A deep learning driven method and system for particle identification in insulating oil

Abstract

The present application relates to a kind of deep learning driven insulating oil particle identification method and system, combine microflow digital imaging technology and the semantic segmentation method of deep learning, construct unified training data set, combine multi-task segmentation model, can realize the automatic identification and classification of particle in insulating oil.Through microflow digital imaging module, the particle image in transformer oil is obtained, and the image is classified and particle is identified using multi-task segmentation model.The segmentation branch in the model is responsible for the extraction of particle boundary, and the classification branch analyzes the morphology of particle, and outputs the category of particle.The model uses multi-scale feature fusion, attention mechanism and deep residual network structure, solves the problems of particle morphology diversity, size variation and image boundary blur.The present application improves the accuracy and efficiency of particle detection in insulating oil through high-precision particle recognition algorithm and system, reduces manual intervention, and improves the safety and long-term stability of transformer.

Description

Technical Field

[0001] This invention belongs to the field of transformers, specifically relating to a deep learning-driven method and system for identifying particles in insulating oil. Background Technology

[0002] With the continuous development of power systems, transformers, as key equipment in these systems, are of paramount importance in terms of safety and stability. Transformer oil not only serves as the insulating medium for power transformers but also plays a crucial role in heat dissipation and protection; its performance is decisive for the transformer's performance. However, during operation, transformers inevitably become contaminated with impurities. When the concentration of particulate matter reaches a certain level, it may affect the dielectric strength of the oil, reduce the transformer's insulation performance, and even lead to partial discharge or equipment failure. Therefore, timely detection and identification of particulate matter in transformer oil has become a research hotspot in the power industry.

[0003] Currently, the main methods for detecting particles in transformer oil include optical image analysis and chemical analysis. While these traditional methods can provide some detection results in certain specific applications, they typically suffer from problems such as poor detection accuracy, inaccurate identification, and significant susceptibility to human error.

[0004] The rise of deep learning technology, especially the rapid development of convolutional neural networks (CNNs) and semantic segmentation, has brought new opportunities to the field of particulate matter detection. Semantic segmentation technology can classify images at the pixel level, significantly improving the accuracy of particulate matter recognition. Deep learning methods, through large amounts of training data and powerful feature learning capabilities, can automatically extract the morphological features of particulate matter, thereby achieving more accurate particulate matter recognition.

[0005] However, despite the significant progress made by deep learning technology in image processing, existing methods still face several challenges. For example, the morphology of particles in transformer oil is complex and varied, with large differences in particle size and blurred boundaries in particle images. These problems make it difficult for existing deep learning models to achieve efficient and accurate particle recognition in practical applications. When processing particulate matter images, existing models often rely on relatively simple feature extraction methods and lack the ability to capture subtle differences between particles, thus requiring further innovation and optimization.

[0006] Furthermore, although some existing studies have attempted to combine microfluidic digital imaging technology with deep learning for particulate matter detection, the diversity and complexity of transformer oil particles mean that current technologies still fall short of the requirements for real-time detection, accurate classification, and efficient processing. Therefore, developing a high-precision, high-efficiency transformer oil particle identification technology based on microfluidic digital imaging technology and incorporating deep learning models remains a current research hotspot and challenge.

[0007] Therefore, this invention proposes a new particle recognition scheme by combining a deep learning multi-task segmentation model to address the diversity and size variations of particle images. This scheme aims to solve the problems of poor particle detection accuracy, inaccurate recognition results, and high susceptibility to human influence in existing technologies, and to provide a more intelligent and efficient technical solution for the safe operation of transformers. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a deep learning-driven method for particle identification in insulating oil. This method solves problems such as the lack of a unified and effective training dataset for insulating oil particles, blurred image segmentation boundaries, diverse particle morphologies and sizes, and poor identification accuracy. This invention combines microfluidic digital imaging technology with deep learning-based semantic segmentation technology to propose a method suitable for transformer oil particle detection. By constructing a unified and effective training dataset for oil particles and utilizing semantic segmentation technology for pixel-level classification of microfluidic imaging images, automated identification and classification of particles are achieved. This method not only overcomes the limitations of traditional technologies in handling particle morphological diversity, size variations, and blurred segmentation edges, but also improves detection accuracy and efficiency, reducing reliance on manual operation. This technical solution provides a more intelligent and efficient solution for online monitoring and fault early warning of power equipment, contributing to improved long-term stability and safety of transformers.

[0009] The present invention adopts the following technical solution.

[0010] This invention provides a deep learning-driven method for particle identification in insulating oil, comprising:

[0011] An image to be detected is acquired and input into a pre-trained multi-task segmentation model. Based on the image, the multi-task segmentation model outputs the detection result of oil particles. The multi-task segmentation model is trained through the following steps: acquiring images of oil particles and constructing a multi-task segmentation dataset; preprocessing the multi-task segmentation dataset; building a multi-task segmentation model using a dual-branch framework of segmentation and classification branches as the basic structure; the segmentation branch is used to segment oil particles and obtain particle boundary information; the classification branch is used for identification, performing morphological analysis based on the information obtained from the segmentation branch, and then combining the image extracted from the classification branch for identification to obtain the types of oil particles; training the constructed multi-task segmentation model using the preprocessed multi-task segmentation dataset; setting network training parameters and using a loss function to train the constructed multi-task segmentation model.

[0012] Preferably, the step of acquiring images of particles in oil and constructing a multi-task segmentation dataset includes:

[0013] Images of particles in transformer oil are acquired using a microfluidic digital imaging module; synthetic particle data and difficult samples are added to form the final multi-task segmentation dataset.

[0014] Preferably, the preprocessing of the multi-task segmentation dataset includes data augmentation of the images in the multi-task segmentation dataset, wherein the data augmentation includes horizontal flipping, vertical flipping, rotation, translation, scaling, cropping, erasing, and color dithering operations.

[0015] Preferably, the multi-task segmentation model trained using the preprocessed multi-task segmentation dataset includes:

[0016] In the segmentation branch, the image features output by the skeleton network are used to guide segmentation feature learning, and the image features include high-order features and low-order features; in the classification branch, the high-order features input from the segmentation branch are used as auxiliary discriminant information to classify confusion items.

[0017] Preferably, the skeleton network is an optimized Xception deep residual network, and the Xception includes stage one, stage two, stage three, stage four and stage five.

[0018] Preferably, the extraction of output features from the skeleton network in the segmentation branch includes:

[0019] After the high-order features output by stage four are fed into the multi-scale context contrast local module, the local features output by stage four, along with the low-order features output by stages one, two, and three, are fed into a two-layer attention convolutional GRU.

[0020] Preferably, the multi-scale contextual contrast local module processes the local features fed in at stage four by connecting them to four 2×2 convolutions with dilated convolution rates of 2, 3, 6 and 8, and then subtracts and fuses the processed local features, connecting them to the input of the channel attention module.

[0021] Preferably, the channel attention module uses global average pooling and two fully connected layers to calculate channel weights.

[0022] This invention also provides a deep learning-driven particle identification system for insulating oil, the system being the same as the aforementioned deep learning-driven particle identification method for insulating oil, comprising:

[0023] The microfluidic digital imaging module is used to acquire the image to be detected and input it into a pre-trained multi-task segmentation model;

[0024] The image recognition module is used by the multi-task segmentation model to output particle image recognition results based on the image to be detected;

[0025] A multi-scale contextual contrast local module is used to enhance the capture and optimization of image boundaries;

[0026] The channel attention module is used to uncover the intrinsic semantic relationships between multi-scale feature maps;

[0027] The pyramid pooling module is used to perform pooling operations at four different scales on the input feature map, generating multi-scale feature maps.

[0028] The present invention also provides a terminal, including a processor and a storage medium;

[0029] The storage medium is used to store instructions;

[0030] The processor is configured to operate according to the instructions to perform the steps according to the aforementioned method.

[0031] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned method.

[0032] The beneficial effects of this invention are that, compared with the prior art,

[0033] 1. Improved Particle Recognition Accuracy: This invention introduces a multi-task segmentation model, combining segmentation and classification branches, and effectively fusing high-order and low-order features to accurately segment and classify particles based on their morphology. Compared to existing technologies, this invention significantly improves particle recognition accuracy, especially when particles have complex shapes and varying sizes. Through the comprehensive application of a global attention upsampling module, a hybrid pooling module, and a feature pyramid attention module, this invention can better capture local details and global information, ensuring high-precision particle recognition and classification.

[0034] 2. Addressing the Problem of Blurred Image Boundaries: This invention further enhances the capture and optimization of image boundaries by employing a multi-scale contextual contrast local module and a two-layer attention convolutional GRU (Gated Recurrent Unit). In traditional methods, due to the complex morphology and blurred boundaries of particles, accurate segmentation is often difficult. By combining global features, local features, and multi-scale features, this invention can effectively improve the clarity and accuracy of image boundaries, especially in cases of blurred particle boundaries, significantly improving segmentation results and generating more accurate particle segmentation results.

[0035] 3. Enhanced handling of inter-class similarity and intra-class dissimilarity: This invention effectively solves the confusion problem caused by particle morphological similarity by combining global contextual information with multi-scale fusion of local features. This method enhances the ability to distinguish different particle types, especially when particles have similar morphologies but significant size variations, still achieving accurate classification and segmentation. Furthermore, the introduction of the pyramid pooling module further optimizes the model's ability to handle particles with varied morphologies and large size differences, ensuring high-precision particle classification and segmentation.

[0036] 4. Constructing a Unified and Effective Training Dataset for Oil Particles: This invention significantly improves the accuracy and robustness of particle recognition by constructing a unified and effective training dataset for oil particles. This dataset consists of real particle images acquired using microfluidic digital imaging technology, and its diversity is expanded through synthetic data and challenging samples. Combined with data augmentation techniques, the model can adapt to variations in particle morphology and size. Through standardized and unified dataset design, the problems of data scarcity and low recognition accuracy in existing technologies are solved, significantly improving the model's ability to distinguish between inter-class similarity and intra-class dissimilarity, thereby ensuring more accurate particle recognition and segmentation results.

[0037] 5. Optimized Training and Loss Calculation: This invention introduces a Balanced Segmentation Loss (BSL) function, which not only focuses on overall segmentation accuracy but also finely optimizes the segmentation boundaries. The loss function combines cross-entropy loss and intersection-over-union (IoU) loss, further improving the model's sensitivity to particle boundaries and effectively reducing missegmentation and misclassification. This optimized loss function enables the model to more accurately identify particle shapes and boundaries during training, improving segmentation accuracy and stability. Attached Figure Description

[0038] Figure 1 This is an overall flowchart of a deep learning-driven method and system for identifying particles in insulating oil provided by the present invention;

[0039] Figure 2 This is a network framework diagram of a deep learning-driven method and system for identifying particles in insulating oil provided by the present invention;

[0040] Figure 3 This is a schematic diagram of the attention convolutional GRU in a deep learning-driven particle identification method and system provided by the present invention.

[0041] Figure 4 This is a schematic diagram of a deep learning-driven particle identification method in insulating oil and a local module of the multi-scale context comparison system provided by the present invention.

[0042] Figure 5 This is a schematic diagram of the channel attention module in a deep learning-driven particle identification method and system for insulating oil provided by the present invention.

[0043] Figure 6 This is a schematic diagram of the pyramid pooling module in a deep learning-driven particle identification method and system for insulating oil provided by the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of this invention.

[0045] like Figure 1 As shown, in a first aspect, the present invention provides a deep learning-driven method for identifying particles in insulating oil, comprising:

[0046] The image to be detected is acquired and input into a pre-trained multi-task segmentation model;

[0047] The multi-task segmentation model outputs the detection results of particles in oil based on the image to be detected.

[0048] The multi-task segmentation model is obtained through the following steps:

[0049] Acquire images of particles in oil and construct a multi-task segmentation dataset;

[0050] Preprocess the multi-task segmentation dataset;

[0051] A multi-task segmentation model is built based on a two-branch framework consisting of segmentation and classification branches.

[0052] The segmentation branch is used to segment particles in the oil and obtain particle boundary information; the classification branch is used for identification. Morphological analysis is performed based on the information obtained from the segmentation branch, and then the image extracted from the classification branch is combined for identification to obtain the types of particles in the oil.

[0053] The multi-task segmentation model was trained using the preprocessed multi-task segmentation dataset.

[0054] Set the network training parameters and use the loss function to train the multi-task segmentation model.

[0055] In conjunction with the first aspect, further, the acquisition of oil particle images and the construction of a multi-task segmentation dataset include:

[0056] Images of particles in transformer oil were acquired using a microfluidic digital imaging module.

[0057] Add synthetic granular data and difficult samples to form the final multi-task segmentation dataset.

[0058] In conjunction with the first aspect, the preprocessing of the multi-task segmentation dataset further includes data augmentation of the images in the multi-task segmentation dataset. The data augmentation of the images in the multi-task segmentation dataset includes performing horizontal flipping, vertical flipping, rotation, translation, scaling, cropping, erasing, and color dithering operations on the images in the multi-task segmentation dataset.

[0059] In conjunction with the first aspect, the multi-task segmentation model trained using the preprocessed multi-task segmentation dataset further includes:

[0060] The segmentation branch guides the learning of segmentation features from the image features output by the skeleton network. These image features include both high-order and low-order features. The classification branch uses the input high-order features as auxiliary discriminant information to classify and resolve confusion terms. This dual-branch network framework (segmentation and classification) completes the classification task, ultimately achieving the identification of particles in insulating oil.

[0061] In conjunction with the first aspect, the skeleton network is further described as an optimized Xception deep residual network, wherein Xception includes stage one, stage two, stage three, stage four and stage five; the improvement of the Xception deep residual network includes replacing the ordinary convolutions in the five stages of the network with dilated convolutions, thereby enhancing the network's ability to process multi-scale features and improving the network's performance when processing particles of different sizes.

[0062] In conjunction with the first aspect, further, the extraction of output features from the skeleton network in the segmentation branch includes:

[0063] like Figure 4 As shown, the high-order features output from stage four are fed into a multi-scale contextual contrast local module for processing. This local module extracts features through four 2×2 convolutions with dilated convolution rates of 2, 3, 6, and 8. After feature processing, the local features and the low-order features output from stages one, two, and three are fed into a dual-layer attention convolutional GRU. This operation effectively combines feature information at different scales, enhancing the ability to capture morphological details of particles, particularly significantly improving the segmentation accuracy of tiny particles.

[0064] like Figure 3As shown, the dual-layer attention GRU network replaces the traditional fully connected layers in this task, processing the generated tensors through convolutional layers, thereby reducing the number of model parameters and improving the model's computational efficiency. Simultaneously, the introduction of GRU enables the model to effectively handle temporal relationships between features and further optimizes the model's decision-making process using an attention mechanism. By adding an attention mechanism to GRU, the model can dynamically adjust the degree of attention given to features at different scales, improving segmentation accuracy.

[0065] like Figure 5 As shown, the attention-based convolutional GRU, by introducing a channel attention module, helps the model uncover the intrinsic semantic relationships between multi-scale feature maps. Specifically, the attention mechanism, by introducing a gating structure, limits the model's influence on specific feature maps and can quickly adjust model parameters according to changes in the current target scene, thereby improving the accuracy and robustness of segmentation.

[0066] Furthermore, in conjunction with the first aspect, after receiving the output of the dual-layer attention convolutional GRU, the channel attention module sums it with the features output by the GRU, upsamples it to the original image size, and then feeds it into a 1×1 convolution to integrate the features. This channel attention module uses global average pooling and two fully connected layers to calculate channel weights, thereby selectively enhancing effective features and suppressing ineffective features, achieving feature recalibration. Global average pooling effectively extracts the spatial information of the features, ensuring spatial consistency, while global max pooling retains the most salient information in the features. These operations help enhance the representativeness and discriminative power of the features.

[0067] like Figure 6 As shown, in conjunction with the first aspect, further, the classification branch processes the output features of stages one, two, three, four, and five of the Xception network through the pyramid pooling module to extract global information from the image, and then fuses it with the result of the segmentation branch to finally output the classification result. This pyramid pooling module generates multi-scale feature maps by performing pooling operations at four different scales on the input feature map, then adjusts its channel count through a 1×1 convolution, and upsamples it to its original size to complete feature fusion. Through this multi-scale pooling and feature fusion method, the network can accurately classify particles from multiple scale perspectives and enhance the global consistency of the segmentation results.

[0068] In conjunction with the first aspect, further, the setting of network training parameters and the training of the constructed semantic segmentation model using a loss function include:

[0069] The images in the preprocessed semantic segmentation image dataset are used as input images, and the size of the input images is standardized.

[0070] The semantic segmentation network model is trained using the stochastic gradient descent algorithm with momentum. The weights in the semantic segmentation network model are initialized using Kaiming, and the activation function is ReLU.

[0071] The degree of discrepancy between the data and the actual data is predicted using a loss function, wherein the loss function is the Balanced Segmentation Loss (BSL).

[0072] The formula for the loss function is:

[0073] BSL = αL CE +βL IoU

[0074] Among them, L CE It is the cross-entropy loss, L IoU It is the IoU loss. α and β are hyperparameters that control the weights of the two parts. Their values ​​range from 0 to 1 and usually need to be tuned through cross-validation.

[0075] The formula for the cross-entropy loss function is:

[0076]

[0077] Where N is the pixel value, C is the number of categories, and y i,c P is the true label (0 or 1) of the i-th pixel belonging to category c. i,c It is the predicted probability that the i-th pixel belongs to category c.

[0078] The crossover ratio loss function formula is:

[0079]

[0080] Where A∪B is the intersection of the predicted region and the true region, and A∪B is the union of the predicted region and the true region.

[0081] like Figure 2 As shown, in a second aspect, further, the present invention provides a deep learning-driven particle identification system for insulating oil, comprising:

[0082] The microfluidic digital imaging module is used to acquire the image to be detected and input it into a pre-trained multi-task segmentation model;

[0083] The image recognition module is used by the multi-task segmentation model to output particle image recognition results based on the image to be detected;

[0084] A multi-scale contextual contrast local module is used to enhance the capture and optimization of image boundaries;

[0085] The channel attention module is used to uncover the intrinsic semantic relationships between multi-scale feature maps;

[0086] The pyramid pooling module is used to perform pooling operations at four different scales on the input feature map, generating multi-scale feature maps.

[0087] The multi-task segmentation model is obtained through the following steps:

[0088] Acquire granular images and construct a multi-task segmentation dataset;

[0089] Preprocess the multi-task segmentation dataset;

[0090] A multi-task segmentation model is built based on a two-branch framework consisting of segmentation and classification branches.

[0091] The multi-task segmentation model was trained using the preprocessed multi-task segmentation dataset.

[0092] Set the network training parameters and use the loss function to train the multi-task segmentation model.

[0093] Thirdly, the present invention provides a deep learning-driven particle identification system in insulating oil, including a processor and a storage medium;

[0094] The storage medium is used to store instructions;

[0095] The processor is configured to operate according to the instructions to perform the steps of the method described in any of the first aspects.

[0096] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the first aspects.

[0097] The beneficial effects of this invention are that, compared with the prior art,

[0098] 1. Improved Particle Recognition Accuracy: This invention introduces a multi-task segmentation model, combining segmentation and classification branches, and effectively fusing high-order and low-order features to accurately segment and classify particles based on their morphology. Compared to existing technologies, this invention significantly improves particle recognition accuracy, especially when particles have complex shapes and varying sizes. Through the comprehensive application of a global attention upsampling module, a hybrid pooling module, and a feature pyramid attention module, this invention can better capture local details and global information, ensuring high-precision particle recognition and classification.

[0099] 2. Addressing the Problem of Blurred Image Boundaries: This invention further enhances the capture and optimization of image boundaries by employing a multi-scale contextual contrast local module and a two-layer attention convolutional GRU (Gated Recurrent Unit). In traditional methods, due to the complex morphology and blurred boundaries of particles, accurate segmentation is often difficult. By combining global features, local features, and multi-scale features, this invention can effectively improve the clarity and accuracy of image boundaries, especially in cases of blurred particle boundaries, significantly improving segmentation results and generating more accurate particle segmentation results.

[0100] 3. Enhanced handling of inter-class similarity and intra-class dissimilarity: This invention effectively solves the confusion problem caused by particle morphological similarity by combining global contextual information with multi-scale fusion of local features. This method enhances the ability to distinguish different particle types, especially when particles have similar morphologies but significant size variations, still achieving accurate classification and segmentation. Furthermore, the introduction of the pyramid pooling module further optimizes the model's ability to handle particles with varied morphologies and large size differences, ensuring high-precision particle classification and segmentation.

[0101] 4. Constructing a Unified and Effective Training Dataset for Oil Particles: This invention significantly improves the accuracy and robustness of particle recognition by constructing a unified and effective training dataset for oil particles. This dataset consists of real particle images acquired using microfluidic digital imaging technology, and its diversity is expanded through synthetic data and challenging samples. Combined with data augmentation techniques, the model can adapt to variations in particle morphology and size. Through standardized and unified dataset design, the problems of data scarcity and low recognition accuracy in existing technologies are solved, significantly improving the model's ability to distinguish between inter-class similarity and intra-class dissimilarity, thereby ensuring more accurate particle recognition and segmentation results.

[0102] 5. Optimized Training and Loss Calculation: This invention introduces a Balanced Segmentation Loss (BSL) function, which not only focuses on overall segmentation accuracy but also finely optimizes the segmentation boundaries. The loss function combines cross-entropy loss and intersection-over-union (IoU) loss, further improving the model's sensitivity to particle boundaries and effectively reducing missegmentation and misclassification. This optimized loss function enables the model to more accurately identify particle shapes and boundaries during training, improving segmentation accuracy and stability.

[0103] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0104] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0105] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0106] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A deep learning-driven method for particle identification in insulating oil, comprising: An image to be detected is acquired and input into a pre-trained multi-task segmentation model. Based on the image to be detected, the multi-task segmentation model outputs the detection result of particles in the oil. The multi-task segmentation model is trained through the following steps: acquiring images of particles in transformer oil using a microfluidic digital imaging module to construct a multi-task segmentation dataset; preprocessing the multi-task segmentation dataset; and building a multi-task segmentation model using a dual-branch framework of segmentation and classification branches as the basic structure. The segmentation branch is used to segment particles in the oil and obtain the boundary information of the particles. The classification branch is used for identification; morphological analysis is performed based on the information obtained from the segmentation branch, and then combined with the image extracted from the classification branch for identification to obtain the types of particles in the oil. The multi-task segmentation model trained using a preprocessed multi-task segmentation dataset includes: extracting the output features of the skeleton network in the segmentation branch, wherein the skeleton network includes stage 1, stage 2, stage 3, stage 4 and stage 5; the high-order features output by stage 4 are fed into the multi-scale context contrast local module, and its output local features are fed into the low-order features output by stage 1, stage 2 and stage 3 respectively into a two-layer attention convolutional GRU; the multi-scale context contrast local module processes the local features fed into stage 4 by connecting them to four 2×2 convolutions with dilated convolution rates of 2, 3, 6 and 8 respectively, and then subtracts and fuses the processed local features, and connects them to the input of the channel attention module; the classification branch processes the output features of stage 1, stage 2, stage 3, stage 4 and stage 5 through the pyramid pooling module, extracts the global information of the image, and fuses it with the result of the segmentation branch, finally outputting the classification result; Set the network training parameters and use the loss function to train the multi-task segmentation model.

2. The deep learning-driven particle identification method in insulating oil according to claim 1, characterized in that: The construction of the multi-task segmentation dataset includes: Add synthetic granular data and difficult samples to form the final multi-task segmentation dataset.

3. The deep learning-driven particle identification method in insulating oil according to claim 2, characterized in that: The preprocessing of the multi-task segmentation dataset includes data augmentation of the images in the multi-task segmentation dataset. The data augmentation includes horizontal flipping, vertical flipping, rotation, translation, scaling, cropping, erasing, and color dithering operations.

4. The deep learning-driven particle identification method in insulating oil according to claim 3, characterized in that: The multi-task segmentation model trained using the preprocessed multi-task segmentation dataset includes: In the segmentation branch, the image features output by the skeleton network are used to guide segmentation feature learning, and the image features include high-order features and low-order features; in the classification branch, the high-order features input from the segmentation branch are used as auxiliary discriminant information to classify confusion items.

5. The deep learning-driven particle identification method in insulating oil according to claim 4, characterized in that: The backbone network is an optimized Xception deep residual network.

6. The deep learning-driven particle identification method in insulating oil according to claim 1, characterized in that: The channel attention module uses global average pooling and two fully connected layers to calculate channel weights.

7. A deep learning-driven particle identification system for insulating oil, wherein the system is the same as the deep learning-driven particle identification method for insulating oil described in any one of claims 1-6, comprising: The microfluidic digital imaging module is used to acquire the image to be detected and input it into a pre-trained multi-task segmentation model; The image recognition module is used by the multi-task segmentation model to output particle image recognition results based on the image to be detected; A multi-scale contextual contrast local module is used to enhance the capture and optimization of image boundaries; The channel attention module is used to uncover the intrinsic semantic relationships between multi-scale feature maps; The pyramid pooling module is used to perform pooling operations at four different scales on the input feature map, generating multi-scale feature maps.

8. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-6.

Images