A transformer diagnosis method and system based on acoustic wave interference

Abstract

The application relates to the technical field of transformer detection, and particularly discloses a transformer diagnosis method and system based on sound wave interference, wherein the method comprises the following steps: collecting multi-channel sound pressure signals through a differential microphone array and calculating a three-dimensional sound pressure gradient tensor; inputting the tensor into a pre-trained space-time gradient convolutional neural network, the network having a parallel time convolution branch and a space gradient convolution branch, and extracting time sequence features and space distribution features of the sound pressure gradient along the time dimension and the space dimension respectively; and inputting feature maps output by the two branches into a multi-task output head after cross-modal attention fusion, and simultaneously outputting a transformer state category probability distribution and three-dimensional space coordinates of a fault source. The application deeply fuses sound source positioning and voiceprint recognition, does not need to transform hardware, and significantly improves the accuracy and robustness of transformer fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of transformer testing technology, and in particular to a transformer diagnostic method and system based on acoustic wave interference. Background Technology

[0002] Transformers are core equipment in power systems, and their operational status monitoring is crucial for grid security. Existing acoustic-based transformer diagnostic methods typically involve two steps: first, using a microphone array to locate the sound source through beamforming or TDOA algorithms; then, performing spectral analysis or deep learning classification on the located signal to identify the fault type. However, this sequential architecture has inherent drawbacks: the location error is directly transmitted and amplified to the subsequent identification stage, and the two steps are optimized independently, unable to reinforce each other. Furthermore, traditional methods often use scalar sound pressure as input, losing the directional information of sound wave propagation (i.e., phase interference information), limiting location accuracy and noise resistance. Although some scholars have attempted to jointly model location and identification in recent years, this often requires complex multimodal fusion strategies or customized hardware modifications, resulting in poor versatility. Therefore, there is an urgent need for an integrated diagnostic method that can achieve high-precision location and high-reliability identification simultaneously without hardware modifications. Summary of the Invention

[0003] To address the problem of low diagnostic accuracy caused by the disconnect between positioning and identification in existing technologies, this invention provides a transformer diagnostic method and system based on acoustic wave interference.

[0004] The technical solution adopted in this invention is:

[0005] A transformer diagnostic method based on acoustic wave interferometry includes the following steps:

[0006] S1. Multi-channel sound pressure signals are synchronously acquired by a differential microphone array arranged around the transformer, and the sound pressure gradient tensor of the multi-channel sound pressure signals is calculated. The sound pressure gradient tensor is a three-dimensional tensor with dimensions represented as T×N×G, where T is the number of time frames, N is the number of sensor units in the microphone array, and G is the number of sound pressure gradient components at each sensor unit.

[0007] S2. Input the sound pressure gradient tensor into a pre-trained spatiotemporal gradient convolutional neural network, which includes a parallel dual-branch feature extraction backbone network and a multi-task output head.

[0008] S3. The parallel dual-branch feature extraction backbone network includes a temporal convolution branch and a spatial gradient convolution branch. The temporal convolution branch uses a one-dimensional causal convolution kernel to extract the temporal features of the sound pressure gradient as a function of time along the dimension of the number of time frames T. The spatial gradient convolution branch uses a two-dimensional convolution kernel to extract the spatial distribution features of the sound pressure gradient along the two-dimensional plane formed by the number of sensor units N and the number of sound pressure gradient components G.

[0009] S4. Perform cross-modal fusion of the feature maps output by the temporal convolution branch and the spatial gradient convolution branch to obtain a fused feature map;

[0010] S5. Input the fused feature map to the multi-task output head. The multi-task output head includes a classification sub-network and a regression sub-network. The classification sub-network outputs the probability distribution of the current state category of the transformer, and the regression sub-network outputs the estimated three-dimensional spatial coordinates of the fault source in a preset coordinate system.

[0011] Preferably, in step S1, calculating the sound pressure gradient tensor further includes:

[0012] For each sensor unit i, which consists of M closely spaced miniature microphones located on a sphere of radius r, calculate the sound pressure gradient vector ∇p at the center point of the sensor unit. i (t), its calculation formula is:

[0013] ;

[0014] Where, p i (t) represents the sound pressure at the center point of sensor unit i, p i,x (t) and p i,-x (t) represents the sound pressure measured by two miniature microphones along the positive and negative x-axis, respectively. i,y (t) and p i,-y (t) represents the sound pressure measured by two miniature microphones along the positive and negative y-axis, respectively. i,z (t) and p i,-z (t) represents the sound pressure measured by two miniature microphones along the positive and negative z-axis, respectively, and c is the current sound velocity in the medium. The sound pressure gradient tensor Grad(t) at time t is an N×3 matrix, which is composed of the stacked sound pressure gradient vectors of all N sensor units, i.e., Grad(t) = [∇p1(t), ∇p2(t), ..., ∇p...]. N (t)] T Finally, the Grad(t) of all T time frames are stacked to form the three-dimensional tensor.

[0015] Preferably, in step S3, the temporal convolution branch is composed of L t The convolutional layer consists of consecutive temporal convolutional blocks, each of which includes a one-dimensional causal convolutional layer, a batch normalization layer, and a ReLU activation function layer; the kernel size of the one-dimensional causal convolutional layer is k. t The number of convolution kernels is f t The step size is 1, and its output depends only on the current time step and previous inputs to ensure temporal causality; the spatial gradient convolution branch consists of L... s It consists of consecutive spatial gradient convolutional blocks, each spatial gradient convolutional block including a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer; the kernel size of the two-dimensional convolutional layer is k. s ×k s The number of convolution kernels is f s The step size is 1, and zero padding is used to keep the size of the feature map constant.

[0016] Preferably, in step S4, the cross-modal fusion is implemented through an attention-based fusion module, which first calculates the first feature map F output by the temporal convolution branch. t The second feature map F output by the spatial gradient convolution branch s The global average pooling vector v t and v s Then, the attention weight vector a is computed through a shared fully connected layer. t and a s Its calculation formula is: a t =σ(W a ⋅v t +b a ), a s =σ(W a ⋅v s +b a ), where σ is the sigmoid activation function, W a Let b be the weight matrix of the shared fully connected layer. a The bias vector is given; finally, the first feature map and the second feature map are weighted and summed according to the attention weights to obtain the fused feature map F. f Its calculation formula is: F f =a t ⊙F t +a s ⊙F s , where ⊙ represents element-wise multiplication.

[0017] Preferably, in step S5, the classification subnetwork is composed of a global average pooling layer, a fully connected layer, and a softmax activation function layer connected in sequence, and its output is a probability vector of dimension C, where C is the total number of transformer state categories, and the index corresponding to the maximum value in the probability vector is the diagnosed transformer state category; the regression subnetwork is composed of a global average pooling layer, a fully connected layer, and a linear activation function layer connected in sequence, and its output is a real vector of dimension 3, which respectively represent the estimated x, y, and z coordinates of the fault source in the preset coordinate system.

[0018] Preferably, the training process of the spatiotemporal gradient convolutional neural network employs a multi-task loss function L. total Its definition is L total =αL cls +βL reg , where L cls For classification loss, the cross-entropy loss function is used, L reg For regression loss, a smoothed L1 loss function is used, where α and β are hyperparameters used to shift the weights of the two loss classes, both being positive real numbers; the Adam optimizer is used during training, with an initial learning rate of l. r0 =0.001, and decays to 0.1 times its original value after every E training cycles, for a total of Epochs training cycles, where E and Epochs are both positive integers.

[0019] Preferably, the differential microphone array is spherically distributed, the sensor units are evenly distributed on the surface of the sphere, the diameter of the sphere is D, and the value of D ranges from 0.3 meters to 0.6 meters; the number of sensor units N ranges from 8 to 32.

[0020] A transformer diagnostic system based on acoustic wave interferometry includes:

[0021] The acoustic pressure gradient tensor construction module is used to perform step S1 as described above;

[0022] The feature extraction and fusion module is used to perform steps S2 to S4 as described above, and it has a pre-trained spatiotemporal gradient convolutional neural network as described above.

[0023] The diagnostic result output module is used to perform step S5 as described above and output the transformer status category and fault source coordinates.

[0024] Preferably, the feature extraction and fusion module further includes a data caching unit for temporarily storing the sound pressure gradient tensor and providing a continuous data stream to the spatiotemporal gradient convolutional neural network in a sliding window manner, wherein the length of the sliding window is T. win In each time frame, the overlap rate of adjacent windows is O.r T win The value range is from 512 to 2048, O r The value ranges from 50% to 75%.

[0025] Preferably, the diagnostic result output module further includes a visualization unit, which combines the estimated three-dimensional spatial coordinates output by the regression subnetwork with a pre-established three-dimensional transformer model to display the location of the fault source in real time on the display interface in the form of marked points or a heat map.

[0026] The beneficial effects of this invention are:

[0027] By constructing a unified sound pressure gradient tensor representation and designing a spatiotemporal gradient convolutional neural network for multi-task learning, the network can perceive the spatial location of the sound source when extracting soundprint features and enhance the signal-to-noise ratio by utilizing the frequency domain characteristics of soundprints when locating the source. This achieves true end-to-end joint learning and overcomes the drawbacks of error accumulation in traditional serial architectures.

[0028] By using a differential microphone array to obtain the sound pressure gradient vector instead of the traditional scalar sound pressure, the sound field information is extended from a scalar field to a vector field. This naturally contains the propagation direction of the sound wave, i.e., the phase interference information, providing richer physical priors for subsequent networks and significantly improving the diagnostic robustness in environments with strong noise and multi-source interference.

[0029] The method is entirely based on algorithmic innovation and does not involve any changes to the microphone array hardware structure. It can be directly deployed on various existing microphone array systems, with good compatibility and low promotion cost. The temporal convolution branch uses causal convolution to ensure that the model output depends only on the current and historical information, which meets the causal constraints of online real-time monitoring and is suitable for continuous monitoring scenarios in substations.

[0030] By automatically learning the relative importance of temporal and spatial features through an attention mechanism, adaptive fusion for different fault types is achieved, further enhancing the model's generalization ability. Attached Figure Description

[0031] Figure 1 This is a schematic flowchart of the transformer diagnosis method based on acoustic wave interference in an embodiment of the present invention;

[0032] Figure 2 This is a structural block diagram of a transformer diagnostic system based on acoustic wave interferometry in an embodiment of the present invention. Detailed Implementation

[0033] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0034] Example 1:

[0035] like Figure 1As shown, a transformer diagnostic method based on acoustic wave interferometry includes the following steps:

[0036] S1. Multi-channel sound pressure signals are synchronously acquired by a differential microphone array arranged around the transformer, and the sound pressure gradient tensor of the multi-channel sound pressure signals is calculated. The sound pressure gradient tensor is a three-dimensional tensor with dimensions represented as T×N×G, where T is the number of time frames, N is the number of sensor units in the microphone array, and G is the number of sound pressure gradient components at each sensor unit.

[0037] S2. Input the sound pressure gradient tensor into a pre-trained spatiotemporal gradient convolutional neural network, which includes a parallel dual-branch feature extraction backbone network and a multi-task output head.

[0038] S3. The parallel dual-branch feature extraction backbone network includes a temporal convolution branch and a spatial gradient convolution branch. The temporal convolution branch uses a one-dimensional causal convolution kernel to extract the temporal features of the sound pressure gradient as a function of time along the dimension of the number of time frames T. The spatial gradient convolution branch uses a two-dimensional convolution kernel to extract the spatial distribution features of the sound pressure gradient along the two-dimensional plane formed by the number of sensor units N and the number of sound pressure gradient components G.

[0039] S4. Perform cross-modal fusion of the feature maps output by the temporal convolution branch and the spatial gradient convolution branch to obtain a fused feature map;

[0040] S5. Input the fused feature map to the multi-task output head. The multi-task output head includes a classification sub-network and a regression sub-network. The classification sub-network outputs the probability distribution of the current state category of the transformer, and the regression sub-network outputs the estimated three-dimensional spatial coordinates of the fault source in a preset coordinate system.

[0041] Sound source localization and voiceprint recognition are typically two separate steps. Localization errors accumulate in subsequent recognition, and the lack of mutual verification and reinforcement between the two leads to insufficient diagnostic robustness. This method integrates the phase information required for localization and the spectral information required for recognition by constructing a unified sound pressure gradient tensor representation. A dedicated spatiotemporal gradient convolutional neural network is designed, with a dual-branch structure to handle the temporal and spatial dimensions separately, avoiding the information loss caused by forcibly reducing multidimensional signals to a one-dimensional sequence in traditional methods. Finally, a multi-task output head enables the network to simultaneously learn classification and regression tasks, achieving deep coupling between localization and recognition.

[0042] This method fundamentally solves the problem of the separation between localization and recognition. Through end-to-end joint learning, the network can perceive the spatial location of the sound source when extracting voiceprint features, and can enhance the signal-to-noise ratio by utilizing the frequency domain characteristics of the voiceprint during localization. This significantly improves the diagnostic accuracy in environments with strong noise and multi-source interference. Moreover, this method does not require any hardware modification and has strong universality.

[0043] To accurately and efficiently obtain the sound pressure gradient tensor from the raw microphone signal, step S1, calculating the sound pressure gradient tensor, further includes:

[0044] For each sensor unit i, which consists of M closely spaced miniature microphones located on a sphere of radius r, calculate the sound pressure gradient vector ∇p at the center point of the sensor unit. i (t), its calculation formula is:

[0045] ;

[0046] Where, p i (t) represents the sound pressure at the center point of sensor unit i, p i,x (t) and p i,-x (t) represents the sound pressure measured by two miniature microphones along the positive and negative x-axis, respectively. i,y (t) and p i,-y (t) represents the sound pressure measured by two miniature microphones along the positive and negative y-axis, respectively. i,z (t) and p i,-z (t) represents the sound pressure measured by two miniature microphones along the positive and negative z-axis, respectively, and c is the current sound velocity in the medium. The sound pressure gradient tensor Grad(t) at time t is an N×3 matrix, which is composed of the stacked sound pressure gradient vectors of all N sensor units, i.e., Grad(t) = [∇p1(t), ∇p2(t), ..., ∇p...]. N (t)] T Finally, the Grad(t) of all T time frames are stacked to form the three-dimensional tensor.

[0047] By measuring the sound pressure difference between microphones positioned symmetrically on a sphere, the sound pressure gradient vector at the center point can be approximated. This vector not only contains the amplitude information of the sound pressure but, more importantly, the direction information of sound wave propagation (i.e., phase interference information), significantly increasing the amount of information compared to the scalar sound pressure collected by a single microphone. Introducing the sound pressure gradient allows subsequent neural networks to directly process vector sound fields containing spatial phase information, rather than scalar sound fields. This is the key physical basis for achieving deep fusion of localization and recognition. It avoids complex direction-of-arrival (DOA) estimation algorithms, simplifying the localization task into a feature extraction problem.

[0048] Regarding the specific structure of the parallel dual-branch feature extraction backbone network, in step S3, the temporal convolution branch consists of L... t The convolutional layer consists of consecutive temporal convolutional blocks, each of which includes a one-dimensional causal convolutional layer, a batch normalization layer, and a ReLU activation function layer; the kernel size of the one-dimensional causal convolutional layer is k. t The number of convolution kernels is f t The step size is 1, and its output depends only on the current time step and previous inputs to ensure temporal causality; the spatial gradient convolution branch consists of L... s It consists of consecutive spatial gradient convolutional blocks, each spatial gradient convolutional block including a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer; the kernel size of the two-dimensional convolutional layer is k. s ×k s The number of convolution kernels is f s The step size is 1, and zero padding is used to keep the size of the feature map constant.

[0049] The temporal branch uses causal convolution to ensure the model doesn't see future information when processing current sound events, which is crucial for real-time diagnostics. The spatial branch uses standard 2D convolution to learn the spatial correlation of sound pressure gradients between different sensor units, i.e., the interference pattern of the sound field. Causal convolution guarantees the model's real-time performance, making it suitable for online monitoring. Batch normalization and ReLU accelerate training convergence and enhance non-linear expressiveness. By specifying the kernel size and number, the granularity and capacity of feature extraction are clearly defined.

[0050] Since simply concatenating or adding the feature maps of two branches can lead to information redundancy or conflict, and lacks the ability to dynamically select and balance features from different modalities, in step S4, the cross-modal fusion is implemented through a fusion module based on an attention mechanism. This module first calculates the first feature map F output by the temporal convolution branch respectively. t The second feature map F output by the spatial gradient convolution branch s The global average pooling vector v t and v s Then, the attention weight vector a is computed through a shared fully connected layer. t and a s Its calculation formula is: a t =σ(W a ⋅v t +b a ), a s =σ(W a ⋅v s +b a ), where σ is the sigmoid activation function, W aLet b be the weight matrix of the shared fully connected layer. a The bias vector is given; finally, the first feature map and the second feature map are weighted and summed according to the attention weights to obtain the fused feature map F. f Its calculation formula is: F f =a t ⊙F t +a s ⊙F s , where ⊙ represents element-wise multiplication.

[0051] This step introduces an attention-based fusion mechanism that generates corresponding weights by learning the global importance of features from two branches. These weights are not fixed but dynamically generated based on the features of the input data. For example, for a stable 100Hz hum, the spatial branch might be more important; while for a transient partial discharge pulse, the temporal branch might be more important. The model automatically adjusts the contributions of both. This step achieves intelligent and dynamic fusion of temporal and spatial features. Compared to simple concatenation or addition, the attention mechanism can suppress useless features and enhance useful features, greatly improving the model's adaptability and generalization ability to different types of faults.

[0052] To ensure stable joint training of the classification and regression tasks, in step S5, the classification subnetwork consists of a global average pooling layer, a fully connected layer, and a softmax activation function layer connected sequentially. Its output is a probability vector of dimension C, where C is the total number of transformer state categories. The index corresponding to the maximum value in the probability vector is the diagnosed transformer state category. The regression subnetwork consists of a global average pooling layer, a fully connected layer, and a linear activation function layer connected sequentially. Its output is a real vector of dimension 3, representing the estimated x, y, and z coordinates of the fault source in the preset coordinate system.

[0053] The classification head uses Softmax to output a probability distribution, while the regression head uses linear activation to output continuous coordinate values. Both heads share previously extracted fused features but have their own parameters. This allows the model to learn from both classification labels ("What is the fault?") and regression labels ("Where is the fault?") during training, mutually reinforcing each other. The shared feature extractor reduces the total number of model parameters, avoiding overfitting. Simultaneously, the spatial information provided by the regression task can serve as a strong regularization term, preventing the classification task from getting stuck in local optima, and vice versa. This collaborative training mechanism is key to improving the overall performance of the model.

[0054] Since the key to multi-task learning lies in balancing the losses of different tasks, otherwise one task might dominate the training, causing another task to fail. Therefore, the training process of the spatiotemporal gradient convolutional neural network adopts a multi-task loss function L. total Its definition is L total =αL cls +βL reg , where L cls For classification loss, the cross-entropy loss function is used, L reg For regression loss, a smoothed L1 loss function is used, where α and β are hyperparameters used to shift the weights of the two loss classes, both being positive real numbers; the Adam optimizer is used during training, with an initial learning rate of l. r0 =0.001, and decays to 0.1 times its original value after every E training cycles, for a total of Epochs training cycles, where E and Epochs are both positive integers.

[0055] Cross-entropy loss is used to optimize classification, and smooth L1 loss is used to optimize regression (smooth L1 is less sensitive to outliers and is more stable). The weights of the two losses are manually adjusted by introducing hyperparameters α and β, so that the model focuses on classification or regression in the early stage of training and reaches a balance in the later stage. The standard learning rate decay strategy helps the model converge.

[0056] To improve the applicability of the differential microphone array in real transformer environments, the differential microphone array is spherically distributed, with the sensor units evenly distributed on the surface of the sphere. The diameter of the sphere is D, and the value of D ranges from 0.3 meters to 0.6 meters. The number of sensor units N ranges from 8 to 32.

[0057] Spherical arrays are isotropic and can capture sound waves uniformly from all directions, making them ideal for polyhedral structures like transformers. The choice of diameter D balances low-frequency response (large diameter) and high-frequency spatial aliasing (small diameter), while the number of sensor elements N determines the spatial resolution.

[0058] Example 2:

[0059] like Figure 2 As shown, a transformer diagnostic system based on acoustic wave interferometry includes:

[0060] A sound pressure gradient tensor construction module is used to perform step S1 as described in Example 1;

[0061] The feature extraction and fusion module is used to perform steps S2 to S4 as described in Example 1, and it has a pre-trained spatiotemporal gradient convolutional neural network as described in Example 1.

[0062] The diagnostic result output module is used to execute step S5 as described in Example 1 and output the transformer status category and fault source coordinates.

[0063] The system abstracts the method steps S1-S5 into three functional modules, which correspond to the method and constitute a complete diagnostic system.

[0064] Since the real-time diagnostic system needs to process continuous data streams, rather than single static data blocks, the feature extraction and fusion module also includes a data caching unit for temporarily storing the sound pressure gradient tensor and providing a continuous data stream to the spatiotemporal gradient convolutional neural network in a sliding window manner, wherein the length of the sliding window is T. win In each time frame, the overlap rate of adjacent windows is O. r T win The value range is from 512 to 2048, O r The value ranges from 50% to 75%.

[0065] A sliding window mechanism was introduced, stipulating that the data input to the neural network consists of continuous, overlapping segments. This overlap ensures a smooth transition between adjacent analysis windows, preventing the omission of transient events occurring at window boundaries. This upgrades the system from offline analysis to online real-time monitoring, clarifies the system's data processing methods, and enables its application in practical substation monitoring scenarios. The overlap rate setting balances computational load and temporal resolution.

[0066] To avoid the problem of outputting only coordinate numbers being unintuitive for maintenance personnel and making it difficult to quickly locate fault points, the diagnostic result output module also includes a visualization unit, which combines the estimated three-dimensional spatial coordinates output by the regression subnetwork with a pre-established three-dimensional transformer model and displays the location of the fault source in real time on the display interface in the form of marked points or heat maps.

[0067] A visualization function has been added, which maps abstract coordinate information onto a real 3D model of the transformer, allowing maintenance personnel to see the location of the fault at a glance.

[0068] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A transformer diagnostic method based on acoustic wave interferometry, characterized in that, Includes the following steps: S1. Multi-channel sound pressure signals are synchronously acquired by a differential microphone array arranged around the transformer, and the sound pressure gradient tensor of the multi-channel sound pressure signals is calculated. The sound pressure gradient tensor is a three-dimensional tensor with dimensions represented as T×N×G, where T is the number of time frames, N is the number of sensor units in the microphone array, and G is the number of sound pressure gradient components at each sensor unit. S2. Input the sound pressure gradient tensor into a pre-trained spatiotemporal gradient convolutional neural network, which includes a parallel dual-branch feature extraction backbone network and a multi-task output head. S3. The parallel dual-branch feature extraction backbone network includes a temporal convolution branch and a spatial gradient convolution branch. The temporal convolution branch uses a one-dimensional causal convolution kernel to extract the temporal features of the sound pressure gradient as a function of time along the dimension of the number of time frames T. The spatial gradient convolution branch uses a two-dimensional convolution kernel to extract the spatial distribution features of the sound pressure gradient along the two-dimensional plane formed by the number of sensor units N and the number of sound pressure gradient components G. S4. Perform cross-modal fusion of the feature maps output by the temporal convolution branch and the spatial gradient convolution branch to obtain a fused feature map; S5. Input the fused feature map to the multi-task output head. The multi-task output head includes a classification sub-network and a regression sub-network. The classification sub-network outputs the probability distribution of the current state category of the transformer, and the regression sub-network outputs the estimated three-dimensional spatial coordinates of the fault source in a preset coordinate system.

2. The transformer diagnostic method based on acoustic wave interferometry according to claim 1, characterized in that, In step S1, calculating the sound pressure gradient tensor further includes: For each sensor unit i, which consists of M closely spaced miniature microphones located on a sphere of radius r, calculate the sound pressure gradient vector ∇p at the center point of the sensor unit. i (t), its calculation formula is: ； Where, p i (t) represents the sound pressure at the center point of sensor unit i, p i,x (t) and p i,-x (t) represents the sound pressure measured by two miniature microphones along the positive and negative x-axis, respectively. i,y (t) and p i,-y (t) represents the sound pressure measured by two miniature microphones along the positive and negative y-axis, respectively. i,z (t) and p i,-z (t) represents the sound pressure measured by two miniature microphones along the positive and negative z-axis, respectively, and c is the current sound velocity in the medium. The sound pressure gradient tensor Grad(t) at time t is an N×3 matrix, which is composed of the stacked sound pressure gradient vectors of all N sensor units, i.e., Grad(t) = [∇p1(t), ∇p2(t), ..., ∇p...]. N (t)] T Finally, the Grad(t) of all T time frames are stacked to form the three-dimensional tensor.

3. The transformer diagnostic method based on acoustic wave interferometry according to claim 1, characterized in that, In step S3, the temporal convolution branch is L t The convolutional layer consists of consecutive temporal convolutional blocks, each of which includes a one-dimensional causal convolutional layer, a batch normalization layer, and a ReLU activation function layer; the kernel size of the one-dimensional causal convolutional layer is k. t The number of convolution kernels is f t The step size is 1, and its output depends only on the current time step and previous inputs to ensure temporal causality; the spatial gradient convolution branch consists of L... s It consists of consecutive spatial gradient convolutional blocks, each spatial gradient convolutional block including a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer; the kernel size of the two-dimensional convolutional layer is k. s ×k s The number of convolution kernels is f s The step size is 1, and zero padding is used to keep the size of the feature map constant.

4. The transformer diagnostic method based on acoustic wave interferometry according to claim 1, characterized in that, In step S4, the cross-modal fusion is implemented through an attention-based fusion module, which first calculates the first feature map F output by the temporal convolution branch. t The second feature map F output by the spatial gradient convolution branch s The global average pooling vector v t and v s Then, the attention weight vector a is computed through a shared fully connected layer. t and a s Its calculation formula is: a t =σ(W a ⋅v t +b a ), a s =σ(W a ⋅v s +b a ), where σ is the sigmoid activation function, W a Let b be the weight matrix of the shared fully connected layer. a The bias vector is given; finally, the first feature map and the second feature map are weighted and summed according to the attention weights to obtain the fused feature map F. f Its calculation formula is: F f =a t ⊙F t +a s ⊙F s , where ⊙ represents element-wise multiplication.

5. The transformer diagnostic method based on acoustic wave interference according to claim 1, characterized in that, In step S5, the classification subnetwork consists of a global average pooling layer, a fully connected layer, and a softmax activation function layer connected sequentially. Its output is a probability vector of dimension C, where C is the total number of transformer state categories. The index corresponding to the maximum value in the probability vector is the diagnosed transformer state category. The regression subnetwork consists of a global average pooling layer, a fully connected layer, and a linear activation function layer connected sequentially. Its output is a real vector of dimension 3, representing the estimated x, y, and z coordinates of the fault source in the preset coordinate system.

6. The transformer diagnostic method based on acoustic wave interferometry according to claim 1, characterized in that, The training process of the spatiotemporal gradient convolutional neural network adopts a multi-task loss function L. total Its definition is L total =αL cls +βL reg , where L cls For classification loss, the cross-entropy loss function is used, L reg For regression loss, a smoothed L1 loss function is used, where α and β are hyperparameters used to shift the weights of the two loss classes, both being positive real numbers; the Adam optimizer is used during training, with an initial learning rate of l. r0 =0.001, and decays to 0.1 times its original value after every E training cycles, for a total of Epochs training cycles, where E and Epochs are both positive integers.

7. The transformer diagnostic method based on acoustic wave interferometry according to claim 1, characterized in that, The differential microphone array is spherically distributed, and the sensor units are evenly distributed on the surface of the sphere. The diameter of the sphere is D, and the value of D ranges from 0.3 meters to 0.6 meters. The number of sensor units N ranges from 8 to 32.

8. A transformer diagnostic system based on acoustic wave interferometry, characterized in that, include: A sound pressure gradient tensor construction module is used to perform step S1 as described in claim 1; The feature extraction and fusion module is used to perform steps S2 to S4 as described in claim 1, and it has a pre-trained spatiotemporal gradient convolutional neural network as described in claim 1 embedded within it. The diagnostic result output module is used to execute step S5 as described in claim 1, and output the transformer status category and fault source coordinates.

9. The transformer diagnostic system based on acoustic wave interferometry according to claim 8, characterized in that, The feature extraction and fusion module further includes a data caching unit for temporarily storing the sound pressure gradient tensor and providing a continuous data stream to the spatiotemporal gradient convolutional neural network in a sliding window manner, wherein the length of the sliding window is T. win In each time frame, the overlap rate of adjacent windows is O. r T win The value range is from 512 to 2048, O r The value ranges from 50% to 75%.

10. The transformer diagnostic system based on acoustic wave interferometry according to claim 8, characterized in that, The diagnostic result output module also includes a visualization unit, which combines the estimated three-dimensional spatial coordinates output by the regression subnetwork with a pre-established three-dimensional transformer model to display the location of the fault source in real time on the display interface in the form of marked points or heat maps.

Images