Self-learning power transmission line fault detection system based on cloud edge cooperation

Abstract

The invention discloses a self-learning power transmission line fault detection system based on cloud-edge cooperation, which comprises an edge end, a cloud end and a user end, and is characterized in that the edge end collects an image sample, obtains a fault identification result and a confidence score by using a lightweight edge end detection model, judges whether an anomaly or a defect exists based on the confidence score, and sends the anomaly or the defect to the user end; when abnormity or defects exist, the side end generates alarm information and transmits the alarm information, a fault identification result and an image sample to the cloud end; the cloud side forwards the alarm information and the image sample to the user side, and if the user side carries out fault label marking and submits the fault label to the cloud side, the cloud side forms training data based on the fault label submitted by the user side; and if the user side does not submit the marked image sample to the cloud side, the cloud side forms training data based on a high-precision integrated model fault identification analysis result, and the cloud side performs incremental training on a lightweight edge detection model stored by the cloud side by using the training data to obtain model parameters and sends the model parameters to the edge side so as to iteratively optimize the model of the edge side.

Description

Technical Field

[0001] The present invention relates to the field of power transmission line fault detection systems, and specifically to a self-learning power transmission line fault detection system based on cloud-edge collaboration. Background Art

[0002] With the continuous growth of power demand and the expansion of power grid infrastructure, the inspection tasks of transmission lines have gradually increased. Due to the problems of low efficiency, high cost, and high risk in the traditional manual inspection method, it has been gradually replaced by an intelligent inspection system based on artificial intelligence (AI) and unmanned aerial vehicle (UAV). The UAV inspection system can obtain real-time image data of transmission lines by carrying sensors such as high-resolution cameras and infrared imaging devices, and use deep learning algorithms for fault detection and analysis, thereby improving the inspection efficiency and accuracy.

[0003] In the power inspection system, the real-time nature of data and the processing ability are crucial. The introduction of edge computing technology enables some data processing to be performed at the inspection device end, thereby reducing the data transmission delay and improving the system response speed. However, the computing power of edge devices is relatively limited and cannot handle complex deep learning models and large-scale data, resulting in a computing bottleneck when dealing with complex scenarios or large amounts of data. To make up for this defect, some computing tasks are transferred to the cloud for processing. Although cloud processing is powerful, it faces high transmission delays and network bandwidth requirements. Especially in a large-scale power inspection system, this delay may affect the real-time nature of fault detection.

[0004] With the continuous development of edge technology, how to efficiently deploy deep learning networks at the edge has become a new challenge. If the deep learning model is fully deployed at the edge, it will result in a huge amount of computation and reduce the inspection efficiency.

[0005] Regarding this problem, in the existing technical literature (Lei Z, Ren S, Hu Y, et al. Latency-aware collaborative perception[C] / / European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022: 316-332.), Lei et al. proposed a partitioned framework for latency awareness and joint accuracy, which reduces the computational pressure at the edge by dividing some computational tasks between the cloud and the edge. However, this method requires manual setting and has a low level of intelligence. In the existing technical literature (Kang Y, Hauswald J, Gao C, et al. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge[J]. ACM SIGARCH Computer Architecture News, 2017, 45(1): 615-629.), Kang et al. proposed the Neurosurgeon system, which can automatically partition neural networks to improve system performance, but it has high hardware requirements, limiting the flexibility and universality of practical applications.

[0006] In addition to the deployment optimization problem, how to ensure high-precision detection of transmission line defects remains one of the core challenges in power line inspection. In the existing technical literature (Li H, Liu L, Du J, et al. An improved YOLOv3 for foreign objects detection of transmission lines[J]. IEEE Access, 2022, 10:45620-45628.), Li et al. proposed an improved detection model based on the YOLOv3 model, which has certain improvement in detection performance. In the existing technical literature (Wu J, Cheng S, Pan S, et al. Detection method based on improved faster R-CNN for pin defect in transmission lines[C] / / E3S Web of Conferences. EDP Sciences, 2021, 300: 01011.), Wu et al. adopted a transmission line defect detection method based on RCNN and achieved good results in small target detection. However, these methods generally have problems such as insufficient detection accuracy and long time consumption, and it is difficult to meet the requirements of high efficiency and accuracy for large-scale power line inspection.

[0007] In addition, the existing power line inspection systems lack flexible intelligent learning and optimization mechanisms. Traditional deep learning models are often fixed and cannot be dynamically adjusted and optimized according to the actual inspection situation. When facing a new inspection environment or a new type of fault, the existing systems usually rely on manual intervention to update the model and cannot achieve fast self-adaptation and automatic optimization. This results in the difficulty of maintaining consistency and stability in the detection accuracy and alarm strategies of the model under different environmental conditions. Summary of the Invention

[0008] The present invention provides a self-learning power transmission line fault detection system based on cloud-edge collaboration to solve the problems of high processing time consumption, large delay, and difficulty in dynamically adjusting the edge model and alarm strategy existing in the edge devices used for power transmission line fault detection in the prior art.

[0009] To achieve the above object, the technical solution adopted by the present invention is as follows: A self-learning power transmission line fault detection system based on cloud-edge collaboration includes an edge end, a cloud end, and a user end, wherein: The edge device collects real-time image samples of the power transmission line, and uses a lightweight edge detection model to detect and identify the image samples to obtain a fault identification result and a confidence score of the fault identification result. When the confidence score of the fault identification result exceeds the preset threshold T, the edge device determines that there is an abnormality or defect in the power transmission line, generates an alarm message, and transmits the alarm message, the fault identification result, and the collected image samples to the cloud; After receiving the alarm message, the fault identification result, and the image samples from the edge device, the cloud forwards the alarm message and the image samples to the user device; If the user device performs fault label annotation on the received image samples and submits them to the cloud, the cloud compares the fault labels submitted by the user with the fault identification results of the edge device; when the comparison result is consistent, the cloud directly adds the image samples from the edge device to the training data to expand the training data; if the comparison result is inconsistent, the cloud uses the fault labels submitted by the user as the true labels of the image samples, adds the image samples with the true labels to the training data to expand the training data, and increases the training weight of the image samples with the true labels; If the user device does not submit annotated image samples to the cloud, the cloud re-performs fault identification analysis on the image samples from the edge device through the high-precision integrated model in the cloud; when the fault identification analysis result of the high-precision integrated model in the cloud is consistent with the fault identification result of the edge device, the cloud directly adds the image samples from the edge device to the training data to expand the training data; if the analysis results are inconsistent, the cloud uses the fault identification analysis result of the high-precision integrated model in the cloud to label the image samples, and then adds the labeled image samples to the training data to expand the training data, and increases the training weight of the image samples; Finally, the cloud uses the expanded training data to perform incremental training on the lightweight edge detection model stored in the cloud. After the training is completed, the cloud transmits the model parameters of the trained lightweight edge detection model back to the edge device, and the edge device loads the model parameters from the cloud, thereby iteratively optimizing the lightweight edge detection model at the edge device.

[0010] Further, the lightweight edge detection model in the edge device extracts features in the power transmission line image samples through depthwise separable convolution and pruning optimization, generates candidate regions based on the region proposal network, then classifies and identifies the fault types, and assigns a confidence score to each fault identification result.

[0011] Further, the lightweight edge detection model in the edge device includes a feature extraction part, a region proposal network, and a classification module.

[0012] Further, the feature extraction part includes ten convolutional kernels and three depthwise separable convolutions.

[0013] Further, a lightweight edge detection model in the edge terminal introduces a lightweight strategy, uses the L1 norm to measure the importance of channel weights, and eliminates channels with channel weights less than a threshold.

[0014] Further, when training the lightweight edge detection model in the edge terminal, Focal Loss is used as the loss function.

[0015] Further, the cloud also dynamically adjusts the preset threshold T of the edge terminal and transmits it to the edge terminal.

[0016] Further, the high-precision integrated model in the cloud is a backbone detection model combined with a voting fusion strategy, where the backbone detection model is composed of RetinaNet, Swin Transformer, DETR, and YOLOv5.

[0017] The present invention designs a lightweight edge detection model for edge devices to improve the efficiency and accuracy of processing inspection images and alarms at the edge.

[0018] The lightweight edge detection model also detects the detection ability and response speed through the cloud and the user terminal. Specifically, the edge device is responsible for real-time sampling of image samples and fault detection and identification. When an anomaly is detected, it sends alarm information, fault identification results, and image sample data to the cloud and the user terminal; after receiving the samples, the cloud uses a high-precision integrated model formed by combining a backbone detection model composed of RetinaNet, Swin Transformer, DETR, and YOLOv5 with a voting fusion strategy for further fault detection and judgment; the user terminal can perform manual annotation on the image sample data to provide reliable manual labels to assist in training the lightweight edge detection model stored in the cloud; the lightweight edge detection model stored in the cloud performs efficient self-learning incremental training in the cloud according to user feedback, and the trained model parameters are transferred back to the edge to update the lightweight edge detection model at the edge.

[0019] The system of the present invention can also dynamically adjust the alarm strategy according to the cloud learning results and user feedback to dynamically adjust the preset threshold T of the edge terminal to ensure high-precision detection in different environments (such as fog, night, etc.) and reduce false alarms and missed alarms.

[0020] In the feature extraction part of the lightweight edge detection model at the edge, the present invention introduces ten convolutional kernels and three layers of depthwise separable convolutions to decompose the traditional convolution operation into two steps: depthwise convolution and pointwise convolution. Compared with the standard convolution, in the feature extraction part of the lightweight edge detection model of the present invention, by introducing depthwise separable convolutions, the number of parameters and the amount of computation are greatly reduced, which helps to reduce the computational burden on edge devices. For each input channel, in the feature extraction part of the lightweight edge detection model of the present invention, a separate convolutional kernel is used for convolution operation without mixing with other channels, thereby reducing the amount of computation. Information fusion between channels is performed through 1x1 convolutional kernels, which further improves the expression ability of the lightweight edge detection model while maintaining computational efficiency.

[0021] The present invention combines edge computing, cloud computing, and self-learning mechanisms, and can dynamically adjust the detection model and alarm strategy in different environments, improve the detection accuracy and real-time performance of the power inspection system, reduce false alarms and missed alarms, thereby improving the intelligent level and efficiency of power transmission line inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a data flow diagram of an embodiment of the present invention.

[0023] Figure 2 is an overall flowchart of the model in an embodiment of the present invention.

[0024] Figure 3 is an overall architecture diagram of the edge model in an embodiment of the present invention.

[0025] Figure 4 is a structure diagram of depthwise separable convolution in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] The present invention will be further described below with reference to the drawings and embodiments.

[0027] As Figure 1 shown, this embodiment discloses a self-learning power transmission line fault detection system based on cloud-edge collaboration, including an edge side, a cloud side, and a user side, where the cloud side is respectively communicatively connected to the edge side and the user side, and through the cooperation of the three, the intelligence and real-time performance of the power transmission line inspection system can be effectively improved. The data processing process of this embodiment is as Figure 2 shown, where: In this embodiment, the edge device is deployed at key positions of the power transmission line, and uses a sampling frequency of 4 times per minute to regularly collect real-time image samples of the power transmission line. The lightweight edge detection model in the edge device is used to detect and identify the image samples to obtain the fault identification result and the confidence score of the fault identification result. Through depthwise separable convolution and pruning optimization, the lightweight edge detection model can efficiently extract features from the image samples of the power transmission line, generate candidate regions based on the Region Proposal Network (RPN), and then classify and identify the fault types, such as wire breakage or foreign object hanging. After the lightweight edge detection model assigns a confidence score to each fault identification result, it is compared with the preset threshold T. If the confidence exceeds the preset threshold T, the edge device determines that there is an abnormality or defect in the power transmission line, generates an alarm message, and transmits the alarm message, the fault identification result, and the collected image samples to the cloud.

[0028] Specifically, the architecture of the lightweight edge detection model in the edge device is as Figure 3 shown, including a feature extraction part, a Region Proposal Network (RPN), and a classification module. The feature extraction part extracts the deep feature information of the picture sample, the Region Proposal Network generates high-quality candidate regions from the input image sample, and the classification module performs further classification and regression of the bounding box.

[0029] The feature extraction part in the lightweight edge detection model uses ten convolutional layers and introduces three depthwise separable convolutional layers. The convolutional layer consists of a convolutional kernel and a Relu activation layer. During data processing, the image sample is first input into three pre-convolutional layers. After depth convolution operation, a preliminary feature map is obtained, which contains the low-level feature information of the image. Then, the preliminary feature map is input into the depthwise separable convolutional layer as Figure 4 shown. The depthwise separable convolution module disassembles the traditional convolution operation into two steps: depth convolution and pointwise convolution. First, depth convolution is used to process each input channel with a separate convolutional kernel to extract the features of each channel and obtain the depth convolution feature map. Subsequently, the depth convolution feature map is processed through pointwise convolution (i.e., 1×1 convolution) to achieve information fusion between different channels and obtain a refined feature map. Subsequently, the features processed by the depthwise separable convolution enter the pooling layer for pooling operation. After pooling, they continue to be sent to subsequent similar feature extraction layers until the last layer of the extraction module. In addition, when performing convolution operations, channel weights are also calculated and pruning operations are carried out.

[0030] After being processed by the multi-layer convolution and depthwise separable convolution of the feature extraction module, deep feature information is finally obtained. This information contains the high-level semantic features of the image and is used by the subsequent detection module. Finally, deep feature information is obtained. Compared with the standard convolution, the feature extraction part in the lightweight edge detection model uses depthwise separable convolution to greatly reduce the number of parameters and the amount of computation, which helps to reduce the computational burden of edge devices. While further improving the model's expression ability, it maintains the computational efficiency.

[0031] The data processing process of the Region Proposal Network (RPN) in the lightweight edge detection model is as follows: Through the sliding window mechanism, high-quality candidate regions are generated in the image samples. Specifically, RPN slides a small window on the feature map output by the convolutional layer, generates multiple anchor boxes with different scales and aspect ratios for each position, and classifies each anchor box to determine whether it contains the target and adjusts the position of the anchor box. Through the region proposal network, multiple candidate regions can be quickly generated in the image, including regions that may contain the target. These candidate regions will be used for subsequent object detection, not only generating high-quality candidate regions, but also greatly improving the efficiency of candidate region generation, thereby optimizing the computational burden and inference speed of edge devices. High-quality candidate regions are generated in the image samples.

[0032] The data processing process of the classification module in the lightweight edge detection model is as follows: The candidate regions generated by the region proposal network are passed through RoI Pooling (Region of Interest Pooling) to extract feature maps of a fixed size, and then input into the fully connected layer in the classification layer for processing, performing object classification and bounding box regression, thereby further classifying and regressing the bounding boxes to obtain the fault recognition result.

[0033] In the lightweight edge detection model, a lightweight strategy is introduced. The L1 norm is used to measure the importance of channel weights, and the scoring function S(c) is used to evaluate the importance of each output channel c. Wc represents the channel weight, and the scoring is shown in the following formula:

[0034] By calculating the scoring function S(c), the channels with channel weights less than the threshold are removed, as shown in the following formula:

[0035] Among them, P represents the set of retained channel indices; represents the given threshold.

[0036] Thus, in the feature extraction module, the design of combining pruning and depthwise separable convolution module enables the lightweight edge detection model of this embodiment to significantly reduce the computational complexity and storage requirements while retaining high accuracy.

[0037] Moreover, when training the lightweight edge detection model, Focal Loss is used as the loss function to replace the commonly used traditional cross-entropy loss. Cross-Entropy Loss is often used to measure the difference between the prediction and the true label. However, the cross-entropy loss pays too much attention to easy-to-separate samples (such as background and non-defect regions), resulting in weak learning of difficult-to-separate samples (such as broken wires, foreign object hanging, etc.), which may affect the detection accuracy of the model. Therefore, this embodiment uses Focal Loss to replace the traditional cross-entropy loss. Focal Loss improves the detection ability of the model for complex faults (such as broken wires, foreign object hanging) by applying a lower weight to easy-to-separate samples and enhancing the attention to difficult-to-separate samples. Focal Loss is shown as follows:

[0038] Where: represents the output value of the loss function.; represents the predicted probability for the sample; represents a weighting factor used to balance the importance between classes; represents a focusing parameter used to control the influence of easy and difficult samples and reduce the loss of samples that have been correctly classified. A larger value will make the model pay more attention to difficult-to-classify samples.

[0039] In this embodiment, after the cloud receives the alarm information, fault recognition result, and image sample from the edge, it forwards the alarm information and the image sample to the user side.

[0040] If the user side annotates the received image sample with a fault label and submits it to the cloud, indicating that the user side has intervened manually, then the cloud compares the fault label submitted by the user with the fault recognition result of the edge; when the comparison result is consistent, the cloud directly adds the image sample from the edge to the training data to expand the training data; if the comparison result is inconsistent, the cloud uses the fault label submitted by the user as the true label of the image sample, adds the image sample with the true label to the training data to expand the training data, and increases the training weight of the image sample with the true label.

[0041] If the client does not submit the labeled image samples to the cloud, it means that there is no manual intervention on the client side. Then the cloud enables the high-precision integrated model to re-perform fault identification and analysis on the image samples from the edge side. In this embodiment, the high-precision integrated model in the cloud is a backbone detection model combined with a voting fusion strategy. The backbone detection model is composed of RetinaNet (), Swin Transformer (Liu Z, Lin Y, Cao Y, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2021: 10012-10022.), DETR (Carion N, Massa F, Synnaeve G, et al. End-to-end object detection with transformers[C] / / European conference on computer vision. Cham: Springer International Publishing, 2020: 213-229.), and YOLOv5. After inputting the target samples into multiple models, the weighted bounding box fusion method is used to vote on the detection results of RetinaNet, Swin Transformer, DETR, and YOLOv5, effectively integrating the prediction advantages of each model, and the voting result is used as the final detection result.

[0042] If the fault identification and analysis results of the high-precision integrated model in the cloud are consistent with the fault identification results of the edge side, the cloud directly adds the image samples from the edge side to the training data to expand the training data; if the analysis results are inconsistent, the fault identification and analysis results of the high-precision integrated model in the cloud are used to label the image samples, and then the labeled image samples are added to the training data to expand the training data, and the training weight of the image samples is increased.

[0043] Based on this, the cloud combines user feedback and its own judgment results to expand the training data. The cloud stores a lightweight edge detection model that is consistent with the edge device. After obtaining new samples, the lightweight edge detection model is incrementally trained on the cloud to achieve iterative optimization. With the computing power and security advantages of the cloud, iterative optimization can be carried out faster. After the lightweight edge detection model is optimized on the cloud, the parameters of the updated lightweight edge detection model obtained after training are transmitted back to the edge device. The edge device loads the model parameters from the cloud, and thus the lightweight edge detection model is iteratively optimized at the edge. This process is efficient and secure, ensuring that the edge device can quickly resume operation after the update. The optimized detection model is more accurate and efficient, enabling the lightweight edge detection model to gradually adapt to different scenarios.

[0044] The cloud also obtains the false alarm rate and missed alarm rate based on the comparison results, and then dynamically adjusts the preset threshold T of the edge device and transmits the adjusted preset threshold T to the edge device to ensure the detection performance of the lightweight edge detection model in the edge device under complex environments (such as at night or in bad weather). Thus, the lightweight edge detection model in the edge device can self-train and adapt, so that it can optimize the detection effect of the edge device according to the actual situation with little or no manual intervention.

[0045] Through the cloud-edge collaborative self-learning defect detection framework of this embodiment, the edge device is responsible for real-time data collection and preliminary fault detection, and uploads the preliminary alarm information and relevant sample data to the cloud and the user terminal. The cloud server further analyzes and verifies the data through a complex deep learning model and optimizes the detection model using the self-learning mechanism. The user terminal can manually annotate the sample data to provide accurate fault labels to assist the cloud in improving the recognition accuracy of the model. Through the feedback from the cloud, the edge device can update the detection model and alarm strategy, so as to maintain high detection ability and stability in various complex environments (such as bad weather, a large amount of noise interference, etc.), achieve fast alarm, reduce false alarms and missed alarms, and significantly improve the overall performance of the system.

[0046] Moreover, through the optimized design of pruning, depthwise separable convolution, and Focal Loss in this embodiment, the computational complexity of the lightweight edge detection model in the edge device is significantly reduced, the detection accuracy for complex faults is enhanced, and at the same time, the detection strategy is dynamically adjusted in combination with the cloud feedback to achieve efficient and intelligent real-time alarm, which can comprehensively improve the detection performance and response efficiency of the power inspection system.

[0047] The preferred embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. The embodiments described herein are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Among the various specific technical features described in the above specific embodiments, they can be combined in any suitable manner without contradiction. As long as such a combination does not violate the idea of the present invention, it should also be regarded as the content disclosed in this disclosure. To avoid unnecessary repetition, the present invention will not separately describe various possible combination methods.

[0048] The present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention and without departing from the design idea of the present invention, various modifications and improvements made by those skilled in the art to the technical solution of the present invention should fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the claims.

Claims

1. A self-learning power transmission line fault detection system based on cloud-edge collaboration, characterized in that: It includes edge, cloud and user side, among which: The edge collects real-time image samples of the power transmission line, and uses a lightweight edge detection model to detect and identify the image samples to obtain a fault identification result and a confidence score of the fault identification result. When the confidence score of the fault identification result exceeds a preset threshold value T, the edge determines that there is an abnormality or defect in the power transmission line, generates an alarm message, and transmits the alarm message, the fault identification result and the collected image samples to the cloud. After receiving the alarm information, fault identification results and image samples from the edge, the cloud forwards the alarm information and image samples to the user end; If the user terminal annotates the received image samples with fault labels and submits them to the cloud, the cloud will compare the fault labels submitted by the user with the fault identification results of the edge. If the comparison results are consistent, the cloud will directly add the image samples from the edge to the training data to expand the training data. If the comparison results are inconsistent, the cloud will use the fault labels submitted by the user as the true labels of the image samples, add the image samples with the true labels to the training data to expand the training data, and increase the training weights of the image samples with the true labels. If the user end does not submit annotated image samples to the cloud, the high-precision integrated model on the cloud will be used to re-perform fault identification analysis on the image samples from the edge. If the fault identification analysis results of the high-precision integrated model on the cloud are consistent with those on the edge, the cloud will directly add the image samples from the edge to the training data to expand the training data. If the analysis results are inconsistent, the image samples will be annotated with the fault identification analysis results of the high-precision integrated model on the cloud, and then the annotated image samples will be added to the training data to expand the training data, and the training weight of the image samples will be increased. Finally, the cloud uses the expanded training data to perform incremental training on the lightweight edge detection model stored in the cloud. After the training, the cloud transmits the model parameters of the trained lightweight edge detection model back to the edge. The edge loads the model parameters from the cloud, thereby iteratively optimizing the lightweight edge detection model at the edge.

2. The self-learning power transmission line fault detection system based on cloud-edge collaboration according to claim 1 is characterized in that: The lightweight edge detection model in the edge extracts features from power transmission line image samples through deep separable convolution and pruning optimization, generates candidate regions based on a region proposal network, and then classifies and identifies the fault type, and assigns a confidence score to each fault identification result.

3. The self-learning power transmission line fault detection system based on cloud-edge collaboration according to claim 2 is characterized in that: The lightweight edge detection model in the edge includes a feature extraction part, a region proposal network, and a classification module.

4. The self-learning power transmission line fault detection system based on cloud-edge collaboration according to claim 3 is characterized in that: The feature extraction part includes ten convolution kernels and three depth-wise separable convolutions.

5. The self-learning power transmission line fault detection system based on cloud-edge collaboration according to claim 3 is characterized in that: The lightweight edge detection model in the edge introduces a lightweight strategy and uses the L1 norm to measure the importance of channel weights to eliminate channels whose channel weights are less than a threshold.

6. The self-learning power transmission line fault detection system based on cloud-edge collaboration according to claim 3 is characterized in that: Focal Loss is used as the loss function when training the lightweight edge detection model in the edge.

7. The self-learning power transmission line fault detection system based on cloud-edge collaboration according to claim 1 is characterized in that: The cloud also dynamically adjusts the preset threshold T of the edge and transmits it to the edge.

8. The self-learning power transmission line fault detection system based on cloud-edge collaboration according to claim 1 is characterized in that: The high-precision integrated model in the cloud is a backbone detection model combined with a voting fusion strategy, wherein the backbone detection model is composed of RetinaNet, Swin Transformer, DETR, and YOLOv5.

Images