the noise-sensing classifier network are adjusted based on the channel-level noise index. The fused feature tensor is input into the noise-aware classifier network and the corresponding fault category or health status label is output. Specifically, when the channel-level noise index is high, the classification threshold or the temperature coefficient of the normalization function is increased; when the channel-level noise index is low, the classification threshold or the temperature coefficient of the normalization function is decreased.

[0055] In a specific application embodiment, step S5 is described in detail as follows: Fusion Feature Tensor After dynamic weighting, the results are input to the fault discrimination module (specifically, a noise perception classifier network) to obtain the score vector for each candidate fault category. This module receives not only the fused features but also the channel-level noise index. This allows the discrimination process to adaptively adjust based on signal and noise levels. At the output layer, a temperature-coefficient adjustable softmax normalization function is used to convert the score vector into a probability distribution for each candidate fault category. The candidate fault category with the highest probability is selected as the fault type, or the probability corresponding to the healthy state in the probability distribution is used as the health state label. Finally, the fault discrimination module outputs the specific fault type (such as winding short circuit, partial discharge, etc.) or health state label.

[0056] The fault detection module includes a noise perception classifier network, which receives the fusion features output by the fusion layer. and channel-level noise figure And based on the channel-level noise index Adjust the classification threshold or normalization function parameters to output fault category or health status label.

[0057] In one embodiment, the noise-aware classifier network can employ a standard classification structure, such as a multilayer perceptron or other feedforward neural network. Its key innovation lies in the fact that the classifier not only uses fused features but also leverages channel-level noise exponents. To adjust the classification process. For example, when When the threshold is high, the classifier can increase the threshold or adjust the normalization function parameters to avoid misclassification caused by noise; when When the threshold is low, the classifier maintains a normal threshold to ensure classification sensitivity.

[0058] The noise perception classifier network includes a multilayer perceptron structure. The multilayer perceptron structure employs a softmax normalization function at the output layer. The temperature coefficient of the softmax normalization function is determined by the channel-level noise index. Dynamic adjustment.

[0059] In one implementation, the noise-aware classifier network consists of several fully connected layers, and a softmax function is used in the output layer to transform the output into a probability distribution for each class. Unlike traditional methods, the softmax function in this embodiment incorporates a temperature coefficient during computation, which varies with the channel-level noise index. It adjusts dynamically according to changes. When When the temperature coefficient is large, the probability distribution becomes smoother, reducing the classifier's oversensitivity to high-noise inputs; when When the temperature coefficient is smaller, the classifier becomes less sensitive to feature differences, thereby improving the discrimination accuracy under low noise conditions.

[0060] Compared with the prior art, the present invention has the following beneficial effects: This invention ensures consistency across different signal sources in the time dimension through synchronous acquisition from multiple types of sensors, avoiding feature misalignment caused by asynchronous sampling. In the preprocessing stage, adaptive filtering and noise reduction, variable resolution time-frequency transformation, and multi-scale feature extraction are combined to enhance the integrity of features across multiple levels and scales. Furthermore, the joint calculation of pixel-level noise maps and channel-level noise indices effectively improves the system's robustness in low signal-to-noise ratio environments. The multi-channel deep learning model models the time, frequency, and sensor domains separately, comprehensively capturing temporal dependencies, spectral patterns, and cross-sensor correlations, ensuring the accuracy and completeness of feature extraction. In the fusion layer, the weighting coefficients are adaptively adjusted based on the channel-level noise index, achieving a dynamic balance of features from different domains and overcoming the diagnostic accuracy degradation caused by fixed fusion in traditional methods. The noise-sensing classifier utilizes both fused features and noise information in the output stage to accurately distinguish fault types and health states.

[0061] Example 2 like Figure 2 As shown, this embodiment provides a transformer fault diagnosis system based on deep learning and sensor fusion. The system is applied to the method described above and includes: The sensor acquisition module is used to acquire raw signals generated by at least two types of multi-source sensors installed on the transformer to be monitored. The preprocessing module is used to calculate the noise parameters of the transformer to be monitored, and to preprocess the original signal based on the noise parameters to generate a preprocessed feature tensor. The deep learning modeling module is used to input the preprocessed feature tensor into a multi-channel deep learning model, and to model the preprocessed feature tensor in each channel to obtain the corresponding time domain features, frequency domain features and sensor domain features. The feature fusion module is used to adaptively adjust the fusion weights based on the noise parameters in the fusion layer of the deep learning model, and fuse the time domain features, frequency domain features and sensor domain features of the preprocessed feature tensor to generate a fused feature tensor. The fault identification module is used to determine the fault type or health status of the transformer under monitoring based on the fused feature tensor and noise parameters.

[0062] In summary, the transformer fault diagnosis method and system based on deep learning and sensor fusion proposed in this invention constructs a complete diagnostic chain from multi-source data acquisition, signal preprocessing, feature modeling to result discrimination, forming a systematic and implementable solution. This solution maintains stable performance under low signal-to-noise ratio and cross-operating condition conditions, possesses good generalization ability and real-time performance, and significantly improves the reliability of transformer operating status monitoring. Due to its technical completeness and adaptability, this invention has broad engineering application prospects and industrialization value in intelligent operation and maintenance, condition assessment, and fault prediction of power equipment.

[0063] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0064] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process. Furthermore, the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functionality involved.

[0065] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A transformer fault diagnosis method based on deep learning and sensor fusion, characterized in that, Including the following steps: S1, Acquire raw signals generated by at least two types of multi-source sensors installed on the transformer to be monitored; S2, calculate the noise parameters of the transformer to be monitored, and preprocess the original signal based on the noise parameters to generate a preprocessed feature tensor; S3, input the preprocessed feature tensor into a deep learning model with multiple channels, and model the preprocessed feature tensor in each channel to obtain the corresponding time domain features, frequency domain features and sensor domain features; S4, in the fusion layer of the deep learning model, the fusion weights are adaptively adjusted based on the noise parameters to fuse the time domain features, frequency domain features and sensor domain features of the preprocessed feature tensor to generate a fused feature tensor; S5. Based on the fused feature tensor and noise parameters, determine the fault type or health status of the transformer to be monitored.

2. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 1, characterized in that, In step S2, the noise parameters of the transformer to be monitored are calculated, specifically including: The pixel-level noise map is calculated based on the local fluctuation and instantaneous anomaly at each time and frequency point using the time-frequency tensor of the original signal, as shown in the following expression: ； In the above formula, This is a pixel-level noise map; The volatility in the frequency dimension; Anomalies in the time dimension; This is a compression function; , These are weight parameters; Indexed by frequency and time; The channel-level noise index is calculated based on the mean of the pixel-level noise map across all time frequency points, as shown in the following expression: ； In the above formula, The noise level is the channel level. , These represent the total number of frequency points and the total number of time points, respectively. This represents summation across all time-frequency points.

3. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 2, characterized in that, The variability in the frequency dimension is determined by the frequency points. The variance of the power series at all time points is obtained by calculating the following expression: ； In the above formula, This represents the time-frequency power value. Representing frequency point Average power at all time points; The anomaly in the time dimension is characterized by comparing the deviation of the power at that point from the mean of the time window, as expressed below: ； In the above formula, Based on time points Average power of the window centered on the center To prevent constants with a denominator of zero.

4. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 1, characterized in that, In step S2, the original signal is preprocessed based on the noise parameters to generate a preprocessed feature tensor, specifically including: Based on the noise parameters, threshold adjustment is performed on the original signal to adjust the filtering coefficient, and nonlinear filtering is performed on the original signal to obtain the filtered signal; The filtered signal is subjected to time-frequency transformation with variable resolution, and time-frequency features are generated. Signal features are extracted at multiple time or frequency scales, and the signal features are weighted in combination with the noise parameters to generate the preprocessed feature tensor.

5. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 4, characterized in that, When performing threshold adjustment on the original signal based on the noise parameters, the threshold is specifically adjusted based on the pixel-level noise map and channel-level noise index in the noise parameters to perform shrinkage processing on the original signal, resulting in a denoised signal, as expressed below: ； In the above formula, For denoising signals; The original signal; The baseline threshold; This is a threshold adjustment factor; It is a positive part function; It is a symbolic function; The noise level is the channel level. This is a pixel-level noise map.

6. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 1, characterized in that, In step S4, when adaptively adjusting the fusion weights based on the noise parameters, the weighting coefficients of the multi-domain features of the preprocessed feature tensor are specifically adjusted based on the channel-level noise index in the noise parameters, and the fusion features are generated, as shown in the following expression: ； In the above formula, These are multi-domain features, including time domain, frequency domain, and sensor domain features; For feature domain indexing, This indicates accumulation over all feature domains; These are weighting coefficients; These are the domain sensitivity coefficients determined during the model training phase. Features of fusion; This is the channel-level noise figure.

7. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 1, characterized in that, Step S5 includes: The noise-aware classifier network receives the fusion feature tensor output by the fusion layer and the channel-level noise index in the noise parameters. The classification threshold or normalization function parameters of the noise-sensing classifier network are adjusted based on the channel-level noise index. The fused feature tensor is input into the noise-aware classifier network and the corresponding fault category or health status label is output. Specifically, when the channel-level noise index is high, the classification threshold or the temperature coefficient of the normalization function is increased; when the channel-level noise index is low, the classification threshold or the temperature coefficient of the normalization function is decreased.

8. The transformer fault diagnosis method based on deep learning and sensor fusion according to any one of claims 1-7, characterized in that, The deep learning model includes a time-domain modeling channel, a frequency-domain modeling channel, and a sensor-domain modeling channel. The time-domain modeling channel includes a multi-layer self-attention network, the frequency-domain modeling channel includes a combination structure of convolutional layers and self-attention layers, and the sensor-domain modeling channel includes a self-attention layer with residual connections. The attention weights of the multi-layer self-attention network, the convolutional kernel weights or the number of attention heads of the combination structure, and the scaling factor of the residual connections of the self-attention layer are adjusted based on the channel-level noise index in the noise parameters.

9. The transformer fault diagnosis method based on deep learning and sensor fusion according to any one of claims 1-7, characterized in that, The multi-source sensor includes two or more of the following: current sensor, voltage sensor, temperature sensor, vibration sensor, acoustic emission sensor, or gas sensor. Each sensor is used to synchronously acquire the original signal under a unified clock.

10. A transformer fault diagnosis system based on deep learning and sensor fusion, the system being applied to the method described in any one of claims 1-9, characterized in that, The system includes: The sensor acquisition module is used to acquire raw signals generated by at least two types of multi-source sensors installed on the transformer to be monitored. The preprocessing module is used to calculate the noise parameters of the transformer to be monitored, and to preprocess the original signal based on the noise parameters to generate a preprocessed feature tensor. The deep learning modeling module is used to input the preprocessed feature tensor into a multi-channel deep learning model, and to model the preprocessed feature tensor in each channel to obtain the corresponding time domain features, frequency domain features and sensor domain features. The feature fusion module is used to adaptively adjust the fusion weights based on the noise parameters in the fusion layer of the deep learning model, and fuse the time domain features, frequency domain features and sensor domain features of the preprocessed feature tensor to generate a fused feature tensor. The fault identification module is used to determine the fault type or health status of the transformer under monitoring based on the fused feature tensor and noise parameters.

Images