Booster station fire smoke detection method based on YOLO11
Through the YOLO11-based fire smoke detection method, combined with knowledge migration and knowledge distillation technology, a lightweight detection model is generated, and a variety of data sources are used for fusion detection, which solves the problems of slow response, high false alarm rate and complex calculations in the existing technology, and achieves high-precision and real-time monitoring of fire smoke detection effects.
Patent Information
- Application Number
- CN202510201518.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-24
- Publication Date
- 2025-06-27
AI Technical Summary
The existing fire smoke detection methods at boost stations are slow to respond, have high false alarm rates, are complex in calculations, and are difficult to apply in real-time monitoring, and it is difficult to achieve high detection accuracy for a single data source.
Using the YOLO11-based fire smoke detection method, a lightweight fire smoke detection model is generated by pre-training the YOLO11 model and combining knowledge migration and knowledge distillation technology, and fusion detection is performed by combining visible light and infrared image data.
It improves the accuracy and reliability of fire smoke detection, improves detection speed and efficiency, reduces false alarm rate, and realizes the application of real-time monitoring.
Smart Images
Figure CN120219794A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technology of fire and smoke detection in a step-up substation, and in particular to a method for fire and smoke detection in a step-up substation based on YOLO11. Background Art
[0002] A step-up substation is an important part of the power system, and its safe operation directly affects the stability of the power system. Traditional methods for fire and smoke detection in step-up substations mainly rely on manual inspections or traditional image processing techniques, but these methods have problems such as slow response, high false alarm rate, and high labor costs. With the rapid development of deep learning technology, object detection models based on deep learning have become the mainstream solution for fire and smoke detection. However, these models are usually computationally complex and difficult to apply in real-time monitoring. In addition, a single data source (such as only using infrared images) is likely to overlook potential fire risks and is difficult to achieve higher detection accuracy. Therefore, how to improve the detection speed and efficiency of the model while ensuring high accuracy and combining multiple data sources for fusion detection has become an urgent problem to be solved. Summary of the Invention
[0003] The purpose of the present invention is to overcome the problems existing in the above-mentioned prior art and provide a method for fire and smoke detection in a step-up substation based on YOLO11.
[0004] To achieve the above purpose, the present invention adopts the following technical solutions:
[0005] A method for fire and smoke detection in a step-up substation based on YOLO11 includes:
[0006] Pre-training the YOLO11 model using a power big data set to obtain an initial fire and smoke detection model;
[0007] Applying the initial fire and smoke detection model to the training of fire and smoke field data through knowledge transfer to generate a professional fire and smoke detection model; generating a lightweight fire and smoke detection model by knowledge distillation for the professional fire and smoke detection model;
[0008] Real-time obtaining visible light image data collected by the monitoring host of the step-up substation, and using the lightweight fire and smoke detection model for preliminary prediction of fire and smoke;
[0009] Based on the result of the preliminary prediction of fire and smoke, collecting real-time visible light image data and infrared image data of the equipment that may have fire and smoke;
[0010] Performing secondary fire and smoke detection on the visible light image data using the professional fire and smoke detection model, and performing heat source detection on the infrared image data, and determining the final fire and smoke detection result based on the secondary fire and smoke detection result and the heat source detection result.
[0011] As a preferred implementation, the power big data set includes visible light scene data from various power facilities; the categories of data tags of the visible light scene data at least include fire smoke.
[0012] As a preferred implementation, when the YOLO11 model is pre-trained, data augmentation uses a smaller random probability and amplitude.
[0013] As a preferred implementation, the method of generating a professional fire smoke detection model through knowledge transfer is as follows:
[0014] Take the initial fire smoke detection model as the transfer network, freeze the weight knowledge of its backbone network, adjust the model input size to be larger than the training image size of the initial fire smoke detection model, and replace the multi-class detection head with a single-class detection head for fire smoke. Use the prior knowledge learned in the training of the power big data set to perform fine-tuning training on the fire smoke data set to obtain a professional fire smoke detection model.
[0015] Furthermore, the method further includes that when training the professional fire smoke detection model, data augmentation uses a larger random probability and amplitude.
[0016] As a preferred implementation, the method of generating a lightweight fire smoke detection model through knowledge distillation includes:
[0017] Take the professional fire smoke detection model as the teacher model, use the probability distribution output of the teacher model as the soft label, use the intermediate layer features of the teacher model to guide the student model, and optimize and train the student model, that is, the lightweight fire smoke detection model, on the fire smoke data set based on decoupled knowledge distillation.
[0018] As a preferred implementation, the fire smoke data set includes the scene data related to fires in the power big data set and the fire smoke data of non-power scenes.
[0019] As a preferred implementation, performing weighted fusion judgment on the secondary fire smoke detection result and the heat source detection result as the final detection result includes:
[0020] For the overlapping target boxes in the heat source detection result and the secondary fire smoke detection result, use the weighted average method to calculate the position and size of the fused target box as the final detection result.
[0021] As a preferred implementation, the method further includes that for the secondary fire smoke detection result and the heat source detection result, if the adjacent heat source areas or prediction boxes in the detection result are less than a preset distance threshold, then merge the adjacent heat source areas or prediction boxes to form a larger heat source area or prediction box.
[0022] As a preferred embodiment, the method further includes generating a task log after determining the final detection result of the fire smoke, recording the information during the detection process, including device information and detection results, and uploading them to the terminal.
[0023] The present invention pre-trains the YOLO11 model by making full use of the power big dataset, and applies the pre-trained model to the fire smoke detection task through the knowledge transfer technology to construct a professional fire smoke detection model YOLO11x, improving the detection accuracy. Then, using the professional fire smoke detection model as the teacher model, the lightweight YOLO11s preliminary detection model is trained by the knowledge distillation technology, enabling it to efficiently and accurately perform the preliminary prediction of fire smoke. Finally, the YOLO11x model is used for secondary verification, and combined with the infrared images of the electrical equipment in the infrared booster station for fusion detection, excluding false detections and improving the overall prediction accuracy, thus improving the detection accuracy and reliability. Description of the Drawings
[0024] Figure 1 is the flowchart of the method of the present invention. Detailed Embodiments
[0025] The technical solutions of the present invention will be further elaborated below in conjunction with the description of the drawings and the detailed embodiments.
[0026] Embodiment 1
[0027] This embodiment provides a method for detecting fire smoke in a booster station, including the following steps:
[0028] 1. Pre-training of the YOLO11x model
[0029] The YOLO11 model with a more complex network structure is pre-supervised trained using the power big dataset to obtain a pre-trained model YOLO11x with more power prior knowledge and fire smoke detection capabilities.
[0030] The input resolution of the image data is set to 640*640, the data augmentation uses a small random probability and amplitude, and the training epoch is set to 500. Using a smaller scale for training helps to quickly extract global features and reduce the training duration, and using a lower data augmentation ensures that the focus of training is on learning the general features of the power scenario and can accelerate the convergence speed.
[0031] The specific operations of the data augmentation include hue jitter (variation range: 1.5%), saturation jitter (variation range: 70%), brightness jitter (variation range: 40%), random rotation (range: 0 - 15 degrees), random flipping (probability: 30%), image translation (variation range: 10%), and mosaic augmentation (probability: 50%, automatically turned off in the last 10 epochs).
[0032] This power big data set mainly comes from the visible light scene data of various domestic power facilities, including substations, step-up stations, distribution stations, etc. The data acquisition methods include but are not limited to video surveillance, drone shooting, and on-site manual collection. The data labels include 123 custom categories such as fire smoke, as well as other power equipment and their components, defects, etc., with approximately 430,000 images. The labels comply with the national grid's power equipment defect image annotation specification.
[0033] 2. Knowledge Transfer and Adaptation to the Fire Smoke Domain
[0034] After pre-training on the power big data set, the trained YOLO11x model is used as the transfer network. The weight knowledge of its backbone network is frozen, the input size of the model is adjusted to 896 * 896, and the original multi-class detection head is replaced with a single-class detection head for fire smoke. Using the prior knowledge learned during the power scene training, fine-tuning training is carried out on the fire smoke data set to obtain the professional fire smoke detection model YOLO11x.
[0035] The described fire smoke data set includes not only the fire-related scene data in the power big data set but also the fire data of other non-power scenes, with a total of more than 13,000 images.
[0036] During the fine-tuning training process, freezing the backbone network can retain the multi-scale features learned during the power scene training, thus avoiding over-adjustment in subsequent fire smoke detection tasks. Switching to a larger-scale input can improve the detection accuracy and detail capture ability.
[0037] At the same time, in the early stage of the fine-tuning training, a larger data augmentation probability and larger-scale augmentation operations will be adopted in the data preprocessing link to better improve the adaptability of the model in the fire smoke scene.
[0038] The specific operations of the data augmentation include hue jitter (variation range: 2%), saturation jitter (variation range: 80%), brightness jitter (variation range: 50%), random rotation (range: 0 - 30 degrees), random flipping (probability: 50%), image translation (variation range: 20%), and mosaic augmentation (probability: 100%, automatically turned off in the last 10 epochs).
[0039] This process ensures that the YOLO11x model can recognize the special features of fire smoke through domain adaptation, improving the accuracy in this field.
[0040] 3. Knowledge Distillation and Training of the Lightweight YOLO11s Model
[0041] Using the principle of decoupled knowledge distillation technology, the trained YOLO11x model is used as the teacher model, and the more lightweight YOLO11s with the same algorithm structure is used as the student model for optimized training on the fire smoke dataset.
[0042] During the decoupled knowledge distillation process, the input resolution is set to 896*896 to ensure that the student model can handle more details and learn high-quality fire smoke features.
[0043] The decoupled knowledge distillation training process will be divided into two independent training objectives. First is the soft label training. The soft label output of YOLO11x is used to help YOLO11s learn the probability distribution of YOLO11x in the fire smoke detection task, guiding the final output of YOLO11s. The soft label loss is as follows:
[0044]
[0045] Among them, and are the output probabilities of the teacher and student models respectively. By minimizing the cross-entropy loss, the student model learns the output probability distribution of the teacher model.
[0046] In addition to the soft label training, there is also the feature map training. In this process, the intermediate layer feature maps of YOLO11x can guide YOLO11s to learn deeper visual features and improve the intermediate layer representation of the YOLO11s model. The feature map loss is as follows:
[0047]
[0048] Among them, and are the feature maps of the teacher and student models respectively. By calculating the difference between the feature maps, the feature consistency of the teacher and student in the intermediate layer is ensured.
[0049] These two training processes are decoupled and do not interfere with each other, enabling the student model to more accurately learn different levels of information from the teacher model. The decoupled loss is as follows:
[0050]
[0051] Among them, λ is the weight hyperparameter that controls the balance between the soft label loss and the feature map loss.
[0052] 4. Fusion Detection
[0053] Infrared images can capture temperature differences under low-light conditions and are more suitable for detecting heat sources under low-light conditions. Therefore, they can identify heat sources or smoke in a timely manner during a fire. Visible light images, on the other hand, are better at identifying the visual characteristics of flames and smoke. In the present invention, the detection results of infrared images and visible light images are processed independently, and integrated judgment is carried out through a weighted average method. The specific process is as Figure 1 shown, including:
[0054] First, use the visible light image data collected in real time by the monitoring host of the booster station as input data (an optional implementation is to obtain the video monitoring data of the equipment at the inspection target location, and intercept the video monitoring data as input data). Use the trained lightweight model to perform preliminary fire and smoke detection to obtain the preliminary fire and smoke detection results. For the electrical equipment of the booster station where fire and smoke may exist, collect the current visible light image and infrared image of the electrical equipment through a visible light camera and an infrared camera respectively, and perform heat source positioning and extraction.
[0055] After preprocessing operations such as denoising, contrast enhancement, and image normalization on the collected infrared images, dynamically set pixel thresholds according to the statistical characteristics of the images to screen heat source pixels, and then use the connected component analysis method and morphological processing to screen out the heat source candidate regions (the heat source candidate regions are presented in the form of target boxes, and the size of the target box is the minimum bounding rectangle of the heat source pixels). Moreover, for adjacent heat source regions, if their distance is less than the set threshold, they are merged into a large heat source region (the minimum bounding rectangle of the adjacent heat source regions). At the same time, the confidence level of the heat source region being a fire can be calculated by taking the difference between the average value of the heat source pixels in the region and the set threshold as the numerator, and the difference between the maximum value that the pixels can reach and the threshold as the denominator, and the calculated value is used to set the confidence level of the heat source region being a fire. Finally, it is presented in the form of a target detection box.
[0056] For visible light images, use the professional fire and smoke detection model YOLO11x constructed in step 2 for secondary detection to more accurately predict the fire and smoke targets. In the prediction results, if the distance of the prediction boxes is less than the set threshold, they are merged into a large fire and smoke region.
[0057] For the overlapping target boxes between the infrared detection results and the visible light detection results, finally use the weighted average method to calculate the position and size of the fused box. The calculation formula is as follows:
[0058]
[0059]
[0060]
[0061]
[0062] Among them, and are the confidence levels of the visible light and infrared image detection result boxes respectively, 、 、 、 are the coordinate information of the visible light image detection result box, 、 、 、 are the coordinate information of the infrared image detection result box, 、 are the coordinate information of the fused target box.
[0063] Use the professional fire smoke detection model YOLO11x and the infrared image heat source detection results for fusion detection to further eliminate false alarms and improve the detection accuracy.
[0064] 5. Task Log and Exception Report Management
[0065] During the entire fire smoke detection process, the system will automatically record important data such as device information, detection time, prediction results, etc., and upload them to the terminal.
[0066] The system will count the abnormal events during the detection process, generate task logs, and summarize the abnormal data to ensure the integrity and accuracy of the task.
Claims
1. A fire smoke detection method for a booster station based on YOLO11, characterized in that: include: The YOLO11 model is pre-trained using the power big data set to obtain the initial fire smoke detection model; Applying the initial fire smoke detection model to fire smoke field data training through knowledge transfer to generate a professional fire smoke detection model; Generating a lightweight fire smoke detection model through knowledge distillation of the professional fire smoke detection model; Obtain visible light image data collected by the booster station monitoring host in real time, and use the lightweight fire smoke detection model to make a preliminary prediction of fire smoke; Based on the preliminary prediction results of fire smoke, real-time visible light image data and infrared image data are collected for equipment that may have fire smoke; The professional fire smoke detection model is used to perform secondary fire smoke detection on the visible light image data, and heat source detection is performed on the infrared image data, and a final fire smoke detection result is determined based on the secondary fire smoke detection result and the heat source detection result.
2. The method according to claim 1, characterized in that The electric power big data set includes visible light scene data from various types of electric power facilities; the categories of data labels of the visible light scene data at least include fire smoke.
3. The method according to claim 1, characterized in that When pre-training the YOLO11 model, data augmentation uses smaller random probabilities and amplitudes.
4. The method according to claim 1, characterized in that: The method of generating a professional fire smoke detection model through knowledge transfer is as follows: The initial fire smoke detection model is used as a migration network, and the weight knowledge of its backbone network is frozen. The model input size is adjusted to be larger than the training image size of the initial fire smoke detection model, and the multi-category detection head is replaced with a single-category detection head for fire smoke. The prior knowledge learned in the training of the large electric power dataset is retained to perform fine-tuning training on the fire smoke dataset to obtain a professional fire smoke detection model.
5. The method according to claim 4, characterized in that It also includes the use of larger random probabilities and amplitudes for data augmentation when training professional fire smoke detection models.
6. The method according to claim 1, characterized in that The method of generating a lightweight fire smoke detection model through knowledge distillation includes: A professional fire smoke detection model is used as the teacher model, the probability distribution output of the teacher model is used as the soft label, the intermediate layer features of the teacher model are used to guide the student model, and the student model is optimized and trained on the fire smoke dataset based on decoupled knowledge distillation, that is, the lightweight fire smoke detection model.
7. The method according to claim 4 or 7, characterized in that: The fire smoke data set includes scene data related to fire in the power data set, and fire smoke data of non-power scenes.
8. The method according to claim 1, characterized in that The secondary fire smoke detection result and the heat source detection result are weighted and fused to form a final detection result, including: For the overlapping target frames in the heat source detection results and the secondary fire smoke detection results, the weighted average method is used to calculate the position and size of the fused target frame as the final detection result.
9. The method according to claim 1 or 9, characterized in that: It also includes, for the secondary fire smoke detection results and heat source detection results, if adjacent heat source areas or prediction boxes in the detection results are less than a preset distance threshold, merging the adjacent heat source areas or prediction boxes to form a larger heat source area or prediction box.
10. The method according to claim 1, characterized in that It also includes, after determining the final fire smoke detection result, generating a task log to record information during the detection process, including equipment information and detection results, and uploading it to the terminal.