Transformer multi-measuring-point working condition identification method based on deep convolution feature fusion

Abstract

The invention discloses a transformer multi-measuring-point working condition identification method based on deep convolution feature fusion, and belongs to the technical field of transformer detection, and the method comprises the steps: obtaining vibration signals of a plurality of preset points of a transformer, carrying out the processing based on a pre-trained transformer working condition identification model, and outputting a transformer working condition identification result, based on a Swin Transformer encoder and a depth deformable MobileNet module, a double-path collaborative architecture is adopted, a first path and a second path of a transformer working condition recognition model are constructed, and the first path and the second path are subjected to feature interaction through a cross-modal attention fusion module. According to the method, the problems of low efficiency and poor accuracy caused by neglecting vibration signal time sequence dependence, local detail features, global structures and spatial features in existing transformer multi-measuring-point working condition identification are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer detection, and particularly to a method for identifying the working conditions of multiple measuring points of a transformer based on deep convolutional feature fusion. Background Art

[0002] In recent years, in the field of identifying the working conditions of multiple measuring points of a transformer, the existing technologies mainly rely on various monitoring means and algorithms. A common technology is infrared camera monitoring, which uses the infrared radiation intensity on the surface of the transformer to reflect the temperature distribution, and judges the working temperature of the transformer and whether there is a fault by presenting the difference between bright and dark in the infrared image. In addition, on-line monitoring technology has also been widely used, which is based on partial discharge detection and vibration analysis. Partial discharge detection is used to detect mechanical faults that may occur during the normal operation of the transformer, and vibration analysis is to check the operating state of the transformer by analyzing the vibration signals during the operation of the transformer. In recent years, artificial intelligence algorithms have also been introduced into this field, such as the artificial neural network fault diagnosis method, which imitates the development process of the human brain, has the ability of self-learning and processing complex problems, and is suitable for the fault diagnosis of large-scale power equipment.

[0003] Although the existing technologies for identifying the working conditions of multiple measuring points of a transformer have made certain progress, there are still problems of low efficiency and poor accuracy caused by ignoring the temporal dependence, local detail features, global structure and spatial features of vibration signals. Specifically, the temporal dependence of vibration signals is crucial for judging the operating state of the transformer, because the changes in vibration signals are often closely related to time, and the existing technologies often do not fully consider this, resulting in the loss of key information during the identification process. In addition, the vibration signals of the transformer present different characteristics at different scales, including local characteristics and global characteristics. Local characteristics reflect the operating state of a certain component or area of the transformer, while global characteristics reflect the overall state of the entire transformer. However, the existing technologies often only focus on the characteristics at a certain scale and ignore the fusion of local detail features, global structure and spatial features, which limits the improvement of the identification efficiency and accuracy. Therefore, how to comprehensively consider the temporal dependence and local detail features, global structure and spatial features of vibration signals to improve the efficiency and accuracy of identifying the working conditions of multiple measuring points of a transformer is an urgent problem to be solved in this field. Summary of the Invention

[0004] The technical problem solved by the present invention is: the problem of low efficiency and poor accuracy in the existing identification of the working conditions of multiple measuring points of a transformer caused by ignoring the temporal dependence, local detail features, global structure and spatial features of vibration signals.

[0005] To solve the above technical problems, the present invention provides the following technical solutions: The present invention provides a method for identifying the operating conditions of a transformer based on deep convolutional feature fusion, including: Obtain the vibration signals of multiple preset points of the transformer; Based on the vibration signals, process them using a pre-trained transformer operating condition identification model, and output the transformer operating condition identification result; Among them, based on the Swin Transformer encoder and the depth deformable MobileNet module, a dual-path collaborative architecture is adopted to construct the first path and the second path of the transformer operating condition identification model, and the first path and the second path perform feature interaction through a cross-modal attention fusion module.

[0006] Furthermore, when constructing the transformer operating condition identification model, the Swin Transformer encoder is improved by optimizing the hierarchical self-attention mechanism into a dynamic window attention.

[0007] Furthermore, the first path is used to capture the temporal dependence and local patterns in the signal using the self-attention mechanism according to the improved Swin Transformer encoder, and extract the local detail features in the vibration signals by stacking multiple layers of Transformer encoding blocks; The second path is used to capture the global patterns and structural information in the signal using the spatial perception ability of the convolutional neural network according to the depth deformable MobileNet module, and extract the global structure and spatial features in the vibration signals.

[0008] Furthermore, the dynamic window attention mechanism includes deformable attention heads, and the deformable attention heads perform non-rectangular local feature sampling by predicting a learnable spatial offset matrix.

[0009] Furthermore, the region shape of the non-rectangular local feature sampling is dynamically determined by the short-time Fourier transform spectral features of the vibration signals.

[0010] Furthermore, the cross-modal attention fusion module includes: A multi-scale feature alignment unit for dynamically matching the resolutions of the feature maps of the first path and the second path according to the convolutional kernels of the depth deformable MobileNet module, and then inputting them into a heterogeneous feature conversion layer; A heterogeneous feature conversion layer for compressing the channels and aligning the semantics of the feature maps of the first path and the second path using 1x1 convolutions with quantization awareness, and then inputting them into a gated fusion mechanism; A gated fusion mechanism for dynamically adjusting the feature weights of the feature maps of the first path and the second path based on the operating condition mode classification confidence.

[0011] Further, before obtaining the vibration signals at multiple preset points of the transformer, it includes: Adopting a sensor array with an irregular hexagonal layout at multiple preset points of the transformer; Performing power frequency period framing on the vibration signals generated by the transformer during operation, setting the frame length to 0.02 s and non-overlapping, and generating a two-dimensional vibration texture map; Stacking the two-dimensional vibration texture maps according to the topological structure of the sensor array to form a three-dimensional tensor with the height dimension equal to the number of preset points; Applying frequency domain enhancement processing to the three-dimensional tensor and amplifying the wavelet packet coefficients of the 100 Hz multiple frequency components.

[0012] Further, the depth deformable MobileNet module includes: A topological relationship modeling sub-module based on a graph neural network for modeling the topological relationship between sensor arrays according to the graph neural network to capture the mutual influence between sensors; An adaptive dilated convolution time series modeling sub-module for modeling time series data according to adaptive dilated convolution to capture feature changes in the time dimension.

[0013] Further, based on a meta-learner driven by a transformer operating condition mode database, the convolution kernel deformation parameters of the depth deformable MobileNet module are dynamically generated.

[0014] Further, the meta-learner adopts the MAML framework to update the convolution kernel deformation parameters of the depth deformable MobileNet module through second-order gradients.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. By obtaining the vibration signals at multiple preset points of the transformer, the present invention can more comprehensively capture the operating state information of the transformer and improve the accuracy of condition recognition; at the same time, adopting a dual-path collaborative architecture based on the Swin Transformer encoder and the depth deformable MobileNet module enables the transformer condition recognition model to extract features from two different paths simultaneously and realize effective interaction and fusion of features through the cross-modal attention fusion module. This not only fully utilizes the temporal dependence, local detail features, global structure and spatial features of the vibration signals, but also enhances the recognition ability of the transformer condition recognition model for complex conditions. Therefore, compared with traditional methods, the present invention shows higher accuracy and efficiency in the condition recognition of multi-measurement points of transformers, solves the problems of low efficiency and poor accuracy in the existing condition recognition of multi-measurement points of transformers caused by ignoring the temporal dependence, local detail features, global structure and spatial features of vibration signals, and provides a strong guarantee for the stable operation of the power system.

[0016] 2. The present invention optimizes the Swin Transformer encoder by introducing a dynamic window attention mechanism, and combines the depthwise deformable MobileNet module to form a dual-path collaborative architecture, which can more accurately capture the temporal dependence and local detail features, global structure and spatial features in the transformer vibration signal. The deformable attention head in the dynamic window attention mechanism can predict a learnable spatial offset matrix to achieve non-rectangular local feature sampling. This flexible sampling method can more accurately capture the key features in the vibration signal, thereby improving the accuracy of working condition recognition. At the same time, the dual-path architecture processes in parallel, accelerating the feature extraction and fusion process and significantly improving the recognition efficiency.

[0017] 3. The present invention also implements dynamic matching and semantic alignment of the feature maps of the first path and the second path in terms of resolution and channel dimension based on the cross-modal attention fusion module, using the multi-scale feature alignment unit and the heterogeneous feature conversion layer. Through the gated fusion mechanism, the feature weights of the feature maps of the first path and the second path are dynamically adjusted based on the working condition mode classification confidence, enabling the transformer working condition recognition model to adaptively fuse features according to different working condition modes, enhancing the generalization ability of the transformer working condition recognition model and its adaptability to different working condition modes. This flexibility enables the transformer working condition recognition model to maintain a high recognition accuracy when facing complex and changing transformer working conditions.

[0018] 4. Before obtaining the vibration signal, the present invention adopts a sensor array with an irregular hexagonal layout, which can better capture the vibration information on the transformer surface and reduce information loss. At the same time, the vibration signal is subjected to power frequency cycle framing and frequency domain enhancement processing to generate a two-dimensional vibration texture map and stack it into a three-dimensional tensor, further improving the data quality and signal-to-noise ratio. In addition, the graph neural network in the depthwise deformable MobileNet module models the topological relationship of the sensor array to capture the mutual influence between sensors, and the adaptive dilated convolutional temporal modeling sub-module captures the feature changes in the time dimension, jointly improving the understanding and analysis ability of the transformer working condition recognition model for the vibration signal and providing a more reliable data basis for working condition recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the basic process of a transformer multi-measurement point working condition recognition method based on deep convolutional feature fusion provided by an embodiment of the present invention; Figure 2 Schematic diagram of the process of a preprocessing method for vibration signals at multiple preset points provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following provides a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments.

[0021] Embodiment 1 As Figure 1 shown, this embodiment introduces a method for identifying the operating conditions of a transformer based on deep convolutional feature fusion from a principle perspective, including: Step 1: Obtain the vibration signals of multiple preset points of the transformer.

[0022] During the operation of the transformer, various components inside and outside it will generate vibrations, and the vibration signals contain important information about the working state of the transformer. By installing vibration sensors at multiple preset points of the transformer, such as key parts like the box body, cooling system, and windings, these vibration signals can be captured in real time.

[0023] The present invention can cover the key parts of the transformer through the setting of multiple preset points, improving the comprehensiveness and accuracy of the operating condition identification, enabling the operating condition identification to be carried out in a timely manner, and helping to detect the abnormal state of the transformer in a timely manner.

[0024] Step 2: Process the vibration signals based on a pre-trained transformer operating condition identification model and output the transformer operating condition identification result.

[0025] Among them, based on the Swin Transformer encoder and the depth deformable MobileNet module, a dual-path collaborative architecture is adopted to construct the first path and the second path of the transformer operating condition identification model, and the first path and the second path perform feature interaction through a cross-modal attention fusion module.

[0026] The transformer operating condition identification model provided by the present invention adopts a dual-path collaborative architecture, including a first path and a second path. In the first path, by using the self-attention mechanism of the Swin Transformer encoder, the long-range dependence relationship in the vibration signals can be captured, and the local detail features in the vibration signals can be extracted. In the second path, through the lightweight convolution and deformable convolution operations of the depth deformable MobileNet module, the global structure and spatial features in the vibration signals can be efficiently extracted.

[0027] The first path and the second path perform feature interaction through a cross-modal attention fusion module, which can fuse local detail features, global structure, and spatial features, enhancing the identification ability of the transformer operating condition identification model for complex operating conditions. The cross-modal attention fusion module dynamically adjusts the proportion of feature fusion by calculating the correlation between the features of different paths, enabling the transformer operating condition identification model to more accurately identify the operating conditions of the transformer.

[0028] After the transformer condition recognition model is constructed, the present invention uses a large amount of labeled transformer vibration signal data for training, so that the transformer condition recognition model can learn the mapping relationship between the vibration signal and the transformer condition.

[0029] Embodiment 2 With the same inventive concept as Embodiment 1, this embodiment introduces the implementation steps of the transformer multi-measurement point condition recognition method based on deep convolutional feature fusion, including: Step 1: Obtain the vibration signals of multiple preset points of the transformer.

[0030] As Figure 2 shown, in some embodiments, before obtaining the vibration signals of multiple preset points of the transformer, it includes: Step 1.1: Adopt a non-regular hexagonal layout sensor array at multiple preset points of the transformer.

[0031] This embodiment breaks through the limitation of the traditional regular grid layout. By enhancing the spatial sampling diversity through the hexagonal geometric structure, it covers more vibration mode directions, including axial / radial vibrations, avoids the spatial aliasing effect caused by regular arrangement, improves the sensitivity to complex vibration modes such as partial discharge and core looseness, and provides a structured data basis for subsequent topological relationship modeling.

[0032] Step 1.2: Perform power frequency period framing on the vibration signals generated by the transformer during operation, with the frame length set to 0.02 s and non-overlapping, to generate a two-dimensional vibration texture map.

[0033] In this embodiment, power frequency period framing is performed on the vibration signals generated by the transformer during operation, with the frame length set to 0.02 s and non-overlapping, to generate a two-dimensional vibration texture map, achieving the optimal balance of time-frequency resolution and avoiding information redundancy through non-overlapping framing. First, the vibration signals generated by the transformer during operation are subjected to short-time Fourier transform to be converted into a two-dimensional spectrum containing time-frequency features, and then a standardized texture image of 256×256 pixels is generated, converting the traditional time series signal into an image format suitable for convolutional neural network processing, laying a foundation for feature extraction and pattern recognition of subsequent deep learning models.

[0034] Step 1.3: Stack the two-dimensional vibration texture maps according to the topological structure of the sensor array to form a three-dimensional tensor with the height dimension equal to the number of preset points.

[0035] In this embodiment, by stacking the two-dimensional vibration texture maps of each preset point according to the topological structure of the sensor array, the spatial relationship between sensors can be retained. This stacking method enables each slice in the three-dimensional tensor, that is, each two-dimensional vibration texture map, to correspond to a specific sensor position, so that this spatial relationship can be utilized for feature extraction and pattern recognition in subsequent processing.

[0036] As a structured data representation method, the three-dimensional tensor can capture the vibration information of the transformer more comprehensively. Compared with processing each two-dimensional vibration texture map separately, stacking multiple texture maps into a three-dimensional tensor allows the transformer condition recognition model to simultaneously consider the mutual influence between different sensor positions during the training process, thereby improving the performance and generalization ability of the model, making better use of spatial information, and improving the efficiency of feature extraction and pattern recognition.

[0037] This embodiment also retains the topological information of the sensor array through the stacking method, which helps the transformer condition recognition model better understand the spatial relationship between different sensor positions, thereby more accurately identifying the conditions of the transformer.

[0038] Step 1.4: Apply frequency-domain enhancement processing to the three-dimensional tensor and amplify the wavelet packet coefficients of the 100Hz multiple-frequency components.

[0039] Frequency-domain enhancement processing is a method for enhancing signals in the frequency domain. In this embodiment, by performing a frequency-domain transformation on the three-dimensional tensor, such as Fourier transform, the representation of the vibration signal in the frequency domain is obtained. The 100Hz multiple-frequency components are closely related to specific conditions of the transformer, such as partial discharge and core looseness. Amplifying the wavelet packet coefficients of the 100Hz multiple-frequency components highlights the importance of the 100Hz multiple-frequency components in the vibration signal, improves the sensitivity of the transformer condition recognition model to the 100Hz multiple-frequency components, enables the transformer condition recognition model to more easily capture features related to specific conditions during the training process, and thus improves the accuracy of condition recognition.

[0040] Step 2: Build a transformer condition recognition model.

[0041] In some embodiments, based on the Swin Transformer encoder and the depth deformable MobileNet module, a dual-path collaborative architecture is adopted to build the first path and the second path of the transformer condition recognition model. The first path and the second path perform feature interaction through the cross-modal attention fusion module.

[0042] The dual-path collaborative architecture combines the advantages of the self-attention mechanism and the convolutional neural network. The self-attention mechanism can capture the temporal dependencies and local patterns in the signal, while the convolutional neural network has strong spatial perception ability and can capture the global patterns and structural information in the signal. By working collaboratively in a dual-path manner, comprehensive extraction and fusion of signal features can be achieved, improving the generalization ability and robustness of the transformer condition recognition model, enabling the transformer condition recognition model to more accurately identify the vibration signals of transformers under different conditions.

[0043] In this embodiment, when constructing the transformer condition recognition model, the Swin Transformer encoder is improved by optimizing the hierarchical self-attention mechanism into a dynamic window attention.

[0044] In this embodiment, the dynamic window attention mechanism includes deformable attention heads, and the deformable attention heads perform non-rectangular local feature sampling by predicting a learnable spatial offset matrix.

[0045] In this embodiment, the region shape of the non-rectangular local feature sampling is dynamically determined by the short-time Fourier transform spectral features of the vibration signal.

[0046] The dynamic window attention mechanism is an improved attention mechanism. By predicting a learnable spatial offset matrix for non-rectangular local feature sampling, it can more flexibly capture the local features in the vibration signal. At the same time, since the sampling region shape is dynamically determined by the short-time Fourier transform spectral features of the vibration signal, it also improves the ability of the transformer condition recognition model to capture local features, enabling the transformer condition recognition model to more accurately identify the subtle changes in the vibration signal.

[0047] In this embodiment, the first path is used to capture the temporal dependencies and local patterns in the signal according to the improved Swin Transformer encoder, and extract the local detailed features in the vibration signal by stacking multiple layers of Transformer encoding blocks. In this embodiment, the second path is used to capture the global patterns and structural information in the signal according to the depth deformable MobileNet module, and extract the global structure and spatial features in the vibration signal by utilizing the spatial perception ability of the convolutional neural network.

[0048] In this embodiment, the depth deformable MobileNet module includes: A sub-module for modeling the topological relationship of the sensor array based on the graph neural network, which is used to model the topological relationship between sensor arrays according to the graph neural network and capture the mutual influence between sensors.

[0049] In this embodiment, a graph neural network is used to model the topological relationship between sensor arrays, capture the mutual influence between sensors, and can more accurately reflect the spatial structure of the sensor array, thereby improving the model's ability to extract spatial features. The ability of the transformer condition recognition model to extract spatial features is improved, enabling the transformer condition recognition model to more accurately identify the vibration signal differences between different sensor positions.

[0050] The adaptive dilated convolution time series modeling sub-module is used to model time series data according to adaptive dilated convolution and capture feature changes in the time dimension.

[0051] Adaptive dilated convolution is a convolution operation with variable dilation rates that can capture feature changes in the time dimension. In this embodiment, by adjusting the dilation rate, the capture of features at different time scales is realized, improving the transformer condition recognition model's ability to extract time series features, enabling the transformer condition recognition model to more accurately capture the time series dependence in vibration signals.

[0052] In this embodiment, based on a transformer condition mode database-driven meta-learner, the convolution kernel deformation parameters of the depth deformable MobileNet module are dynamically generated.

[0053] In this embodiment, the meta-learner adopts the MAML framework and updates the convolution kernel deformation parameters of the depth deformable MobileNet module through second-order gradients in the few-shot condition mode.

[0054] The meta-learner used in this embodiment adopts the MAML framework and dynamically generates the convolution kernel deformation parameters of the depth deformable MobileNet module through second-order gradient updates in the few-shot condition mode, enabling the transformer condition recognition model to more flexibly adapt to the signal features under different condition modes, improving the flexibility and adaptability of the transformer condition recognition model, and enabling the transformer condition recognition model to also achieve good recognition results in the few-shot condition mode.

[0055] In this embodiment, the cross-modal attention fusion module includes: The multi-scale feature alignment unit is used to input the feature maps of the first path and the second path into the heterogeneous feature conversion layer after dynamically matching the resolutions according to the convolution kernels of the depth deformable MobileNet module.

[0056] In this embodiment, the resolutions of the feature maps of the first path and the second path are dynamically matched according to the convolution kernels of the depth deformable MobileNet module, ensuring the compatibility between the feature maps of the first path and the second path, laying a foundation for subsequent heterogeneous feature conversion and gated fusion, and facilitating subsequent fusion processing.

[0057] The heterogeneous feature conversion layer is used to compress the feature maps of the first path and the second path in the channel dimension and align the semantics by using quantization-aware 1x1 convolution, and then input them into the gated fusion mechanism. In this embodiment, the quantization-aware 1x1 convolution is used to compress the feature maps of the first path and the second path in the channel dimension and align the semantics, which helps to reduce the dimension of the feature maps, improve the computational efficiency, and retain key information at the same time, providing high-quality feature input for the gated fusion mechanism.

[0058] The gated fusion mechanism is used to dynamically adjust the feature weights of the feature maps of the first path and the second path based on the confidence of the working condition mode classification. In this embodiment, based on the confidence of the working condition mode classification, the feature weights of the feature maps of the first path and the second path are dynamically adjusted, so that the transformer working condition recognition model can more flexibly utilize the feature information of different paths, dynamically adjust the feature weights according to the classification confidence of the current working condition mode, realize adaptive feature fusion, and improve the accuracy and robustness of working condition recognition. Step 3: Process the vibration signal based on the pre-trained transformer working condition recognition model and output the transformer working condition recognition result.

[0059] In this embodiment, the vibration signals of multiple preset points of the transformer are input into the pre-trained transformer working condition recognition model, and the vibration signals are processed and analyzed through the forward propagation process.

[0060] The dual-path collaborative architecture in the transformer working condition recognition model extracts the local detail features, global structure and spatial features of the vibration signal respectively, and then performs feature fusion through the cross-modal attention fusion module to obtain a comprehensive feature representation.

[0061] The transformer working condition recognition model makes a classification decision according to the comprehensive feature representation and outputs the transformer working condition recognition result.

[0062] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. Among them, the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk. These computer program instructions can also be stored in a computer-readable memory capable of guiding a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device, and the instruction device implements the functions specified in one process Figure 1 one process or multiple processes and / or boxes Figure 1 specified in one box or multiple boxes.

[0063] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for identifying the operating conditions of a transformer with multiple measurement points based on deep convolutional feature fusion, characterized in that, Including: Obtain the vibration signals of multiple preset points of the transformer; Based on the vibration signals, process them using a pre-trained transformer operating condition recognition model and output the transformer operating condition recognition result; Among them, based on the Swin Transformer encoder and the depth deformable MobileNet module, a dual-path collaborative architecture is adopted to construct the first path and the second path of the transformer operating condition recognition model, and the first path and the second path perform feature interaction through a cross-modal attention fusion module.

2. The method for identifying the operating conditions of multiple measuring points of a transformer based on deep convolutional feature fusion according to claim 1, wherein It also includes that when constructing the transformer operating condition recognition model, the Swin Transformer encoder is improved by optimizing the hierarchical self-attention mechanism into a dynamic window attention.

3. The method for identifying the operating conditions of multiple measuring points of a transformer based on deep convolutional feature fusion according to claim 2, wherein The first path is used to capture the temporal dependence and local patterns in the signal using the self-attention mechanism according to the improved Swin Transformer encoder, and extract the local detailed features in the vibration signal by stacking multiple layers of Transformer encoding blocks; The second path is used to capture the global patterns and structural information in the signal using the spatial perception ability of the convolutional neural network according to the depth deformable MobileNet module, and extract the global structure and spatial features in the vibration signal.

4. The method for identifying the operating conditions of multiple measuring points of a transformer based on deep convolutional feature fusion according to claim 2, wherein The dynamic window attention mechanism includes deformable attention heads, and the deformable attention heads perform non-rectangular local feature sampling by predicting a learnable spatial offset matrix.

5. The method for identifying the operating conditions of multiple measuring points of a transformer based on deep convolutional feature fusion according to claim 4, characterized in that The region shape of the non-rectangular local feature sampling is dynamically determined by the short-time Fourier transform spectral features of the vibration signal.

6. The method for identifying operating conditions of a transformer with multiple measuring points based on deep convolutional feature fusion according to claim 2, characterized in that, The cross-modal attention fusion module includes: A multi-scale feature alignment unit, which is used to perform dynamic resolution matching on the feature maps of the first path and the second path according to the convolutional kernel of the depth deformable MobileNet module and then input them into the heterogeneous feature conversion layer; The heterogeneous feature conversion layer is used to perform channel dimension compression and semantic alignment on the feature maps of the first path and the second path using 1x1 convolution with quantization awareness and then input them into the gated fusion mechanism; The gated fusion mechanism is used to dynamically adjust the feature weights of the feature maps of the first path and the second path based on the operating condition mode classification confidence.

7. The method for identifying the operating conditions of multiple measuring points of a transformer based on deep convolutional feature fusion according to claim 1, wherein Before obtaining the vibration signals of multiple preset points of the transformer, it includes: Adopt a sensor array with an irregular hexagonal layout at multiple preset points of the transformer; Perform power frequency period framing on the vibration signals generated by the transformer during operation, set the frame length to 0.02s and non-overlapping, and generate a two-dimensional vibration texture map; Stack the two-dimensional vibration texture maps according to the sensor array topology structure to form a three-dimensional tensor with the height dimension equal to the number of preset points; Apply frequency domain enhancement processing to the three-dimensional tensor and amplify the wavelet packet coefficients of the 100Hz multiple frequency components.

8. The method for identifying the operating conditions of multiple measuring points of a transformer based on deep convolutional feature fusion according to claim 7, characterized in that, The depth deformable MobileNet module includes: A sub-module for modeling the topological relationship of the sensor array based on the graph neural network, which is used to model the topological relationship between sensor arrays using the graph neural network and capture the mutual influence between sensors; The adaptive dilated convolution time series modeling sub-module is used to model time series data according to the adaptive dilated convolution and capture the feature changes in the time dimension.

9. The method for identifying the operating conditions of multiple measuring points of a transformer based on deep convolutional feature fusion according to claim 8, wherein, The meta-learner driven by the transformer working condition mode database dynamically generates the convolution kernel deformation parameters of the depth deformable MobileNet module.

10. The method for identifying the operating conditions of multiple measurement points of a transformer based on deep convolutional feature fusion according to claim 9, wherein, The meta-learner adopts the MAML framework and updates the convolution kernel deformation parameters of the depth deformable MobileNet module through second-order gradients.

Images