Real-time monitoring system and method for high-risk animals around a substation based on deep learning

Abstract

The application discloses a substation surrounding high-risk animal real-time monitoring system based on deep learning, which combines computer vision and improved YOLOV8 high-risk animal detection model real-time processing technology, accurately detects and identifies high-risk animals appearing near the substation and alarms, guarantees normal operation and safety of the substation, and improves work efficiency of staff; the substation surrounding high-risk animal real-time monitoring system based on deep learning comprises an image acquisition module, a control module, a monitoring alarm module, a data storage module and a user interface, the control module is connected with the image acquisition module, the monitoring alarm module and the storage module in data respectively; the image acquisition module is used for acquiring image information around the substation and transmitting the image information to the control module, the monitoring information is analyzed and processed through the control module, data information is obtained, and the data information is transmitted to the monitoring alarm module and the storage module respectively.

Description

Technical Field

[0001] This invention relates to the field of substation monitoring technology, specifically to a real-time monitoring system for high-risk animals around substations based on deep learning. Background Technology

[0002] Electricity supply is a vital cornerstone for maintaining urban operations, and substations, as a core component of the power system, play a crucial role in converting and distributing electrical energy. However, the high-voltage equipment in substations poses a danger to the surrounding environment, making the areas around substations typically high-risk zones. In such cases, dangerous animals, such as small animals and birds, may unfortunately wander into the hazardous areas of the substation, causing equipment malfunctions, power outages, or even animal deaths. These animals may also cause accidents such as short circuits and fires in power equipment, thereby affecting the stability and security of the power supply. Therefore, establishing a real-time monitoring system to control and prevent the potential dangers posed by these dangerous animals is essential.

[0003] To address this issue, we propose a deep learning-based real-time monitoring system for high-risk animals around substations. This system aims to accurately detect and identify high-risk animals near substations by combining computer vision and real-time processing technologies, and to promptly issue an alarm when they enter dangerous areas. The system can promptly detect and report the activity of high-risk animals around substations, thereby helping staff take timely measures to ensure the normal operation and safety of the substation. Furthermore, the system can improve staff efficiency, reduce waste of human and material resources, and lower substation management costs. Summary of the Invention

[0004] This invention aims to address the technical deficiencies of existing technologies by providing a real-time monitoring system for high-risk animals around substations based on deep learning. By combining computer vision and real-time processing technology with an improved YOLOv8 high-risk animal detection model, the system can accurately detect and identify high-risk animals appearing near substations and issue alarms, thereby ensuring the normal operation and safety of substations and improving the work efficiency of staff.

[0005] This invention provides the following technical solution: a real-time monitoring system for high-risk animals around substations based on deep learning, including an image acquisition module, a control module, a monitoring and alarm module, a data storage module, and a user interface. The control module is connected to the image acquisition module, the monitoring and alarm module, and the storage module respectively.

[0006] The image acquisition module is used to acquire image information around the substation and transmit the image information to the control module. The control module analyzes and processes the monitoring information to obtain data information, which is then transmitted to the monitoring alarm module and the storage module respectively.

[0007] The storage module stores the data information for user interface data monitoring;

[0008] The control module includes an image preprocessing module, an embedded training model algorithm unit, and an improved YOLOV8 high-risk animal detection model. The control module is installed on an embedded computing device.

[0009] Preferably, it also includes a web terminal and a mobile terminal. When the system detects the target model, it triggers the monitoring and alarm module and transmits the alarm signal to the web terminal and the mobile terminal.

[0010] Preferably, the monitoring and alarm module includes multiple sets of audible and visual alarm devices, each of which is installed around the substation and covers a key area, including the safety distance boundary area of ​​high-voltage equipment.

[0011] The image acquisition module includes an image acquisition device, which is a camera network consisting of multiple sets of cameras. Each camera is installed around the substation and covers a key area, including the safety distance boundary area of ​​high-voltage equipment.

[0012] A deep learning-based method for real-time monitoring of high-risk animals around substations includes the following steps:

[0013] S1. The image acquisition module acquires image data and transmits it to the control module;

[0014] S2. The embedded computer device is equipped with a control module. The control module builds a high-risk animal image sample library around the substation through the embedded training model algorithm unit. The image preprocessing module processes the input image information around the substation to achieve a fixed-size image as required by YOLOv8, and then outputs it to the improved YOLOv8 high-risk animal detection model, and finally outputs the location information of the detected animals.

[0015] S3. If the YOLOV8 high-risk animal detection model outputs location information, the monitoring alarm module will be triggered, the audible and visual alarm device will be activated to drive away the animals, and the alarm signal will be transmitted to the Web terminal and mobile terminal at the same time.

[0016] S4. The detected data is stored in the data storage module, and users can view camera footage, receive alarm information, and query historical data in web and / or mobile applications.

[0017] Preferably, the improved YOLOV8 high-risk animal detection model includes a feature extraction network, a feature fusion network, and a head prediction network.

[0018] Preferably, the feature extraction network of the improved YOLOV8 high-risk animal detection model adopts CSPDarknet, which consists of a CSP structure and an SPPF module. The CSP divides the input into two branches: one branch performs basic CBS module convolution operations, and the other branch uses multiple CBS or a combination of CBS and residual structures.

[0019] Preferably, the feature fusion network of the improved YOLOV8 high-risk animal detection model adopts a path aggregation network, which performs upsampling fusion and downsampling fusion of feature maps simultaneously.

[0020] Preferably, the head prediction network of the improved YOLOV8 high-risk animal detection model includes three output layers of different scales. Each output layer consists of three convolutional layers, and each output layer is responsible for predicting targets of different scales. The output layer finally outputs the prediction result, which has the dimension of H×W×(3×(4+1+N)).

[0021] Preferably, the embedded training model algorithm unit employs semi-supervised learning and transfer learning.

[0022] Preferably, the semi-supervised learning and transfer learning include the following steps:

[0023] S100: Collect a large amount of image and video data, including animals and the environment near the substation, through monitoring cameras, drones or manual patrols in the substation.

[0024] S200. Use a pre-trained improved YOLOV8 high-risk animal detection model to identify animals near the substation. Perform inference on unlabeled data and set a confidence threshold. Then, filter the confidence of the prediction boxes according to the threshold and find the prediction results with high confidence, and use them as pseudo-labels.

[0025] S300. Use existing labeled data and generated pseudo-label data to train an initial object detection model;

[0026] S400. Select a suitable source domain to train the model. The source domain includes a large dataset of humans and different animals that is highly similar to the animals that may appear near the substation.

[0027] S500: Fine-tune the model using data from the target domain, using a small learning rate to take advantage of the features of the target domain data and avoid overfitting, or expand the training set through data augmentation techniques such as random scaling, flipping, and rotation.

[0028] S600. Evaluate the trained model and use mean average precision to assess its performance.

[0029] S700, regularly updates the model to adapt to new data and environmental changes.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] In the above technical solution, the real-time monitoring system for high-risk animals around substations based on deep learning includes an image acquisition module, a control module, a monitoring and alarm module, a data storage module, and a user interface. The control module is connected to the image acquisition module, the monitoring and alarm module, and the storage module respectively.

[0032] The image acquisition module is used to acquire image information around the substation and transmit the image information to the control module. The control module analyzes and processes the monitoring information to obtain data information, which is then transmitted to the monitoring alarm module and the storage module. The storage module stores the data information for user interface data monitoring. The control module includes an image preprocessing module, a model training module, and an improved YOLOv8 high-risk animal detection model. The control module is installed on an embedded computing device.

[0033] This invention discloses a real-time monitoring system for high-risk animals around substations based on deep learning. The entire system consists of an image acquisition module, a control module, a monitoring and alarm module, a data storage module, and a user interface. The image acquisition module acquires image information about the area around the substation and transmits it to the control module. The image preprocessing module in the control module fixes the size of the input image. An image sample library of high-risk animals around the substation is built through an embedded training model algorithm unit. The input image information is used to perform target detection through an improved YOLOv8 high-risk animal detection model to obtain data information, which is then transmitted to the monitoring and alarm module and the storage module. The storage module stores the data information for data monitoring in the user interface. Users can monitor the system status, view alarm information, and detection results in real time. This system combines computer vision and real-time processing technology of the improved YOLOv8 high-risk animal detection model to accurately detect and identify high-risk animals appearing near the substation and issue alarms, ensuring the normal operation and safety of the substation and improving the work efficiency of staff. Attached Figure Description

[0034] Figure 1 A flowchart illustrating a specific embodiment of the present invention;

[0035] Figure 2 The feature extraction network structure diagram for improving the YOLOv8 high-risk animal detection model;

[0036] Figure 3 A security status diagram for the user interface;

[0037] Figure 4 This is a warning status diagram for the user interface.

[0038] Figure 5 This is a diagram of dangerous states for the user interface. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0040] like Figures 1-5 As shown, it illustrates a specific embodiment of the present invention: as Figures 1-5 As shown, the real-time monitoring system for high-risk animals around substations based on deep learning disclosed in this invention includes an image acquisition module, a control module, a monitoring and alarm module, a data storage module, and a user interface. The control module is connected to the image acquisition module, the monitoring and alarm module, and the storage module respectively.

[0041] The image acquisition module is used to acquire image information around the substation and transmit the image information to the control module. The control module analyzes and processes the monitoring information to obtain data information, which is then transmitted to the monitoring alarm module and the storage module respectively.

[0042] The storage module stores the data information for user interface data monitoring;

[0043] The control module includes an image preprocessing module, an embedded training model algorithm unit, and an improved YOLOV8 high-risk animal detection model. The control module is installed on an embedded computing device.

[0044] Preferred, such as Figure 1 As shown, it also includes a web terminal and a mobile terminal. When the system detects a target model, it triggers a monitoring alarm module and transmits the alarm signal to both the web terminal and the mobile terminal. By improving the YOLOv8 high-risk animal detection model for target detection, when the system detects a high-risk animal, such as a small animal or bird that has mistakenly entered a dangerous area of ​​the substation, it will trigger an alarm mechanism to notify staff immediately. The alarm can be viewed at any time on the web terminal and the mobile terminal, improving the work efficiency of the staff.

[0045] Preferably, the monitoring and alarm module includes multiple sets of audible and visual alarm devices. Each audible and visual alarm device is installed around the substation and covers a key area, which includes the safety distance boundary area of ​​high-voltage equipment. The audible and visual alarm devices are installed around the substation and cover the key area to achieve comprehensive audible and visual alarms for the substation. Animals can be driven away by sound, and at night, animals can be driven away by light.

[0046] The image acquisition module includes an image acquisition device, which is a camera network consisting of multiple cameras. Each camera is installed around the substation and covers a key area, including the safety distance boundary area of ​​high-voltage equipment. The cameras, installed around the substation to cover the key area, enable comprehensive video surveillance of the substation. The coverage data is transmitted to a data module for target detection. High-definition cameras should be used to provide clear image features and improve monitoring effectiveness.

[0047] A deep learning-based method for real-time monitoring of high-risk animals around substations includes the following steps:

[0048] S1. The image acquisition module acquires image data and transmits it to the control module;

[0049] S2. The embedded computer device is equipped with a control module. The control module builds a high-risk animal image sample library around the substation through the embedded training model algorithm unit. The image preprocessing module processes the input image information around the substation to achieve a fixed-size image as required by YOLOv8, and then outputs it to the improved YOLOv8 high-risk animal detection model, and finally outputs the location information of the detected animals.

[0050] S3. If the YOLOV8 high-risk animal detection model outputs location information, the monitoring alarm module will be triggered, the audible and visual alarm device will be activated to drive away the animals, and the alarm signal will be transmitted to the Web terminal and mobile terminal at the same time.

[0051] S4. The detected data is stored in the data storage module, and users can view camera footage, receive alarm information, and query historical data in web and / or mobile applications.

[0052] The input image preprocessing module is used because YOLOv8 takes fixed-size images as input. The images need to be preprocessed to meet the model's input requirements. Preprocessing includes operations such as scaling, padding, and data augmentation.

[0053] Preferably, the improved YOLOV8 high-risk animal detection model includes a feature extraction network, a feature fusion network, and a head prediction network.

[0054] Preferably, the feature extraction network of the improved YOLOv8 high-risk animal detection model adopts CSPDarknet, which consists of a CSP structure and an SPPF module. The CSP divides the input into two branches: one branch performs basic CBS module convolution operations, and the other branch uses multiple CBS modules or a combination of CBS and residual structures. YOLOv8 consists of a CSP (Cross Stage Partial) structure and an SPPF (Spatial Pyramid Pooling-Fast) module. The CSP structure improves detection accuracy while maintaining low computational complexity. It increases the receptive field of the network and improves detection accuracy by dividing the input feature map into two parts, passing each part through different convolutional layers, and then concatenating them. The SPPF module fuses multi-scale receptive field information through three cascaded pooling operations. It divides the input feature map into multiple sub-maps of different scales, performs pooling operations on each sub-map, and obtains feature maps of different scales. Finally, the SPPF module concatenates these feature maps to obtain a feature map with multi-scale receptive field information, thereby improving the network's ability to detect targets at different scales.

[0055] Preferably, the feature fusion network of the improved YOLOv8 high-risk animal detection model adopts a path aggregation network. The path aggregation network performs upsampling fusion and downsampling fusion simultaneously with feature map upsampling fusion. This upsampling fusion followed by downsampling fusion enriches semantic information and is beneficial for multi-scale target detection.

[0056] Preferably, the improved YOLOv8 high-risk animal detection model's head prediction network includes three output layers at different scales. Each output layer consists of three convolutional layers, and each output layer is responsible for predicting targets at different scales. The output layers ultimately output the prediction result, which has a dimension of H×W×(3×(4+1+N)). The detection head part of YOLOv8 adopts a structure similar to YOLOv3, including three output layers at different scales. Each output layer consists of three convolutional layers, and each output layer is responsible for predicting targets at different scales. The output layers ultimately output the prediction result, which has a dimension of H×W×(3×(4+1+N)). Here, H and W represent the length and width of the feature map input to the detection head, respectively; 4 represents the four coordinates of the bounding box in the prediction result; 1 represents the confidence level of the prediction; N represents the total number of categories in the dataset; and finally, 3 indicates that the detection head can detect a maximum of 3 targets in each grid cell. This structure can predict multiple targets simultaneously, with high efficiency and accuracy, making it suitable for real-time target detection in this system. At the same time, due to the use of a multi-scale output layer structure, it can adapt to targets of different sizes, improving the robustness of detection.

[0057] Preferably, the embedded training model algorithm unit employs semi-supervised learning and transfer learning. Labeling large amounts of data for training deep learning models is time-consuming and expensive. We address the difficulty of data labeling through semi-supervised learning and transfer learning. Traditional supervised learning requires a large amount of labeled data to train the model, while semi-supervised learning and transfer learning can utilize unlabeled data or data from other domains to assist training, thereby improving data utilization. Semi-supervised learning uses a small amount of labeled data and a large amount of unlabeled data to train the model, thus reducing labeling costs. Transfer learning allows partial knowledge from a model trained on one task to be transferred to another related task, reducing the need for labeling on new tasks. This significantly reduces the manual labeling required for traditional model training, resulting in high data utilization and improved work efficiency.

[0058] Preferably, the semi-supervised learning and transfer learning include the following steps:

[0059] S100. Data Collection and Preparation: Collect a large amount of image and video data, including animals and the environment near the substation. This stage can be achieved through methods such as surveillance cameras within the substation, drones, or manual patrols. During data collection, it is important to ensure the diversity and representativeness of the data. By considering factors such as different times, weather, lighting conditions, and seasons, more comprehensive data can be collected. Depending on the needs of the actual application scenario, additional data with significant background interference, complex lighting, or large variations in animal posture can be collected to improve the robustness of the model.

[0060] S200. Generating Pseudo-Labels (Semi-Supervised Learning): Using a pre-trained object detection model, YOLOv8, to identify animals near a substation. Pre-training on large-scale datasets, such as COCO and ImageNet, improves the detection accuracy of YOLOv8 and significantly increases the model's convergence speed. Inference is performed on unlabeled data, and a confidence threshold is set. Predictions with high confidence are then filtered based on the threshold and used as pseudo-labels. To improve the accuracy of pseudo-labels, manual review can be incorporated to confirm or adjust predictions with low confidence, thereby reducing erroneous pseudo-labels.

[0061] S300, Supervised Learning: An initial object detection model is trained using existing labeled data and generated pseudo-label data. During model training, common optimization algorithms such as stochastic gradient descent (SGD) or adaptive optimization algorithms such as Adam can be employed to adjust the model weights by minimizing the loss function, enabling it to better predict the location and category of animals.

[0062] S400, Transfer Learning: Select a suitable source domain to train the model. The source domain should be a large dataset including humans and various animals, with high similarity to animals likely to appear near the substation. This effectively improves the accuracy of animal monitoring systems near substations. Initialize the initial model parameters using the weights of the source domain, which can be achieved by loading some or all of the weights from a pre-trained model. Through transfer learning, the model inherits the features and knowledge of the source domain, enabling it to learn and adapt to data in the target domain more quickly and effectively.

[0063] S500, Fine-Tuning: This involves fine-tuning the model using data from the target domain to further improve its performance on animal monitoring tasks near substations. During fine-tuning, a smaller learning rate can be used to leverage the features of the target domain data and avoid overfitting. Data augmentation techniques such as random scaling, flipping, and rotation can also be used to expand the training set, improving the model's generalization ability and robustness.

[0064] S600, Model Evaluation and Deployment: The trained model is evaluated using standard metrics such as mean average precision (mAP) to assess its performance. During evaluation, techniques such as cross-validation can be used to evaluate the model's robustness in different environments. The model is deployed to a monitoring system near the substation to achieve real-time animal detection. Dedicated hardware such as embedded devices or GPU servers can be used for acceleration.

[0065] S700, Model Updates and Continuous Monitoring: Regularly update the model to adapt to new data and environmental changes. Animal behavior and appearance may change over time, requiring timely model updates to maintain accuracy. Use feedback data from the system to improve model performance. Compare and analyze actual animal data collected by the monitoring system with model predictions to further improve model accuracy and robustness. Simultaneously, animal protection and privacy issues must be considered. During model updates and continuous monitoring, relevant laws, regulations, and ethical guidelines must be followed to ensure compliance. Semi-supervised learning and transfer learning are complex tasks requiring careful parameter tuning and performance monitoring. Furthermore, ensuring proper data collection and preprocessing is crucial, as high-quality data directly impacts model performance. Privacy and ethical issues must also be considered within the monitoring system to ensure compliance.

[0066] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention. These changes involve related technologies well known to those skilled in the art, and all of them fall within the protection scope of the present invention.

[0067] Many other changes and modifications can be made without departing from the concept and scope of this invention. It should be understood that this invention is not limited to the specific embodiments, and the scope of this invention is defined by the appended claims.

Claims

1. A real-time monitoring system for high-risk animals around substations based on deep learning, characterized in that, It includes an image acquisition module, a control module, a monitoring and alarm module, a data storage module, and a user interface. The control module is connected to the image acquisition module, the monitoring and alarm module, and the storage module. The image acquisition module is used to acquire image information around the substation and transmit the image information to the control module. The control module analyzes and processes the monitoring information to obtain data information, which is then transmitted to the monitoring alarm module and the storage module respectively. The storage module stores the data information for user interface data monitoring; The control module includes an image preprocessing module, an embedded training model algorithm unit, and an improved YOLOv8 high-risk animal detection model, which is installed on an embedded computing device. The improved YOLOv8 high-risk animal detection model includes a feature extraction network, a feature fusion network, and a head prediction network. The feature extraction network uses CSPDarknet, which consists of a CSP structure and an SPPF module. The CSP divides the input into two branches: one branch performs basic CBS module convolution operations, and the other branch uses multiple CBS or a combination of CBS and residual structures. The embedded training model algorithm unit employs semi-supervised learning and transfer learning, which includes the following steps: S100: Collect a large amount of image and video data, including animals and the environment near the substation, through monitoring cameras, drones or manual patrols within the substation. S200. Use a pre-trained improved YOLOV8 high-risk animal detection model to identify animals near the substation. Perform inference on unlabeled data and set a confidence threshold. Then, filter the confidence of the prediction boxes according to the threshold and find the prediction results with high confidence, and use them as pseudo-labels. S300. Use existing labeled data and generated pseudo-label data to train an initial object detection model; S400. Select a suitable source domain to train the model. The source domain includes a large dataset of humans and different animals that is highly similar to the animals that may appear near the substation. S500: Fine-tune the model using data from the target domain to take advantage of the features of the target domain data and avoid overfitting, or expand the training set using data augmentation techniques. S600. Evaluate the trained model and use mean average precision to assess its performance. S700, regularly updates the model to adapt to new data and environmental changes.

2. The real-time monitoring system for high-risk animals around substations based on deep learning according to claim 1, characterized in that, It also includes a web terminal and a mobile terminal. When the system detects the target model, it triggers the monitoring and alarm module and transmits the alarm signal to the web terminal and the mobile terminal.

3. The real-time monitoring system for high-risk animals around substations based on deep learning according to claim 2, characterized in that, The monitoring and alarm module includes multiple sets of audible and visual alarm devices. Each audible and visual alarm device is installed around the substation and covers a key area, which includes the safety distance boundary area of ​​high-voltage equipment. The image acquisition module includes an image acquisition device, which is a camera network consisting of multiple sets of cameras. Each camera is installed around the substation and covers a key area, including the safety distance boundary area of ​​high-voltage equipment.

4. A method for real-time monitoring of high-risk animals around substations based on deep learning, using the real-time monitoring system for high-risk animals around substations based on deep learning as described in any one of claims 1-3, characterized in that... Includes the following steps: S1. The image acquisition module acquires image data and transmits it to the control module; S2. The embedded computer device is equipped with a control module. The control module builds a high-risk animal image sample library around the substation through the embedded training model algorithm unit. The image preprocessing module processes the input image information around the substation to achieve a fixed-size image as required by YOLOv8, and then outputs it to the improved YOLOv8 high-risk animal detection model, and finally outputs the location information of the detected animals. S3. If the YOLOV8 high-risk animal detection model outputs location information, the monitoring alarm module will be triggered, the audible and visual alarm device will be activated to drive away the animals, and the alarm signal will be transmitted to the Web terminal and mobile terminal at the same time. S4. The detected data is stored in the data storage module, and users can view camera footage, receive alarm information, and query historical data in web and / or mobile applications.

5. The method for real-time monitoring of high-risk animals around substations based on deep learning according to claim 4, characterized in that, The improved YOLOV8 high-risk animal detection model uses a path aggregation network for its feature fusion network, which performs upsampling fusion and downsampling fusion of feature maps simultaneously.

6. The method for real-time monitoring of high-risk animals around substations based on deep learning according to claim 4, characterized in that, The improved YOLOV8 high-risk animal detection model's head prediction network consists of three output layers at different scales. Each output layer is composed of three convolutional layers, and each output layer is responsible for predicting targets at different scales. The output layer finally outputs the prediction result, which has the dimension of H×W×(3×(4+1+N)).

Images