Epoxy potting defect identification method, system and equipment based on multi-modal convolution and medium
By fusing the time-domain, frequency-domain, and time-frequency-domain features of epoxy resin defect signals using a multimodal convolutional neural network, the problem that single feature domain analysis methods cannot fully capture defect information is solved, and efficient and accurate identification of internal defect types in epoxy resin is achieved.
Patent Information
- Application Number
- CN202510934035.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-08
- Publication Date
- 2025-12-05
AI Technical Summary
Existing single-feature-domain analysis methods cannot fully capture the complete feature information of epoxy resin defects, resulting in insufficient defect identification accuracy.
A multimodal convolutional neural network is used to fuse the time-domain, frequency-domain, and time-frequency-domain features of partial discharge signals. The signal is acquired through an ultra-wideband antenna and the model is trained using the multimodal convolutional neural network to identify epoxy resin defects.
By constructing a multimodal convolutional neural network recognition model, the time-domain, frequency-domain, and time-frequency-domain features of partial discharge signals are organically integrated. With the help of the powerful learning ability of convolutional neural networks, the complex relationships between these features are automatically learned, thus achieving efficient and accurate determination of the internal defect type of epoxy resin.
Smart Images
Figure CN121069109A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of partial discharge detection of power equipment, and particularly relates to an epoxy potting defect recognition method, system and device based on multi-modal convolution and a medium. BACKGROUND
[0002] Epoxy resin is widely used in various electrical equipment due to its excellent electrical properties, mechanical properties and chemical stability. However, during the preparation process and service process of the epoxy resin, various complex factors can easily cause typical defects such as bubbles, cracks and delamination. The existence of such defects can seriously deteriorate the insulation performance of the epoxy resin, greatly shorten its service life, and thus bring potential risks to the safe and reliable operation of the electrical equipment. Therefore, accurately identifying the defect type of the epoxy resin is of great practical significance to ensure the stable operation of the electrical equipment. Partial discharge (PD) is a common physical phenomenon when an insulating material has a defect, and the partial discharge signal generated thereby contains rich information closely related to the defect type and characteristics. In the past research, researchers have tried to analyze the characteristics of the partial discharge signal in different characteristic domains to effectively identify the defect type of the epoxy resin.
[0003] When using time domain features to carry out defect recognition on the partial discharge signal, although basic information such as the amplitude and occurrence time of the partial discharge pulse can be obtained, the time domain information dimension is relatively single and the content is relatively scarce. In view of the complexity of the internal structure of the epoxy resin and the diversity of the defect characteristics, it is difficult to comprehensively and accurately map the internal microstructure and defect state of the epoxy resin by relying solely on time domain information. When the partial discharge signal is analyzed from the frequency domain, the frequency composition of the signal can be clearly revealed, and some defect-related characteristic frequencies can be extracted. However, this analysis method loses the time dimension information of the signal during processing. Time-frequency domain analysis methods, such as wavelet transform (Wavelet Transform), can take into account the time information and frequency information of the partial discharge signal to some extent, providing a more comprehensive perspective for defect recognition. However, it cannot be ignored that a single time-frequency domain analysis method has inherent limitations. For example, the selection of the wavelet basis function has a decisive influence on the analysis result, but there is no universal optimal basis function selection strategy for different types of epoxy resin defects, which greatly limits the effectiveness and accuracy of the method in practical applications. SUMMARY
[0004] In view of the above existing problems, the present application is proposed.
[0005] Therefore, the application provides an epoxy pouring defect recognition method, system, device and medium based on multi-modal convolution, which solves the problem that a single feature domain analysis method (time domain, frequency domain or time-frequency domain) cannot comprehensively capture complete feature information of epoxy resin defects, resulting in insufficient defect recognition accuracy.
[0006] To solve the above technical problems, the application provides the following technical solutions.
[0007] In a first aspect, the application provides an epoxy pouring defect recognition method based on multi-modal convolution, comprising:
[0008] Using a super wideband antenna to collect discharge signals of three defects of epoxy resin bubbles, cracks and delamination;
[0009] Extracting time domain, frequency domain and time-frequency features of a single partial discharge pulse signal, and inputting the features of the three modalities into a multi-modal convolutional neural network for training to obtain an epoxy pouring defect recognition model;
[0010] After preprocessing, inputting the partial discharge signal of unknown defect type into the epoxy pouring defect recognition model to obtain the recognition result of the defect type.
[0011] As a preferred scheme of the epoxy pouring defect recognition method based on multi-modal convolution, the extraction of the time domain, frequency domain and time-frequency features of the single partial discharge pulse signal comprises:
[0012] The time domain feature of the signal is a time domain waveform diagram of a single discharge pulse signal;
[0013] The frequency domain feature of the signal is a frequency spectrum distribution diagram of a single discharge pulse signal, and the frequency spectrum distribution diagram is obtained by performing fast Fourier transform on the original discharge signal;
[0014] The time-frequency feature of the signal is a time-frequency diagram of a single discharge pulse signal, and the time-frequency diagram is obtained by performing wavelet transform on the original discharge signal, and the wavelet function is a bump wavelet.
[0015] The preferred technical scheme has the beneficial effects that more comprehensive, detailed and deep information about the defect partial discharge signal can be obtained, and the ability to distinguish different defect types such as bubbles, cracks and delamination in the epoxy resin is significantly improved.
[0016] As a preferred scheme of the epoxy pouring defect recognition method based on multi-modal convolution, the multi-modal convolutional neural network comprises:
[0017] When processing the time domain, frequency domain and time-frequency domain diagrams of the discharge signals of the three defects of the epoxy resin, an independent input layer is set for the data of each modality;
[0018] The independent pre-processing of each modal data includes: adjusting the image size to a unified specification, performing data normalization processing, and applying image enhancement operations;
[0019] The pre-processed modal data is multiplied by two different preset scaling factors to generate corresponding dual-channel feature representations, and the dual-channel feature representations are spliced according to the channel dimension to form a six-channel fusion feature input.
[0020] As a preferred scheme of the epoxy pouring defect recognition method based on multi-modal convolution, the multi-modal convolutional neural network comprises:
[0021] A three-level convolutional network architecture is constructed, a first level is configured with a convolutional layer with 64 channels for extracting edge and texture features, a second level is configured with a convolutional layer with 128 channels for extracting geometric shape features, and a third level is configured with a convolutional layer with 256 channels for identifying overall morphological features of defects;
[0022] A feature map down-sampling operation is arranged between each level of convolutional layer;
[0023] A batch normalization layer is inserted between each level of network to standardize the input data, and a Dropout layer is arranged after a specific network layer to randomly shield part of the neuron connections;
[0024] A global average pooling layer is arranged before the output layer to compress the feature map of each channel into a single numerical feature.
[0025] The beneficial effects of the preferred technical scheme are: with the powerful learning ability of the convolutional neural network, the complex relationship between features can be automatically learned, and efficient and accurate determination of the type of internal defects of the epoxy resin can be realized.
[0026] As a preferred scheme of the epoxy pouring defect recognition method based on multi-modal convolution, the multi-modal convolutional neural network comprises:
[0027] An Adam optimization algorithm is used to train the model, and the parameter learning rate of each network layer is dynamically adjusted to avoid falling into a local optimal solution;
[0028] The size of the training batch is set, and the gradient is calculated using the samples of the training batch size when updating the parameters;
[0029] A high-frequency verification strategy is implemented, including:
[0030] The verification set is evaluated once every 2 training cycles;
[0031] The accuracy and loss function value of each verification are recorded;
[0032] The performance trend of the monitoring model on the validation set is monitored;
[0033] The training parameters are dynamically adjusted according to the verification result, and the learning rate is automatically reduced or the training is terminated in advance when the performance of the verification set is detected to decrease.
[0034] As a preferred scheme of the epoxy pouring defect recognition method based on multi-modal convolution, the local discharge signal of the unknown defect type is preprocessed according to the same data processing and storage mode as the model training, so as to be directly input into the epoxy pouring defect recognition model for defect type recognition.
[0035] As a preferred scheme of the epoxy pouring defect recognition method based on multi-modal convolution, the response bandwidth of the ultra-wideband antenna is 50 kHz-6.5 GHz.
[0036] In a second aspect, the present application provides an epoxy pouring defect recognition system based on multi-modal convolution, comprising:
[0037] The signal acquisition module is used for acquiring the discharge signals of the three defects of epoxy resin bubbles, cracks and delamination by using the ultra-wideband antenna.
[0038] The model training module is used for extracting the time domain, frequency domain and time-frequency features of a single partial discharge pulse signal, and inputting the features of the three modalities into a multi-modal convolution neural network for training to obtain an epoxy pouring defect recognition model.
[0039] The defect type recognition module is used for inputting the preprocessed local discharge signal of the unknown defect type into the epoxy pouring defect recognition model to obtain the recognition result of the defect type.
[0040] In a third aspect, the present application provides an electronic device comprising a memory and a processor; the memory is used for storing computer executable instructions, and the processor implements the steps of the epoxy pouring defect recognition method based on multi-modal convolution when executing the computer executable instructions.
[0041] In a fourth aspect, the present application provides a computer readable storage medium storing computer executable instructions, which implement the steps of the epoxy pouring defect recognition method based on multi-modal convolution when executed by a processor.
[0042] Compared with the prior art, the present application has the beneficial effects that: the present application provides an epoxy pouring defect recognition method, system, device and medium based on multi-modal convolution, which organically fuses the time domain, frequency domain and time-frequency domain characteristics of partial discharge signals by constructing a multi-modal convolution neural network recognition model, automatically learns the complex relationship between these characteristics with the powerful learning ability of the convolution neural network, discovers some patterns and rules that cannot be perceived by single feature recognition, enables the model to learn more rich and complete feature representation, and finally realizes efficient and accurate determination of the internal defect type of the epoxy resin. The present application can obtain more comprehensive, detailed and deep information about the defect partial discharge signal, significantly improves the distinguishing ability of different defect types such as bubbles, cracks and delamination in the epoxy resin, greatly improves the accuracy and reliability of defect recognition, and provides strong technical support for the whole life cycle health monitoring and efficient operation of the epoxy resin insulated electrical equipment. BRIEF DESCRIPTION OF DRAWINGS
[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0044] Figure 1 The overall flow logic diagram of the epoxy pouring defect recognition method based on multi-modal convolution provided by the embodiment of the present application.
[0045] Figure 2 The three defect sample schematic diagram of the epoxy pouring defect recognition method based on multi-modal convolution provided by the embodiment of the present application.
[0046] Figure 3 The time domain waveform diagram of the partial discharge signal of the three defects of the epoxy pouring defect recognition method based on multi-modal convolution provided by the embodiment of the present application.
[0047] Figure 4 The frequency spectrum distribution diagram of the partial discharge signal of the three defects of the epoxy pouring defect recognition method based on multi-modal convolution provided by the embodiment of the present application.
[0048] Figure 5 The time-frequency diagram of the partial discharge signal of the three defects of the epoxy pouring defect recognition method based on multi-modal convolution provided by the embodiment of the present application.
[0049] Figure 6 The multi-modal convolution neural network design idea schematic diagram of the epoxy pouring defect recognition method based on multi-modal convolution provided by the embodiment of the present application.
[0050] Figure 7 A confusion matrix diagram generated based on a test set of the epoxy pouring defect recognition method based on multi-modal convolution provided by the embodiment of the present application is shown. DETAILED DESCRIPTION
[0051] In order to make the above-mentioned objects, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.
[0052] Embodiment 1, refer to Figure 1 For an embodiment of the present application, an epoxy pouring defect recognition method based on multi-modal convolution is provided, as shown in Figure 1 which specifically comprises the following steps:
[0053] S100: using an ultra-wideband antenna to collect discharge signals of three defects of epoxy resin bubbles, cracks and delamination;
[0054] S200: extracting time domain, frequency domain and time-frequency features of a single partial discharge pulse signal, and inputting the features of the three modalities into a multi-modal convolution neural network for training to obtain an epoxy pouring defect recognition model;
[0055] S300: inputting the preprocessed partial discharge signal of unknown defect type into the epoxy pouring defect recognition model to obtain the recognition result of the defect type.
[0056] It should be noted that, in order to solve the problem that the single feature domain analysis method (time domain, frequency domain or time-frequency domain) in the prior art cannot comprehensively capture the complete feature information of the epoxy resin defect, resulting in insufficient defect recognition accuracy, the above steps S100-S300 organically fuse the time domain, frequency domain and time-frequency domain features of the partial discharge signal by constructing a multi-modal convolution neural network recognition model, and by means of the powerful learning ability of the convolution neural network, the complex relationship between these features can be automatically learned, and some patterns and rules that cannot be perceived by single feature recognition can be discovered, so that the model can learn more rich and complete feature representation, and finally realize efficient and accurate determination of the defect type in the epoxy resin. The present application can obtain more comprehensive, detailed and deep information about the defect partial discharge signal, significantly improve the distinguishing ability of different defect types such as bubbles, cracks and delamination in the epoxy resin, greatly improve the accuracy and reliability of defect recognition, and provide strong technical support for the whole life cycle health monitoring and efficient operation of the epoxy resin insulated electrical equipment.
[0057] Embodiment 2, refer toFigures 2 to 7 Based on the previous embodiment, the embodiment provides a specific implementation of the epoxy pouring defect recognition method based on multi-modal convolution, which is used to illustrate the technical means adopted in the method.
[0058] In the embodiment of the present application, the step S100 of collecting the discharge signals of the three defects of epoxy resin bubbles, cracks and delamination using a super wideband antenna comprises:
[0059] Specifically, the three types of defects selected in the experiment sample of the embodiment are epoxy resin bubble defects, crack defects and surface delamination defects. The diameters of the bubbles inside the bubble defects include the following sizes: 0.2mm, 0.35mm, 0.5mm, 0.65mm, 1.1mm, 1.4mm and 1.6mm; the width of the crack defect inside the crack defect is 2-10μm; the delamination depth between the epoxy and the metal is 10-50μm, and the delamination area is 0-3cm 2 . The three samples are shown in Figure 2 .
[0060] Specifically, the discharge signals of the three defects of epoxy resin bubbles, cracks and delamination are collected using a super wideband antenna, wherein the response bandwidth of the super wideband antenna is 50kHz-6.5GHz.
[0061] It should be noted that the step S100 can comprehensively cover the electromagnetic signal characteristic frequency bands generated by various defects of the epoxy resin, ensuring that the obtained partial discharge signals contain complete bubble, crack and delamination defect characteristic information, and providing a high-quality data basis for subsequent multi-modal feature analysis.
[0062] In the embodiment of the present application, the step S200 of extracting the time domain, frequency domain and time-frequency features of a single partial discharge pulse signal and inputting the features of the three modalities into a multi-modal convolution neural network for training to obtain an epoxy pouring defect recognition model comprises the following sub-steps B1-B2:
[0063] In B1, the time domain, frequency domain and time-frequency features of a single partial discharge pulse signal are extracted, comprising:
[0064] The time domain feature of the signal is the time domain waveform graph of the single discharge pulse signal;
[0065] The frequency domain feature of the signal is the frequency spectrum distribution graph of the single discharge pulse signal, which is obtained by performing fast Fourier transform on the original discharge signal;
[0066] The time-frequency feature of the signal is the time-frequency graph of the single discharge pulse signal, which is obtained by performing wavelet transform on the original discharge signal, and the wavelet function is bump wavelet.
[0067] In an optional embodiment, in addition to using the bump wavelet transform to extract the time-frequency features, a short-time Fourier transform can also be used to generate a time-frequency distribution by performing a Fourier transform on the segmented and windowed signal. The short-time Fourier transform method has high computational efficiency and is suitable for analyzing stationary or slowly varying signals, but the time-frequency resolution is limited by the window function.
[0068] In another optional embodiment, an S transform can also be used to combine the advantages of the short-time Fourier transform and the wavelet transform, providing frequency-dependent resolution and more accurately depicting the non-stationary characteristics of the signal, especially for transient feature analysis of partial discharge signals.
[0069] Specifically, Figure 3 The time-domain waveform graphs of partial discharge signals of the three defects were obtained. In the experiment, it was found that the time-domain waveforms of the same defect had certain differences, which could be used as picture input for defect type recognition. For a single discharge signal pulse, a fast Fourier transform was performed on about 1000 points containing the complete discharge pulse to obtain its frequency spectrum. Figure 4 The frequency spectrum waveform graphs of partial discharge signals of the three defects were obtained. Similar to the time-domain features, the frequency spectrum distribution of the same defect had certain differences, which could be used as picture input for defect type recognition. A wavelet transform was performed on about 1000 points containing the complete discharge pulse to draw the time-frequency graphs of discharge signals of different defects. The bump wavelet was selected as the wavelet function. The time-frequency graphs of partial discharge signals of the three defects are shown in Figure 5 .
[0070] It should be noted that, in the field of epoxy resin defect discharge signal recognition, compared with single feature recognition, combined recognition of time domain, frequency domain and time-frequency domain graphs using a convolutional neural network has significant advantages. Combined recognition can automatically learn the complex relationships between these features with the powerful learning ability of the convolutional neural network, and can discover patterns and rules that are difficult to detect by single feature recognition. The mining of these potential relationships is of great significance for in-depth understanding of the nature of epoxy resin defect discharge signals. Therefore, in the classification and recognition of epoxy resin three defect discharge signals, a synchronous recognition strategy of time domain, frequency domain and time-frequency domain graphs will be adopted. By integrating the feature information contained in the three pictures, a more accurate recognition model can be constructed with the powerful learning ability of the convolutional neural network to achieve efficient and accurate determination of the defect type.
[0071] In B2, the features of the three modalities are input into a multi-modal convolutional neural network for training to obtain an epoxy potting defect recognition model. The design scheme is shown in Figure 6 .
[0072] In the embodiments of the present application, the data processing of the multi-modal convolutional neural network includes:
[0073] In the process of processing the time domain, frequency domain and time-frequency domain graphs of the three defect discharge signals of the epoxy resin, an independent input layer is set for the data of each mode;
[0074] Each mode data is independently preprocessed, including adjusting the image size to a unified specification, performing data normalization processing, and applying image enhancement operations;
[0075] The preprocessed mode data is multiplied by two different preset scaling factors to generate corresponding dual-channel feature representations, and the dual-channel feature representations are spliced by channel dimension to form six-channel fusion feature input.
[0076] It should be noted that the independent input layer can ensure that each data enters the network in the most suitable way for its own characteristics, avoiding interference between different modal data in the initial stage. Combining different modalities into multi-channel input, the preprocessing of each modality uses different scaling factors, and the data of different modalities provides different perspective information about the target object. Combining these different modal data into multi-channel input can achieve complementary features, so that the model can analyze and judge defects from multiple angles, thereby obtaining more rich and comprehensive information, which helps to improve the recognition accuracy of the model for defect types.
[0077] It should be noted that a special preprocessing method is adopted according to the characteristics of different modal data. For image data, including adjusting the size, cropping, normalizing the color, and performing image enhancement (such as rotation, flipping, etc.). Through modality-specific preprocessing, the data can be converted into a form more suitable for model processing, highlighting the key features in the data and reducing noise and irrelevant information interference. The data set is strictly divided according to the 70-15-15 ratio, and random permutation index is used to avoid data order bias. Random permutation index can give each sample an equal chance to be allocated to each subset, ensuring that each subset can fully cover all types of samples and fully represent the feature distribution of the original data set.
[0078] In the embodiments of the present application, the structure configuration of the multi-modal convolutional neural network includes:
[0079] A three-level convolutional network architecture is constructed, the first level is configured with a convolutional layer with 64 channels to extract edge and texture features, the second level is configured with a convolutional layer with 128 channels to extract geometric shape features, and the third level is configured with a convolutional layer with 256 channels to identify overall morphological features of defects;
[0080] Feature map down-sampling operations are set between each convolutional layer;
[0081] Batch normalization layers are inserted between the levels of the network to standardize the input data, and Dropout layers are set after specific network layers to randomly mask some neuron connections;
[0082] The output layer is provided with a global average pooling layer to compress the feature map of each channel into a single numerical feature.
[0083] Specifically, the channel increment strategy (64→128→256) is adopted, the shallow network (first level) captures basic features: edges, textures, etc. (corresponding to 64 channels); the middle network (second level) extracts composite features: geometric shapes (128 channels); and the deep network (third level) identifies abstract features: overall morphology of defects (256 channels), which balances the local features and the number of parameters through multi-scale perception, and gradually reduces the size of the feature map (224→112→56→28).
[0084] It should be noted that the joint adaptability between neurons is broken through Dropout position optimization to prevent over-reliance on specific features; and the data input to ReLU is standardized through batch normalization position to prevent gradient disappearance. In a deep learning model, too many parameters can easily cause overfitting problems and greatly increase the computational load and storage requirements of the model. The original feature map of each channel with a size of 28x28 contains a large amount of spatial information, and the number of parameters is relatively large. Through spatial information compression, the average value of each channel feature map is calculated to convert it into a single value, and finally a 256-dimensional vector is obtained. This greatly reduces the number of parameters required for feature representation. The reduction of the number of parameters reduces the complexity of the model, thereby reducing the risk of overfitting and improving the generalization ability of the model, so that the model can also have good performance on unknown data.
[0085] In an optional embodiment, the structural configuration of the multi-modal convolutional neural network can also adopt a residual connection structure, which introduces a cross-layer skip connection in the network to directly pass the shallow features to the deep network, alleviate the gradient disappearance problem, and at the same time preserve the multi-scale feature information and enhance the recognition ability of the model for defects of different sizes. Specifically, an identity mapping path can be added after each convolutional block to enable the network to learn residual features.
[0086] In another optional embodiment, the structural configuration of the multi-modal convolutional neural network can also adopt an attention mechanism module to embed channel attention or spatial attention mechanism in the feature fusion stage to automatically weight important feature channels or regions. For example, channel weights are generated through global average pooling to dynamically adjust the contribution of each modal feature and improve the expression ability of key features while suppressing noise interference.
[0087] In the embodiments of the present application, the training process of the multi-modal convolutional neural network includes:
[0088] The Adam optimization algorithm is used to train the model to dynamically adjust the learning rate of each network layer to avoid falling into a local optimal solution.
[0089] a training batch size is set, and the gradient is calculated using the samples of the training batch size at the time of parameter update;
[0090] A high-frequency validation strategy is implemented, including:
[0091] The validation set is evaluated once every 2 training cycles;
[0092] The accuracy and loss function value of each validation are recorded;
[0093] The performance change trend of the model on the validation set is monitored;
[0094] According to the validation result, the training parameters are dynamically adjusted, and when the performance of the validation set is detected to decrease, the learning rate is automatically reduced or the training is terminated in advance.
[0095] It should be noted that the adaptive learning rate scheduling is adopted, and the learning speed of different feature layers is balanced through the dynamic adjustment mechanism of Adam. The adaptive learning rate scheduling helps the model to avoid falling into a local optimal solution in the training process. In the training process of the deep learning model, batch processing optimization is crucial. Setting the batch size to 16, the golden value, can achieve an optimal balance between memory occupation and gradient stability. In addition, high-frequency validation can timely capture the overfitting signs of the model. High-frequency validation can frequently use the validation set to evaluate the model at different stages of training, and compared with low-frequency validation, it can more finely understand the performance change of the model. For example, by validating once every few training cycles, the model's accuracy on the validation set can be timely detected to no longer improve or even decrease, and the loss function can be detected to no longer decrease but increase, and other overfitting signs. The training state of the model is continuously monitored.
[0096] It should be noted that the above step S200 adopts a multi-modal convolutional neural network for feature fusion training, which can automatically mine the internal correlation between time domain, frequency domain and time-frequency domain features, establish a more comprehensive and accurate defect feature representation model, and significantly improve the feature expression ability of the model.
[0097] In the embodiment of the present application, the above step S300 inputs the preprocessed partial discharge signal of unknown defect type into the epoxy pouring defect recognition model to obtain the recognition result of the defect type, including:
[0098] Specifically, the preprocessing of the partial discharge signal of unknown defect type is to process the partial discharge signal of unknown defect according to the same data processing and storage method as during model training, so as to directly input into the epoxy pouring defect recognition model for defect type recognition.
[0099] It should be noted that the above step S300 breaks through the limitation of single feature domain analysis, and judges comprehensively multi-dimensional feature information, so that the recognition accuracy of different types of defects such as bubbles, cracks and delamination is greatly improved, and reliable technical support is provided for the insulation state evaluation of power equipment.
[0100] In one possible embodiment, as shown in Figure 7 The confusion matrix generated based on the test set is shown. The accuracy of the validation set and the test set is about 95%, indicating that the defect classification model has high recognition accuracy.
[0101] Therefore, according to the above embodiment, the method provided in the embodiment organically fuses the time domain, frequency domain and time-frequency domain features of the partial discharge signal by constructing a multi-modal convolution neural network recognition model, and by means of the powerful learning ability of the convolution neural network, the complex relationship between the features can be automatically learned, and some patterns and rules that cannot be perceived by single feature recognition can be discovered, so that the model can learn more rich and complete feature representation, and finally realize efficient and accurate judgment of the internal defect type of the epoxy resin. The present application can obtain more comprehensive, detailed and deep information about the defect partial discharge signal, significantly improve the discrimination ability of different defect types such as bubbles, cracks and delamination in the epoxy resin, greatly improve the accuracy and reliability of defect recognition, and provide strong technical support for the whole life cycle health monitoring and efficient operation and maintenance of the epoxy resin insulated electrical equipment.
[0102] Embodiment 3 provides an epoxy pouring defect recognition system based on multi-modal convolution, comprising:
[0103] The signal acquisition module is configured to use an ultra-wideband antenna to collect discharge signals of three types of defects of the epoxy resin, i.e., bubbles, cracks and delamination.
[0104] The model training module is configured to extract time domain, frequency domain and time-frequency features of a single partial discharge pulse signal, and input the features of the three modalities into a multi-modal convolution neural network for training to obtain an epoxy pouring defect recognition model.
[0105] The defect type recognition module is configured to input a preprocessed partial discharge signal of an unknown defect type into the epoxy pouring defect recognition model to obtain a recognition result of the defect type.
[0106] It should be noted that the technical scheme of the epoxy pouring defect recognition system based on multi-modal convolution belongs to the same concept as the technical scheme of the epoxy pouring defect recognition method based on multi-modal convolution described above. The technical scheme of the epoxy pouring defect recognition system based on multi-modal convolution in the present embodiment is not described in detail, and can be referred to the description of the technical scheme of the epoxy pouring defect recognition method based on multi-modal convolution.
[0107] The above-mentioned each unit module can be embedded in or independent of the processor in the electronic device in hardware form, or can be stored in the memory in the electronic device in software form, so as to be called and executed by the processor to perform the operation corresponding to each module.
[0108] The embodiment further provides an electronic device including a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the electronic device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, operator network, NFC (Near Field Communication) or other technologies. The computer program is executed by the processor to implement the epoxy pouring defect identification method based on multi-modal convolution. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen. The input device of the electronic device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the electronic device, or an external keyboard, touchpad or mouse, etc.
[0109] The embodiment further provides a computer readable storage medium having a computer program stored thereon. The program is executed by the processor to implement the method proposed in the above embodiment.
[0110] The storage medium proposed in the embodiment belongs to the same inventive concept as the method proposed in the above embodiment. The technical details not described in detail in the embodiment can be referred to the above embodiment, and the embodiment has the same beneficial effects as the above embodiment.
[0111] From the above description of the embodiments, those skilled in the art can clearly understand that the present application can be realized by means of software and necessary general hardware, and of course can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a floppy disk, a ROM, a RAM, a FLASH, a hard disk or an optical disc, etc., and includes a number of instructions to make an electronic device (which can be a personal computer, a server, or a network device, etc.) execute the method of the embodiments of the present application.
[0112] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.
Claims
1. A method for epoxy potting defect recognition based on multi-modal convolution, characterized in that, The method comprises the steps of: Collecting discharge signals of epoxy resin bubbles, cracks and delamination defects using an ultra-wideband antenna; Extracting time domain, frequency domain and time-frequency features of a single partial discharge pulse signal, and inputting the features of the three modalities into a multi-modal convolutional neural network for training to obtain an epoxy pouring defect recognition model; Inputting a partial discharge signal of an unknown defect type into the epoxy pouring defect recognition model after preprocessing to obtain a recognition result of the defect type.
2. The multi-modal convolution based epoxy potting defect recognition method of claim 1, wherein, The extraction of the time domain, frequency domain and time-frequency features of the single partial discharge pulse signal comprises: The time domain feature of the signal is a time domain waveform diagram of a single discharge pulse signal; The frequency domain feature of the signal is a frequency spectrum distribution diagram of a single discharge pulse signal, which is obtained by performing fast Fourier transform on the original discharge signal; The time-frequency feature of the signal is a time-frequency diagram of a single discharge pulse signal, which is obtained by performing wavelet transform on the original discharge signal, and the wavelet function is a bump wavelet. 3.The multi-modal convolution based epoxy potting defect recognition method of claim 2, wherein, The multi-modal convolutional neural network comprises: When processing time domain, frequency domain and time-frequency domain diagrams of discharge signals of three epoxy defects, an independent input layer is set for the data of each modality; Each modality data is independently preprocessed, including adjusting the image size to a uniform specification, performing data normalization processing, and applying image enhancement operations; The preprocessed modality data is multiplied by two different preset scaling factors to generate corresponding double-channel feature representations, and the double-channel feature representations are spliced by channel dimension to form six-channel fusion feature inputs.
4. The multi-modal convolution-based epoxy potting defect recognition method of claim 3, wherein, The multi-modal convolutional neural network comprises: A three-level convolutional network architecture is constructed, a first level is configured with a convolutional layer with 64 channels for extracting edge and texture features, a second level is configured with a convolutional layer with 128 channels for extracting geometric shape features, and a third level is configured with a convolutional layer with 256 channels for identifying overall morphological features of defects; A feature map down-sampling operation is set between each convolutional layer; A batch normalization layer is inserted between each level of the network to standardize the input data, and a Dropout layer is set after a specific network layer to randomly shield part of the neuron connections; A global average pooling layer is set before the output layer to compress the feature map of each channel into a single numerical feature.
5. The multi-modal convolution-based epoxy potting defect recognition method of claim 4, wherein, The multi-modal convolutional neural network comprises: An Adam optimization algorithm is used to train the model, and the learning rate of each network layer is dynamically adjusted to avoid falling into a local optimal solution; The size of the training batch is set, and the gradient is calculated using the training batch size of samples when updating the parameters; A high-frequency verification strategy is implemented, including: The validation set is evaluated once every 2 training cycles; The accuracy and loss function value of each validation are recorded; The performance change trend of the model on the validation set is monitored; According to the validation result, the training parameters are dynamically adjusted, and when the performance of the validation set is detected to decrease, the learning rate is automatically reduced or the training is terminated in advance.
6. The multi-modal convolution-based epoxy potting defect recognition method of claim 5, wherein, The preprocessing of the partial discharge signal of the unknown defect type is to process the partial discharge signal of the unknown defect according to the same data processing and storage method as during model training, so as to directly input the signal into the epoxy pouring defect recognition model for defect type recognition.
7. The multi-modal convolution-based epoxy potting defect recognition method of claim 1, wherein, The response bandwidth of the ultra-wideband antenna is 50 kHz-6.5 GHz.
8. An epoxy potting defect recognition system based on multi-modal convolution, applying the epoxy potting defect recognition method based on multi-modal convolution according to any one of claims 1 to 7, characterized in that, Comprise: The signal acquisition module is used for collecting discharge signals of three defects of epoxy resin bubbles, cracks and delamination by using an ultra-wideband antenna; The model training module is used for extracting time domain, frequency domain and time-frequency characteristics of a single partial discharge pulse signal, and inputting the characteristics of the three modalities into a multi-modal convolutional neural network for training to obtain an epoxy pouring defect recognition model; The defect type recognition module is used for inputting the preprocessed partial discharge signal of an unknown defect type into the epoxy pouring defect recognition model to obtain a recognition result of the defect type. 9.An electronic device comprising a memory and a processor, the electronic device characterized by: The memory is used for storing computer executable instructions, and the processor executes the computer executable instructions to realize the steps of the epoxy pouring defect recognition method based on multi-modal convolution in any one of claims 1-7.
10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: The computer executable instructions are executed by the processor to realize the steps of the epoxy pouring defect recognition method based on multi-modal convolution in any one of claims 1-7.