Target detection method, system and equipment for automatic transformer disassembling system
Patent Information
- Application Number
- CN202411706039.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-26
- Publication Date
- 2025-03-11
Smart Images
Figure CN119600268A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of target detection of distribution transformers, and in particular relates to a target detection method, system and electronic equipment for an automated transformer disassembly system. Background Art
[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.
[0003] The structure of distribution transformers is complex and diverse, and the accurate identification and disassembly of its components places extremely high demands on the automation system. Traditional robotic arms and automation equipment mainly rely on pre-programmed paths and simple visual systems, which are often unable to adapt to the complex structure and changing appearance characteristics of transformers, thus limiting the flexibility and accuracy of automated operations.
[0004] Before the popularization of deep learning technology, target detection tasks mainly relied on traditional computer vision and machine learning methods. The limitations of traditional methods are that they rely heavily on feature engineering, have low detection accuracy, and are difficult to handle the complex and diverse structures and changing operating environments of distribution transformers. With the rapid development of deep learning technology, target detection algorithms such as YOLO, Faster R-CNN, and SSD have greatly improved the accuracy and robustness of detection. However, these models are usually complex, computationally intensive, and have high hardware requirements. They perform poorly when applied in resource-constrained environments (such as edge devices or embedded systems).
[0005] In summary, current target detection is not suitable for the complex and diverse structures and changeable operating environments of distribution transformers, resulting in poor detection accuracy and stability of distribution transformers. Summary of the invention
[0006] In order to solve the above technical problems, the present invention provides a target detection method, system and electronic equipment for a transformer automated disassembly system, which are suitable for the complex and diverse structures and changeable operating environments of distribution transformers and can improve the detection accuracy and stability of distribution transformers.
[0007] In order to achieve the above object, the present invention adopts the following technical solution:
[0008] A first aspect of the present invention provides a target detection method for a transformer automated disassembly system.
[0009] In one or more embodiments, a target detection method for a transformer automated disassembly system is provided, comprising:
[0010] Obtain the number of the current process of the distribution transformer and its corresponding process actual image set;
[0011] According to the number of the current process, a corresponding target detection model is retrieved from the model library, and the actual image set of the process is identified by using the target detection model to identify whether the actual image set of the current process and the standard image set of the process are consistent with each other;
[0012] Among them, each process in the model library is associated with the target detection model through the process number;
[0013] The target detection model corresponding to each process is trained using the process standard image set corresponding to each process of the distribution transformer.
[0014] As an implementation method, the target detection model is a CrossKD target detection model.
[0015] As an implementation mode, the process standard image set corresponding to each process includes a standard initial state image of the distribution transformer in each process and a standard component image that needs to be assembled or disassembled in each process.
[0016] As an implementation mode, the standard initial state image and the standard component image both include a first sample image with a label and a second sample image without a label; using the process standard image set of each process to train the CrossKD target detection model of each process, including:
[0017] A process standard image set of a process is arbitrarily selected, and based on the first sample image, the teacher model is used to perform knowledge distillation training on the student model, and the trained student model is determined as the initial target detection model of each process;
[0018] Based on the second sample image, using the teacher model to perform knowledge distillation training processing on the initial target detection model of each process, and determining the trained initial target detection model as the CrossKD target detection model of each process;
[0019] The teacher model is a model pre-trained based on the first sample image, and the knowledge distillation training process includes feature layer distillation training and logic layer distillation training.
[0020] As an implementation mode, the teacher model includes a first feature extraction module and a first segmentation module;
[0021] The student model includes a second feature extraction module, and based on the first sample image, performs knowledge distillation training processing on the student model using the teacher model, including:
[0022] Inputting the first sample image into the first feature extraction module and the second feature extraction module respectively to obtain the first image feature output by the teacher model and the second image feature output by the student model;
[0023] Inputting the first image feature into the first segmentation module to obtain a first target detection result output by the teacher model;
[0024] Inputting the second image feature into the first segmentation module to obtain a second object detection result output by the teacher model with respect to the student model;
[0025] Calculating a first loss between the first image feature and the second image feature with respect to feature layer distillation training, and calculating a second loss between the first object detection result and the second object detection result with respect to logic layer distillation training;
[0026] Inversely update the model parameters of the student model based on the first loss and the second loss.
[0027] As an implementation manner, the first sample image is respectively input into the first feature extraction module and the second feature extraction module to obtain the first image feature output by the teacher model and the second image feature output by the student model, including:
[0028] Inputting the first sample image into the first feature extraction module and the second feature extraction module respectively to obtain a first feature map output by the first feature extraction module and a second feature map output by the second feature extraction module;
[0029] Normalizing the pixel values in the first feature map and the second feature map respectively to obtain the first image feature and the second image feature.
[0030] As an implementation manner, in the process of reversely updating the model parameters of the student model based on the first loss and the second loss, the method further includes:
[0031] Determine whether the first loss and the second loss have reached convergence;
[0032] If so, it is determined that the training of the student model is completed, and the trained student model is determined as the initial target detection model.
[0033] As an implementation mode, the actual process image set of the current process is consistent with the process standard image set as follows:
[0034] The actual parts that need to be assembled or disassembled in the current process identified are consistent with the standard parts images that need to be assembled or disassembled in the current process, and the actual initial state image of the distribution transformer in the current process is consistent with the standard initial state of the current process;
[0035] The actual process image set of the current process does not match the process standard image set:
[0036] The actual parts identified to be assembled or disassembled in the current process are inconsistent with the standard parts images to be assembled or disassembled in the current process, and / or the actual initial state image of the distribution transformer in the current process is inconsistent with the standard initial state of the current process.
[0037] A second aspect of the present invention provides a target detection system for a transformer automated disassembly system.
[0038] In one or more embodiments, a target detection system for a transformer automated disassembly system includes:
[0039] A process information acquisition module, which is used to obtain the number of the current process of the distribution transformer and its corresponding process actual image set;
[0040] The target detection and recognition module is used to retrieve the target detection model corresponding to the current process from the model library according to the number of the current process, use the target detection model to recognize the actual image set of the process, and identify whether the actual image set of the current process and the standard image set of the process are consistent with each other;
[0041] Among them, each process in the model library is associated with the target detection model through the process number;
[0042] The target detection model corresponding to each process is trained using the process standard image set corresponding to each process of the distribution transformer.
[0043] A third aspect of the present invention provides an electronic device.
[0044] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the target detection method for a transformer automated disassembly system as described above are implemented.
[0045] Compared with the prior art, the present invention has the following beneficial effects:
[0046] (1) The present invention provides a target detection technology for automatic loading and unloading of distribution transformers. By identifying the actual process image set, it is determined whether the actual process image set of the current process is consistent with the process standard image set. This solves the shortcomings of the prior art in terms of recognition accuracy, computing resource requirements, and environmental adaptability, and realizes target detection of distribution transformers with complex and diverse structures and changeable operating environments, thereby improving the detection accuracy and stability of distribution transformers.
[0047] (2) The present invention transfers the knowledge of a high-precision model to a lightweight model through a knowledge distillation method, thereby achieving high-precision and high-robustness target detection in complex scenarios; and significantly compresses the model size and computational complexity, so that the detector can run efficiently in an embedded system or edge device, thereby reducing dependence on high-performance hardware.
[0048] (3) The lightweight detector of the present invention has strong environmental adaptability through optimized model design and knowledge distillation strategy, can cope with various changes in the production process, and improve the operational flexibility and reliability of the system.
[0049] (4) The present invention detects the initial state of the distribution transformer in each process and the image of the standard parts that need to be assembled or disassembled, and determines whether the initial state of the distribution transformer in each process and the actual parts that need to be assembled or disassembled are correct. If they are correct, the present invention outputs a prompt signal indicating that the actual initial state of the distribution transformer in the current process and the actual parts that need to be assembled or disassembled are normal. If they are incorrect, the present invention outputs an alarm signal indicating that the actual initial state of the distribution transformer in the current process and the actual parts that need to be assembled or disassembled are abnormal, thereby realizing effective monitoring of each process of automatic loading and unloading of the distribution transformer.
[0050] (5) The present invention reduces the demand for high-performance computing resources by achieving lightweight detection models, thereby providing a low-cost, high-efficiency automation solution, which helps to improve overall production efficiency and reduce operating costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0052] Figure 1 1 is a flow chart of a target detection method for a transformer automated disassembly system according to an embodiment of the present invention;
[0053] Figure 2 is a teacher-student guidance framework diagram of the CrossKD target detection model according to an embodiment of the present invention;
[0054] Figure 3 is a flowchart of the training process of the CrossKD target detection model according to an embodiment of the present invention;
[0055] Figure 4 is a schematic diagram of the structure of a target detection system for a transformer automated disassembly system according to an embodiment of the present invention;
[0056] Figure 5 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0057] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0058] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.
[0059] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.
[0060] Distribution transformers play a key role in power transmission and distribution, and the automation of their production, maintenance, and disassembly processes has become an urgent need in the industry. Through automated operations, production efficiency can be significantly improved, labor costs can be reduced, and safety hazards in manual operations can be reduced. Traditional computer vision and machine learning methods, such as sliding window and classifier-based methods, identify targets by moving a fixed-size sliding window in the image one by one and using a classifier (such as support vector machine or naive Bayes) to classify the contents in the window. However, this method has low computational efficiency, especially poor performance in multi-scale detection, and is not accurate enough in locating transformer components in complex scenes. Feature matching-based visual guidance locates and identifies targets by extracting feature points in the image (such as SIFT, SURF, HOG, etc.) for matching, but these manually designed features perform poorly under illumination changes, perspective changes, and occlusion, and are difficult to meet the needs of complex transformer structures. In addition, rule-based image segmentation and traditional pattern recognition methods rely on manual segmentation of images or segmentation based on rules such as color and texture, combined with simple pattern recognition algorithms (such as decision trees or random forests) to achieve target detection. This type of method is highly dependent on predefined rules and manually designed features, has poor scalability, and is not robust enough to complex backgrounds. Another type of method is motion detection based on background modeling, which detects moving objects in the video through background subtraction techniques (such as Gaussian mixture models), which is suitable for target detection in dynamic environments. However, in the task of automatic disassembly of distribution transformers, many components are stationary or move slowly, so this method has limited detection capabilities for static targets. Template matching-based methods detect transformer components by searching for areas in the image that are similar to predefined templates, but these methods are only applicable to objects with fixed shapes and are difficult to effectively identify components in transformers that have different shapes and are easily deformed.
[0061] In order to solve the problem that current target detection is not suitable for the complex and diverse structures and changeable operating environments of distribution transformers, resulting in poor detection accuracy and stability of distribution transformers, the present invention provides a target detection method, system and electronic equipment for a transformer automatic disassembly system, which are suitable for the complex and diverse structures and changeable operating environments of distribution transformers and can improve the detection accuracy and stability of distribution transformers.
[0062] Figure 1 FIG. 1 is a flow chart of a target detection method for a transformer automatic disassembly system according to an embodiment of the present invention. Figure 1 The target detection method for the transformer automated disassembly system in the present embodiment shown may include:
[0063] S101, obtaining the number of the current process of the distribution transformer and its corresponding process actual image set.
[0064] In this embodiment, the actual process image set of the current process of the distribution transformer includes: an actual initial state image of the distribution transformer in the current process and images of actual parts that need to be assembled or disassembled in the current process.
[0065] S102, according to the number of the current process, retrieve the corresponding target detection model from the model library, use the target detection model to identify the actual image set of the process, and identify whether the actual image set of the current process and the standard image set of the process are consistent with each other;
[0066] Among them, each process in the model library is associated with the target detection model through the process number;
[0067] The target detection model corresponding to each process is trained using the process standard image set corresponding to each process of the distribution transformer.
[0068] In the specific implementation process, the actual process image set of the current process is consistent with the process standard image set:
[0069] The actual parts that need to be assembled or disassembled in the current process identified are consistent with the standard parts images that need to be assembled or disassembled in the current process, and the actual initial state image of the distribution transformer in the current process is consistent with the standard initial state of the current process;
[0070] The actual process image set of the current process does not match the process standard image set:
[0071] The actual parts identified to be assembled or disassembled in the current process are inconsistent with the standard parts images to be assembled or disassembled in the current process, and / or the actual initial state image of the distribution transformer in the current process is inconsistent with the standard initial state of the current process.
[0072] In one or more embodiments, it is also determined based on the judgment result whether to output a prompt signal regarding the actual initial state of the distribution transformer in the current process and whether the actual parts to be assembled or disassembled are detected normally.
[0073] If the actual process image set of the current process matches the process standard image set, a prompt signal is outputted indicating that the actual initial state of the distribution transformer in the current process and the actual parts to be assembled or disassembled are detected normally;
[0074] If the actual process image set of the current process does not match the process standard image set, an alarm signal is output regarding the actual initial state of the distribution transformer in the current process and abnormal detection of actual parts that need to be assembled or disassembled.
[0075] In this embodiment, the target detection model is a CrossKD target detection model, such as Figure 2 shown.
[0076] The process standard image set corresponding to each process includes the standard initial state image of the distribution transformer in each process and the standard component images that need to be assembled or disassembled in each process.
[0077] Specifically, the standard initial state image and the standard component image both include a first sample image with a configuration label and a second sample image without a configuration label; the CrossKD target detection model of each process is trained using the process standard image set of each process, such as Figure 3 As shown, including:
[0078] A process standard image set of a process is arbitrarily selected, and based on the first sample image, the teacher model is used to perform knowledge distillation training on the student model, and the trained student model is determined as the initial target detection model of each process;
[0079] Based on the second sample image, using the teacher model to perform knowledge distillation training processing on the initial target detection model of each process, and determining the trained initial target detection model as the CrossKD target detection model of each process;
[0080] The teacher model is a model pre-trained based on the first sample image, and the knowledge distillation training process includes feature layer distillation training and logic layer distillation training.
[0081] It should be noted that the acquisition module of the present invention can simultaneously acquire a first sample image with configured labels and a second sample image without configured labels; based on the first sample image, the teacher model is used to perform knowledge distillation training on the student model, and the trained student model is determined as the initial target detection model; based on the second sample image, the teacher model is used to perform knowledge distillation training on the initial target detection model to obtain the target detection model. By training the target detection model using knowledge distillation, the feature extraction capability of the target detection model can be improved, so that the target detection model can ensure the accuracy of target detection while being successfully applied on the vehicle platform. In addition, during the training process, by adopting a semi-supervised training method, the unlabeled image data can be accurately used to improve the model performance, and there is no need to label a large amount of image data, which can reduce the labeling cost of the training samples.
[0082] Specifically, the process standard image sets corresponding to each process include multiple standard initial state images and multiple standard component images. The number of standard initial state images and standard component images can be selected according to actual needs. In the present invention, the number of standard initial state images and standard component images is preferably 200. Among the 200 standard initial state images, the shooting angle of each standard initial state image varies.
[0083] Among them, a preset proportion of images are labeled images, recorded as the first sample images, and the rest are unlabeled images, recorded as the second sample images. Preferably, the preset proportion is 50%. After obtaining the sample images, a small number of sample images can be pre-labeled. Specifically, the classification label of each pixel in the sample image can be marked. The classification label can be the number of each process. Among the 200 standard parts images, the shooting angle of each standard part image varies. Among them, a preset proportion of images are labeled images, recorded as the first sample images, and the rest are unlabeled images, recorded as the second sample images.
[0084] Preferably, the preset ratio is 50%. After acquiring the sample images, a small number of sample images can be pre-labeled. Specifically, the classification labels of each pixel in the sample images can be labeled. The classification labels can be the names of various components. Then, the labeled sample images are determined as the first sample images, and the remaining unlabeled sample images are determined as the second sample images. In the subsequent training of the target detection model, semi-supervised training can be performed based on the first sample images and the second sample images, and the unlabeled image data can be accurately used to improve the model performance. There is no need to label a large amount of image data, which can reduce the labeling cost of training samples.
[0085] Specifically, the teacher model includes a first feature extraction module and a first segmentation module;
[0086] The student model includes a second feature extraction module, and based on the first sample image, performs knowledge distillation training processing on the student model using the teacher model, including:
[0087] Inputting the first sample image into the first feature extraction module and the second feature extraction module respectively to obtain the first image feature output by the teacher model and the second image feature output by the student model;
[0088] Inputting the first image feature into the first segmentation module to obtain a first target detection result output by the teacher model;
[0089] Inputting the second image feature into the first segmentation module to obtain a second object detection result output by the teacher model with respect to the student model;
[0090] Calculating a first loss between the first image feature and the second image feature with respect to feature layer distillation training, and calculating a second loss between the first object detection result and the second object detection result with respect to logic layer distillation training;
[0091] Inversely update the model parameters of the student model based on the first loss and the second loss.
[0092] In some optional embodiments, the CWD method can be used to input the first sample image into the first feature extraction module in the teacher model and the second feature extraction module in the student model respectively. The first feature extraction module will output multiple first feature maps under the preset feature layer number, and then the softmax normalization processing can be performed on each pixel value in each of the multiple first feature maps, so that each pixel value is distributed between 0-1, and the first image feature T1 is obtained; at the same time, the second feature extraction module will output multiple second feature maps under the preset feature layer number, and then the softmax normalization processing can be performed on each pixel value in each of the multiple second feature maps, so that each pixel value is distributed between 0-1, and the second image feature S1 is obtained. Among them, the preset feature layer number can be set according to the actual application scenario, such as 256 layers, 512 layers, etc.
[0093] The CrossKD method can be used to forward propagate the teacher model, input the first image feature into the first segmentation module, and generate a category probability map for each pixel, that is, the first target detection result; the second image feature output by the student model is taken out, and the second image feature is input into the segmentation head (that is, the first segmentation module) of the teacher model. The first segmentation module is used to process the second image feature into the number of categories to be segmented. For example, the number of segmentation categories is 19, and the shape of the feature vector at this time is [2,256,224,224]. After being processed by the segmentation head, the vector shape is [2,19,224,224], where 2 is the number of batches, 19 is the number of channels, and 224 is the width and height of the image. Finally, the category probability map for each pixel output by the first segmentation module is obtained, that is, the second target detection result of the student model output by the teacher model.
[0094] The first sample image is input into the first feature extraction module and the second feature extraction module respectively to obtain the first image feature output by the teacher model and the second image feature output by the student model, including:
[0095] Inputting the first sample image into the first feature extraction module and the second feature extraction module respectively to obtain a first feature map output by the first feature extraction module and a second feature map output by the second feature extraction module;
[0096] Normalizing the pixel values in the first feature map and the second feature map respectively to obtain the first image feature and the second image feature.
[0097] In some other optional embodiments, the DIST method can be used. After obtaining the first image feature and the second image feature, the Pearson correlation coefficient can be used as a loss function to calculate the first loss of the second image feature and the first image feature, and the student model parameters are updated by back propagation based on the first loss. The first loss is used to represent the completion of the feature layer distillation training under the model training with labeled data. The closer the first loss is to 0, the higher the completion of the feature layer distillation training. After obtaining the first target detection result and the second target detection result, the Pearson correlation coefficient can be used as a loss function to calculate the second loss of the second target detection result and the first target detection result, and the student model parameters are updated by back propagation based on the second loss. The second loss is used to represent the completion of the logic layer distillation training under the model training with labeled data. The closer the second loss is to 0, the higher the completion of the logic layer distillation training.
[0098] In the process of reversely updating the model parameters of the student model based on the first loss and the second loss, the method further includes:
[0099] Determine whether the first loss and the second loss have reached convergence;
[0100] If so, it is determined that the training of the student model is completed, and the trained student model is determined as the initial target detection model.
[0101] In some other optional embodiments, the teacher model includes a first feature extraction module and a first segmentation module, the initial target detection model includes a third feature extraction module and a second segmentation module, and the initial target detection model is subjected to the knowledge distillation training process by using the teacher model based on the second sample image, including:
[0102] Inputting the second sample image into the first feature extraction module and the third feature extraction module respectively to obtain the third image feature output by the teacher model and the fourth image feature output by the initial object detection model;
[0103] Inputting the third image feature into the first segmentation module to obtain a third object detection result output by the teacher model;
[0104] Inputting the fourth image feature into the second segmentation module to obtain a fourth object detection result output by the initial object detection model;
[0105] Calculating a third loss between the third image feature and the fourth image feature with respect to feature layer distillation training, and calculating a fourth loss between the third object detection result and the fourth object detection result with respect to logic layer distillation training;
[0106] The model parameters of the initial target detection model are reversely updated based on the third loss and the fourth loss.
[0107] Specifically, the method further includes:
[0108] The DIST method can be used. After obtaining the third image feature and the fourth image feature, the Pearson correlation coefficient can be used as the loss function to calculate the third loss of the fourth image feature and the third image feature, and the initial target detection model parameters can be updated by back propagation based on the third loss. The third loss is used to represent the completion of the feature layer distillation training under the model training of unlabeled data. The closer the third loss is to 0, the higher the completion of the feature layer distillation training. After obtaining the third target detection result and the fourth target detection result, the Pearson correlation coefficient can be used as the loss function to calculate the fourth loss of the fourth target detection result and the third target detection result, and the initial target detection model parameters can be updated by back propagation based on the fourth loss. The fourth loss is used to represent the completion of the logic layer distillation training under the model training of unlabeled data. The closer the fourth loss is to 0, the higher the completion of the logic layer distillation training.
[0109] Accordingly, when back propagation is used to update the student model parameters, the model gradient corresponding to the third loss or the fourth loss can be calculated based on the back propagation algorithm, and the model parameters of the initial target detection model can be updated according to the model gradient. In a specific application scenario, the number of training rounds of the initial target detection model can be set according to the data size of the second sample image, and after it is determined that the iterative training of the initial target detection model reaches the number of training rounds, it is determined whether the third loss and the fourth loss have reached convergence.
[0110] In summary, the technical solution in this application can improve the feature extraction capability of the target detection model by training the target detection model through the method of knowledge distillation, so that the target detection model can have the same performance as the large model when it can be successfully applied to the target monitoring scheme of automatic loading and unloading of distribution transformers, and ensure the accuracy of target detection. In addition, during the training process, by adopting a semi-supervised training method, the unlabeled image data is accurately used to improve the model performance, and a large amount of image data annotation is not required, which can reduce the annotation cost of training samples.
[0111] When the third loss and the fourth loss reach a convergence state, it can be determined that the initial target detection model training is completed, and the trained initial target detection model is determined as the final target detection model. Among them, the preset thresholds corresponding to the feature layer distillation training and the logic layer distillation training can be the same or different, and can be a value between 0 and 1. The closer the preset threshold is to 0, the higher the training accuracy requirement of the corresponding feature layer distillation training or logic layer distillation training. The specific value can be set according to the actual application scenario, and no specific limitation is made here.
[0112] Specifically, the step of inputting the second sample image into the first feature extraction module and the third feature extraction module respectively to obtain the third image feature output by the teacher model and the fourth image feature output by the initial target detection model includes:
[0113] Inputting the second sample image into the first feature extraction module and the third feature extraction module respectively to obtain a third feature map output by the first feature extraction module and a fourth feature map output by the third feature extraction module;
[0114] Normalization is performed on the pixel values in the third feature map and the fourth feature map, respectively, to obtain the third image feature and the fourth image feature.
[0115] Specifically, the method further includes:
[0116] Determine whether the third loss and the fourth loss have reached convergence; if so, determine that the initial target detection model training is completed, and determine the trained initial target detection model as the target detection model.
[0117] In traditional knowledge distillation methods, target conflict is a key issue that affects the effectiveness of knowledge transfer. Specifically, classification tasks usually assign fixed category labels to each image, while in object detection tasks, label assignment is dynamic and relies on a manually designed allocator to determine the label of each detection position. This dynamic assignment will cause target conflict between the teacher model and the student model, which is manifested as inconsistency in the prediction results of the two. Especially in complex scenarios, the automatic loading and unloading and disassembly tasks of distribution transformers require precise positioning and classification, and the problem of target conflict is particularly prominent.
[0118] In classification tasks, each image is assigned a fixed category label, and the distillation goal between the teacher and student models is usually unified, that is, to simulate the category probability distribution of the teacher model as much as possible. However, in object detection tasks, label assignment is dynamic and relies on manually designed allocators. These allocators are responsible for assigning labels to each detection location, usually based on distance, overlap, or other specific rules. This dynamic assignment method will cause a conflict of goals between the teacher model and the student model, which is manifested as inconsistent prediction results between the two.
[0119] In complex scenarios, such as the automatic loading and unloading and disassembly tasks of distribution transformers, environmental factors (such as light, obstacles) and device diversity increase the uncertainty of label assignment, which intensifies the goal conflict problem. The main impact of goal conflict is that the student model cannot accurately simulate the prediction of the teacher model, resulting in insufficient or inaccurate knowledge transfer during the distillation process.
[0120] To quantify this impact, this example counts the ratio of conflicting areas to positive areas under different teacher-student combinations in the COCO minival dataset. The results show that even under the same label assignment strategy, there are still significant differences between the true labels and the distillation targets, especially when the teacher and the student use different allocators, the conflict area increases significantly. This shows that even with a reasonable design of the assignment strategy, the problem of target conflict is still difficult to completely avoid, which seriously affects the learning effect of the student model and, in turn, affects the accuracy of the automated operation of the distribution transformer.
[0121] When the conflict area increases (where the difference between the true label and the distillation target is greater than 0.5), the prediction performance drops significantly. Traditional prediction imitation methods usually directly minimize the prediction difference between the teacher and the student, as shown below:
[0122]
[0123] Among them, p s and p t are the prediction vectors of the student and teacher models respectively; D pred (·) indicates the function used to calculate p s and p t The difference between the loss function, such as KL divergence in classification tasks, L1 loss or LD loss in regression tasks; S(·) is the region selection principle, which is used for image regions. A weight is generated for each position r in .
[0124] In the distribution transformer task, due to the diversity of equipment and the complexity of the environmental background, ignoring high uncertainty areas may cause the model to lose key information and weaken the detection performance. Therefore, it is necessary to design a strategy that can gracefully handle target conflicts rather than simply reducing the weight.
[0125] In order to deal with the above-mentioned goal conflict problem, this paper proposes a cross-head knowledge distillation method (CrossKD), which aims to optimize the knowledge transfer process and enhance the performance of the student model in the automatic loading and unloading task of distribution transformers. The core idea of CrossKD is to generate cross-head predictions by passing the intermediate feature maps of the student model to the detection head of the teacher model. This method increases the consistency between the teacher and student models and effectively reduces the impact of goal conflicts.
[0126] Specifically, for a dense detector (such as RetinaNet), its detection head consists of a series of convolutional layers. Assume that the detection head has a total of n convolutional layers, of which the first n-1 layers generate the intermediate feature map f i , and the last layer generates the prediction result p. The operation of CrossKD is to pass the feature map of the middle layer of the student model to the corresponding layer of the teacher detection head, for example, the feature map of the i-th layer of the student Passed to the teacher's i+1th convolutional layer Thus generating cross-head predictions This process can be expressed as:
[0127]
[0128] Among them, S(·) and They are the region selection principle and the normalization factor. This loss calculation method avoids the trap of direct prediction differences between teachers and students, allowing the student's partial detection head to better optimize the real target. This distillation method is particularly suitable for complex task scenarios, such as automatic loading and unloading and disassembly tasks of distribution transformers, because it can effectively handle multi-target and multi-scale detection requirements.
[0129] Through this cross-head prediction approach, CrossKD effectively reduces the inconsistency between the teacher and student models, thereby alleviating the negative impact of target conflict on the distillation effect. Compared with traditional prediction imitation methods, CrossKD is unique in that it does not simply minimize the direct prediction difference between the teacher and the student, but enhances the collaborative work between the teacher and the student through the sharing of cross-head features, so that the student model can better learn the deep features and prediction strategies of the teacher model.
[0130] CrossKD is designed to be particularly suitable for complex task scenarios, such as automatic loading and unloading and disassembly of distribution transformers. Such tasks usually have multi-target and multi-scale detection requirements, and are also accompanied by problems of equipment diversity and environmental complexity. By sharing feature maps, CrossKD can better handle multi-target and multi-scale detection requirements and improve the generalization and robustness of the model.
[0131] In addition, since the automated operation of distribution transformers requires high-precision target localization and classification, CrossKD can show significant performance improvement in such scenarios. It can effectively reduce the prediction error caused by target conflicts, thereby improving the accuracy and efficiency of the system in practical applications.
[0132] The overall training loss function consists of the weighted sum of the detection loss and the distillation loss:
[0133]
[0134] in, and Represents classification loss and regression loss, respectively, calculated between the student model prediction and the true target. The additional CrossKD loss and Used for distillation between cross-head predictions and teacher predictions.
[0135] In the automatic loading and unloading task of distribution transformers, the classification loss can use quality focus loss to process soft label prediction, and the regression loss uses GIoU loss or KL divergence according to the regression form of the detector. Through the optimization of CrossKD, the model can complete the detection, positioning and classification of transformers more accurately and efficiently, thereby improving the accuracy and robustness of the automatic loading and unloading system.
[0136] In the automatic loading and unloading task of distribution transformers, the classification loss can use Quality Focal Loss to process soft label prediction. This loss function can better balance the ratio between positive and negative samples, and is particularly suitable for dense detection tasks. The regression loss uses GIoU loss or KL divergence according to the regression form of the detector. These loss functions can provide more accurate bounding box adjustments during the regression process and improve the model's prediction accuracy for the target location.
[0137] CrossKD's distillation loss aligns the cross-head predictions with the teacher model's predictions, thus avoiding the pitfall of direct teacher-student prediction differences in traditional distillation methods. This design can better maintain the prediction strategy of the teacher model while allowing the student model to more accurately capture the true target through the optimization of part of the detection head.
[0138] This optimization method is particularly suitable for multi-target and multi-scale detection in complex scenarios. For example, the automated operating system of a distribution transformer needs to accurately detect, locate, and classify targets in different environments and equipment types. CrossKD improves the design of distillation loss to enable the model to better cope with these challenges, thereby improving the accuracy and robustness of the automated system.
[0139] CrossKD combines the optimization of detection loss and distillation loss to improve the performance of the student model in a more comprehensive way. By adjusting the weights between different loss terms, CrossKD is able to achieve a balance between accuracy and computational efficiency. This is particularly important for distribution transformer automation operating systems, as it needs to minimize computational overhead while ensuring high accuracy.
[0140] The comprehensive optimization design of CrossKD can not only show excellent performance improvement in experiments, but also improve the operational efficiency of automatic loading and unloading and disassembly tasks of distribution transformers in practical applications. With the rapid development of smart grid and automation technology, CrossKD is expected to become an important tool for improving the detection performance of industrial automation systems and provide strong support for more efficient and intelligent distribution transformer management.
[0141] Figure 4 is a schematic diagram of the structure of a target detection system for a transformer automatic disassembly system according to an embodiment of the present invention. Figure 1 The target detection method for transformer automatic disassembly system corresponds to Figure 2 As shown, the target detection system for the transformer automated disassembly system in this embodiment may include:
[0142] A process information acquisition module 401 is used to obtain the number of the current process of the distribution transformer and its corresponding process actual image set;
[0143] The target detection and recognition module 402 is used to retrieve the target detection model corresponding to the current process from the model library according to the number of the current process, use the target detection model to recognize the process actual image set, and identify whether the process actual image set of the current process and the process standard image set are consistent with each other;
[0144] Among them, each process in the model library is associated with the target detection model through the process number;
[0145] The target detection model corresponding to each process is trained using the process standard image set corresponding to each process of the distribution transformer.
[0146] It should be noted here that Figure 4 The various modules in the target detection system for transformer automatic disassembly system are Figure 1 The various steps in the target detection method for the transformer automated disassembly system correspond one to one, and the specific implementation process is the same, which will not be repeated here.
[0147] Reference Figure 5 , a schematic diagram of an electronic device is given. It should be noted that, Figure 5 The electronic device 500 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.
[0148] like Figure 5 As shown, electronic device 500 includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 502 or a program loaded from a storage part 508 into a random access memory (RAM) 503. In RAM 503, various programs and data required for system operation are also stored. Central processing unit 501, ROM 502 and RAM 503 are connected to each other via a bus 504. Input / output (I / O) interface 505 is also connected to bus 504.
[0149] The following components are connected to the I / O interface 505: an input section 506 including a keyboard, a mouse, etc.; an output section 507 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a local area network (LAN) card, a modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the I / O interface 505 as needed. A removable medium 511, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 510 as needed, so that a computer program read therefrom is installed into the storage section 508 as needed.
[0150] When the central processing unit 501 in the electronic device of this embodiment executes the program, the following is achieved: Figure 1 The steps in the target detection method for the transformer automated disassembly system are shown.
[0151] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer readable medium, the computer program including a computer program for executing Figure 1In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 509, and / or installed from the removable medium 511. When the computer program is executed by the central processing unit 501, various functions defined in the apparatus of the present application are executed.
[0152] in, Figure 1 The computer program instructions corresponding to the method shown may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0153] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.
[0154] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A target detection method for a transformer automated disassembly system, characterized in that: include: Obtain the number of the current process of the distribution transformer and its corresponding process actual image set; According to the number of the current process, a corresponding target detection model is retrieved from the model library, and the actual image set of the process is identified by using the target detection model to identify whether the actual image set of the current process and the standard image set of the process are consistent with each other; Among them, each process in the model library is associated with the target detection model through the process number; The target detection model corresponding to each process is trained using the process standard image set corresponding to each process of the distribution transformer.
2. The target detection method for the transformer automated disassembly system according to claim 1, characterized in that: The target detection model is a CrossKD target detection model.
3. The target detection method for the transformer automated disassembly system according to claim 2, characterized in that: The process standard image set corresponding to each process includes the standard initial state image of the distribution transformer in each process and the standard component images that need to be assembled or disassembled in each process.
4. The target detection method for the transformer automated disassembly system according to claim 3, characterized in that: The standard initial state image and the standard component image both include a first sample image with a configuration label and a second sample image without a configuration label; using the process standard image set of each process to train the CrossKD target detection model of each process, including: A process standard image set of a process is arbitrarily selected, and based on the first sample image, the teacher model is used to perform knowledge distillation training on the student model, and the trained student model is determined as the initial target detection model of each process; Based on the second sample image, using the teacher model to perform knowledge distillation training processing on the initial target detection model of each process, and determining the trained initial target detection model as the CrossKD target detection model of each process; The teacher model is a model pre-trained based on the first sample image, and the knowledge distillation training process includes feature layer distillation training and logic layer distillation training.
5. The target detection method for the transformer automated disassembly system according to claim 4, characterized in that: The teacher model includes a first feature extraction module and a first segmentation module; The student model includes a second feature extraction module, and based on the first sample image, performs knowledge distillation training processing on the student model using the teacher model, including: Inputting the first sample image into the first feature extraction module and the second feature extraction module respectively to obtain the first image feature output by the teacher model and the second image feature output by the student model; Inputting the first image feature into the first segmentation module to obtain a first target detection result output by the teacher model; Inputting the second image feature into the first segmentation module to obtain a second object detection result output by the teacher model with respect to the student model; Calculating a first loss between the first image feature and the second image feature with respect to feature layer distillation training, and calculating a second loss between the first object detection result and the second object detection result with respect to logic layer distillation training; Inversely update the model parameters of the student model based on the first loss and the second loss.
6. The target detection method for the transformer automated disassembly system according to claim 5, characterized in that: Inputting the first sample image into the first feature extraction module and the second feature extraction module respectively to obtain the first image feature output by the teacher model and the second image feature output by the student model, including: Inputting the first sample image into the first feature extraction module and the second feature extraction module respectively to obtain a first feature map output by the first feature extraction module and a second feature map output by the second feature extraction module; Normalizing the pixel values in the first feature map and the second feature map respectively to obtain the first image feature and the second image feature.
7. The target detection method for a transformer automated disassembly system according to claim 5, characterized in that: In the process of reversely updating the model parameters of the student model based on the first loss and the second loss, the method further includes: Determine whether the first loss and the second loss have reached convergence; If so, it is determined that the training of the student model is completed, and the trained student model is determined as the initial target detection model.
8. The target detection method for a transformer automated disassembly system according to claim 1, characterized in that: The actual process image set of the current process is consistent with the process standard image set: The actual parts that need to be assembled or disassembled in the current process identified are consistent with the standard parts images that need to be assembled or disassembled in the current process, and the actual initial state image of the distribution transformer in the current process is consistent with the standard initial state of the current process; The actual process image set of the current process does not match the process standard image set: The actual parts identified to be assembled or disassembled in the current process are inconsistent with the standard parts images to be assembled or disassembled in the current process, and / or the actual initial state image of the distribution transformer in the current process is inconsistent with the standard initial state of the current process.
9. A target detection system for a transformer automated disassembly system, characterized in that: include: A process information acquisition module, which is used to obtain the number of the current process of the distribution transformer and its corresponding process actual image set; The target detection and recognition module is used to retrieve the target detection model corresponding to the current process from the model library according to the number of the current process, use the target detection model to recognize the actual image set of the process, and identify whether the actual image set of the current process and the standard image set of the process are consistent with each other; Among them, each process in the model library is associated with the target detection model through the process number; The target detection model corresponding to each process is trained using the process standard image set corresponding to each process of the distribution transformer.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the target detection method for a transformer automated disassembly system are implemented as described in any one of claims 1 to 8.