Transformer fault diagnosis method based on combination of multi-wavelet convolution and channel attention

Abstract

The invention discloses a transformer fault diagnosis method based on combination of multi-wavelet convolution and channel attention, and the method comprises the steps: enabling a collected time domain signal to be converted into an equivalent time-frequency diagram through the multi-wavelet convolution operation, and carrying out the self-adaption searching of a specific frequency band feature for a diagnosis object through a data driving mode; a digital signal processing method and a deep learning method are effectively combined, attention and a full connection layer are used for carrying out feature extraction and representation on faults to finally achieve the classification purpose, the method does not depend on expert experience, complex data preprocessing and feature extraction are not needed, the fault recognition accuracy is high, training parameters are few, the algorithm complexity is low, and the method is suitable for large-scale popularization and application. And the fault identification speed is high.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault detection, and in particular to a transformer fault diagnosis method based on multi-wavelet convolution combined with channel attention mechanism. Background Technology

[0002] As a critical hub in the power system, the safe and stable operation of transformers is of paramount importance. During long-term operation, transformers are exposed to complex and diverse environments and operating conditions, making them prone to problems such as material deterioration, fatigue wear of local components, and decreased insulation performance. This can induce defects such as abnormal vibration and partial discharge, and in severe cases, lead to tripping and other faults, causing large-scale power outages. Due to factors such as magnetostriction, when there are potential mechanical defects in a transformer, the vibration state of components such as the core windings changes, often accompanied by abnormal acoustic signatures. Therefore, conducting fault diagnosis research based on transformer acoustic signature data is of great significance for timely identification of potential transformer defects and prevention of large-scale power outages.

[0003] For the problem of transformer fault diagnosis, researchers both domestically and internationally have conducted extensive theoretical and applied research. Currently, there are two main technical approaches: expert experience methods based on digital signal processing and data-driven artificial intelligence methods.

[0004] The main process of expert-based signal processing methods involves first preprocessing the acquired signal, including noise reduction and source selection. Next, the preprocessed signal undergoes transformation and feature extraction operations, including frequency domain transformation, time-frequency domain transformation, fundamental and harmonic frequency analysis, harmonic component analysis, and kurtosis extraction. This process relies heavily on expert experience for manual parameter tuning and method selection, making it complex and costly. Finally, the extracted features are combined into feature vectors or matrices for pattern recognition and diagnosis. While these methods are theoretically sound and highly reliable, they require diagnostic personnel to possess extensive theoretical knowledge and digital signal processing skills, increasing the cost of routine transformer noise analysis and identification.

[0005] The shortcomings of existing technologies are as follows: First, intelligent diagnostic algorithms employ deep learning algorithms such as convolution and attention, and their model reasoning process lacks interpretability, failing to provide reliable theoretical basis for on-site inspection personnel. Second, feature extraction and classification selection after feature representation are also crucial. Currently, methods based on expert experience consume significant manpower, relying on experts to observe feature maps for diagnosis, making reliable intelligentization difficult to achieve.

[0006] In conclusion, how to construct interpretable intelligent diagnostic models and improve the accuracy and efficiency of fault diagnosis has become an urgent technical problem to be solved. Summary of the Invention

[0007] To solve the above-mentioned technical problems, the present invention adopts the following solution.

[0008] A transformer fault diagnosis method based on multi-wavelet convolution combined with channel attention mechanism includes the following steps: S1: Collect the acoustic signature signal of the transformer and obtain time-domain acoustic signal samples; S2: Divide the time-domain acoustic signal samples into a training set and a validation set; S3: Use multiple wavelet convolutional layers to extract features from the temporal acoustic signals of the training set; The multi-wavelet convolutional layer includes three wavelet convolutional kernels with different waveform features: Laplace wavelet convolutional kernel, Morlet wavelet convolutional kernel and MexHat wavelet convolutional kernel, with each wavelet convolutional kernel corresponding to one channel; The trainable parameters of the Laplace wavelet convolution kernel include: frequency parameter and attenuation coefficient; the trainable parameters of the Morlet wavelet convolution kernel include: frequency parameter and phase parameter; and the trainable parameters of the MexHat wavelet convolution kernel include: scale parameter. S4: Update the trainable parameters using a data-driven approach to convert the time-domain acoustic signal into a multi-channel equivalent time-frequency diagram; S5: The multi-channel equivalent time-frequency map is weighted using a channel attention mechanism to obtain a weighted feature map; S6: Input the weighted feature map into the fully connected layer and output the fault classification results of the transformer.

[0009] Furthermore, in step S1, acoustic fingerprint signals from the surface of the transformer tank are collected by an acoustic sensor. The horizontal distance between the acoustic sensor and the surface of the transformer tank is 0.5m to 2m, and the height is 1 / 3 to 2 / 3 of the height of the transformer tank.

[0010] Furthermore, in step S2, the time-domain acoustic signal samples are divided into a training set and a validation set in an 8:2 ratio.

[0011] Furthermore, in S3, the Laplace wavelet convolution kernel uses the real part of the Laplace wavelet to process the voiceprint signal, and the Morlet wavelet convolution kernel uses the real part of the Morlet wavelet to process the voiceprint signal.

[0012] Furthermore, the transformer acoustic signature signals collected in S1 include acoustic signature signals under normal conditions, acoustic signature signals under winding deformation fault conditions, acoustic signature signals under discharge fault conditions, and acoustic signature signals under DC bias fault conditions.

[0013] Furthermore, the MexHat wavelet convolution kernel is used to detect transient pulse signals, and the Morlet wavelet convolution kernel and Laplace wavelet convolution kernel are used to detect decaying pulse signals.

[0014] This invention also provides a transformer fault diagnosis system based on multi-wavelet convolution combined with channel attention mechanism, comprising: The acoustic signature acquisition module is used to acquire the acoustic signature signal of the transformer and obtain time-domain acoustic signal samples. The data preprocessing module is used to divide the acquired time-domain acoustic signal samples into training and validation sets; The multi-wavelet convolution module, including the Laplace wavelet convolution unit, the Morlet wavelet convolution unit, and the MexHat wavelet convolution unit, is used to extract features from time-domain acoustic signals and convert them into multi-channel equivalent time-frequency maps. The trainable parameters of the Laplace wavelet convolutional unit include frequency parameters and attenuation coefficients; the trainable parameters of the Morlet wavelet convolutional unit include frequency parameters and phase parameters; and the trainable parameters of the MexHat wavelet convolutional unit include scale parameters. The parameter update module is used to update the trainable parameters in a data-driven manner. The channel attention module is used to perform weighted processing on the multi-channel equivalent time-frequency map to obtain a weighted feature map; The classification output module is used to input the weighted feature map into the fully connected network and output the fault classification results of the transformer.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: 1. Improve fault diagnosis accuracy: By using three wavelet convolution kernels (Laplace, Morlet, MexHat) with different waveform features for multidimensional feature extraction, the time-frequency features of transformer acoustic signals can be captured more comprehensively. Combined with the channel attention mechanism to adaptively select the optimal feature representation, the accuracy can reach more than 90% after one round of training, which significantly improves the accuracy of fault diagnosis.

[0016] 2. Reduce reliance on expert experience: By using an end-to-end deep learning approach, feature extraction and classification decisions are unified into a single model, avoiding the complex manual feature engineering and reliance on expert experience for feature selection found in traditional methods. This reduces diagnostic costs and improves diagnostic efficiency. Attached Figure Description

[0017] The accompanying drawings illustrate exemplary embodiments of the invention and, together with the description thereof, serve to explain the principles of the invention. These drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification.

[0018] Figure 1 This is a transformer fault diagnosis method based on multi-wavelet convolution combined with channel attention mechanism in one embodiment of the present invention; Figure 2 is a Laplace wavelet waveform diagram under different parameters in one embodiment of the present invention; Figure 2(a) Laplace wavelet waveform diagram under different frequencies, Figure 2(b) Laplace wavelet waveform diagram under different damping ratios; Figure 3 These are Morlet wavelet waveforms with different parameters in one embodiment of the present invention; Figure 4 These are Mexhat wavelet waveforms with different parameters in one embodiment of the present invention; Figure 5 This is a diagram showing the model training effect in one embodiment of the present invention; Figure 6 This is a schematic diagram of a specific single-channel Laplace convolution process in one embodiment of the present invention; Figure 7 This is a multi-channel layer in one embodiment of the present invention; Figure 8 This is a confusion matrix in one embodiment of the present invention; Figure 9 This is a schematic diagram of an acoustic signal acquisition system in one embodiment of the present invention. Detailed Implementation

[0019] The following is in conjunction with the appendix Figures 1 to 9 The present invention will be further described in detail below with reference to the embodiments. It is to be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present invention are shown in the accompanying drawings.

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. The technical solution of this invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise stated, the exemplary embodiments / exemplifications shown are to be understood as providing exemplary features of various details that provide ways in which the technical concept of the invention can be implemented in practice. Therefore, unless otherwise stated, the features of the various embodiments / exemplifications may be additionally combined, separated, interchanged and / or rearranged without departing from the technical concept of the invention.

[0022] The use of crosshairs and / or shading in the accompanying drawings is generally used to clarify the boundaries between adjacent components. Thus, unless otherwise stated, the presence or absence of crosshairs or shading does not convey or indicate any preference or requirement for the specific material, material properties, dimensions, proportions, commonalities between the illustrated components, or any other characteristics, properties, etc., of the components. Furthermore, in the accompanying drawings, the dimensions and relative dimensions of components may be exaggerated for clarity and / or descriptive purposes. When exemplary embodiments can be implemented differently, a specific process sequence may be performed in a different order than that described. For example, two consecutively described processes may be performed substantially simultaneously or in the reverse order of their description. Furthermore, the same reference numerals denote the same components.

[0023] When a component is referred to as being "on" or "above" another component, "connected to," or "joined to" another component, the component may be directly on, directly connected to, or directly joined to the other component, or there may be intermediate components. However, when a component is referred to as being "directly on" another component, "directly connected to," or "directly joined to" another component, there are no intermediate components. Therefore, the term "connection" can refer to a physical connection, an electrical connection, etc., and may or may not have intermediate components.

[0024] For descriptive purposes, the present invention may use spatial relative terms such as “below,” “under,” “below,” “down,” “above,” “above,” “higher,” and “side (e.g., in a “sidewall”)” to describe the relationship between one component and another component as shown in the accompanying drawings. In addition to the orientations depicted in the drawings, the spatial relative terms are also intended to encompass different orientations of the device during use, operation, and / or manufacture. For example, if the device in the drawings is flipped, a component described as “below” or “under” another component or feature would subsequently be positioned “above” said other component or feature. Thus, the exemplary term “below” can encompass both “above” and “below” orientations. Furthermore, the device may be otherwise positioned (e.g., rotated 90 degrees or in other orientations), thus interpreting the spatial relative descriptive terms used herein accordingly.

[0025] The terminology used herein is for the purpose of describing particular embodiments and is not intended to be limiting. As used herein, unless the context clearly indicates otherwise, the singular forms “a” and “the” are intended to include the plural forms as well. Furthermore, when the terms “comprising” and / or “including” and variations thereof are used in this specification, it indicates the presence of the stated features, integrals, steps, operations, parts, components, and / or groups thereof, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, parts, components, and / or groups thereof. It should also be noted that, as used herein, the terms “substantially,” “about,” and other similar terms are used as approximate terms rather than as terms of degree, thus explaining the inherent biases in measurements, calculated values, and / or provided values ​​that would be recognized by one of ordinary skill in the art.

[0026] In one embodiment, the present invention provides a transformer fault diagnosis method based on multi-wavelet convolution combined with channel attention mechanism, comprising the following steps: S1: Collect the acoustic signature signal of the transformer and obtain time-domain acoustic signal samples; S2: Divide the time-domain acoustic signal samples into a training set and a validation set; S3: Use multiple wavelet convolutional layers to extract features from the temporal acoustic signals of the training set; The multi-wavelet convolutional layer includes three wavelet convolutional kernels with different waveform features: Laplace wavelet convolutional kernel, Morlet wavelet convolutional kernel and MexHat wavelet convolutional kernel, with each wavelet convolutional kernel corresponding to one channel; The trainable parameters of the Laplace wavelet convolution kernel include: frequency parameter and attenuation coefficient; the trainable parameters of the Morlet wavelet convolution kernel include: frequency parameter and phase parameter; and the trainable parameters of the MexHat wavelet convolution kernel include: scale parameter. S4: Update the trainable parameters using a data-driven approach to convert the time-domain acoustic signal into a multi-channel equivalent time-frequency diagram; S5: The multi-channel equivalent time-frequency map is weighted using a channel attention mechanism to obtain a weighted feature map; S6: Input the weighted feature map into the fully connected layer and output the fault classification results of the transformer.

[0027] This invention transforms the acquired time-domain signal into an equivalent time-frequency map through multiple wavelet convolution operations, and adaptively searches for frequency band features specific to the diagnostic object through a data-driven approach. It effectively combines digital signal processing methods and deep learning methods, and uses attention and fully connected layers to extract and represent fault features to ultimately achieve classification. This method does not rely on expert experience, does not require complex data preprocessing and feature extraction, has a high accuracy rate in fault identification, requires few training parameters, has low algorithm complexity, and a fast fault identification speed.

[0028] Furthermore, in step S1, acoustic fingerprint signals from the surface of the transformer tank are collected by an acoustic sensor. The horizontal distance between the acoustic sensor and the surface of the transformer tank is 0.5m to 2m. Under the premise of ensuring stable acquisition, a suitable sensor height is selected, which is 1 / 3 to 2 / 3 of the height of the transformer tank.

[0029] Furthermore, in step S2, the time-domain acoustic signal samples are divided into a training set and a validation set in an 8:2 ratio.

[0030] Furthermore, in S3, the Laplace wavelet convolution kernel uses the real part of the Laplace wavelet to process the voiceprint signal, and the Morlet wavelet convolution kernel uses the real part of the Morlet wavelet to process the voiceprint signal.

[0031] Furthermore, the transformer acoustic signature signals collected in S1 include acoustic signature signals under normal conditions, acoustic signature signals under winding deformation fault conditions, acoustic signature signals under discharge fault conditions, and acoustic signature signals under DC bias fault conditions.

[0032] Furthermore, the MexHat wavelet convolution kernel is used to detect transient pulse signals, and the Morlet wavelet convolution kernel and Laplace wavelet convolution kernel are used to detect decaying pulse signals.

[0033] In another embodiment, the present invention also provides a transformer fault diagnosis system based on multi-wavelet convolution combined with channel attention mechanism, comprising: The acoustic signature acquisition module is used to acquire the acoustic signature signal of the transformer and obtain time-domain acoustic signal samples. The data preprocessing module is used to divide the acquired time-domain acoustic signal samples into training and validation sets; The multi-wavelet convolution module, including the Laplace wavelet convolution unit, the Morlet wavelet convolution unit, and the MexHat wavelet convolution unit, is used to extract features from time-domain acoustic signals and convert them into multi-channel equivalent time-frequency maps. The trainable parameters of the Laplace wavelet convolutional unit include frequency parameters and attenuation coefficients; the trainable parameters of the Morlet wavelet convolutional unit include frequency parameters and phase parameters; and the trainable parameters of the MexHat wavelet convolutional unit include scale parameters. The parameter update module is used to update the trainable parameters in a data-driven manner. The channel attention module is used to perform weighted processing on the multi-channel equivalent time-frequency map to obtain a weighted feature map; The classification output module is used to input the weighted feature map into the fully connected network and output the fault classification results of the transformer.

[0034] In another embodiment, the present invention provides a transformer fault diagnosis method based on multi-wavelet convolution combined with channel attention mechanism, comprising the following steps: S1: Collect the acoustic signature signal of the transformer and obtain time-domain acoustic signal samples; Specifically, acoustic fingerprint signals are collected from the surface of the transformer tank using acoustic sensors to obtain time-domain acoustic signal samples; the acoustic fingerprint signals include acoustic fingerprint signals under normal conditions, acoustic fingerprint signals under winding deformation fault conditions, acoustic fingerprint signals under discharge fault conditions, and acoustic fingerprint signals under DC bias fault conditions. S2: Divide the time-domain acoustic signal samples into a training set and a validation set; Specifically, the training set and validation set are set up in an 8:2 ratio; S3: Use multiple wavelet convolutional layers to extract features from the temporal acoustic signals of the training set; The multi-wavelet convolutional layer includes three wavelet convolutional kernels with different waveform features: Laplace wavelet convolutional kernel, Morlet wavelet convolutional kernel and MexHat wavelet convolutional kernel, with each wavelet convolutional kernel corresponding to one channel; The trainable parameters of the Laplace wavelet convolution kernel include: frequency parameter f and attenuation coefficient. The Laplace wavelet kernel function is defined as follows: (1) in, It is the attenuation coefficient, which determines the rate of Laplace attenuation; f is the frequency parameter. The real part of the Laplace wavelet is used to process the voiceprint signal for detecting attenuation pulse signals; The trainable parameters of the Morlet wavelet convolution kernel include: frequency parameter and phase parameter; the Morlet wavelet kernel function is defined as: (2) Where f is the frequency parameter, It is a phase parameter; The real part of the Morlet wavelet is used to process the voiceprint signal for detecting attenuation pulse signals; The trainable parameters of the MexHat wavelet convolution kernel include: scale parameter. The Mexhat wavelet kernel function is defined as follows: (3) Used to detect transient pulse signals.

[0035] Specifically, this patent sets the truncated signal length to 2560, the kernel size to 1025, the stride to 32, and the number of channels for each wavelet convolution to 32. The specific convolution operation for a single wavelet convolution is as follows:

[0036] In the formula, h c [ m ] is by the first c The output feature vector corresponding to the i-th wavelet function is the th... m One value, L in For the input sound signal, g c For the first c Several wavelet functions. The amplitude of the convolution result represents the similarity to the convolution kernel, indicating the frequency band distribution in the signal. The specific process is as follows: Figure 6 As shown.

[0037] This patent's convolution is based on one-dimensional signal convolution. After performing the above operation on all three wavelet functions, the two-dimensional time-frequency matrix of each wavelet is concatenated again to form a multi-channel layer. This layer ultimately serves as the multi-channel input feature matrix for subsequent models. (Refer to...) Figure 7 .

[0038] S4: Update the trainable parameters using a data-driven approach to convert the time-domain acoustic signal into a multi-channel equivalent time-frequency diagram; S5: The multi-channel equivalent time-frequency map is weighted using a channel attention mechanism to obtain a weighted feature map; S6: Input the weighted feature map into the fully connected layer, and output the transformer fault classification results. The classification results for each category after model processing are as follows: Figure 8 As shown.

[0039] In S1, acoustic signal acquisition equipment is used to acquire acoustic signals from the transformer under normal and preset fault conditions, respectively. Figure 9 As shown; In S4, during the training process, the waveform of the wavelet function is continuously updated with the data-driven approach. The learnable parameters of the Laplace wavelet are frequency and damping ratio. The changes in the time-domain waveform and frequency-domain characteristics of different parameters are shown in Figures 2(a) to 2(b).

[0040] In this context, the Morlet wavelet sets learnable parameters as frequency and phase. The changes in time-domain waveforms and frequency-domain characteristics with different parameters are as follows: Figure 3 As shown.

[0041] In the mexhat wavelet, the learnable parameter is the scale parameter (frequency). The changes in the time-domain waveform and frequency-domain characteristics of different parameters are as follows: Figure 4 As shown.

[0042] The result after convolution is the equivalent time-frequency map of different wavelets at different frequencies. The two-dimensional time-frequency map of the multi-channel is used as the input of the channel attention. The channel attention can adaptively weight the acoustic signal of the power transformer according to different wavelet basis functions and search for the optimal wavelet basis function suitable for the power transformer.

[0043] An attention model is used to mine and represent features from the time-frequency graph, and finally a fully connected layer is used to output the classification category.

[0044] In actual training, the convergence speed of this patented model is extremely fast, referring to... Figure 5 After one round of training, the model achieves an accuracy rate of over 90% without overfitting, indicating that it can extract feature parameters that indicate health status. This patented model requires fewer training parameters, significantly saving computational resources without sacrificing recognition accuracy, making it highly practical in real-world applications.

[0045] In another embodiment, the present invention provides a transformer fault diagnosis method based on multi-wavelet convolution combined with channel attention mechanism, comprising the following steps: S1: Arrange acoustic fingerprint sensors around a transformer in normal health condition. The sensors are 1m horizontally and 1 / 2 of the height from the surface of the transformer tank to collect acoustic fingerprint data on the transformer's operating status. S2: Set up winding deformation fault, discharge fault and DC bias fault for the transformer, which correspond to the main fault categories of the transformer winding and iron core, and collect the soundprint of the transformer operating status. S3: Select and divide the collected signals into training and validation sets in an 8:2 ratio; S4: Wavelet transform can be obtained by calculating the inner product with the wavelet function. The wavelet kernel function family is defined as follows: (4) In the formula, s and These represent the scaling parameter and the translation parameter, respectively. The scaling parameter controls the length of the wavelet signal, causing it to expand or compress. A small scaling parameter means a short wavelet signal, which is sensitive to high-frequency signals. A large scaling parameter means the signal is stretched to detect low-frequency components. The translation parameter controls the translation speed of the wavelet function in the time domain.

[0046] Different types of transformer faults generate different vibration signals, including steady-state and transient signals. Transient signals are classified into three basic types: transient pulses, two-sided decaying pulses, and right-sided unilateral decaying pulses. Therefore, different wavelet basis functions can fully utilize their pulse characteristics to identify different transient signals.

[0047] This invention selects the mexhat wavelet kernel function to detect transient pulses; and selects the Morlet and Laplace wavelet functions to detect bilaterally decaying pulses and right-side unilaterally decaying pulses.

[0048] The Mexhat wavelet kernel function is defined as follows: (5) in, These are scale parameters that can be learned and updated.

[0049] The Morlet wavelet kernel function is defined as: (6) Where f is the frequency parameter, These are phase parameters, and they can all be updated through learning.

[0050] The Laplace wavelet kernel function is defined as follows: (7) in, It is the attenuation coefficient, which determines the rate of Laplace attenuation; f is the frequency parameter. The real parts of Morlet and Laplace wavelets are used to process the voiceprint signal.

[0051] Specifically, the data was acquired at 25600Hz, therefore, the number of data points selected during data slicing was 2560, or 100ms. For data with different sampling rates, the sampling rate was aligned to 25600Hz using either downsampling or upsampling methods to ensure spectral stability and accuracy during data processing.

[0052] Regarding the choice of convolution kernel size, to ensure good frequency resolution and range during data processing, and since the kernel length must be an odd number, a kernel size of 1025 was chosen; the stride was 32, and the number of channels was 32. Finally, the multi-channel wavelet convolutions were concatenated using the concat function.

[0053] The specific calculation formula is as follows: (8) The specific convolution process for one of the single-channel Laplace channels is as follows: Figure 6 As shown.

[0054]

[0055] In this embodiment, the learning rate is set to a maximum of 30 iterations and a batch size of 32 during training. To accelerate model convergence, different learning rates are applied to the sinc convolutional filtering and attention parts, with initial learning rates of 0.1 and 0.001 respectively. Dynamic learning rates are adaptively adjusted during training, using an exponential decay strategy with a multiplier of 0.99. Layer normalization and dropout regularization methods are introduced to avoid overfitting, and the final loss function is cross-entropy.

[0056] Based on the mathematical theory and methodology of convolution, the formal description of convolution is as follows: (9) h is the result of the convolution operation. Based on this convolution form, the wavelet transform can be designed as a neural network layer with learnable parameters and embedded into a standard CNN. Specifically, the scale parameter can be equivalent to frequency information as training parameters, while the translation parameter corresponds to the stride in the convolution operation and does not need to be trained. To update the parameters of the wavelet filter, backpropagation can be performed using the following formula: (10) In the formula For loss function For scale parameters The gradient; The loss function; This indicates the mapping to subsequent network layers; This is the output feature map of the current layer; This is a wavelet filter function; is the learning rate. Parameters can be updated using gradient descent.

[0057] S5: The channel attention mechanism focuses on the weighting of each channel in the input feature layer, adaptively calibrating the feature responses of each channel, which can significantly improve the performance of convolutional neural networks in accurately locating key features. The principle of the channel attention mechanism is shown in the figure below: Assuming the input is a multi-channel feature map First, deform A to ,calculate Matrix multiplication is performed, and finally, a softmax function is used to obtain the relationship matrix between each channel. (11) In the formula The influence relationships between different channels were measured. Finally, the output matrix is ​​obtained by multiplying the relationship matrices X and A and restoring them to their initial shape. : (12) Finally, the classification objective is achieved through a fully connected network.

[0058] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment / mode or example is included in at least one embodiment / mode or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.

[0059] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0060] Those skilled in the art should understand that the above embodiments are merely for illustrating the present invention and are not intended to limit the scope of the invention. Those skilled in the art can make other changes or modifications based on the above disclosure, and these changes or modifications still fall within the scope of the present invention.

Claims

1. A transformer fault diagnosis method based on multi-wavelet convolution combined with channel attention, characterized in that, The method comprises the following steps: S1: Collecting the voiceprint signal of the transformer to obtain a time-domain acoustic signal sample; S2: Dividing the time-domain acoustic signal sample into a training set and a validation set; S3: Extracting features of the time-domain acoustic signal of the training set by using a multi-wavelet convolution layer; S4: Updating trainable parameters in a data-driven manner to convert the time-domain acoustic signal into a multi-channel equivalent time-frequency graph; S5: Weighting the multi-channel equivalent time-frequency graph by using a channel attention mechanism to obtain a weighted feature graph; S6: Inputting the weighted feature graph into a fully connected layer to output a fault classification result of the transformer.

2. The method of claim 1, wherein, Preferably, in S1, the voiceprint signal of the transformer tank surface is collected by an acoustic sensor.

3. The method of claim 2, wherein, The horizontal distance between the acoustic sensor and the transformer tank surface is 0.5-2 m, and the height is 1 / 3-2 / 3 of the height of the transformer tank.

4. The method of claim 1, wherein, In S2, the time-domain acoustic signal sample is divided into the training set and the validation set according to a ratio of 8:

2.

5. The method of claim 1, wherein, In S3, the multi-wavelet convolution layer comprises three wavelet convolution kernels of different waveforms, i.e., a Laplace wavelet convolution kernel, a Morlet wavelet convolution kernel and a MexHat wavelet convolution kernel, each of which corresponds to a channel.

6. The method of claim 1, wherein, In S3, the Laplace wavelet convolution kernel processes the voiceprint signal by using the real part of the Laplace wavelet.

7. The method of claim 1, wherein, The transformer voiceprint signal collected in S1 comprises normal state voiceprint signals, winding deformation fault state voiceprint signals, discharge fault state voiceprint signals and direct current magnetic bias fault state voiceprint signals.

8. The method of claim 1, wherein, The MexHat wavelet convolution kernel is used to detect transient pulse signals.

9. The method of claim 1, wherein, The Morlet wavelet convolution kernel and the Laplace wavelet convolution kernel are used to detect decaying pulse signals.

10. A transformer fault diagnosis system based on multiwavelet convolution combined with channel attention mechanism, characterized in that, The method comprises: A voiceprint collection module is configured to collect the voiceprint signal of the transformer to obtain a time-domain acoustic signal sample; A data preprocessing module is configured to divide the collected time-domain acoustic signal sample into a training set and a validation set; A multi-wavelet convolution module comprises a Laplace wavelet convolution unit, a Morlet wavelet convolution unit and a MexHat wavelet convolution unit, and is configured to extract features of the time-domain acoustic signal and convert the time-domain acoustic signal into a multi-channel equivalent time-frequency graph; The trainable parameters of the Laplace wavelet convolution unit comprise a frequency parameter and an attenuation coefficient, the trainable parameters of the Morlet wavelet convolution unit comprise a frequency parameter and a phase parameter, and the trainable parameters of the MexHat wavelet convolution unit comprise a scale parameter; A parameter updating module is configured to update the trainable parameters in a data-driven manner; A channel attention module is configured to weight the multi-channel equivalent time-frequency graph to obtain a weighted feature graph; A classification output module is configured to input the weighted feature graph into a fully connected network to output a fault classification result of the transformer.

Images