A fire alarm system based on edge computing

Abstract

The present invention belongs to the field of fire alarm technology, and provides a fire alarm system based on edge computing, including an image acquisition module, an Internet of Things cloud and an information monitoring center. The image acquisition module is used to acquire fire images; the Internet of Things cloud acquires fire images from the image acquisition module based on the MQTT protocol; the image acquisition module sets a topic on the Internet of Things cloud, and uploads the acquired images to the Internet of Things cloud, and the Internet of Things cloud sends the images to all subscribers; the information monitoring center subscribes to the topic established by the image acquisition module, and receives data distributed by the Internet of Things cloud and associated with the topic. The present invention uses the MQTT protocol for data transmission and device interaction, which can be easily integrated with other sensors, enhance the functionality and stability of the system, and solve the problem of insufficient flexibility and scalability of traditional limited systems.

Description

Technical Field

[0001] The present invention relates to the field of fire alarm technology, and in particular to a fire alarm system based on edge computing. Background Art

[0002] In modern society, cameras play an increasingly important role in security warning and disaster alarm. The traditional early warning alarm mechanism is to monitor the camera images manually, which is often inefficient, has high labor costs, and usually cannot be monitored around the clock.

[0003] In recent years, with the development of artificial intelligence, more and more manual work has been transferred to artificial intelligence for automated control. Security and disaster warning alarms are no exception, that is, the image data of the camera is transmitted to the background through the network, and the background uses data model reasoning to determine whether the current picture needs to issue an alarm. By replacing manual identification with artificial intelligence, all-weather warning alarms are achieved. However, due to the large amount of image data transmission, a stable network environment and high bandwidth are required, which has certain requirements for the hardware environment. Secondly, when multiple images are connected to the background, the background needs to infer the results at the same time. On the one hand, the hardware requirements for the background are harsh and the cost is high. On the other hand, the delay caused by the reasoning process is large, which is not conducive to emergency response to disasters.

[0004] Traditional fire alarm systems are mainly based on centralized architecture and rarely consider the implementation of remote data transmission or complex distributed alarms. The commonly used RS485 and industrial bus protocols are suitable for short-distance communication and are difficult to meet the needs of large-scale Internet of Things. In addition, early fire alarm systems were not deeply integrated with Internet of Things technology. They were more of an independent security monitoring subsystem and did not consider intelligent interaction with remote servers or multiple devices. Summary of the invention

[0005] In view of the defects in the prior art, the present invention provides a fire alarm system based on edge computing to solve the problem that image data currently needs to be transmitted to the back end through a medium for inference, and fire alarm cannot be performed in some environments without a network.

[0006] The present invention provides a fire alarm system based on edge computing, comprising:

[0007] An image acquisition module, used for acquiring fire images;

[0008] The IoT cloud obtains the fire image from the image acquisition module based on the MQTT protocol; the image acquisition module sets a topic on the IoT cloud and uploads the acquired image to the IoT cloud, and the IoT cloud sends the image to all subscribers;

[0009] The information monitoring center subscribes to the topic established by the image acquisition module and receives data associated with the topic distributed by the Internet of Things cloud.

[0010] It can be seen from the above technical solution that the fire alarm system based on edge computing provided by the present invention performs data transmission and device interaction through the MQTT protocol, which can be easily integrated with other sensors, enhance the functionality and stability of the system, and solve the problem of insufficient flexibility and scalability of traditional limited systems.

[0011] Optionally, before transmitting the fire image to the IoT cloud based on the MQTT protocol, the image acquisition module further pre-processes the fire image, including:

[0012] Classifying the fire images according to resolution, the classification categories including low-resolution images and medium-high-resolution images;

[0013] Preprocessing the low-resolution image to enhance image quality, thereby obtaining a first target image;

[0014] Performing a cropping operation on the medium and high resolution image to obtain a target area and obtain a second target image;

[0015] The first target image and the second target image are fused respectively, and image fusion data is output.

[0016] It can be seen from the above technical solution that the image acquisition module performs edge computing, reduces the data transmission rate, and alleviates the background computing burden, which is suitable for scenarios with limited network conditions.

[0017] Optionally, the image acquisition module performs preprocessing on the low-resolution image, including:

[0018] Processing the low-resolution image based on the FSRCNN model to improve the image resolution;

[0019] Perform edge enhancement based on the Sobel operator to highlight the information contour;

[0020] Simulate images in different scenes to obtain diverse image samples.

[0021] Optionally, the image acquisition module is further used to perform model reasoning based on the diverse image samples, including:

[0022] Extract features of the preprocessed low-resolution image at multiple scales based on the YOLOv5n model;

[0023] Predicting a position target in the image based on the first detection frame;

[0024] Based on NMS suppression, the first detection box whose confidence is higher than a preset threshold is retained, and the coordinates, category label and confidence associated with the first detection box are output.

[0025] Optionally, the image acquisition module performs a cropping operation on the high-resolution image, including:

[0026] Performing target detection on the high-resolution image and marking the image based on a second detection frame;

[0027] Extracting the fire feature area in the second detection frame and performing high-resolution reasoning;

[0028] The fire feature areas derived through high-resolution reasoning are refined and classified to determine the fire type and intensity.

[0029] Optionally, the image acquisition module fuses the first target image and the second target image, including:

[0030] Deduplication of the first detection frame and the second detection frame that are overlapped;

[0031] When the first detection frame and the second detection frame detect different targets, output a weighted fusion classification result;

[0032] When the first detection box and the second detection box detect the same target but have different associated classification results, the classification result with the highest confidence is output.

[0033] Optionally, the image acquisition module allocates QoS levels according to data types, specifically:

[0034] QoS 0: Periodic update of environmental variables and device health status;

[0035] QoS 1: alarm data issued by the image acquisition module after detecting a fire, equipment failure status, and video clips or key frames uploaded by the image acquisition module when detecting a fire;

[0036] QoS 2: Fire confirmation command.

[0037] Optionally, the image acquisition module is further used to dynamically adjust the QoS level according to the network status and device load, including:

[0038] When the network load is high or the bandwidth is insufficient, the data type of QoS 1 is downgraded to QoS 0;

[0039] When a critical event occurs, the QoS 1 level of the alarm data is upgraded to QoS 2.

[0040] Optionally, the information monitoring center is also used to realize big data visualization, including:

[0041] Determine image risk assessment levels for multiple image acquisition modules;

[0042] Obtain risk scores based on image confidence and image risk assessment levels;

[0043] Risk score = w1×image confidence + w2×image risk assessment level; where w1 and w2 are predetermined weights.

[0044] By adopting the above technical solution, the present application has the following beneficial effects:

[0045] The fire alarm system based on edge computing provided by the present invention performs data transmission and device interaction through the MQTT protocol, which can be easily integrated with other sensors, enhance the functionality and stability of the system, and solve the problem of insufficient flexibility and scalability of traditional limited systems;

[0046] The present invention performs edge computing through an image acquisition module, reduces the data transmission rate, and alleviates the background computing burden, and is suitable for scenarios with limited network conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0048] Figure 1 A schematic diagram of a fire alarm system based on edge computing provided by an embodiment of the present invention is shown;

[0049] Figure 2 A flow chart showing the preprocessing of fire images by the image acquisition module provided in an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of an AI mainboard of an image acquisition module provided in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0051] The following embodiments of the technical solution of the present invention are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and cannot be used to limit the protection scope of the present invention.

[0052] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0053] In one embodiment, Figure 1 As shown, a local edge computing fire alarm system is provided, including an image acquisition module, an Internet of Things cloud and an information monitoring center.

[0054] The image acquisition module is used to collect fire images.

[0055] The IoT cloud obtains the fire image from the image acquisition module based on the MQTT protocol; the image acquisition module sets a topic on the IoT cloud and uploads the acquired image to the IoT cloud, which sends the image to all subscribers. As the master station of MQTT, the IoT cloud follows the MQTT3.0 protocol, receives the registration of devices, establishes topics, and provides a WEB interface for master station management. The lightweight MQTT protocol is applied to the fire alarm system for transmitting reasoning results and alarm instructions, which solves the problem of insufficient flexibility and scalability of traditional wired systems.

[0056] The information monitoring center subscribes to the topic established by the image acquisition module and receives data associated with the topic distributed by the Internet of Things cloud.

[0057] The image acquisition module uses a smart camera, which adopts the artificial intelligence edge computing RV1126 platform. The hardware design needs to be designed according to the application scenario. On the one hand, the size structure adopts the IPC standard of 50MM×50MM; on the other hand, the peripherals mainly include Gigabit network interface, POE interface, WIFI module, and MIPI CSI interface. The hardware receives the data of the camera sensor from MIPI CSI, processes it through RV1126, and finally sends it to the background data link hardware channel through the network.

[0058] The AI ​​motherboard is the core part of the smart camera. On the one hand, we need to consider the realization of the required functional applications and facilitate debugging to a certain extent. On the other hand, we need to consider the size and structure to achieve product standardization and facilitate application installation.

[0059] Figure 3 This is a schematic diagram of the AI ​​motherboard. It is a standard 50MM×50MM IPC structure, which is convenient for installation into common cavities on the market. The following is a description of its internal modules.

[0060] Gigabit network interface 1, connected to a router or switch through this interface, connected to the Internet of Things cloud of the public network through the router or connected to the Internet of Things cloud of the community through the switch;

[0061] POE port 2: When there is no power supply device in the application scenario, power can be supplied through the POE network port, thus reducing the power supply environment requirements;

[0062] WIFI chip 3, when there is no network cable routing condition in the application scenario, the network can be accessed through WIFI wirelessly;

[0063] The core chip 4 of the AI ​​motherboard is the computing chip RV1126; the CPU used in this chip has a 4-core ARMCortex-A7, an internal integrated 2.0TOPS NPU unit, supports INT8 / INT16, supports TensorFlow, TF-lite, Pytorch, Caffe, ONNX, MXNet, Keras and Darknet. 2 MIPI CSI interfaces can be connected to MIPI CSI cameras. Each MIPI CSI interface supports 4 Lanes mode, and the maximum rate of MIPI CIS is 2.5Gbps / lane. This system uses the MIPI CSI camera to capture images, performs board-level inference on the captured images through the NPU unit, and finally transmits the inference results through the network. The scheduling of the entire drive task requires a high-performance CPU to process; TF card slot 5 can store on-site pictures and operation logs;

[0064] MIPI CSI interface 6 can be adapted to monocular or binocular cameras through this interface;

[0065] OTG port 7 ADB debugging is performed through this port;

[0066] USB HOST interface 8 can be used to expand functions, such as USB to CAN.

[0067] MQTT is a lightweight protocol, but the hardware performance and network environment were insufficient in the early days, so the real-time requirements of the fire alarm system could not be fully met. At the same time, fire alarm equipment in the past was mostly embedded microcontrollers with limited performance, which made it difficult to support the implementation of complex protocols and computing tasks (such as real-time data processing and model reasoning). Therefore, in the field of fire alarm, which requires extremely high security, the relevant standardization and security solutions of MQTT are not mature enough, resulting in slow industry adoption.

[0068] The data matching and real-time issues between smart cameras and MQTT. The image information or video data generated by the camera is usually large in size, while MQTT is a lightweight protocol suitable for transmitting small data. Transmitting the high-capacity data generated by the camera directly through MQTT will increase latency and overload the network bandwidth.

[0069] Therefore, on the one hand, the raw image or video data collected by the smart camera is large in size. Direct transmission through the MQTT protocol without processing will cause delays and network congestion, so the acquired image or video data needs to be adjusted. On the other hand, by performing artificial intelligence reasoning on the camera device side, the amount of data transmission is reduced and the background computing burden is reduced, which is suitable for scenarios with limited network conditions. Figure 2 As shown in FIG. 1 , before transmitting the fire image to the IoT cloud based on the MQTT protocol, the image acquisition module also pre-processes the fire image, including:

[0070] S101. Classify the fire images according to the resolution, and the classification categories include low-resolution images and medium-high-resolution images.

[0071] S102. Preprocess the low-resolution image to enhance image quality and obtain a first target image.

[0072] Specifically, step S102 includes:

[0073] S102.1. Process the low-resolution image based on the FSRCNN model to improve the image resolution; use the FSRCNN model in deep learning super-resolution to improve the image resolution and restore image details, and use Gaussian filtering to reduce image noise to improve image quality, improve image resolution and denoise.

[0074] S102.2. Perform edge enhancement based on the Sobel operator to highlight the outlines of information such as flames and smoke, which is conducive to subsequent model reasoning to obtain key information.

[0075] S102.3. Simulate images in different scenes to obtain diverse image samples, including:

[0076] (1) Geometric transformation and random cropping are used to crop the area where the fire target is located and randomly rotate the image at a certain angle to simulate different viewing angles;

[0077] (2) Lighting and environment simulation: brightness adjustment is used to simulate flame brightness; noise is added by introducing Gaussian noise or salt and pepper noise to simulate low light or harsh environments; random blurring is used to randomly blur the image to simulate rain and fog environments;

[0078] (3) Random occlusion: randomly occlude areas in the image to simulate the situation where flames and smoke are occluded;

[0079] (4) Adjust resolution by reducing the image to different resolutions to simulate low-resolution scenarios in edge devices.

[0080] Based on the above methods, we increase the diversity of samples and simulate changes in different scenarios (angle, lighting, environment and image resolution). The purpose is to provide the model with a wide range of learning opportunities and help the model understand more feature distributions through multi-angle and multi-changing methods.

[0081] Specifically, after obtaining the diversity image samples in step S102, the image acquisition module is also used to perform model inference on the diversity image samples, normalize the pre-processed image information to adapt to the model input requirements, and then upload it to the YOLOv5n model to perform the following processing:

[0082] S102.1. Extract features of preprocessed low-resolution images at multiple scales based on the YOLOv5n model;

[0083] CSPDarknet extracts image features through continuous convolution layers, enhances details and semantic information, and then after multiple convolution layers, the feature map will go through FPN and PANet structures for multi-scale feature extraction. The FPN structure is used to process targets of different sizes and extract multi-scale feature maps through different convolution layers. The PANet structure is mainly to further enhance the fusion process of the extracted feature maps, and strengthens small targets at the detail level. In this way, fire and smoke targets at different scales can be captured. Regardless of the size of the target, YOLOv5n can improve detection accuracy through multi-scale feature maps, and at the same time, attention enhancement (such as edge enhancement) helps to highlight the target area and help the model extract more accurate features.

[0084] S102.2. Predict the position of the target in the image based on the first detection frame; use the Anchor-based method to predict the targets at various positions in the image at different scales. The image after edge enhancement processing helps the model to better identify small target boundaries.

[0085] S102.3. Based on NMS, the first detection frame with a confidence higher than a preset threshold is suppressed and retained, and the coordinates, category label and confidence score associated with the first detection frame are output; among multiple overlapping candidate first detection frames, NMS suppresses those frames with a high degree of overlap and only retains frames with a high confidence level to ensure the accuracy of the output frame, and further optimizes the detection results through NMS to reduce redundant frames.

[0086] Based on the final frame after non-maximum suppression (NMS) processing, the final detection result after removing duplicate frames is obtained. At the model level, the object detection model (YOLOv5n) is trained on a low-resolution dataset to enhance the feature extraction capability under high-resolution input. The specific steps are as follows:

[0087] Data enhancement and model training: The following methods are used for model training: The enhanced data is merged with the original data, with the ratio of original samples to enhanced samples set to 50%-80% to balance data diversity and authenticity. The merged data is divided into training set, validation set and test set in a ratio of 7:2:1. Then the model is preliminarily trained, and the model weights pre-trained on the COCO large-scale dataset are used to fine-tune the fire target detection task. During the training process, the model is prevented from overfitting on the validation set by using methods such as occlusion, and enhancement methods such as blur or noise are introduced to stabilize the performance of the model on the test set.

[0088] Multi-level resolution reasoning: In fire alarm scenarios, system response time is an important indicator of system performance, so multi-level resolution reasoning can detect potential risks in a short time. When inputting low-resolution images, the fast reasoning speed of low-resolution images is used to initially detect targets. The YOLOv5-Tiny model is used to reason about low-resolution images, return the rough position and confidence of the target, and judge whether the threshold is reached based on the returned confidence, so as to screen out possible fire areas.

[0089] The above is a method for processing image data after reducing the resolution during image preprocessing, ensuring that image data information is not lost.

[0090] S103. Perform a cropping operation on the medium and high resolution images to obtain a target area and obtain a second target image.

[0091] Specifically, step S103 uses the target detection capability of the camera to extract key areas related to the fire, such as smoke and flames, from the image information after the resolution adjustment, and discards irrelevant background areas, specifically including:

[0092] S103.1. Perform target detection on the high-resolution image and mark it based on the second detection frame; perform target detection on the medium- and high-resolution images through the target detection framework YOLOv5n model, obtain the candidate frame area, and mark the target by means of a bounding box after detecting the target area.

[0093] S103.2. Extract the fire feature area in the second detection frame and perform high-resolution reasoning; extract the part with fire or smoke through area extraction, and perform subsequent analysis, use high-resolution reasoning for the extracted key areas to more accurately identify the target and reduce the missed detection rate, wherein the output resolution of the camera is dynamically adjusted according to the results of the camera target detection to balance the bandwidth and detection accuracy.

[0094] S103.3. Refine the classification of the fire characteristic area determined by high-resolution reasoning to determine the fire type and intensity.

[0095] S103.4. Perform result fusion on the first target image and the second target image respectively, and output image fusion data.

[0096] The overlapping detection frames of the first target image and the second target image are deduplicated, and the frame with the highest confidence is retained. If the same target is detected in both the low-resolution and high-resolution images, but the positions or sizes of the frames are different, NMS will merge these frames to avoid duplicate detection.

[0097] Box fusion: When the same target is detected, different boxes may be obtained from images with different resolutions. They are weighted according to the confidence level to obtain a more accurate target location.

[0098] Specifically, step S104 includes:

[0099] S104.1. De-duplicate the first detection frame and the second detection frame that are overlapped;

[0100] Non-maximum suppression (NMS): Deduce overlapping detection boxes from different resolutions and retain the most confident boxes. If the same object is detected in both low-resolution and high-resolution images, but the positions or sizes of the boxes are different, NMS will merge these boxes to avoid duplicate detections.

[0101] Box fusion: When the same target is detected, different boxes may be obtained from images with different resolutions. They are weighted according to the confidence level to obtain a more accurate target location.

[0102] S104.2. When the first detection box and the second detection box detect different targets, output a weighted fusion classification result.

[0103] The confidence levels of target detection results at different resolutions are different. Generally, the confidence level at high resolution is higher. However, in order to ensure the accuracy of detection in the case of low resolution, the confidence levels of each resolution can be combined during fusion.

[0104] Weighted fusion: Different weights are assigned to the detection results of each resolution according to the resolution. Results with higher resolution are generally given higher weights.

[0105] Confidence Threshold Fusion: If two resolutions of detection results have high confidence for the same target, the two boxes can be merged into one result. If the confidence of one resolution is too low, the box is discarded.

[0106] S104.3. When the first detection box and the second detection box detect the same target but the associated classification results are different, output the classification result with the highest confidence.

[0107] If the first detection box and the second detection box both detect the same object, but may be associated with different classification results, the object needs to be fused. The highest confidence or more detailed high-resolution classification result is used as the output result.

[0108] If the first detection frame and the second detection frame detect different targets, the target regions detected in images with different resolutions are merged to ensure that the final target region is as accurate as possible, especially when the target boundary is blurred, and the high-resolution image can provide more accurate information.

[0109] After processing by S104.2-S104.3, the fused results will provide the final target detection frame, target category and confidence level, ensuring that there is no missed detection or false detection. By fusing the information of images with different resolutions, the final target position will be more accurate, avoiding positioning errors caused by insufficient resolution.

[0110] After the image acquisition module obtains the processed output results, it also assigns QoS levels according to the data type, and then uploads the detection results to the IoT cloud through the MQTT protocol for remote monitoring and alarm. The specific allocation of QoS levels is shown in Table 1.

[0111] Table 1

[0112]

[0113] MQTT provides three levels of QoS, which are reasonably allocated according to the importance and real-time requirements of the data:

[0114] QoS 0: At most once, no retransmission, suitable for non-critical data;

[0115] QoS 1: At least once transmission. Messages may be repeated, but they are guaranteed to arrive.

[0116] QoS 2: Exactly once, which ensures that messages arrive without duplication and is suitable for critical data with high reliability requirements.

[0117] QoS 1 or QoS 2 is used to transmit high-priority data including critical alarms and control instructions, QoS 1 is used to transmit medium-priority data such as status updates and real-time monitoring information, and QoS 0 is used to transmit device health status and periodic reporting data.

[0118] The image acquisition module is also used to dynamically adjust the QoS level according to the network status and device load, including:

[0119] When the network load is high or the bandwidth is insufficient, the QoS of some data is reduced (the data type of QoS 1 is reduced to QoS 0);

[0120] When a critical event occurs, the QoS 1 level of the alarm data is upgraded to QoS 2.

[0121] As the client of the IoT cloud, the information monitoring center subscribes to the topic set up by the image acquisition module. When the image acquisition module sends data to the topic, the information monitoring center can obtain the data on the topic distributed by the IoT cloud. Then, the received data is diagnosed and evaluated, and the confidence data of the reasoning results of the image acquisition module are compared. Then, combined with the on-site pictures with corresponding confidence levels, the disaster situation and risk assessment level are obtained, as well as whether there are safety issues around the disaster site. Based on the above results, it is judged whether the threshold of safety warning has been reached, and a safety warning is issued through alarm equipment, such as a photoelectric alarm.

[0122] The information monitoring center displays the distribution of regional image acquisition modules on a display screen. The risk assessment level of the data sent back by each image acquisition module (below 60: dangerous; 60-80: normal; above 80: good) can be observed in real time, helping staff to quickly understand relevant conditions.

[0123] The information monitoring center is also used to realize risk assessment visualization, including:

[0124] Determine image risk assessment levels for multiple image acquisition modules;

[0125] Obtain risk scores based on image confidence and image risk assessment levels;

[0126] Risk score = w1×image confidence + w2×image risk assessment level; where w1 and w2 are predetermined weights.

[0127] Based on this, the intelligent monitoring center displays the inference confidence and risk assessment level, and displays the fire risk situation in different areas in real time, which is convenient for rapid disposal. By combining the edge device inference data and sensor data (such as temperature and smoke concentration), dynamic assessment and classification of fire risk (low, medium and high risk) are achieved. The risk threshold is adjusted dynamically according to the environment to reduce false alarms.

[0128] The above embodiments are only used to introduce the technical solutions of the present application in detail, but the description of the above embodiments is only used to help understand the methods of the embodiments of the present invention and should not be understood as limiting the embodiments of the present invention. Any changes or substitutions that can be easily thought of by those skilled in the art should be included in the protection scope of the embodiments of the present invention.

Claims

1. A fire alarm system based on edge computing, characterized in that: include: An image acquisition module, used for acquiring fire images; The Internet of Things cloud obtains the fire image from the image acquisition module based on the MQTT protocol; The image acquisition module sets a topic on the IoT cloud and uploads the acquired image to the IoT cloud, and the IoT cloud sends the image to all subscribers; An information monitoring center subscribes to the topic established by the image acquisition module and receives data associated with the topic distributed by the IoT cloud; Before transmitting the fire image to the IoT cloud based on the MQTT protocol, the image acquisition module also pre-processes the fire image, including: Classifying the fire images according to resolution, the classification categories including low-resolution images and medium-high-resolution images; Preprocessing the low-resolution image to enhance image quality, thereby obtaining a first target image; Performing a cropping operation on the medium and high resolution image to obtain a target area and obtain a second target image; fusing the first target image and the second target image respectively, and outputting image fusion data; The image acquisition module performs preprocessing on the low-resolution image, including: Processing the low-resolution image based on the FSRCNN model to improve the image resolution; Perform edge enhancement based on the Sobel operator to highlight the information contour; Simulate images in different scenes to obtain diverse image samples; The image acquisition module is also used to perform model reasoning based on the diverse image samples, including: Extract features of the preprocessed low-resolution image at multiple scales based on the YOLOv5n model; Predicting a position target in the image based on the first detection frame; Based on NMS suppression, retain the first detection frame whose confidence is higher than a preset threshold, and output the coordinates, category label and confidence associated with the first detection frame; The image acquisition module performs a cropping operation on the medium and high resolution images, including: The YOLOv5n model is used to detect medium and high resolution images and mark them based on the second detection frame. Extracting the fire feature area in the second detection frame and performing high-resolution reasoning; Refining the classification of the fire characteristic area through high-resolution reasoning to determine the fire type and intensity; The image acquisition module fuses the first target image and the second target image, including: Deduplication of the first detection frame and the second detection frame that are overlapped; When the first detection frame and the second detection frame detect different targets, output a weighted fusion classification result; When the first detection box and the second detection box detect the same target but have different associated classification results, the classification result with the highest confidence is output.

2. The fire alarm system based on edge computing according to claim 1, characterized in that: The image acquisition module allocates QoS levels according to the data type, specifically: QoS 0: Periodic update of environmental variables and device health status; QoS 1: alarm data issued by the image acquisition module after detecting a fire, equipment failure status, and video clips or key frames uploaded by the image acquisition module when detecting a fire; QoS 2: Fire confirmation command.

3. The fire alarm system based on edge computing according to claim 2, characterized in that: The image acquisition module is also used to dynamically adjust the QoS level according to the network status and device load, including: When the network load is high or the bandwidth is insufficient, the data type of QoS 1 is downgraded to QoS 0; When a critical event occurs, the QoS 1 level of the alarm data is upgraded to QoS 2.

4. The fire alarm system based on edge computing according to claim 1, characterized in that: The information monitoring center is also used to realize risk assessment visualization, including: Determine image risk assessment levels for multiple image acquisition modules; Obtain risk scores based on image confidence and image risk assessment levels; Risk score = w1·image confidence + w2·image risk assessment level; where w1 and w2 are predetermined weights.

Images