Fire extinguisher monitoring management method based on deep learning and Internet of Things technology

Abstract

The invention relates to a fire extinguisher monitoring management method based on deep learning and Internet of Things technology, which adopts visual identification and multi-sensor fusion technology based on deep learning as a core detection means of a system, and takes a pressure state, a safety bolt state and a position state of a fire extinguisher as system monitoring indexes. The detection comprehensiveness and reliability are effectively improved, and the false alarm rate is reduced; the improved YOLOv5 model is combined with the Internet of Things communication technology, installation and deployment are easy and convenient, the state of the fire extinguisher can be accurately monitored in real time, an alarm is triggered immediately under the abnormal condition, and the effectiveness of fire fighting equipment is ensured. Aiming at the intelligent monitoring of the state of the fire extinguisher, the diversity of fire fighting equipment and the environmental complexity are fully considered, and the system has the characteristics of high precision and low power consumption.

Description

Technical Field

[0001] The present invention belongs to the technical field of fire-fighting equipment monitoring and management, and specifically relates to a fire extinguisher monitoring and management method based on deep learning and Internet of Things technologies. Background Art

[0002] Fire extinguishers are essential firefighting equipment for extinguishing early-stage fires, and their integrity is directly linked to emergency response capabilities in the event of a fire. However, traditional fire extinguisher management methods present numerous challenges. Traditional methods require manual visual inspections of each extinguisher, including pressure gauge indicators, safety pin integrity, and placement compliance. In densely populated environments like large commercial complexes and underground parking lots, manual inspection cycles are lengthy, the risk of missed inspections is high, and inspection results are easily influenced by personnel experience. Furthermore, manual inspections lack real-time performance, making it impossible to monitor the status of fire extinguishers in real time. If problems such as insufficient pressure, missing safety pins, or displacement occur during the inspection cycle, these issues are difficult to detect and address in a timely manner, potentially preventing the extinguisher from functioning properly in the event of a fire, resulting in serious consequences.

[0003] In recent years, the rapid development of the Internet of Things (IoT) and artificial intelligence (AI) technologies has provided new solutions for intelligent fire extinguisher management. While some IoT monitoring devices have been put into use, they still suffer from several drawbacks. First, their detection methods are limited. Most solutions rely solely on pressure sensors to determine device status, lacking multi-dimensional awareness of safety latch status and device displacement. Second, existing monitoring and management methods lack accurate monitoring of fire extinguisher location, making them ineffective in addressing issues such as lost or misplaced extinguishers.

[0004] To address the above issues, the present invention proposes a fire extinguisher monitoring and management method based on deep learning and Internet of Things technology, aiming to achieve real-time monitoring, intelligent analysis and remote management of fire extinguisher status, and solve the drawbacks of traditional management methods. Summary of the Invention

[0005] In view of the current technological deficiencies, the present invention provides a fire extinguisher monitoring and management method based on deep learning and Internet of Things technology. The present invention can effectively solve the problems of low efficiency and easy omissions in traditional fire extinguisher management methods, and realize real-time monitoring, automatic alarm and remote management of fire extinguisher status.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: A fire extinguisher monitoring and management method based on deep learning and Internet of Things technology includes the following steps: S1, the multi-state detection of the deep learning recognition module, is divided into three parts: image acquisition and preprocessing, deep learning model construction and deployment, and anomaly determination. These three parts enable real-time monitoring of the fire extinguisher pressure status and safety latch; S2, pressure sensor senses the position of the fire extinguisher; S3: The limit switch assists in identifying whether the fire extinguisher is correctly placed. If the limit switch and the pressure sensor status conflict, the image acquisition in step S1 is started for manual review; S4, wireless communication module for data transmission; S5. The alarm module has multi-level response, with local sound and light alarm, cloud alarm and self-check feedback mechanism.

[0007] Preferably, step S1 includes: S1-1. Image Acquisition and Preprocessing: A miniature camera is mounted on the side of the fire extinguisher mounting bracket, with the lens aimed at the fire extinguisher's pressure gauge and safety latch area. A built-in infrared fill light automatically activates in low-light environments. Images are taken every four hours. When the pressure sensor or limit switch detects an anomaly, the camera is immediately activated to capture real-time images. A bilateral filtering algorithm is used to eliminate image noise, retain edge details, and perform geometric correction. Based on a pre-set fire extinguisher template, SIFT feature matching is used to locate the pressure gauge and latch areas, and the pressure gauge and latch ROIs are cropped to reduce subsequent computational effort. S1-2. Build a data set to collect images of the pressure gauge and latch under different lighting and angles, and annotate the coordinates of the pointer tip and the area it belongs to, as well as the latch status; To optimize the model, we inserted an angle regression branch before the detection head of the YOLOv5s model. This branch outputs the polar angle θ between the pointer and the center of the dial. Based on θ, we calculate whether the pointer is in the green area. We also embedded an SE module at the end of the Backbone model to adaptively weight channel features, allowing the network to focus on the metallic reflective features of the latch. S1-3. When an abnormal pressure or latch disappearance is detected, the log and the associated JPEG image are uploaded immediately via the MQTT protocol with a service quality level of 1 on the 4G module.

[0008] Preferably, step S2 is: a pressure sensor is selected and placed on the base of the fire extinguisher, and the output signal is connected to the 24-bit high-precision A / D converter HX711 through the I2C interface, and then the converted data is input into the STM32 chip; the STM32 chip is set to collect data at a fixed time, and enters the standby state after the collection is completed; the set reference value is P0, which is the normal weight of the fire extinguisher after system correction. When the single sampling value P<0.9P0, it is marked as suspected displacement; when the single sampling value P<0.5P0, it is marked as the fire extinguisher is lost; the above two abnormal situations are uploaded to the cloud, and the real-time picture is captured in conjunction with the camera, and the target detection is used to confirm whether the fire extinguisher is still in the field of view.

[0009] Preferably, in step S3, a waterproof limit switch with a normally closed contact is installed on the inner side of the fire extinguisher bracket. When the fire extinguisher is correctly placed, its bottom presses the switch trigger arm to keep the circuit closed.

[0010] Preferably, in step S4, a Cat1 communication module is used to configure a deep sleep mode, which wakes up once every 4 hours. After waking up, the data including the device ID, pressure value, latch status, and battery level are packaged and sent to the cloud; at the same time, an event trigger is reported. If any module in steps S1-S3 detects an abnormality, the sleep mode is interrupted immediately and an alarm code is uploaded.

[0011] Preferably, in step S5, after the alarm is triggered, the local buzzer and red LED will first be driven to flash, and will continue until manually reset. At the same time, the alarm information will be pushed to the management platform through the MQTT protocol. The platform automatically generates a work order and notifies the responsible person by SMS. After the alarm is triggered, the system automatically starts the camera to capture the scene image and attaches it to the alarm information for the management personnel to remotely confirm whether it is a false alarm.

[0012] Preferably, in step S1-2, the pointer position is obtained by the following steps: S1-2-1, coordinate conversion, detecting the center point of the pressure gauge dial pointer tip ,in is the coordinate of the center point of the dial on the x-axis, is the coordinate of the center point of the dial on the y-axis, is the coordinate of the tip of the dial pointer on the x-axis, is the coordinate of the tip of the dial pointer on the y-axis; S1-2-2, Angle calculation, first calculate the horizontal difference and vertical difference ,pass Get the radian value, then convert the radian system to the degree system, and finally output : according to The value determines the pressure status of the fire extinguisher. If If it falls in the preset green area, it is judged as normal pressure, and all other areas are judged as abnormal pressure and reported to the cloud; S1-2-3. Design of the loss function: assign different weights to the three loss terms to achieve a balance among multiple tasks: in, This is the positioning loss for target detection. It optimizes bounding box regression accuracy and assigns a high weight of 0.5 to prioritize the positioning accuracy of the pressure gauge dial and latch. is the angle regression loss function, used to constrain the predicted angle The mean square error with the true value is assigned a medium weight of 0.3 to accurately control the pointer direction; It is the target confidence loss, which ensures the reliability of the model for the detection results and is given a low weight of 0.2 to avoid overfitting.

[0013] As a preferred option, the SE module is embedded at the end of Backbone, with the following structure: (1) Squeeze operation: perform global average pooling to compress the feature map to 1×1×C: Among them, C represents the channel dimension of the feature map, is the activation value of the cth channel of the input feature map at the spatial position (i, j), is the height of the feature map, width; (2) Excitation operation: Generate channel weights through two layers of full connection: in, It is the channel-level global average pooling result of the input feature map, with a dimension of 1×1×C; is the weight matrix of the first fully connected layer, with dimension , r is the reduction ratio, is the weight matrix of the second fully connected layer, with dimension C .

[0014] Compared with the prior art, the present invention has the following beneficial effects: This intelligent fire extinguisher status monitoring system takes into account the diversity of firefighting equipment and the complexity of the environment, offering high precision and low power consumption. It utilizes deep learning-based visual recognition and multi-sensor fusion technology as the system's core detection method, using the fire extinguisher's pressure, safety latch, and position as system monitoring indicators. This effectively improves the comprehensiveness and reliability of detection and reduces false alarm rates. This system utilizes an improved YOLOv5 model combined with IoT communication technology, making it easy to install and deploy. It can accurately monitor fire extinguisher status in real time, triggering alarms immediately in abnormal situations to ensure the effectiveness of firefighting equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the system framework of the present invention; Figure 2 This is a schematic diagram of the improved YOLOv5 network structure of the present invention. DETAILED DESCRIPTION

[0016] The technical solutions of the present invention are further specifically described below through examples. These examples are provided for the purpose of illustrating the present invention and are not intended to limit the present invention. All other examples obtained by persons of ordinary skill in the art based on the examples in this application without creative work are intended to fall within the scope of protection of this application.

[0017] Reference Figure 1 , a fire extinguisher monitoring and management method based on deep learning and Internet of Things technology, including the following steps: S1. The multi-state detection (state detection) of the deep learning recognition module is divided into three parts: image acquisition and preprocessing, deep learning model construction and deployment, and anomaly judgment. These three parts enable real-time monitoring of the fire extinguisher pressure status and safety pin.

[0018] S1-1. Image Acquisition and Preprocessing: A miniature camera is mounted on the side of the fire extinguisher mounting bracket, with the lens focused on the fire extinguisher's pressure gauge and safety latch area. A built-in infrared fill light (wavelength 850nm) automatically activates in low-light environments. Images are recorded every four hours. When the pressure sensor or limit switch detects an anomaly, the camera immediately activates to capture real-time images. A bilateral filtering algorithm is used to eliminate image noise, preserve edge details, and perform geometric correction. Based on a pre-set fire extinguisher template, SIFT feature matching is used to locate the pressure gauge and latch areas. The pressure gauge and latch ROIs are then cropped to reduce subsequent computational effort.

[0019] S1-2. Construct a dataset, collecting 2,000 images of pressure gauges and latches under various lighting conditions and angles. The coordinates of the pointer tip, its corresponding region (green safety zone / red abnormal zone), and the latch status are annotated. The model is optimized by inserting an angle regression branch before the YOLOv5s detection head. This branch outputs the polar angle θ between the pointer and the center of the dial. Based on θ, the pointer is calculated to determine whether it is within the green region. To address the narrow detection target of the safety latch, a Squeeze-and-Excitation (SE) module is embedded at the end of the Backbone. This module adaptively weights channel features, allowing the network to focus on the metallic reflective characteristics of the latch.

[0020] The deep learning recognition module realizes the real-time monitoring of the fire extinguisher pressure status and safety pin. It is necessary to identify whether the fire extinguisher pressure is within the safe range and whether the safety pin is detached. Figure 2 The shown part is based on yolov5 and the backbone and head parts are improved.

[0021] Add a pressure gauge pointer angle regression branch, add an angle regression branch before the YOLOv5 detection head, and output it in parallel with the original classification and bounding box branches. This branch outputs the polar angle , calculate the pointer position by the following steps: S1-2-1, coordinate conversion, detecting the center point of the pressure gauge dial pointer tip ; S1-2-2, Angle calculation, first calculate the horizontal difference and vertical difference ,pass Get the radian value, then convert the radian system to the degree system, and finally output The calculation formula is as follows: according to The value determines the pressure status of the fire extinguisher. If If the pressure falls within the preset green area (such as 120°-240°), it is judged as normal. Otherwise, it is judged as abnormal pressure and reported to the cloud.

[0022] S1-2-3. Design of loss function, assigning different weights to the three loss terms to achieve a balance among multiple tasks. The calculation formula is as follows: in, It is the positioning loss of target detection, which optimizes the bounding box regression accuracy and is assigned a high weight of 0.5, giving priority to ensuring the positioning accuracy of the pressure gauge dial and latch. is the angle regression loss function used to constrain the predicted angle The mean square error with the true value is assigned a medium weight of 0.3 to accurately control the direction of the pointer. It is the target confidence loss, which ensures the reliability of the model for the detection results and is given a low weight of 0.2 to avoid overfitting.

[0023] The security bolt occupies a small proportion in the image (usually less than 5% of the pixels), and the metal reflection is easily affected by ambient light, making feature extraction difficult. To solve the problem of small target detection, we use the attention mechanism to let the network focus on the metal reflection features of the bolt. Figure 2 As shown, the SE module is inserted after the C3 module at the end of the Backbone, which is after the SPP module. The structure is as follows: (1) Squeeze operation: Perform global average pooling to compress the feature map to 1×1×C. The calculation formula is as follows: in, The activation value of the c-th channel of the input feature map at the spatial position (i, j), are the height and width of the feature map.

[0024] (2) Excitation operation: Generate channel weights through two layers of full connection. The calculation formula is as follows: in, It is the channel-level global average pooling result of the input feature map, with a dimension of 1×1×C. is the weight matrix of the first fully connected layer, with dimension , r is the reduction ratio (the default is 16). is the weight matrix of the second fully connected layer, with dimension C .

[0025] S1-3. When an abnormal pressure or latch disappearance is detected, the log and the associated JPEG image are uploaded immediately via an MQTT message with a QoS level of 1 via the 4G module.

[0026] S2. A pressure sensor senses the position of the fire extinguisher (position detection). A pressure sensor (range 0-50kg) is used and placed on the base of the fire extinguisher. The output signal is connected to a 24-bit high-precision HX711 A / D converter via the I2C interface. The converted data is then input into the STM32 chip. The STM32 chip is set to collect data at a fixed time and enters standby mode after the collection is completed. To accurately measure the weight of the fire extinguisher, the STM32 is designed to calculate the current null value and automatically tare it after the system is first powered on. If the single-sample value P < 0.9P0, it is marked as "suspected displacement"; if the single-sample value P < 0.5P0, it is marked as "fire extinguisher missing." Both of these abnormal conditions are uploaded to the cloud and linked to the camera to capture real-time footage. Object detection confirms whether the fire extinguisher is still in view.

[0027] S3: A limit switch assists in verifying the correct placement of the fire extinguisher. A waterproof limit switch (normally closed contacts) is installed inside the fire extinguisher bracket. When the fire extinguisher is correctly placed, its bottom presses against the switch trigger arm, keeping the circuit closed. The limit switch and pressure sensor complement each other. If there is a conflict between the two states (e.g., normal pressure but the limit switch is open), the camera in step S1 is activated for manual verification.

[0028] In order to detect whether the fire-fighting facilities are lost, a dual detection mode is adopted with pressure sensors as the main and limit switches as the auxiliary. The overall structural layout is as follows: 1. The pressure weighing module adopts a high-precision strain gauge sensor, which is precisely installed at the geometric center of the PCB board. It is heightened by a precision CNC-machined aluminum alloy bracket to ensure that the load-bearing surface maintains the best contact plane with the fire extinguisher base.

[0029] 2. Special stainless steel compression springs are configured at the four corners of the PCB board. The top of the spring supports a 5mm thick optical grade acrylic board with an anti-slip texture treatment on the surface.

[0030] 3. Use IP67 protection grade limit switches (OMRON SS series) symmetrically arranged on both sides of the long side of the PCB board. Use IP67 protection grade limit switches (OMRON SS series) symmetrically arranged on both sides of the long side of the PCB board.

[0031] The operating state mechanism is divided into normal and abnormal states. In the normal state, the weight of the fire extinguisher compresses the spring by 6mm, and 90% of the load is transmitted to the pressure weighing module through the heightened structure. The limit switch contacts are compressed and maintain the normally closed state. In the abnormal state, the spring rebounds after the fire extinguisher is removed, and the pressure sensor load drops below 5% of the range. The limit switch contacts release and the circuit is disconnected.

[0032] Dual logic verification of the pressure sensor and limit switch. The limit switch, serving as the primary wakeup source for the STM32, is directly connected to the MCU's dedicated wakeup pin and configured in dual-edge trigger mode. This generates an immediate interrupt signal when the fire extinguisher moves, instantly waking the STM32 from deep sleep to full-speed operation. Upon wakeup, the system simultaneously activates a high-precision pressure sensor for real-time weighing. The 24-bit ADC collects the current weight data and compares it with a preset threshold.

[0033] In normal monitoring mode, the system maintains a four-hour wake-up cycle via the real-time clock (RTC) module. Upon each wakeup, the system checks the electrical status of the limit switch and collects pressure sensor data. When the pressure is within the normal range and the limit switch remains closed, the system uploads this status data to the cloud management platform via the 4G network. If pressure anomalies or a disconnected switch occur, a priority alarm mechanism is immediately triggered, sending an abnormality alert to the cloud. If a logical inconsistency between the pressure data and the switch status arises, the system initiates a visual verification process, capturing the scene with a camera and running a deep learning model for target confirmation. The final verification result is marked as "verification abnormal."

[0034] The S4 and 4G wireless communication modules provide low-power data transmission. A Cat1 communication module is configured for deep sleep mode (wake-up every four hours). Upon wake-up, data such as device ID, pressure value, latch status, and battery level are packaged and sent to the cloud. In addition to scheduled wake-up, event triggering is also implemented. If any module detects an anomaly in steps 1-3, sleep is immediately interrupted and an alarm code is uploaded.

[0035] The 4G communication module uses the Quectel EC200T Cat.1 module, which supports the TCP / IP protocol stack and has a default DRX cycle of 4 hours. Upon wakeup, the device sends a heartbeat packet containing information such as the device ID, fire extinguisher weight, limit switch status, battery level, fire extinguisher pressure, and safety latch status. If an anomaly is detected, the device immediately exits sleep mode and issues an alarm message via MQTT.

[0036] The S5 alarm module features a multi-level response mechanism, including local audio and visual alarms, cloud-based alarms, and self-diagnosis feedback. When an alarm is triggered, a local buzzer (sound pressure ≥85dB@1m) and a red LED flash (at a frequency of 2Hz) are activated, which persists until manually reset. Simultaneously, the alarm is pushed to the management platform via the MQTT protocol, which automatically generates a work order and sends a text message to the responsible personnel. Once the alarm is triggered, the system automatically activates a camera to capture the scene and appends it to the alarm message, allowing management personnel to remotely confirm whether it is a false alarm.

[0037] The alarm system utilizes a hierarchical response mechanism, enabling coordinated local and cloud-based alerting. The local alarm unit incorporates a high-pitched piezoelectric buzzer (operating voltage 3.3V, drive current ≤20mA). This buzzer utilizes PWM modulation (50% duty cycle, frequency 2kHz±5%) to generate an audible warning sound of at least 85dB at 1m. Simultaneously, a high-brightness red LED (wavelength 620-625nm) flashes synchronously at a 0.5Hz frequency, providing a dual audio and visual alert. For cloud-based collaboration, the system utilizes an event-driven architecture. When an abnormal signal is triggered, an alert is pushed to the management platform in real time via the MQTT protocol (QoS=1). The platform engine automatically analyzes the event type and generates a maintenance work order, which is then instantly sent to the designated responsible person's mobile device via an SMS gateway. Simultaneously, the system initiates image forensics, uploading JPEG-compressed on-site images to a cloud storage server via an HTTPS encrypted channel. A link to the image is embedded in the work order for visual review by management personnel.

[0038] S6. System-level low-power optimization measures. Regarding microcontroller power management, an intelligent peripheral power control system was designed. A dynamic power switch matrix using N-channel MOSFETs was constructed to precisely sequence power to the pressure sensor and camera module, activating power only during the data acquisition window and completely shutting off power during non-operating periods. The STM32 main control chip utilizes the Stop mode interface in the HAL library to achieve ultra-low-power standby while maintaining SRAM data integrity. A four-hour periodic wake-up mechanism was configured using the RTC alarm interrupt. The wireless communication module's energy-saving design utilizes the enhanced power-saving technology of the LTE Cat.1bis chip (EC200A series). Power Saving Mode (PSM) and extended discontinuous reception (DDR) work together to reduce the module's standby current to 0.85mA (@3.8V) during non-transmission periods, while maintaining an average network reachability of 5 seconds.

[0039] In order to verify the actual application effect of the system of the present invention in fire extinguisher monitoring and management, a systematic test was conducted, mainly evaluating the detection accuracy, false alarm rate, power consumption, response time and other aspects. The specific data are as follows: Detection Accuracy: Utilizing an improved YOLOv5 visual recognition algorithm, this system identifies fire extinguisher pressure gauge readings, safety latch status, and position in complex lighting and occlusion environments. In actual testing, the system achieved 98.7% accuracy in fire extinguisher status recognition, 98.2% in pressure status recognition, 99.1% in safety latch detection, and 98.8% in position recognition. While the traditional YOLOv5 system only recognizes safety latches with an accuracy of approximately 95%, this new system not only significantly improves this accuracy but also accurately identifies pressure status and position.

[0040] False alarm rate: In 300 inspections, the system only reported two false alarms, keeping the overall false alarm rate at 0.67%, far lower than the 2% to 5% false alarm rate of traditional visual inspection systems caused by occlusion, reflection, and blur.

[0041] Low power consumption: The system adopts a phased power supply strategy and deep sleep mode design, reducing power consumption to less than 50μW during non-detection cycles. The average power consumption during the detection and communication active phases is less than 80mW, reducing overall system energy consumption by approximately 72% compared to traditional continuous operation mode.

[0042] Real-time performance and response speed: The average response time from fire extinguisher status detection to alarm signal triggering is less than 1.2 seconds, which can achieve rapid response to sudden situations and effectively support immediate handling in emergency situations.

[0043] Deployment and operational stability: The system can be fixed near the fire extinguisher using standard screws, eliminating the need for complex installation. The system has been tested for over 1,200 hours of continuous operation without any abnormal restarts or communication interruptions, demonstrating excellent stability.

[0044] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A fire extinguisher monitoring and management method based on deep learning and Internet of Things technology, characterized in that The steps include: S1, the multi-state detection of the deep learning recognition module, is divided into three parts: image acquisition and preprocessing, deep learning model construction and deployment, and anomaly determination. These three parts enable real-time monitoring of the fire extinguisher pressure status and safety latch; S2, pressure sensor senses the position of the fire extinguisher; S3: The limit switch assists in identifying whether the fire extinguisher is correctly placed. If the limit switch and the pressure sensor status conflict, the image acquisition in step S1 is started for manual review; S4, wireless communication module for data transmission; S5. The alarm module has multi-level response, with local sound and light alarm, cloud alarm and self-check feedback mechanism.

2. The fire extinguisher monitoring and management method based on deep learning and Internet of Things technology according to claim 1 is characterized in that Step S1 includes: S1-1. Image Acquisition and Preprocessing: A miniature camera is mounted on the side of the fire extinguisher mounting bracket, with the lens aimed at the fire extinguisher's pressure gauge and safety latch area. A built-in infrared fill light automatically activates in low-light environments. Images are taken every four hours. When the pressure sensor or limit switch detects an anomaly, the camera is immediately activated to capture real-time images. A bilateral filtering algorithm is used to eliminate image noise, retain edge details, and perform geometric correction. Based on a pre-set fire extinguisher template, SIFT feature matching is used to locate the pressure gauge and latch areas, and the pressure gauge and latch ROIs are cropped to reduce subsequent computational effort. S1-2. Build a data set to collect images of the pressure gauge and latch under different lighting and angles, and annotate the coordinates of the pointer tip and the area it belongs to, as well as the latch status; To optimize the model, we inserted an angle regression branch before the detection head of the YOLOv5s model. This branch outputs the polar angle θ between the pointer and the center of the dial. Based on θ, we calculate whether the pointer is in the green area. We also embedded an SE module at the end of the Backbone model to adaptively weight channel features, allowing the network to focus on the metallic reflective features of the latch. S1-3. When an abnormal pressure or latch disappearance is detected, the log and the associated JPEG image are uploaded immediately via the MQTT protocol with a service quality level of 1 on the 4G module.

3. The fire extinguisher monitoring and management method based on deep learning and Internet of Things technology according to claim 2 is characterized in that Step S2 is: a pressure sensor is placed on the base of the fire extinguisher, and the output signal is connected to the 24-bit high-precision A / D converter HX711 through the I2C interface, and the converted data is input into the STM32 chip; the STM32 chip is set to collect data at a fixed time and enters the standby state after the collection is completed; The set benchmark value is P0. When the single sampling value P<0.9P0, it is marked as suspected displacement; when the single sampling value P<0.5P0, it is marked as the fire extinguisher is lost. The above two abnormal situations are uploaded to the cloud and linked with the camera to capture real-time images. Through target detection, it is confirmed whether the fire extinguisher is still in the field of view.

4. The fire extinguisher monitoring and management method based on deep learning and Internet of Things technology according to claim 1 is characterized in that In step S3, a waterproof limit switch with normally closed contacts is installed on the inner side of the fire extinguisher bracket. When the fire extinguisher is correctly placed, its bottom presses the switch trigger arm to keep the circuit closed.

5. The fire extinguisher monitoring and management method based on deep learning and Internet of Things technology according to claim 1 is characterized in that In step S4, the Cat1 communication module is used to configure the deep sleep mode, which wakes up once every 4 hours. After waking up, the data including device ID, pressure value, latch status, and battery level are packaged and sent to the cloud; at the same time, event trigger reporting is used. If any module in steps S1-S3 detects an abnormality, the sleep mode is immediately interrupted and an alarm code is uploaded.

6. The fire extinguisher monitoring and management method based on deep learning and Internet of Things technology according to claim 1 is characterized in that In step S5, after the alarm is triggered, the local buzzer and red LED will first be driven to flash, and will continue until manually reset. At the same time, the alarm information will be pushed to the management platform through the MQTT protocol. The platform will automatically generate a work order and notify the responsible personnel via SMS. After the alarm is triggered, the system will automatically start the camera to capture the scene image and attach it to the alarm information for the management personnel to remotely confirm whether it is a false alarm.

7. The fire extinguisher monitoring and management method based on deep learning and Internet of Things technology according to claim 2 is characterized in that In step S1-2, the pointer position is obtained by the following steps: S1-2-1, coordinate conversion, detecting the center point of the pressure gauge dial pointer tip ,in is the coordinate of the center point of the dial on the x-axis, is the coordinate of the center point of the dial on the y-axis, is the coordinate of the tip of the dial pointer on the x-axis, is the coordinate of the tip of the dial pointer on the y-axis; S1-2-2, Angle calculation, first calculate the horizontal difference and vertical difference ,pass Get the radian value, then convert the radian system to the degree system, and finally output : according to The value determines the pressure status of the fire extinguisher. If If it falls in the preset green area, it is judged as normal pressure, and all other areas are judged as abnormal pressure and reported to the cloud; S1-2-3. Design of the loss function: assign different weights to the three loss terms to achieve a balance among multiple tasks: in, This is the positioning loss for target detection. It optimizes bounding box regression accuracy and assigns a high weight of 0.5 to prioritize the positioning accuracy of the pressure gauge dial and latch. is the angle regression loss function, used to constrain the predicted angle The mean square error with the true value is assigned a medium weight of 0.3 to accurately control the pointer direction; It is the target confidence loss, which ensures the reliability of the model for the detection results and is given a low weight of 0.2 to avoid overfitting.

8. The fire extinguisher monitoring and management method based on deep learning and Internet of Things technology according to claim 2 is characterized in that: The SE module is embedded at the end of Backbone, and the structure is as follows: (1) Squeeze operation: perform global average pooling to compress the feature map to 1×1×C: Among them, C represents the channel dimension of the feature map, is the activation value of the cth channel of the input feature map at the spatial position (i, j), is the height of the feature map, width; (2) Excitation operation: Generate channel weights through two layers of full connection: in, It is the channel-level global average pooling result of the input feature map, with a dimension of 1×1×C; is the weight matrix of the first fully connected layer, with dimension , r is the reduction ratio, is the weight matrix of the second fully connected layer, with dimension C .

Images