Subway fire detection method based on YOLOv8

Abstract

The invention discloses a metro fire detection method based on YOLOv8. The metro fire detection method is used for improving the accuracy and response speed of an alarm system. According to the invention, a YOLOv8 target detection algorithm is adopted, a network structure is optimized in combination with deep learning, and a weighted focusing loss function and multi-modal data fusion method is designed, so that the detection precision is improved. A data set covers images and sensor data of different subway environments, an attention mechanism and multi-scale detection are introduced, and the flame and smoke recognition capability is improved. And by combining temperature sensor data, false alarms are reduced. The optimization strategy aims at the carriage and platform environment, the robustness of the model under the conditions of low illumination, reflection and shielding is enhanced, and real-time detection is realized by means of TensorRT accelerated reasoning. The system can operate independently, does not need a network, can be linked with a subway FAS system, and improves the early warning efficiency. Experiments show that the detection performance of the method is better than that of a traditional method, the reasoning speed exceeds 50 frames per second, and the method has wide application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of subway fire detection systems, and in particular to a subway fire detection method based on YOLOv8. Background Art

[0002] The Subway Fire Alarm System (FAS) is a fire detection and alarm system used in subway cars and stations. Its purpose is to improve the ability to quickly identify and respond to fire incidents during subway operation and ensure the safety of life and property of passengers and staff. However, traditional subway fire alarm systems mainly rely on rule-based methods to detect fires through pre-set thresholds and rules. Such methods are effective in simple and static environments, but they are less adaptable to complex and dynamic operating environments such as subways. They are prone to false alarms or missed alarms due to environmental noise, changes in light or other interference factors, thus affecting the reliability of the system.

[0003] In addition, existing image-based fire detectors mainly rely on traditional image processing algorithms and simple threshold judgment models to identify fires by capturing the apparent characteristics of flames, smoke or temperature. However, this method has the following significant defects: (1) Lack of adaptability to dynamic environmental changes in complex scenarios. For example, smoke or light reflections generated by equipment operation in subway platforms may be similar to fire characteristics and easily trigger false alarms; (2) Insufficient use of multispectral images makes it difficult to effectively handle scenes where flames are obscured or smoke is light; (3) Lack of accurate recognition of characteristics of different fire types limits the system's detection performance in complex fire scenarios. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] The present invention discloses a subway fire identification method based on YOLOv8, aiming to solve the problems of high false alarm rate, slow response speed and difficulty in realizing efficient real-time detection in the existing subway fire alarm system. The current fire alarm system fails to fully adapt to the special environment of the subway, and lacks effective solutions in the face of diverse fire characteristics and complex subway scenes (such as low light, reflection, smoke occlusion, etc.). The present invention adopts the latest YOLOv8 target detection algorithm, combines the advantages of deep learning technology, optimizes the network model structure, designs data balancing strategies and integrates efficient reasoning methods to construct an efficient and accurate subway fire detection system. In particular, the present invention is targeted at special scenes such as subway cars and platforms, can accurately capture fire characteristics such as flames and smoke, and supports real-time fire image detection and positioning in a network-free environment, significantly reducing false alarms and missed alarms, and improving the efficiency and reliability of the subway safety monitoring system.

[0006] (II) Technical solution

[0007] To achieve the above objectives, the present invention is implemented by the following technical scheme: a subway fire detection method based on YOLOv8 comprises the following steps:

[0008] The process of constructing the dataset includes:

[0009] The collection of Dataset A is based on a timeline design, obtaining sample data from different subway lines, carriages, and weather conditions. At the same time, it collects video and sensor data during the normal operation of the subway and the occurrence of fire records, including temperature, smoke, and gas concentration data.

[0010] The Labelling tool is used to accurately label dataset A, marking the timestamp, location, scale, flame, smoke and other characteristic areas of the fire. At the same time, the normal operation data is labeled to generate a standardized dataset in XML or COCO format.

[0011] After screening and data enhancement, invalid samples are eliminated, and fire samples are expanded through rotation, scaling, etc. to balance the ratio of normal and fire data and improve the quality of the data set.

[0012] The model training process includes:

[0013] A deep convolutional neural network (CNN) architecture is used to extract important features in the image (such as flames, smoke, etc.) through multi-layer convolution and pooling operations.

[0014] The attention mechanism (such as SE module and CBAM module) is adopted to enable the model to better focus on the fire area, suppress the interference of background noise, and enhance the model's attention to fire characteristics.

[0015] By adopting multi-scale feature detection, it is possible to simultaneously detect fire targets of different sizes (such as small flames, large-scale smoke, etc.).

[0016] The loss function of YOLOv8 is optimized, the weight balance of positive and negative samples is increased, the false positives and negatives are reduced, and the detection accuracy and recall rate are improved.

[0017] The YOLOv8 pre-trained model is used, and transfer learning is performed on this basis, using the existing weight initialization to improve training efficiency and ensure model stability.

[0018] The model optimization process includes:

[0019] Combined with the fire characteristics of subway environment, a weighted focal loss function is proposed to balance the weights of positive and negative samples and improve the model's detection ability for minority class fire characteristics. The specific expression of the weighted focal loss function is as follows:

[0020]

[0021] In the formula, is the model’s predicted probability for a certain category, is the weight coefficient of positive and negative samples, is the focus parameter, which is usually set to 2 to strengthen the focus on difficult samples. To adapt to the minority samples (such as small flames and smoke) in subway fire detection, we and Further weighted adjustments were made to enable the model to pay more attention to difficult-to-detect fire targets, especially small targets such as flames and smoke, thereby improving the detection accuracy of minority classes.

[0022] In addition to the weighted focus loss function, the present invention also specially designs a weighting strategy for minority class samples (such as small flames and smoke) to further improve the model's sensitivity to these features. This optimization process plays a crucial role in model training and improves the detection accuracy of fire targets in complex environments.

[0023] During the training process, the Adam optimizer is used, combined with the learning rate decay strategy, to automatically adjust the learning rate to further improve the training efficiency and convergence speed.

[0024] During the training process, the present invention adopts an adaptive learning rate adjustment strategy and accelerates the convergence of the model through transfer learning. An optimized learning rate decay strategy is introduced to dynamically adjust the training at different stages to avoid overfitting and accelerate the training process.

[0025] Mosaic data enhancement technology is used to increase the diversity of data samples through random cropping, splicing, scaling, etc., and regularization methods such as Dropout are used to prevent overfitting and ensure the generalization ability of the model.

[0026] The fusion detection method of temperature sensor data and image data is introduced to integrate multimodal data. When identifying flames or smoke, the false alarm rate is reduced by the constraint of temperature threshold. The temperature data in the subway car is collected in real time through the temperature sensor, and the temperature threshold T is set. When the temperature value detected by the sensor exceeds the threshold, the system determines that the ambient temperature is abnormal and triggers the fire detection process. If the temperature value does not exceed the threshold, the system maintains normal monitoring.

[0027] The YOLOv8 model detects flames or smoke in images and generates a corresponding fire confidence value C. Combined with the temperature sensor data, when the temperature value exceeds the threshold T, the system positively weights C to improve the reliability of fire detection; if the temperature value does not exceed the threshold, it applies negative weighting to C to reduce the risk of false alarms. Finally, the system compares the weighted confidence value with the preset flame recognition threshold C_th. When the confidence value is greater than or equal to C_th, a fire alarm is triggered; otherwise, it is determined to be a non-fire state and monitoring continues.

[0028] The Anchor Box mechanism is used to accurately locate fire targets. Through multi-scale detection, fire targets of different sizes can be identified, including small flames and large-scale smoke, ensuring that early warning of fire is not affected by the size of the target.

[0029] It supports independent operation without a network environment, and all computing processes are completed locally, which is suitable for the subway intranet environment.

[0030] Build a real-time alarm system. When a fire is detected, the system will immediately trigger an alarm signal and mark the fire area with a rectangular box.

[0031] A dedicated optimization strategy for subway car and platform fire detection is specially designed to increase the robustness of the model to flame reflection and smoke diffusion in the car. At the same time, a real-time alarm system is integrated to link the detection results with the subway FAS system to improve the warning efficiency.

[0032] (III) Beneficial effects

[0033] Through the above technical solution, the following technical effects can be achieved.

[0034] This paper introduces the YOLOv8 target detection algorithm and deep learning technology, and designs a data enhancement method and a weighted focus loss function in combination with the characteristics of the subway environment, thereby solving the problem of imbalanced fire data sets. It performs particularly well in the detection of small targets such as flames and smoke, with the false alarm rate reduced to less than 3% and the missed alarm rate reduced to 1%, significantly improving the accuracy and robustness of fire detection.

[0035] The present invention introduces a multimodal fusion strategy of temperature sensor data and image data in fire detection. By utilizing temperature threshold constraints and dual-modal joint reasoning methods, the detection capability of fire characteristics in complex scenes is enhanced. This strategy not only effectively reduces false alarms caused by complex backgrounds in subways, but also further improves the fire detection system's ability to identify early fires.

[0036] In view of the low light, occlusion and reflection characteristics of subway platforms and carriages, this paper designs a specific optimization strategy to enable the model to accurately identify flame and smoke characteristics under complex conditions. Experiments show that after adopting this optimization strategy, the detection performance (mAP) of the model in subway scenes is improved by more than 15% compared with traditional methods.

[0037] The lightweight reasoning strategy proposed in this invention, combined with the TensorRT acceleration framework, enables the YOLOv8 model to have efficient real-time detection capabilities, with a reasoning speed of more than 50 frames per second. At the same time, the invention can be deployed and operated in a network-free environment, providing a rapid response capability for the fire warning of the subway FAS system in a closed network environment, thereby improving the applicability of the system.

[0038] The present invention effectively optimizes the imbalance problem of fire categories by dynamically adjusting the data sampling strategy and loss function design. At the same time, the system can timely discover performance deficiencies and adjust model parameters by monitoring performance indicators such as false alarm rate and response time, ensuring that the system maintains high reliability and high accuracy during operation.

[0039] Other features and advantages of the present disclosure will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the flow of fire detection by the YOLOv8 model structure of the method of the present invention; Figure 2 is a schematic diagram of a method for fusing temperature and image data according to an example of the present invention; Figure 3 It is a fire detection system architecture diagram of an example of the present invention. DETAILED DESCRIPTION

[0041] The specific embodiments of the present disclosure are described in detail below in conjunction with the accompanying drawings. It should be understood that although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein.

[0042] The subway fire detection method based on YOLOv8 of the present invention realizes accurate detection of fire events by fusing and analyzing the image data and temperature sensor data in the subway environment. The technical scheme and implementation process of the present invention are described in detail below, covering the construction of data sets, model training and optimization, algorithm implementation process, fusion method of temperature sensor data and image data, reasoning acceleration strategy and its practical application.

[0043] The construction of the data set is one of the core parts of the present invention, which determines the effect and accuracy of model training. The data set of the present invention includes image data of multiple scenes such as subway cars and platforms, and combines sensor data such as temperature, smoke, and gas concentration to achieve multimodal data fusion and enhance the accuracy and robustness of fire detection.

[0044] During the data collection process, high-quality camera equipment and sensors were first used to collect videos on different subway lines, carriages, platforms, and in various weather conditions. All collected images include not only images of normal subway operation, but also images of fires, to ensure that the data set covers a variety of complex scenes.

[0045] The image data is high-resolution images collected from different locations such as subway cars, platforms, tunnels, etc., covering different time periods (such as day and night), weather conditions (such as sunny days, rainy days, and haze weather), and different perspectives and distances to ensure data diversity.

[0046] Sensor data uses temperature sensors, smoke sensors, gas concentration sensors and other devices to record real-time environmental information in carriages, platforms and other areas. Each image is equipped with corresponding sensor data, including temperature changes, smoke concentration, gas concentration, etc.

[0047] Data labeling is done by labeling each frame of the image through the Labelling tool to ensure the high quality of the data set. The annotation content includes the timestamp of the fire, the location of the fire area, the location of the bounding box of the flame and smoke, etc.

[0048] The specific annotation process includes: Accurately mark the fire feature areas such as flames and smoke in the image. Each fire event will be marked as a bounding box, and its timestamp and fire size will be recorded; Annotate the normal operating state images without fire to ensure that the training data contains enough positive and negative samples; After labeling, all labeled data will be stored in XML or COCO format to ensure the standardization and compatibility of the dataset for subsequent processing.

[0049] In order to enhance the diversity of data and avoid overfitting, the present invention performs data enhancement processing on the collected image data. Specifically, the dataset is expanded by rotating, scaling, mirroring and other methods, especially for the enhancement of fire samples. These technologies balance the ratio of normal and fire images and improve the training effect.

[0050] In one embodiment, the dataset A is divided into a training set, a validation set, and a test set in proportion, wherein the training set accounts for 70%, the validation set accounts for 15%, and the test set accounts for 15%. Ensure that the training set contains sufficiently diverse samples, the validation set is used for parameter adjustment, and the test set is used to evaluate the generalization ability of the model.

[0051] The present invention adopts the YOLOv8 target detection algorithm, which is based on the deep convolutional neural network (CNN) architecture and optimized for subway fire detection. During the training process, the existing pre-trained model is first used for transfer learning to avoid training from scratch, saving computing resources and time.

[0052] In one implementation, the backbone network (Backbone) of YOLOv8 uses a deep convolutional layer to extract important features in the image, such as flames, smoke, etc., through multi-layer convolution and pooling operations. The process includes forward propagation of the convolutional layer and feature extraction. For the recognition of these features, the present invention uses an attention mechanism (such as the SE module and the CBAM module) so that the model can focus on the fire area more effectively and ignore background noise.

[0053] After image feature extraction, the model fuses multi-scale features through the neck network (Neck), further enhancing the model's adaptability to the size of fire targets. Then, the model performs classification prediction, bounding box regression, and target confidence prediction through the head network (Head). During the training process, in order to improve the detection capability of small targets (such as tiny flame areas), the present invention adopts a weighted focal loss function (Weighted Focal Loss). This loss function gives higher weights to small targets that are difficult to detect during the training process to reduce the model's bias towards large targets, thereby improving the model's detection accuracy for small flame targets. The specific expression of the weighted focal loss function is as follows:

[0054]

[0055] In the formula, is the model’s predicted probability for a certain category, is the weight coefficient of positive and negative samples, is the focus parameter, which is usually set to 2 to strengthen the focus on difficult-to-distinguish samples. To adapt to minority class samples (such as small flames and smoke) in subway fire detection, we further weighted α and γ so that the model can pay more attention to difficult-to-detect fire targets, especially small targets such as flames and smoke, thereby improving the detection accuracy of minority classes.

[0056] In order to improve the model's ability to detect multi-scale small targets, the present invention further adopts a multi-scale feature extraction strategy, which can ensure accurate identification of small flame areas while effectively reducing the interference of background noise.

[0057] The present invention introduces a specific small target detection module. Based on the YOLOv8 model, a denser feature pyramid network (FPN) or path aggregation network (PAN) is used to enhance the fusion capability of features of different scales, thereby effectively improving the detection effect of small flame targets. At the same time, increasing the resolution of the shallow feature map can more clearly express the small flame area in the low-resolution image, ensuring the accurate recognition of small targets.

[0058] In the target classification process, the present invention further introduces an IoU (Intersection over Union) loss optimization strategy based on small target areas to ensure that the model can accurately locate small flame targets. In order to enhance the model's ability to learn small target features, the present invention also uses a variety of data enhancement techniques, including cropping, scaling, rotation, and random noise superposition. These techniques enhance the model's adaptability to small targets and further improve the model's ability to recognize small flame targets. In addition, through oversampling technology, the present invention adds small flame target samples to the training set and adjusts the ratio of small targets to other target categories to balance the ratio of small targets to other target categories, further improving the detection accuracy of the model.

[0059] In the post-processing stage, the present invention introduces a small target priority non-maximum suppression (NMS) algorithm, and improves the confidence of the small flame area by weighting to ensure that small targets can be stably identified in complex backgrounds. In order to further improve the detection accuracy, especially in reducing the underreporting and false alarms of small flame targets, the present invention also introduces a multi-target verification strategy. Through these optimizations, the performance of the YOLOv8 model in subway fire detection is significantly improved, especially in the identification of small flame targets, further improving the detection accuracy and response speed.

[0060] During the training process, the present invention adopts transfer learning technology and uses existing pre-trained models for initial training, which significantly reduces the computing resources and time required for training from scratch. In order to further improve the performance and stability of the model, the present invention combines the Adam optimizer and the learning rate decay strategy to automatically adjust the learning rate to improve the training efficiency and the generalization ability of the model. At the same time, by carefully designing hyperparameters such as the initial learning rate, batch size, and training rounds, the present invention ensures rapid convergence and high-precision detection of the model.

[0061] Among them, the dynamic adjustment of the learning rate is an important part of the optimization process. The initial learning rate of the present invention is set to 0.001, and the cosine annealing strategy is used to gradually decay the learning rate during the training process. When the detection task gradually converges, the stability of the model and the final detection accuracy are significantly improved by reducing the decay amplitude of the learning rate.

[0062] For the batch size setting, the present invention selects 16 as the default value to balance the video memory usage and training time. If the training resources allow, the batch size can be expanded to 32, thereby accelerating the training process while ensuring the stability of the gradient update and improving the overall performance of the model.

[0063] In terms of the selection of training rounds, the present invention sets the training rounds to 50 by default and introduces an early stopping mechanism. By real-time monitoring of the changes in the validation set loss function, when the validation set loss no longer decreases significantly, the system automatically stops training to avoid overfitting and ensure the optimal balance between training efficiency and model performance.

[0064] Through the above optimization strategies, the present invention realizes efficient training under limited computing resources, so that the model can achieve the best combination of accuracy and speed in subway fire detection tasks, and meet the requirements of practical applications for real-time and reliability.

[0065] In order to optimize the training effect of the model and enhance its robustness and generalization ability, the present invention introduces a variety of data enhancement techniques and sample balancing strategies to further improve the performance of the model in complex scenarios.

[0066] This paper adopts Mosaic data enhancement technology, which simulates complex fire scenes in subway environments by randomly cropping and splicing multiple images, effectively increasing the diversity of samples. At the same time, combined with the CutMix data enhancement method, the features of different fire samples are fused together, so that the model can learn more diverse fire scene features, thereby improving the model's ability to recognize fire samples.

[0067] In addition, in order to further optimize the learning ability of the model, the present invention introduces a positive and negative sample weight adjustment strategy. By dynamically adjusting the ratio of fire and non-fire samples, the model can more accurately distinguish between fire and non-fire scenes during the training process. Especially when dealing with areas prone to false alarms, this strategy significantly improves the detection accuracy of the model.

[0068] The combination of the above optimization strategies not only effectively increases the model's adaptability to complex scenarios, but also significantly reduces the model performance bottleneck caused by uneven sample distribution.

[0069] According to the implementation of the present invention, in order to further improve the accuracy of fire detection, a method of fusing temperature sensor data with image data is proposed. Sensor data (such as temperature, gas concentration, smoke, etc.) can provide additional fire information for image data, thereby enhancing the system's ability to identify fire characteristics in complex environments. On this basis, a rule-based determination method is adopted to combine temperature threshold determination with flame detection confidence weighting to improve the accuracy and robustness of the system and reduce the probability of false alarms and missed alarms. Please refer to Figure 2 , Figure 2 It is a schematic diagram of a method for fusing temperature and image data according to an example of the present invention.

[0070] Specifically, the system collects temperature data in the subway car in real time through the temperature sensor and sets a temperature threshold T. When the temperature value detected by the sensor exceeds the threshold T, the system determines that the ambient temperature is abnormal and further triggers the fire detection process; if the temperature value does not exceed the threshold T, it maintains the normal state for continuous monitoring. In this way, the system can strengthen fire detection in a timely manner when the ambient temperature is abnormal.

[0071] At the same time, the YOLOv8 model detects flames or smoke in the image data and generates the corresponding flame detection confidence value C. At this time, the confidence value C is weighted and adjusted in combination with the results of the temperature sensor. When the temperature value exceeds the threshold T, the confidence value C is positively weighted to improve the credibility of fire detection; if the temperature value does not exceed the threshold T, the confidence value C is negatively weighted to reduce the risk of false alarms.

[0072] The system further compares the adjusted confidence value with the preset flame recognition threshold C_th. When the confidence value is greater than or equal to the threshold C_th, the system outputs a fire alarm; if the confidence value is less than C_th, the system determines it as a non-fire state and continues regular monitoring.

[0073] In one embodiment, a temperature sensor and a camera are arranged in a subway car for fire detection. The temperature sensor monitors the temperature change in the car in real time, and the camera is responsible for capturing the image data in the car. In order to achieve accurate fire detection, the present invention sets the temperature threshold T to 60°C and the flame recognition threshold C_th to 0.8.

[0074] The temperature sensor collects temperature data in the car once a second. In one collection, the temperature sensor detected that the temperature in the car was 65°C, which exceeded the set threshold T. According to the system settings, when the temperature value exceeds the threshold, the system will trigger further fire detection processing.

[0075] The camera captures the image inside the car and feeds it into the YOLOv8 model for fire detection. YOLOv8 successfully detects the presence of flames in the image, generates a bounding box for the flame area, and outputs a confidence value C of 0.85 for the area, indicating that the model has a high degree of confidence in the flame detection result.

[0076] Based on the temperature data and the YOLOv8 detection results, the system makes a weighted adjustment to the confidence value C. Since the temperature value exceeds the threshold T (65°C), the system applies a positive weighting. Assuming the weighting coefficient is 1.2, the adjusted confidence value is 0.85 * 1.2 = 1.02. In this way, the system enhances the credibility of fire detection.

[0077] The adjusted confidence value 1.02 is greater than the set flame recognition threshold C\_th (0.8), so the system outputs a fire alarm to remind people and staff in the car to pay attention to the fire.

[0078] If the temperature value detected by the temperature sensor does not exceed the threshold T (for example, the temperature is 50°C), the system will negatively weight the detection result of YOLOv8. Assuming the weighting coefficient is 0.8, the adjusted confidence value is 0.85 * 0.8 = 0.68. In this case, since the adjusted confidence value is lower than the flame recognition threshold C\_th (0.8), the system will judge it as a non-fire state and continue with regular monitoring.

[0079] In one example, the temperature threshold T can be dynamically adjusted according to environmental conditions. In certain specific subway environments (such as high car temperatures due to hot weather or air conditioning system failure), the temperature threshold T can be appropriately increased to prevent the system from triggering fire detection too early. In this case, the system will automatically adjust the threshold T based on real-time temperature monitoring data to adapt to the actual environment, thereby ensuring the efficiency and accuracy of the system. For example, in the hot summer, the temperature in the car often exceeds 60°C. Therefore, the temperature threshold T can be dynamically set to 65°C to avoid false alarms caused by temperature fluctuations. At this time, the system obtains data in real time through the temperature sensor and makes dynamic threshold adjustments to ensure that fire detection is always within a reasonable range.

[0080] In another example, in addition to the temperature sensor, a gas concentration sensor and a smoke sensor are added, and their detection data are combined to make a comprehensive judgment on the fire. Specifically, the system weights and fuses the sensor data, assigning different weights to each sensor data (such as temperature, gas concentration, and smoke concentration). For example, when the gas concentration and smoke concentration are abnormal at the same time, the probability of fire is higher, so the system gives a larger weight to these data, thereby improving the accuracy of fire judgment.

[0081] In order to achieve real-time fire detection in practical applications, the present invention uses the TensorRT framework to quantize and accelerate the YOLOv8 model. Through this optimization, the inference speed of the model is greatly improved, and real-time detection of more than 50 frames per second can be achieved on embedded devices such as NVIDIA Jetson.

[0082] In one embodiment, the model is quantized to FP16 and deployed on an NVIDIA Jetson embedded device to achieve an inference speed of 60 frames per second. This optimization enables the system to identify fire characteristics in real time and trigger an immediate warning during peak hours in the morning and evening when there is high passenger density.

[0083] The present invention is specially designed to support independent operation without network environment. All calculation processes are completed locally, ensuring that there is no need to rely on external network resources in the subway intranet environment. Through the optimization of embedded devices, the system can achieve efficient and stable fire detection in the subway environment.

[0084] In another implementation, the present invention also combines a lightweight design strategy to optimize the model's memory usage and computing requirements, ensuring that it can run efficiently on embedded devices of the subway FAS system and adapt to independent operation in a network-free environment.

[0085] In order to adapt to the complex environment of subway cars and platforms, the present invention has specially designed a dedicated optimization strategy to improve the accuracy and robustness of fire detection. First, the detection ability of the model in low light and complex backgrounds is optimized for common phenomena such as reflected light, illumination changes, and smoke diffusion in subway cars. By introducing enhanced convolutional neural networks and multi-scale feature fusion technology, the model can effectively filter out the interference of reflected light on the detection results and accurately identify the boundaries of diffused smoke, thereby ensuring high-precision positioning of the fire area.

[0086] Secondly, in view of the special enclosed space and dynamic changes of smoke diffusion in subway cars, the system has specially strengthened its ability to adapt to these environmental characteristics during the training stage, making the model's recognition of fire events more reliable, especially in complex environmental lighting and large-area smoke diffusion conditions.

[0087] In addition, in order to further improve the efficiency of fire warning, the present invention deeply links the fire detection results with the subway FAS (fire alarm system). By transmitting the detected fire area information to the subway FAS system in real time, the system can automatically start the early warning mechanism when a fire is detected, and quickly transmit key information such as the fire location and scale to the monitoring center and the car display screen. In this way, not only the speed of fire response is improved, but also passengers can receive the alarm and evacuate in the shortest time, further ensuring the safety of the subway system and the safety of passengers' lives and property. Please refer to Figure 3 , Figure 3 It is an architecture diagram of a fire detection system of an example of the present invention, showing the connection between the temperature sensor, camera, YOLOv8 model and FAS system.

[0088] In one embodiment, in a subway car, when the YOLOv8 model identifies a fire, the system will immediately send an alarm signal through the FAS system and display the fire location on the display screen in the car to help passengers evacuate in time. In the platform scenario, the system can locate the fire area and provide alarm information through the station broadcasting system to reduce evacuation time.

[0089] In another example, the present invention introduces a multi-target verification strategy. When multiple fire targets are detected, the system further verifies through overlap analysis, the correlation of adjacent areas, and the temperature change trend to reduce missed reports and false alarms. Especially in subway cars and platform environments, fires may occur in multiple places at the same time. Through the multi-target verification strategy, the system can accurately locate multiple fire occurrence points and provide immediate warnings and evacuation instructions through the station broadcasting system or car display screens.

[0090] The preferred embodiments of the present disclosure are described in detail above in conjunction with the accompanying drawings; however, the present disclosure is not limited to the specific details in the above embodiments. Within the technical concept of the present disclosure, a variety of simple modifications can be made to the technical solution of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

Claims

1. A subway fire detection method based on YOLOv8, characterized in that: The following steps are involved: Construction of the dataset: Dataset A is collected based on the timeline design, and sample data is obtained from different subway lines, carriages, and weather conditions. Video and sensor data during the normal operation of the subway and fire records are also collected, including temperature, smoke, and gas concentration data; Use the Labelling tool to accurately label dataset A, record the timestamp, location, scale, flame, smoke and other characteristic areas of the fire, and label the normal operation data to generate a standardized dataset in XML or COCO format; Screen and enhance the data set, remove invalid samples, expand fire samples by rotation and scaling, balance the ratio of normal and fire data, and improve the quality of the data set; Model training: Build a fire detection model based on a deep convolutional neural network (CNN) architecture to extract important features such as flames and smoke through multi-layer convolution and pooling operations; Introducing attention mechanism modules (SE module, CBAM module) to focus on the fire area, suppress background noise, and enhance the model's ability to detect fire characteristics; Realize multi-scale feature detection and simultaneously detect fire targets of different sizes (including small flames and large-scale smoke); Optimize the loss function of the YOLOv8 model, increase the weight balance of positive and negative samples, reduce false positives and false negatives, and improve detection accuracy and recall rate; Use the YOLOv8 pre-trained model for transfer learning and use existing weight initialization to improve training efficiency and ensure model stability; Model optimization: A weighted focus loss function is proposed to balance the weights of positive and negative samples and improve the model's ability to detect minority fire features. Fusion of temperature sensor data and image data, through temperature threshold constraints, to reduce the false alarm rate of fire detection; Use the Adam optimizer and learning rate decay strategy to automatically adjust the learning rate to improve training efficiency and model convergence speed; Mosaic data enhancement technology is used to expand the diversity of data samples through random cropping, splicing, and scaling, and Dropout regularization is used to prevent overfitting. Use the Anchor Box mechanism to accurately locate the fire target and ensure the detection capability of fire targets of different sizes through multi-scale detection; Use the TensorRT framework to quantize and accelerate the YOLOv8 model to achieve efficient reasoning on embedded devices; Model deployment and application: Deploy the optimized YOLOv8 model on NVIDIA Jetson devices, achieving an inference speed of over 50FPS; Combine lightweight design strategies to optimize memory usage and computing requirements to ensure that the model runs efficiently in embedded devices; Supports independent operation without network environment, and the calculation process is completed locally, which is suitable for subway intranet environment; Build a real-time alarm system. When a fire is detected, the system triggers an alarm signal and marks the fire area with a rectangular box. Design a dedicated optimization strategy for subway car and platform fire detection to improve the model's robustness to flame reflection and smoke diffusion in the car; Link the detection results with the subway FAS system to improve early warning efficiency.

2. The subway fire detection method based on deep learning according to claim 1 is characterized in that: The dataset A consists of an image dataset A1, a sensor dataset A2 and historical fire data A3, and is specifically constructed as follows: By installing high-definition cameras and infrared sensors in subway cars, we collect image data of flames, smoke, and environmental changes inside the cars, and also cover image information of different time periods, passenger flow densities, and lighting conditions, forming image data set A1. Using temperature, smoke and gas sensors, we collect information about temperature changes, smoke concentration and flammable gas leakage in the cabin environment, which is classified as sensor data set A2. Based on the fire history records disclosed by the subway operation system and relevant departments, the fire causes, spread process, alarm response and emergency handling data are extracted and structured to form historical fire data A3.

3. The subway fire detection method based on deep learning according to claim 2 is characterized in that: The dataset A is divided into training set B, validation set C and test set D with a ratio of 70%, 15% and 15% respectively for training, validating and testing the deep learning model; The key features of the fire are extracted from the training set B, including the intensity, color and dynamic change characteristics of the flame and smoke, the deformation characteristics of the objects inside the car due to heat, the temperature rise rate, the trend of smoke concentration change, the characteristics of combustible gas leakage, and the time series characteristics of the fire spread.

4. The subway fire detection method based on YOLOv8 according to claim 1, characterized in that: In the steps, the fusion of sensor data and YOLOv8 detection results adopts a rule-based judgment method, and the fusion rules include temperature threshold judgment and flame detection confidence weighting, which are as follows: The temperature data in the subway car is collected in real time through the temperature sensor, and the temperature threshold T is set. When the temperature value detected by the sensor exceeds the temperature threshold T, the system determines that the ambient temperature is abnormal and triggers further fire detection processing; If the temperature value does not exceed the threshold value T, the system maintains normal status monitoring; Based on the YOLOv8 model, flame detection is performed on the image data to obtain the confidence value C of the flame. By combining the detection results of the temperature sensor, the confidence value C is weighted and adjusted. When the temperature value exceeds the threshold T, positive weighting is applied to the confidence value C to improve the credibility of flame detection. If the temperature value does not exceed the threshold T, a negative weight is applied to the confidence value C to reduce the risk of false fire alarms; The adjusted confidence value is compared with the flame recognition threshold C\_th set by the system. When the confidence value is greater than or equal to the flame recognition threshold C\_th, the system outputs a fire alarm; if the confidence value is less than the flame recognition threshold C\_th, the system determines that the current state is a non-fire state.

5. The subway fire detection method based on YOLOv8 according to claim 1, characterized in that: The YOLOv8 model combines the small target detection optimization strategy when detecting fires, and improves the recognition ability of small flame targets by adjusting the network structure or loss function, as follows: Based on the YOLOv8 model, a specific small target detection module is introduced, including the use of a denser feature pyramid network (FPN) or path aggregation network (PAN) to enhance the fusion capability of features of different scales; by increasing the resolution of shallow feature maps, the model's detection effect on small flame targets is improved; the model adopts a multi-scale feature extraction strategy to ensure accurate recognition of small targets (including tiny flame areas) while reducing interference with background noise; Define a weighted focal loss function to give higher weights to small target detection errors and reduce the model's bias towards large targets. In the target classification process, the IoU (Intersection over Union) loss optimization strategy based on small target areas is introduced to ensure that the model can accurately locate and identify small flame targets; During the training process, targeted data enhancement is performed on small target areas, including cropping, scaling, rotation, random noise superposition and other methods, to enhance the model's ability to learn small target features; small flame instances are added to flame samples, and the ratio of small targets to other target categories is balanced through oversampling technology; The small target priority non-maximum suppression (NMS) algorithm is adopted to ensure the stable identification of small target flames in complex backgrounds by weighting the confidence of small flame areas. At the same time, a multi-target verification strategy is introduced into the detection results in the post-processing stage to reduce the underreporting and false alarms of small target flames.

Images