A method, device, medium and product for detecting particles in transformer oil

Abstract

The application discloses a transformer oil particle detection method, device, medium and product, relates to the field of power transformer detection, and the method comprises the following steps: acquiring a particle image in transformer oil, and constructing a multi-task segmentation data set; according to the multi-task segmentation data set, a dual-branch framework comprising a segmentation branch and a classification branch is adopted to build a multi-task segmentation model; the segmentation branch comprises a global attention upsampling module, a hybrid pool module, a feature pyramid attention module and a self-attention module; the classification branch comprises a full attention module; the multi-task segmentation model takes the particle image as input and takes the detection result of the particle image as output; a to-be-detected particle image is acquired; and the trained multi-task segmentation model is used to determine the detection result of the to-be-detected particle image. The application can realize accurate and efficient automatic identification and classification of particles.

Description

Technical Field

[0001] This application relates to the field of power transformer testing, and in particular to a method, equipment, medium, and product for detecting particulate matter in transformer oil. Background Technology

[0002] Transformer oil, as a critical insulation and cooling medium in power transformers, has a significant impact on the normal operation and service life of the equipment due to its quality and purity. During long-term operation, various tiny particles are generated and accumulate in the insulating oil. These particles mainly originate from the aging of insulation materials, wear of metal components, and the intrusion of external contaminants. Studies have shown that when particulate matter accumulates to a certain concentration in the oil, it may reduce the oil's dielectric strength, increase the risk of partial discharge, and even accelerate the aging of the equipment's insulation, ultimately leading to transformer failure. Therefore, the accurate identification and quantitative detection of particulate matter in transformer oil has become an important task for ensuring the stable operation of power equipment.

[0003] Currently, particle detection technology mainly relies on traditional methods such as optical microscopy and laser particle size distribution. However, optical microscopy depends on manual operation, is highly subjective, and has low processing efficiency, making it difficult to handle large numbers of samples and affecting the accuracy and repeatability of the detection. While laser particle size distribution can provide particle size distribution information, its accuracy is insufficient when dealing with complex or composite particles. Furthermore, these traditional methods are difficult to meet the real-time requirements of transformer oil particulate matter detection.

[0004] Microfluidic imaging technology has become an innovative approach in particle detection in recent years. It uses microfluidic systems to perform high-resolution imaging of liquid samples, enabling high-throughput, non-destructive real-time detection, and has shown broad application prospects in fields such as biomedicine and environmental monitoring. However, its application in transformer oil particle detection is still in the exploratory stage, facing challenges such as blurred image segmentation boundaries, complex and diverse particle morphologies, and insufficient detection accuracy.

[0005] To address the aforementioned issues, there is an urgent need for a method suitable for detecting particles in transformer oil. This method should overcome the limitations of traditional techniques in handling particle morphology diversity, size variations, and blurred segmentation edges, thereby improving detection accuracy and efficiency while reducing reliance on manual operation. Summary of the Invention

[0006] The purpose of this application is to provide a method, equipment, medium, and product for detecting particulate matter in transformer oil, which can achieve accurate and efficient automated identification and classification of particulate matter.

[0007] To achieve the above objectives, this application provides the following solution:

[0008] In a first aspect, this application provides a method for detecting particulate matter in transformer oil, the method comprising:

[0009] Images of particulate matter in transformer oil were acquired, and a multi-task segmentation dataset was constructed.

[0010] Based on a multi-task segmentation dataset, a two-branch framework including a segmentation branch and a classification branch is adopted to build a multi-task segmentation model. The segmentation branch is used to segment particulate matter images and identify the shape and boundaries of the particles. The segmentation branch includes a global attention upsampling module, a mixing pooling module, a feature pyramid attention module, and a self-attention module. The classification branch is used to classify the type of particulate matter based on image segmentation. The classification branch includes a full attention module. The multi-task segmentation model takes particulate matter images as input and outputs the detection results of the particulate matter images.

[0011] Acquire images of the particulate matter to be detected; and use a trained multi-task segmentation model to determine the detection results of the particulate matter images.

[0012] Optionally, the step of acquiring particulate matter images in transformer oil and constructing a multi-task segmentation dataset specifically includes:

[0013] Using a microfluidic system to acquire images of particulate matter in transformer oil;

[0014] The image of particulate matter in transformer oil is preprocessed; the preprocessing includes: data augmentation and blurring.

[0015] A multi-task segmentation dataset is constructed based on the preprocessed images.

[0016] Optionally, the skeleton network of the multi-task segmentation model adopts the ResNet50 deep residual network; in the ResNet50 deep residual network, the five stages of convolution are dilated convolutions.

[0017] Optionally, in the multi-task segmentation model, the low-order features output from Stage 1, Stage 2, and Stage 3 are input into the global attention upsampling module, and the high-order features output from Stage 4 are input into the mixing pool module to extract local information. These local information is then input into the global attention upsampling module along with the low-order features output from Stage 1, Stage 2, and Stage 3, respectively. Simultaneously, the local information is output to the feature pyramid attention module. The high-order features output from Stage 5 are processed by the full attention module to extract global information and output classification results. Global classification is then performed based on these results. The features output from the self-attention module and the global information output from the full attention module are upsampled layer by layer and fused to generate the detection results for the particulate matter image.

[0018] Optionally, the full attention module employs a dual attention mechanism.

[0019] Optionally, the multi-task segmentation model is trained using a stochastic gradient descent algorithm with momentum.

[0020] Optionally, the loss function of the multi-task segmentation model is a boundary-sensitive balance loss function;

[0021] The boundary-sensitive equilibrium loss function BABL is determined using the formula BABL = α·DPCE Loss + β·Sensitivity-Specificity Loss;

[0022] Where, DPCE Loss=-∑ i (y i log(p i )+(1-y i log(1-p) i Distance(i), DPCE Loss is the distance graph penalized cross-entropy loss function, and Sensitivity-Specificity Loss is the sensitivity-specificity loss function. i p is the true label of the i-th pixel. i is the predicted probability of the i-th pixel, Distance(i) is the distance from the i-th pixel to the target boundary; TP, FN and FP are the true positives, false positives and false positives of the multi-task segmentation model, respectively; α and β are numbers ranging from 0 to 1.

[0023] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for detecting particulate matter in transformer oil.

[0024] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for detecting particulate matter in transformer oil.

[0025] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the method for detecting particulate matter in transformer oil.

[0026] According to the specific embodiments provided in this application, this application has the following technical effects:

[0027] This application provides a method, device, medium, and product for detecting particulate matter in transformer oil. By acquiring images of particulate matter in transformer oil and constructing a multi-task segmentation dataset, a unified and effective training dataset for oil particles is built. Semantic segmentation technology in the multi-task segmentation model is used to perform pixel-level classification of microfluidic imaging images, achieving automated identification and classification of particulate matter. This application overcomes the limitations of traditional technologies in handling particle morphological diversity, size variations, and blurred segmentation edges. It also improves detection accuracy and efficiency, reducing reliance on manual operation. This application 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. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a schematic flowchart of a method for detecting particulate matter in transformer oil according to one embodiment of this application;

[0030] Figure 2 This is an overall flowchart of a method for detecting particulate matter in transformer oil according to one embodiment of this application;

[0031] Figure 3 This is a structural diagram of a multi-task segmentation model in a method for detecting particulate matter in transformer oil according to an embodiment of this application;

[0032] Figure 4 This is a schematic diagram of the mixing pool module in a multi-task segmentation model;

[0033] Figure 5 This is a schematic diagram of the global attention upsampling module in a multi-task segmentation model;

[0034] Figure 6 This is a schematic diagram of the feature pyramid attention module in a multi-task segmentation model.

[0035] Figure 7 This is a schematic diagram of the full attention module in a multi-task segmentation model;

[0036] Figure 8 This is a schematic diagram of the self-attention module in a multi-task segmentation model. Detailed Implementation

[0037] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] In one exemplary embodiment, such as Figure 1 and Figure 2 As shown, a method for detecting particulate matter in transformer oil is provided, the method comprising the following steps S101 to S103. Wherein:

[0040] S101, acquire images of particulate matter in transformer oil and construct a multi-task segmentation dataset;

[0041] S101 specifically includes:

[0042] S1, using a microfluidic system to acquire images of particulate matter in transformer oil; the microfluidic system is based on existing microfluidic imaging technology and applied to the detection of particles in transformer oil to achieve real-time, non-destructive detection of particles;

[0043] S2, Preprocessing the image of particulate matter in transformer oil; the preprocessing includes: data augmentation and blurring.

[0044] Data augmentation processing was performed on particulate matter images in transformer oil, including operations such as horizontal flipping, vertical flipping, rotation, translation, scaling, cropping, erasing, adjusting contrast, saturation, and color jitter, and a synthetic dataset was created by combining real and simulated particulate features.

[0045] In adding difficult samples, the first step is to target morphologically similar particle samples by increasing their proportion and labeling them as SimilarParticles. <type>(e.g., Similar_Particles_Carbon_Metal) to enhance the multi-task segmentation model's ability to identify subtle differences between similar particles. Secondly, some small-sized particle images (a few micrometers in size) are randomly distributed throughout the image and labeled as Tiny_Particles_. <size>(e.g., Tiny_Particles_1um) to improve the detection sensitivity of the multi-task segmentation model for tiny particles. Furthermore, Gaussian blurring is used to blur some particle boundaries, generating samples with indistinct edges, named Blurred_Edge_. <degree>(e.g., Blurred_Edge_High) can improve segmentation accuracy when dealing with particles with blurred boundaries.

[0046] S3: Construct a multi-task segmentation dataset based on the preprocessed images.

[0047] The generated synthetic granular data and hard samples are added to the multi-task segmentation dataset for further data augmentation and annotation. The dataset format is standardized to ensure a high degree of consistency between the synthetic and hard samples and the real samples, thereby constructing the final multi-task segmentation dataset.

[0048] S102, based on a multi-task segmentation dataset, employs a two-branch framework including segmentation and classification branches to build a multi-task segmentation model, and as follows... Figure 3 As shown; the segmentation branch is used to segment particulate matter images and identify the shape and boundaries of particles; the segmentation branch includes: a global attention upsampling module, a mixing pool module, a feature pyramid attention module, and a self-attention module; the classification branch is used to classify particulate matter types based on segmentation to distinguish different types of particles; the classification branch includes: a full attention module; the multi-task segmentation model takes particulate matter images as input and outputs the detection results of particulate matter images;

[0049] 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.

[0050] The skeleton network of the multi-task segmentation model adopts the ResNet50 deep residual network; the five stages of convolution in the ResNet50 deep residual network adopt dilated convolution.

[0051] In the multi-task segmentation model, the low-order features output from stages one, two, and three are input into the global attention upsampling module, while the high-order features output from stage four are input into the mixing pool module to extract local information. These local features are then input into the global attention upsampling module along with the low-order features output from stages one, two, and three, respectively. Simultaneously, the local information is output to the feature pyramid attention module. The high-order features output from stage five are processed by the full attention module to extract global information and output classification results. Global classification is then performed based on these results. The features output from the self-attention module and the global information output from the full attention module are upsampled layer by layer and fused to generate the detection results for the particulate matter image.

[0052] The global attention upsampling module achieves global context weighting of features by combining high-order and low-order features. First, low-order features extract local information through a 3×3 convolutional layer, while high-order features are directly introduced into the global attention upsampling module via skip connections. Next, high-order features undergo global pooling to extract global context information, and are then weighted with the low-order features (after a 1×1 convolution) through pointwise multiplication. By guiding the weighted calculation of low-order features with high-order features, local details are not only interpreted within a broader context but also effectively filter out potential background noise. Finally, the weighted low-order features are added to the high-order features, integrating useful global and local information, thereby enhancing feature representation, improving the expression of target features, and reducing the interference of background noise on the segmentation task.

[0053] like Figure 4 As shown, the hybrid pooling module enhances the expressive power of the multi-task segmentation model by simultaneously processing short-term and long-term dependencies and combining multi-scale features. The hybrid pooling module is divided into two branches, which respectively process... Figure 4 The short-term dependence shown in part (a) and as... Figure 4 The long-term dependency is shown in part (b). The short-term dependency input feature x1 is used to extract fine-grained local features through two 2D convolutions at different scales, generating feature maps of different resolutions. Then, an upsampling (UP) operation is used to upsample the low-resolution features to the same size as the input features, and further processing is performed through 2D convolution to capture local dependency information, thereby accurately distinguishing local details and addressing morphological differences between similar objects. The long-term dependency input feature x2 is used to extract long-term dependency features of different dimensions through 1D convolution, processing rows and columns in space respectively. Subsequently, an upsampling operation is used to restore the original feature size, and 2D convolution is used to further integrate the fused long-term dependency features, ensuring that global morphological differences are captured. Finally, the features generated by the two branches are extracted through multi-scale convolution, integrating short-term and long-term dependency information in space. Through this structure, short-term dependencies preserve local details, while long-term dependencies enhance the capture of global morphological features, ultimately integrating the multi-scale features generated by both to solve the problem of inter-class similarity. For example, although carbon particles, metal particles, and bubbles are similar in shape, through this multi-scale fusion of local and global data, the multi-task segmentation model can better distinguish their subtle differences, enhance its ability to identify various objects, and reduce misclassification.

[0054] like Figure 6 As shown, the Feature Pyramid Attention Module enhances the feature extraction capability of the multi-task segmentation model by fusing global information and multi-scale local features. The input features are first reduced in dimensionality using a 1×1 convolution, then processed along two paths: In the global branch, the input undergoes global pooling to extract global contextual information, capturing the overall semantic structure, followed by 1×1 convolution for dimensionality reduction and upsampling back to the original resolution; in the multi-scale branch, the input is processed using 7×7, 5×5, and 3×3 convolutions to extract local features at different scales (16×16, 8×8, 4×4), generating multi-scale feature maps to capture local morphological and size variations. Next, the local features at each scale are multiplied point-by-point with the global features (i.e., weighted fusion), allowing local details to be re-evaluated within the global context. In this way, global contextual information guides the detailed representation of local features, helping the multi-task segmentation model better understand the diversity of local morphology. Finally, an addition operation is performed to add all the fused multi-scale features to the global features, outputting a feature map that contains both global dependencies and multi-scale details. This combination approach enables the multi-task segmentation model to effectively address the differences in morphology and size of the same type of particles. It ensures that even if particles of the same type vary greatly in shape and size, the multi-task segmentation model can still accurately identify and classify them through the integration of global and local multi-scale information, thereby improving the robustness of classification and segmentation.

[0055] like Figure 5 As shown, the global attention upsampling module first receives feature maps from each stage of ResNet50, extracts global contextual information, and restores the feature map of each stage to the same resolution as the previous stage through upsampling operations. Multi-scale features are fused together, enhancing the global dependencies in the feature maps, and these fused features are input to the next layer through pointwise addition. Feature maps of specific stages are also processed by the pyramid attention module, which extracts contextual information at different spatial scales through multi-scale convolution and weights and fuses these multi-scale features with global information, thereby strengthening the ability to capture different spatial details. The features output by the global attention upsampling module and the pyramid attention module are fused together through addition operations to form a feature map containing global contextual information and multi-scale features. The fused feature map is input to the self-attention module. In the self-attention module, higher-order semantic information is further extracted to capture key details in the input features, especially salient features of image boundaries, strengthening the focus on image details. The features processed by the self-attention module combine global, local, and higher-order semantic information, thereby improving the clarity and accuracy of segmentation edges. Ultimately, multi-level and multi-scale fusion processing enables the multi-task segmentation model to effectively solve the problem of edge blurring and achieve more accurate boundary segmentation.

[0056] like Figure 8 As shown, the self-attention module generates the final output feature 4 by integrating information from the global attention upsampling module and the pyramid attention module (features 1, 2, and 3). First, features 1, 2, and 3 undergo 1×1 convolution for dimensionality reduction and feature extraction, followed by dimensionality transformation to convert them into a shape suitable for matrix multiplication. Next, attention weights are calculated between features 1 and 2 through matrix multiplication (i.e., pointwise multiplication) to obtain their correlation, generating a spatial weight matrix. This weight matrix is ​​then multiplied pointwise with feature 3 to form a weighted feature representation. Afterward, the result of the dimensionality transformation is weighted again with feature 3, and finally, the output is fused with features 1, 2, and 3, obtaining the final feature 4 through addition. Through the self-attention mechanism, a spatial dependency weight matrix is ​​generated based on the correlation between input features, thereby weighting and fusing different features, enhancing the correlation between features and global perception capabilities, and ultimately generating output features with greater semantic expressiveness.

[0057] like Figure 7 As shown, the full attention module employs a dual attention mechanism. It includes a positional attention submodule and a channel attention submodule to enhance feature representation capabilities. First, the input image undergoes preliminary feature extraction using ResNet, generating a feature map of size C×H×W, which is then input into the positional and channel attention modules respectively. In the positional attention module, the feature map is reshaped, and matrix multiplication is used to calculate the correlation between different spatial locations, generating a spatial attention matrix to capture the dependencies between spatial locations in the image. This further helps the multi-task segmentation model better understand the spatial distribution differences of particles and performs weighted fusion with the original features to enhance local feature information. In the channel attention module, the feature map is also reshaped, and matrix multiplication is used to calculate the correlation between different spatial locations, generating a spatial attention matrix to capture the dependencies between spatial locations in the image. Matrix multiplication is then used to calculate the dependencies between channels, generating a channel attention matrix, which is subsequently weighted and fused with the original feature map to output enhanced channel features. Through this dual attention mechanism, the multi-task segmentation model can simultaneously capture spatial and channel differences, especially among morphologically similar particles, more accurately identifying subtle local feature differences. Finally, the feature maps from the positional attention and channel attention modules are added element-wise to generate a single feature map that includes spatial context and channel dependency information. This fused feature map enhances the multi-task segmentation model's ability to perceive subtle differences between particles, enabling it to more accurately distinguish between different types of particles and thus improving recognition and classification accuracy.

[0058] The features output by the self-attention module and the global information output by the full attention module are fused through a layer-by-layer upsampling process to generate the final segmentation result. First, the features output by the self-attention module have already obtained enhanced representations of global and local semantic information through positional and channel attention mechanisms, while the full attention module enhances the overall feature representation by capturing the global contextual information of the input image. Next, before fusion, the output features of these two modules undergo an upsampling operation, gradually restoring them to the same resolution as the input image. This is typically achieved through upsampling techniques such as bilinear interpolation or deconvolution, increasing the feature map size layer by layer. During the upsampling process, the resolution of the feature map is continuously improved while preserving the semantic information of the features. The upsampled self-attention and full attention features are combined through element-wise addition or cascaded fusion to ensure that both global semantic and local detail information are effectively integrated. The fused feature map is then processed by a 1×1 convolutional layer and a sigmoid activation function to output the final segmentation result.

[0059] The multi-task segmentation model is trained using a stochastic gradient descent algorithm with momentum. The training process specifically includes:

[0060] The images in the preprocessed semantic segmentation image dataset are used as input images, and the size of the input images is standardized.

[0061] The semantic segmentation network multi-task segmentation model is trained using the stochastic gradient descent algorithm with momentum. The weights in the semantic segmentation network multi-task segmentation model are initialized using Kaiming, and the activation function is ReLU.

[0062] The degree of discrepancy between the data and the actual data is predicted using a loss function, wherein the loss function is the Boundary-Aware Balanced Loss (BABL).

[0063] The boundary-sensitive equilibrium loss function BABL is determined using the formula BABL = α·DPCE Loss + β·Sensitivity-Specificity Loss;

[0064] Where, DPCE Loss=-∑ i (y i log(p i )+(1-y i log(1-p) i Distance(i), DPCE Loss is the distance graph penalized cross-entropy loss function, and Sensitivity-Specificity Loss is the sensitivity-specificity loss function. i p is the true label of the i-th pixel. i is the predicted probability of the i-th pixel, Distance(i) is the distance from the i-th pixel to the target boundary; TP, FN and FP are the true positives, false positives and false positives of the multi-task segmentation model, respectively, as shown in Table 1; α and β are numbers ranging from 0 to 1.

[0065] Table 1

[0066] Predicting the positive class. Predict the negative class. Actual positive class True positive (TP) False negative (FN) Actual negative class False positive (FP) True negative (TN)

[0067] S103, acquire the image of the particulate matter to be detected; and use the trained multi-task segmentation model to determine the detection result of the particulate matter image.

[0068] This application addresses the issue of poor recognition accuracy in existing technologies by combining segmentation and classification branches in a multi-task segmentation model, utilizing both high-order and low-order features to accurately segment and classify particle morphology. Specifically, the application of a global attention upsampling module, a hybrid pooling module, and a feature pyramid attention module enhances the capture of local details and global information, ensuring high-precision particle recognition and classification. A self-attention module and a dual attention mechanism strengthen the capture and optimization of image boundaries. Combining global, local, and multi-scale features improves the clarity and accuracy of image boundaries, effectively handling the problem of blurred particle boundaries and generating more accurate segmentation results. By combining global contextual information and multi-scale fusion of local features, the multi-task segmentation model can solve the confusion problem caused by particle morphological similarity, improving the classification ability of different types of particles. Simultaneously, the method effectively addresses the variability in morphology and size of the same type of particle through the feature pyramid attention module and the hybrid pooling module, ensuring accurate classification and segmentation even when particle shape and size differ significantly. By constructing a unified and effective training dataset for oil particles, the accuracy and robustness of particle recognition are significantly improved. The dataset comprises real particulate images acquired using microfluidic imaging technology, with its diversity expanded through synthetic data and challenging samples. Combined with data augmentation techniques, this enables the multi-task segmentation model to adapt to particle variations in shape and size. A unified and standardized dataset design addresses the issues of data scarcity and low recognition accuracy in existing technologies, enhancing the multi-task segmentation model's ability to distinguish between inter-class similarity and intra-class dissimilarity, ensuring more accurate particle recognition and segmentation. By using a boundary-sensitive balanced loss function, the training process not only focuses on overall segmentation accuracy but also finely optimizes the segmentation boundaries. The loss function incorporates distance graph-penalized cross-entropy and sensitivity-specificity loss, further improving the multi-task segmentation model's sensitivity to particle boundaries and reducing missegmentation and misclassification.

[0069] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for detecting particulate matter in transformer oil.

[0070] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0071] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0072] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0073] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0074] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0075] In this application, all actions to acquire signals, information, or data are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with the authorization granted by the owner of the relevant device.

[0076] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0077] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.< / degree> < / size> < / type>

Claims

1. A method for detecting particulate matter in transformer oil, characterized in that, The method for detecting particulate matter in transformer oil includes: Images of particulate matter in transformer oil were acquired, and a multi-task segmentation dataset was constructed. Based on a multi-task segmentation dataset, a two-branch framework including a segmentation branch and a classification branch is adopted to build a multi-task segmentation model. The segmentation branch is used to segment particulate matter images and identify the shape and boundaries of the particles. The segmentation branch includes a global attention upsampling module, a mixing pooling module, a feature pyramid attention module, and a self-attention module. The classification branch is used to classify the type of particulate matter based on image segmentation. The classification branch includes a full attention module. The multi-task segmentation model takes particulate matter images as input and outputs the detection results of the particulate matter images. Acquire images of the particulate matter to be detected; and use a trained multi-task segmentation model to determine the detection results of the particulate matter images. The backbone network of the multi-task segmentation model adopts the ResNet50 deep residual network; the five stages of convolution in the ResNet50 deep residual network adopt dilated convolution. In the multi-task segmentation model, the low-order features output from stages one, two, and three are input into the global attention upsampling module, while the high-order features output from stage four are input into the mixing pool module to extract local information. These local features are then input into the global attention upsampling module along with the low-order features output from stages one, two, and three, respectively. Simultaneously, the local information is output to the feature pyramid attention module. The high-order features output from stage five are processed by the full attention module to extract global information and output classification results. Global classification is then performed based on these results. The features output from the self-attention module and the global information output from the full attention module are upsampled layer by layer and fused to generate the detection results for the particulate matter image.

2. The method for detecting particulate matter in transformer oil according to claim 1, characterized in that, The process of acquiring particulate matter images in transformer oil and constructing a multi-task segmentation dataset specifically includes: Using a microfluidic system to acquire images of particulate matter in transformer oil; The image of particulate matter in transformer oil is preprocessed; the preprocessing includes: data augmentation and blurring. A multi-task segmentation dataset is constructed based on the preprocessed images.

3. The method for detecting particulate matter in transformer oil according to claim 1 or claim 2, characterized in that, The full attention module employs a dual attention mechanism.

4. The method for detecting particulate matter in transformer oil according to claim 1, characterized in that, The multi-task segmentation model is trained using a stochastic gradient descent algorithm with momentum.

5. The method for detecting particulate matter in transformer oil according to claim 1, characterized in that, The loss function of the multi-task segmentation model is a boundary-sensitive balance loss function; The boundary-sensitive equilibrium loss function BABL is determined using the formula BABL = α·DPCE Loss + β·Sensitivity-Specificity Loss; Where, DPCE Loss=-∑ i (y i log(p i )+(1-y i log(1-p) i Distance(i), DPCELoss is the distance graph penalized cross-entropy loss function, and Sensitivity-Specificity Loss is the sensitivity-specificity loss function. i p is the true label of the i-th pixel. i is the predicted probability of the i-th pixel, Distance(i) is the distance from the i-th pixel to the target boundary; TP, FN and FP are the true positives, false positives and false positives of the multi-task segmentation model, respectively; α and β are numbers ranging from 0 to 1.

6. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the method for detecting particulate matter in transformer oil according to any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for detecting particulate matter in transformer oil as described in any one of claims 1-5.

8. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the method for detecting particulate matter in transformer oil as described in any one of claims 1-5.

Images