Fire detection edge node and multi-probe linkage control system and method

Abstract

The application discloses a fire detection edge node and a multi-probe linkage control system and method, which comprises an air-inhale smoke detection unit, a multi-gas sensing collection interface, an AI inference unit, a main control unit, a gas sensing data public bus, a multi-port PoE exchange unit and a fire control upper computer, and a double-channel architecture of a SPI management channel and an RMII data channel is adopted for physically isolating the main control unit and the multi-port PoE exchange unit; a hardware mapping table pre-solidified in a nonvolatile memory is used to uniquely bind a physical port number, a detection terminal MAC address and fire point coordinates; after smoke early warning triggering, the main control unit configures a static address forwarding table through the SPI channel, sends a wake-up frame through the RMII channel, and realizes sub-second wake-up of the detection terminal through the multi-port PoE exchange unit. The application replaces a traditional software protocol stack scheduling with a hardware-level linkage link, the actual end-to-end linkage delay is less than or equal to 0.8s, point position misplacement is eliminated through physical port binding, and a multi-mode weighted fusion decision reduces the false alarm rate in a complex scene.

Description

Technical Field

[0001] This invention relates to the field of fire detection technology, and in particular to a fire detection edge node and multi-probe linkage control system and method. Background Technology

[0002] Aspirating smoke detection systems, with their high sensitivity, are widely used in high-value protection scenarios such as data centers, industrial plants, precision instrument workshops, and archives. However, according to statistics on the operational quality of fire protection facilities released by the Ministry of Emergency Management in recent years, traditional fire protection systems suffer from fire location errors and emergency response delays accounting for 27% of failures due to misaligned wiring and software binding failures. This is a core cause of fire protection failures in industrial and computer room settings. Furthermore, according to GB50116-2013 "Code for Design of Automatic Fire Alarm Systems" and GB16806-2006 "Fire Linkage Control System," the fire linkage controller should issue an activation signal within 3 seconds of receiving a fire alarm signal. Traditional software linkage solutions, due to protocol stack overhead, generally have an actual linkage delay greater than 3 seconds, making it difficult to meet the needs of high-level protection scenarios for rapid verification of very early fire conditions, posing serious compliance risks and safety hazards.

[0003] The existing traditional fire detection and linkage architecture has the following inherent defects: First, the linkage response latency is too high. The existing mainstream solution adopts a three-level software linkage architecture of "detector-host computer-terminal device". The detection signal needs to go through multiple software processing steps such as operating system TCP / IP protocol stack encapsulation, network transmission, platform parsing, and command issuance. The overall linkage latency is generally greater than 3 seconds, which cannot meet the needs of rapid verification and handling of fires in the very early stage of smoldering.

[0004] Secondly, the problem of misaligned locations on site is frequent. The location of detectors and cameras in traditional fire protection systems relies on IP addresses and device serial numbers for software binding. Misaligned wiring on construction sites, port changes during later maintenance, and equipment replacements can all lead to a mismatch between the actual physical location of the equipment and the location registered in the CAD drawings. This location deviation after a fire occurs directly delays emergency response work.

[0005] Finally, the false alarm rate remains high in complex scenarios. Single smoke detection methods are highly susceptible to interference from industrial dust, welding fumes, and ambient moisture. Existing multimodal detection solutions mostly rely on loose data fusion at the software level, lacking hardware-level synchronous acquisition and standardized weighted decision mechanisms. As a result, the algorithms have poor robustness and cannot effectively distinguish between real fires and environmental interference.

[0006] Chinese patent publication number CN120356294A discloses an image fusion-based edge-cloud collaborative smart fire monitoring system. It adopts a dual-light fusion algorithm and an edge-cloud collaborative architecture. However, its linkage relies on cloud-based collaborative scheduling. This technical solution does not involve a hardware-level MAC forwarding mechanism for the switching chip. The linkage latency is constrained by the network protocol stack, and it cannot achieve sub-second hardware wake-up.

[0007] Chinese Patent Publication No. CN117095505A discloses a flame detection alarm method for edge AI devices. This technical solution only implements flame recognition alarm based on software algorithms and does not involve physical port hardware point binding, dual-channel physical isolation and MAC layer hardware high-speed wake-up mechanism, and cannot solve the problems of misalignment of engineering points and high latency.

[0008] Chinese patent publication number CN221841481U discloses a novel early-stage gas-breathing pyrolysis particle detector that focuses on improving single-point detection sensitivity, but this technical solution does not involve multimodal fusion and edge computing integration.

[0009] Chinese Patent Publication No. CN118658121A discloses a smart fire monitoring method, medium, and terminal based on YOLOv8 and Involution modules. This technical solution only involves algorithm-level improvements and does not have a hardware-level linkage architecture.

[0010] The aforementioned existing technologies all involve single-aspect technical problems and do not comprehensively consider the shortcomings of early fire detection in high-value asset environments, such as linkage lag, location misalignment, high false alarm rate, and inability to identify early lithium battery fires. Therefore, it is necessary to improve the existing technologies. Summary of the Invention

[0011] To address the shortcomings of existing technologies, such as delayed linkage, misaligned detection points, high false alarm rates, and inability to identify early lithium battery fires, this invention provides an integrated fire detection edge node and multi-probe linkage control system and method. Through hardware integration architecture optimization, physical layer point binding, MAC layer hardware forwarding, dynamic weighted fusion judgment, and a high-risk scenario-specific alarm mechanism, it achieves high-precision intelligent fire detection linkage with sub-second linkage, zero point misalignment, and a low false alarm rate. In a first aspect, the present invention provides a fire detection edge node and multi-probe linkage control system, comprising: The aspirating smoke detection unit has a built-in auxiliary control MCU, which is used to collect ambient smoke concentration signals and transmit smoke warning signals to the main control unit via the SPI bus; Multiple gas sensor acquisition interfaces, compatible with I2C and SPI dual buses, connected to the gas sensor data common bus, used to collect ambient gas concentration data and transmit it to the main control unit; The AI ​​inference unit runs an improved YOLOv8n dual-stream neural network model with an embedded ESCFM-RS attention module to identify early fires; The main control unit is connected to the aspirating smoke detection unit via a gas sensing data common bus, receives smoke warning signals and performs fire judgment and linkage control; The multi-port PoE switching unit is provided with a physically isolated SPI management channel and RMII data channel between itself and the main control unit. The SPI management channel is used for configuring the static address forwarding table, and the RMII data channel is used for hardware forwarding of service data frames. The downlink PoE port of the multi-port PoE switching unit is connected to the probe terminal. The fire alarm control unit is connected to the multi-port PoE switching unit to receive alarm information and location data.

[0012] Secondly, the present invention provides a fire detection edge node and multi-probe linkage control method, comprising the following steps: Step 401: Smoke threshold warning triggered. When the ambient smoke concentration rises and exceeds the preset warning threshold, the auxiliary control MCU triggers an SPI interrupt, wakes up the main control unit (104) and enters the linkage handling process; Step 402: Hardware addressing and probe wake-up. The main control unit (104) retrieves the hardware mapping table pre-stamped in the non-volatile memory, determines the target physical port and the corresponding MAC address, writes the static address forwarding table entry to the multi-port PoE switching unit (106) through the SPI management channel, constructs and sends the Magic Packet wake-up frame through the RMII data channel, and the multi-port PoE switching unit (106) performs Layer 2 hardware forwarding to the target port according to the static address forwarding table to wake up the corresponding camera. Step 403: Multimodal data synchronous acquisition. The awakened camera completes the preset position scan and transmits back the visible light and thermal infrared dual-channel video streams. The AI ​​inference unit outputs the smoke confidence C and the temperature rise slope T. The gas sensor synchronously outputs the hydrogen concentration and CO concentration. The value of C is in the range of [0, 1], and the unit of T is ℃ / min. Step 404: Determine the weight of normal operating conditions according to the normalization formula. Calculate the overall score, where W1, W2, and W3 are all conventional weighting coefficients, satisfying the following conditions: Under normal operating conditions High-risk working conditions T_sat is the upper limit of temperature rise saturation, in °C / min; H2_base and CO_base are the normalized bases for hydrogen and carbon monoxide, respectively. It is a comprehensive fire situation score, with a value range of [0, +∞); Step 405: Determine whether the concentration of high-risk gas exceeds the dynamic weight trigger threshold. Determine whether the current hydrogen concentration exceeds the dynamic weight trigger threshold. If H2 is less than or equal to the dynamic weight trigger threshold, the system maintains the normal operating weight (0.3, 0.3, 0.4) and calculates the normal comprehensive score according to the normalization formula described in step 404. Then proceed to step 408 to perform the normal fire threshold determination. If H2 is greater than the dynamic weight trigger threshold, the system immediately switches to the high-risk operating dynamic weight (0.3, 0.1, 0.6), recalculates the dynamic comprehensive score according to the normalization formula described in step 404, and then proceeds to step 406 to further determine whether the hydrogen-specific alarm threshold is exceeded. Step 406: Further determine whether the hydrogen-specific alarm threshold has been exceeded? If yes, proceed to step 407 to perform a lithium battery-specific mandatory alarm determination; if no, proceed to step 408 to perform a regular fire threshold determination. Step 407: A lithium battery-specific mandatory alarm is triggered, directly identifying it as a lithium battery thermal runaway fire, and proceeding to step 409 to execute coordinated response. Step 408: Standard fire threshold determination. Compare the comprehensive score calculated in step 404 according to the current effective weight with the preset fire judgment threshold. If the comprehensive score is greater than or equal to the preset fire judgment threshold, it is determined to be a real fire and proceeds to step 409 to execute the linkage response. If the comprehensive score is less than the preset fire judgment threshold, return to step 401 to continue monitoring. Step 409: The system activates the audible and visual alarms and fire extinguishing devices. Step 410: Upload the location information to the host computer.

[0013] Furthermore, the preset warning threshold range is 0.001~0.8%obs / m.

[0014] Furthermore, the center value of the dynamic weight trigger threshold is 50ppm, and a ±2ppm hysteresis anti-jitter range is configured, that is, the rising edge trigger point is 52ppm and the falling edge backoff point is 48ppm; the hydrogen-specific alarm threshold is 100ppm.

[0015] Furthermore, the normalized base for hydrogen is 160 ppm, and the normalized base for carbon monoxide is 200 ppm.

[0016] This invention addresses the pain points of traditional fire alarm linkage, such as high latency, easy misalignment of alarm points, and high false alarm rate in complex scenarios. The measured end-to-end linkage latency is as low as 0.8s, achieving sub-second ultra-low latency linkage, with physical-level misalignment prevention. The false alarm rate in industrial scenarios is significantly reduced to 5.2%, a decrease of 33.4 percentage points compared to the traditional solution's 38.6%. The overall technical solution is highly integrated, has strong anti-interference capabilities, and achieves dynamic weight adaptive decision-making, breaking through the limitations of traditional algorithms. It is suitable for high-level fire alarm scenarios, especially lithium battery thermal runaway scenarios, demonstrating outstanding engineering practicality and innovation. Attached Figure Description

[0018] Figure 1 : Connection block diagram of the edge node hardware module of the present invention; Figure 2 : A schematic diagram of the hardware mapping table data structure of this invention; Figure 3 : Flowchart of the multi-probe collaborative wake-up timing of this invention; Figure 4 This invention provides a flowchart of the fire situation judgment and closed-loop handling logic. Detailed Implementation

[0020] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0021] The core of this invention lies in providing a dual-channel physically isolated fire protection edge node hardware architecture, such as... Figure 1 As shown in the edge node hardware module connection diagram of this invention, the main control unit 104 and the multi-port PoE switching unit 106 abandon the traditional single network channel and instead construct a physically separated architecture of SPI management channel (control plane) and RMII data channel (data plane). The SPI management channel is dedicated to register reading and writing and static ATU table configuration, and only transmits control commands; the RMII data channel is dedicated to carrying Magic Packet and service data. This structure physically eliminates data congestion between control commands and video streams, providing a hardware foundation for sub-second linkage.

[0022] refer to Figure 1 The hardware module connection diagram of the edge node of the present invention includes a suction-type smoke detection unit 101, a multi-gas sensor acquisition interface 102, an AI inference unit 103, a main control unit 104, a gas sensor data common bus 105, a multi-port PoE exchange unit 106, and a fire-fighting host computer 107.

[0023] The inhalation-type smoke detection unit 101 and the multi-gas sensor acquisition interface 102 transmit the collected environmental signals to the gas sensor data common bus 105. The multi-gas sensor acquisition interface 102 is compatible with both I2C and SPI buses and is equipped with a MEMS gas sensor array. The hydrogen detection range is 0-1000ppm with a resolution of 1ppm, and the carbon monoxide detection range is 0-500ppm with a resolution of 1ppm. It can collect high-frequency and high-precision data on the concentration of hazardous gases. The sensor signal lines are fully wrapped with an aluminum foil shielding layer and grounded at one end. To avoid industrial electromagnetic interference, the aspirating smoke detection unit 101 is equipped with a 375nm ultraviolet laser light source and a high-precision photoelectric receiving circuit. Combined with a high-quiet air pump and a labyrinth-type air chamber, the smoke detection accuracy can reach 0.001%obs / m, which is precisely adapted to the very early smoldering fire detection scenarios in computer rooms and industrial plants. The device is equipped with an industrial-grade auxiliary control MCU, which is dedicated to the PWM speed regulation of the air pump, laser drive, weak signal acquisition and SPI interrupt triggering, so as to realize the rapid early warning and reporting of smoke anomalies. Preferably, the auxiliary control MCU adopts STM32H743.

[0024] The AI ​​inference unit 103 relies on the computing power of the main control NPU to run an improved YOLOv8n dual-stream neural network model embedded with an ESCFM-RS attention module. Its input resolution is downsampled to 640×640 and compressed by 8-bit quantization. It optimizes feature extraction capabilities for early diffuse smoke characteristics and supports time-division multiplexing of NPU resources for alternating inference of visible light + thermal infrared dual-channel 1080P video streams. The single-channel single-frame inference latency is ≤50ms, and the dual-channel alternating integrated frame rate is ≥20fps. It can quickly identify early fires with low contrast and weak features. The inferred signal is then transmitted to the main control unit 104. The main control unit 104 processes the signal and transmits it to the gas sensor data common bus 105 and the multi-port PoE exchange unit 106. The fire-fighting host computer 107 further receives the signal from the multi-port PoE exchange unit 106.

[0025] The main control unit 104 uses an industrial-grade processor that supports NPU parallel inference, integrating a multi-core CPU and an NPU artificial intelligence processing unit. The NPU computing power can stably support the parallel real-time inference requirements of dual 1080P video streams, meeting the computing power requirements of local AI fire identification at the edge. Preferably, a Rockchip RK3566 processor can be used. The main control unit 104 and the switching chip construct a dual physical channel isolation architecture, with the two channels operating independently and without interference, completely avoiding crosstalk between data and control commands: Channel 1 serves as a management channel, with the processor's SPI controller pin connected to the switching chip's SPI management interface, dedicated to register read / write, static ATU table configuration, and port status query. It only transmits control commands and does not carry any business data. The SPI interface clock frequency is configured to 10MHz to ensure the real-time writing of ATU table entries. Channel 2 serves as a data channel, with the processor's RMII interface connected to the switching chip's dedicated management port, operating in 100Mbps high-speed mode, dedicated to carrying business data transmission such as Magic Packet wake-up frames, video streams, and alarm data. The bandwidth can meet the synchronous backhaul requirements of multiple 1080P video streams.

[0026] The detection node of this invention adopts a six-layer PCB board integrated design. The core components are universally selected and can be adapted for iterative upgrades. The overall architecture realizes the integrated integration of detection, switching, computing power and acquisition functions. All components are selected based on functionality and are not limited to a single model. The layout spacing between the analog sampling circuit and the digital core circuit is strictly controlled at more than 35mm, with a slotted isolation strip in the middle to achieve physical isolation between the digital and analog circuits. The analog ground and digital ground are connected at a single point using a 100MHz high-frequency ferrite bead to completely eliminate ground loop interference. Preferably, the ferrite bead model for analog ground and digital ground is BLM18PG121SN1. The power module adopts a wide voltage input design of 19.2V~28.8V and stably outputs 48V / 1.5A for the PoE port. The full load efficiency is ≥88%. It integrates all-dimensional protection against overvoltage, undervoltage, overcurrent and short circuit, and is suitable for complex industrial power supply environments.

[0027] The multi-port PoE switching unit 106 is composed of a switching chip and an industrial-grade PoE PSE controller, which work together via an I2C bus. The switching chip is an industrial-grade PoE switching chip that supports static MAC address forwarding table (address TranslationUnit) configuration and Layer 2 hardware forwarding, and supports IEEE 802.3af / at standard PoE power supply, providing stable standby and operating power to the downlink camera. Preferably, the switching chip is a Microchip KSZ8863RLL series, and the PoE PSE controller is a TPS23861. The downlink PoE port is paired with a camera that supports IEEE 802.3af standard PoE standby and standard WOL. Devices using the (Wake-on-LAN) protocol typically maintain low-power standby operation with a standby power consumption of ≤1.5W. They do not undergo a complete power-off hibernation process, providing the hardware foundation for sub-second wake-up. Camera wake-up latency is precisely broken down: sensor initialization ≤80ms, lens auto-calibration ≤120ms, thermal infrared and visible light frame synchronous output ≤300ms, and stable link outflow ≤300ms. This invention uses an internal high-precision timer for latency testing, completing 30 valid repeated tests. The end-to-end total latency is consistently ≤0.8s, with a typical runtime latency of 0.65-0.75s. The data is reproducible and verifiable. The standard test procedure is as follows: smoke concentration exceeds the limit → SPI interrupt trigger → MAC table hardware configuration → Magic Packet transmission → camera wake-up outflow → NPU inference completion → comprehensive score exceeds the limit → GPIO triggers relay linkage.

[0028] The end-to-end linkage latency of this invention is obtained by summing the extreme latencies of each hardware component, with no software scheduling redundancy. The specific latency composition details are shown in Table 1 below:

[0029] Referring to Figure 2, which illustrates the hardware mapping table data structure of this invention, it includes a device identity information data table 201, a non-volatile storage fixed area 202, and a physical port point binding field 203. The hardware mapping table of this invention uses a fixed configuration. During the device manufacturing process, an 8-bit DIP switch is used to configure a unique HostID for the entire network, preventing device identity conflicts. After system power-on, the main control unit automatically generates a static IP address based on the HostID, requiring no manual configuration. The main control unit 104 actively detects camera devices connected to each PoE port via the ONVIF protocol, accurately obtains the device MAC address, and generates a standardized hardware mapping table. A single table entry has the format: [HostID, Physical Port Number, Camera MAC Address, Gas Sensor Channel Number, CAD Point Coordinates]. An example entry is: [HostID=3, Port=1, AM_MAC=00:1A:2B:3C:4D:5E, Gas_CH=1, CAD_Pt=(120, 250)]. All mapping table data is stored in SPI Flash. Non-volatile memory ensures data retention even after power failure and automatic reloading upon power-up, eliminating the need for repeated on-site calibration and permanently establishing a unique binding relationship between the port and physical location. The device identity physical binding logic is as follows: the logical fire point identity of a camera is uniquely determined by the physical port number it connects to, regardless of the device's IP address or model. Even if the camera is replaced or the device's IP address is modified on-site, as long as the device is connected to the original physical port, the system automatically matches the original CAD location information, completely avoiding location misalignment issues caused by maintenance operations. The camera MAC self-learning mechanism works as follows: After the camera connects to the network, it periodically sends ARP requests or DHCPDiscover broadcast frames. The switching chip listens in real time and learns the MAC addresses of each port device, dynamically updating the port MAC learning table. This ensures that the main control unit can verify the port device access status in real time. The periodic error prevention verification process is as follows: During system operation, the main control unit reads the dynamic MAC learning table of the switching chip port through the SPI interface every 5 seconds, comparing it with the pre-fixed hardware mapping table port by port. If the actual access device MAC does not match the bound MAC, it is immediately determined to be an abnormal condition such as physical link misalignment or device replacement. The host computer is immediately uploaded with a 0xE1 fault code (defined as: 0xE1 fault code is a port point binding abnormality fault). At the same time, the forwarding permission of the fire alarm linkage control frame of the faulty port is locked. Non-linkage services such as video surveillance are transmitted normally, which not only eliminates the risk of erroneous linkage but also ensures that real-time on-site monitoring is not interrupted until the administrator calibrates and resets the mapping table on-site to restore the port linkage function. When the system detects a 0xE1 fault code (port point binding abnormality), the following complete recovery process is executed: S-B1: The main control unit immediately locks the fire alarm linkage control frame forwarding permission of the faulty port, prohibiting the port from triggering any audible and visual alarms and fire extinguishing device linkages. At the same time, non-linkage business data such as video surveillance are transmitted normally to ensure uninterrupted real-time on-site monitoring. S-B2: The main control unit continuously compares the actual MAC address of the faulty port with the bound MAC address in the fixed mapping table once per second, and periodically uploads the 0xE1 fault code and faulty port information to the fire control computer until the fault is cleared. S-B3: The administrator confirms the physical link status on-site. If it is a legitimate change such as device replacement or port swapping, the mapping table is reset according to the aforementioned "Hardware Mapping Table Update / Rewrite Process". If it is an illegal access or line fault, the physical abnormality is eliminated and the power is restarted or a manual verification command is triggered. S-B4: After the main control unit confirms that the actual MAC address of the faulty port matches the fixed mapping table in the next periodic verification, it will automatically clear the 0xE1 fault code, unlock the linkage permission of the port, and restore the normal fire alarm linkage function. S-B5: The main control unit uploads a 0xE1 fault recovery confirmation frame to the host computer, records the fault duration, recovery time, and operator information to the local log storage for post-event traceability and auditing.

[0030] The recovery process adopts a three-stage security mechanism of "lock-verify-unlock" to ensure that the linkage will not be accidentally triggered before the fault is completely eliminated. At the same time, it can automatically recover without manual intervention after the fault is eliminated, thus balancing security and ease of operation and maintenance.

[0031] The standardized configuration and anti-collision mechanism of the SPI static MAC table is as follows: The main control unit 104 writes ATU table entries using a three-level timing sequence of "clearing the dynamic table - writing to the static table - locking the table entry" to thoroughly eliminate interference from old data and ensure binding accuracy. In multi-device concurrent configuration scenarios, priority arbitration is achieved based on the unique HostID. The smaller the HostID value, the higher the configuration priority, preventing table entry overwriting and configuration failure caused by simultaneous writing by multiple nodes. The hardware forwarding latency is stable and controllable, with a single ATU table write latency ≤0.1ms and a Layer 2 frame forwarding latency ≤0.2ms. There are no random software latency fluctuations, and the entire process is quantifiable and reproducible.

[0032] In concurrent configuration scenarios where multiple edge nodes share the same switching chip's SPI management bus, to prevent table entry overwriting or bus conflicts caused by multiple nodes simultaneously writing to the ATU table, this invention adopts a deterministic priority arbitration mechanism based on HostID. The specific implementation timing is as follows: S-C1: During the system power-on initialization phase, the master control unit of each edge node reads its own HostID (value range 0~255) through an 8-bit DIP switch. The smaller the HostID value, the higher the configuration priority. S-C2: When multiple nodes simultaneously detect a smoke warning and initiate an ATU table write request, each node first sends an arbitration probe frame via the SPI bus, and the frame contains its own HostID; S-C3: After receiving the arbitration probe frame, the SPI controller of the switching chip compares the HostID in the frame with the HostID of the currently authorized write node, and only responds to the node with the smallest HostID value, returning an ACK response to it; the other nodes receive a NACK or do not receive a response within the preset timeout window (preferably 50μs), and then enter the backoff waiting state. S-C4: The node authorized to write executes the three-level configuration timing sequence of "clear dynamic table → write static table → lock table entry", and releases the SPI bus occupancy flag after completion; S-C5: After the backoff waiting node detects that the SPI bus has been released, it will re-initiate the arbitration probe frame in ascending order of HostID, repeating the process from S-C3 to S-C4 until all nodes to be configured have completed the ATU table writing.

[0033] The arbitration mechanism is implemented purely in hardware timing, without relying on software scheduling or random backoff algorithms. In the worst case, the total waiting latency for N concurrent nodes is... This includes the time spent in a single arbitration. This represents the time consumed for a single ATU table write. For a typical 4-node concurrent scenario, the maximum additional waiting latency is ≤0.6ms, which is negligible compared to the 0.8s end-to-end linkage latency, ensuring that the real-time linkage performance is not affected in multi-node concurrent scenarios. At the same time, since the HostID is fixed by a physical DIP switch, there is no risk of priority conflict or deadlock caused by software allocation, and the arbitration result is completely deterministic, predictable, and reproducible.

[0034] When replacing cameras, adjusting port wiring, or adding detection terminals on-site, a hardware mapping table update / rewrite process must be performed. The specific steps are as follows: S-A1: The administrator sends a mapping table update command to the main control unit through the fire alarm supervisory computer or local debugging interface. The command carries the target physical port number and the operation type (add / replace / delete). S-A2: After receiving the instruction, the main control unit first reads the current dynamic MAC learning table entry of the corresponding port of the switching chip through the SPI management channel to obtain the actual MAC address of the newly connected device; S-A3: The main control unit assembles the new MAC address with the physical port number, gas sensor channel number and CAD point coordinates in the instruction into a new entry, writes it to the corresponding position in the hardware mapping table of the SPI Flash non-volatile memory, and overwrites the original old entry; if the operation type is deletion, the corresponding entry is marked as invalid. S-A4: The main control unit executes a three-level configuration sequence of "clearing the dynamic address table → writing a new static address table → locking the table entry" to the switching chip through the SPI management channel, so that the new binding relationship takes effect immediately; S-A5: The main control unit returns an update result confirmation frame to the host computer, which includes the new table entry content and the written verification code; after the host computer verifies that the verification code is consistent, it updates the local CAD drawing filing information and completes end-to-end synchronization.

[0035] The mapping table update process supports single-port online hot updates without requiring a system power outage and restart. Normal monitoring and linkage functions on other ports remain unaffected during the update process. The SPI Flash employs a wear leveling algorithm, ensuring a single-sector erase / write lifespan of ≥100,000 cycles, meeting full lifecycle maintenance requirements.

[0036] Referring to Figure 3, the timing flowchart of the multi-probe collaborative wake-up of the present invention is as follows: at time T0, the smoke concentration exceeds the standard and triggers the SPI interrupt; at time T1, the main control unit completes the static MAC address write lock; at time T2, the RMII interface sends the first Magic Packet byte; at time T3, the switching chip completes the hardware layer 2 forwarding according to the static address forwarding table; at time T4, the camera outputs the first frame of valid video stream.

[0037] Referring to Figure 4, which is a flowchart of the fire detection and closed-loop response logic of this invention, taking a data center cabinet fire monitoring scenario as an example, the test environment is: temperature 22-26℃, humidity 40%-60%, and no strong electromagnetic interference, which is suitable for the rapid response requirements of early-stage fires in computer rooms. The system's preset core parameter is: the first warning threshold. T1=0.005%obs / m, fire judgment threshold Score=0.5, temperature rise saturation upper limit 35℃ / min, hydrogen normalized baseline 160ppm, carbon monoxide normalized baseline 200ppm; conventional weighting ratio (0.3, 0.3, 0.4), dynamic weighting trigger threshold 50ppm, configured ±2ppm hysteresis anti-shake range, specifically including the following steps: Step 401: Smoke threshold warning triggered. When the ambient smoke concentration rises to 0.03%obs / m, exceeding the preset warning threshold, the auxiliary control MCU immediately triggers an SPI interrupt, waking up the main control unit to enter the linkage handling process. Step 402: Hardware addressing and probe wake-up. The main control unit 104 retrieves the hardware mapping table, accurately matches the target camera port and MAC address, and writes a static ATU binding entry to the switching chip via the SPI interface to solidify the correspondence between the port and the device. The standard Magic Packet wake-up frame is constructed and sent via the RMII interface. The switching chip performs high-speed forwarding at hardware layer 2 to complete the camera wake-up and send back dual-channel video streams of visible light and thermal infrared. The measured total wake-up latency is ≤0.8s. Step 403: Multimodal data synchronous acquisition, the camera quickly completes the preset position scan and transmits back the visible light and thermal infrared dual-channel video streams; NPU inference outputs smoke confidence C=0.72, temperature rise slope T=13.35℃ / min; synchronous gas acquisition data: H2=120ppm, CO=45ppm; Step 404: Determine the weight of normal operating conditions according to the normalization formula.

[0038] Calculate the overall score, where W1, W2, and W3 are all conventional weighting coefficients, satisfying the following conditions: Under normal operating conditions High-risk working conditions T_sat is the upper limit of temperature rise saturation, in °C / min; H2_base and CO_base are the normalized bases for hydrogen and carbon monoxide, respectively. It is a comprehensive fire situation score, with a value range of [0, +∞); The algorithm score under normal operating conditions is calculated as follows: Score = 0.3 × 0.72 + 0.3 × (13.35 / 35) + 0.4 × max(120 / 160, 45 / 200) = 0.216 + 0.1144 + 0.3 = 0.6304 ≈ 0.63. The upper limit of temperature rise saturation of this algorithm is 35℃ / min, which is set with reference to the critical threshold of smoldering to open flame in indoor fire test. The weight coefficient is optimized based on 1200 sets of real fire and interference samples. After sensitivity test verification, it can achieve a false alarm rate of <2%, which balances detection sensitivity and anti-interference. Step 405: Determine whether the concentration of high-risk gas exceeds the dynamic weight threshold, and whether the current hydrogen concentration exceeds the dynamic weight trigger threshold. If H2 is less than or equal to the dynamic weight trigger threshold, the system maintains the normal operating weight (0.3, 0.3, 0.4) and calculates the normal comprehensive score according to the normalization formula described in step 404, then proceeds to step 408 to perform the normal fire threshold determination; if H2 is greater than the dynamic weight trigger threshold, the system immediately switches to the high-risk operating dynamic weight (0.3, 0.1, 0.6), recalculates the dynamic comprehensive score according to the normalization formula described in step 404, then proceeds to step 406 to further determine whether the hydrogen-specific alarm threshold is exceeded. Step 406: Further determine whether the hydrogen concentration exceeds the hydrogen-specific alarm threshold. If yes, proceed to step 407 to perform the lithium battery-specific mandatory alarm judgment; if no, proceed to step 408 to determine whether the conventional fire threshold has been reached. Step 407: A lithium battery-specific mandatory alarm is triggered, directly identifying it as a lithium battery thermal runaway fire, and proceeding to step 409 to execute coordinated response. Step 408: Standard fire threshold determination. The main control unit 104 compares the comprehensive score calculated in step 404 according to the current effective weight with the preset fire judgment threshold. If the comprehensive score is greater than or equal to the preset fire judgment threshold, it is determined to be a real fire and proceeds to step 409 to execute the linkage response. If the comprehensive score is less than the preset fire judgment threshold, it returns to step 401 to continue monitoring. Step 409: The system activates the audible and visual alarms and fire extinguishing devices. Step 410: Upload the location information to the host computer.

[0039] In the aforementioned data center embodiment, step 405 determines that H2 = 120ppm > 50ppm, switches to high-risk operating condition dynamic weighting, and calculates a dynamic comprehensive score of 0.70; step 406 further determines that H2 = 120ppm ≥ 100ppm (hydrogen-specific alarm threshold), triggering a lithium battery-specific mandatory alarm (step 407), skipping the scoring and judgment logic, directly determining it as a real very early-stage fire, and proceeding to step 409 to execute coordinated response. Both the normal operating condition score of 0.63 and the dynamic score of 0.70 exceed the 0.5 judgment threshold, further verifying the reliability of the judgment conclusion.

[0040] The performance comparison between the implementation effect of the technical solution of the present invention and the existing technologies 1-2 is shown in Table 2 below:

[0041] The comparison of the implementation effects of the hardware edge linkage scheme of the present invention and the software linkage scheme of the prior art is shown in Table 3 below:

[0042] In another embodiment of the present invention, when the test environment is in an industrial plant setting, the temperature is 18-30℃, humidity is 30%-70%, and industrial electromagnetic interference level is 3, to verify the system's multimodal fusion anti-interference and false alarm suppression capabilities. Ten typical industrial interference scenarios, including welding fumes, equipment dust, ambient moisture, high-temperature hot air, lighting interference, and electromagnetic interference, are selected. 100 sets of valid test samples are set for each scenario, for a total of 1000 sets of control test samples. Comparative tests are carried out simultaneously using traditional software fusion solutions.

[0043] Test Condition: Welding operations generated diffuse fumes that briefly raised the ambient fume concentration to 0.008%obs / m³, triggering a Level 1 system warning. The equipment then initiated multimodal synchronous data acquisition and AI inference. System Acquisition Parameters: AI image recognition identified the issue as welding fume interference, with a fume confidence level C=0.22; the thermal infrared image showed no temperature fluctuations, with a temperature rise slope T=0℃ / min; the gas sensors detected H₂=0ppm and CO=10ppm, indicating no high-risk gas leakage characteristics. Substituting these parameters into a conventional algorithm calculation: Score = 0.3 × 0.22 + 0.3 × 0 + 0.4 × max(0 / 160, 10 / 200) = 0.066 + 0 + 0.02 = 0.086. The current hydrogen concentration is 0 ppm, far below the dynamic weighting trigger threshold of 50 ppm (including the -2 ppm hysteresis fallback point of 48 ppm). The system maintains the conventional weighting mode, and the overall score of 0.086 is far below the 0.5 decision threshold. Test results show that the average false alarm rate of this invention is only 5.2%, while the average false alarm rate of the traditional software fusion scheme is 38.6%. The false alarm rate is significantly reduced to 5.2%, a decrease of 33.4 percentage points compared to the 38.6% of the traditional software fusion scheme. It can accurately distinguish between industrial interference and real fires, and is suitable for long-term stable operation in complex industrial production environments.

[0044] As another embodiment of the present invention, in a specific scenario of thermal runaway in a lithium battery room, this embodiment verifies the system's ability to identify critical fire situations without open flames or visible smoke in the ultra-early detection scenario of thermal runaway in a data center lithium battery room. The system activates the lithium battery-specific detection mode, lowers the smoke warning threshold T1=0.003%obs / m, and adds a single-item mandatory alarm threshold of 100ppm for hydrogen, adapting to the fire situation characteristics where gas leakage characteristics take precedence over smoke and temperature in the early stage of lithium battery thermal runaway. In the critical stage of thermal runaway of a single lithium battery cell, there is no visible smoke, no open flame, and no obvious flame characteristics in the environment. The system collects data from multiple dimensions: the smoke concentration of the aspirating smoke detection unit is 0.002%obs / m, which does not reach the warning threshold, and the smoke confidence level C=0; the ambient temperature rise slope T=8.2℃ / min; and the gas detection CO=45ppm and H2=120ppm. The normalized score is calculated as follows: Score = 0.3 × 0 + 0.3 × (8.2 / 35) + 0.4 × max(120 / 160, 45 / 200) = 0 + 0.0703 + 0.3 = 0.3703 ≈ 0.37. The base score is 0.37 < 0.5, which does not reach the normalized threshold. Simultaneously, the detected hydrogen concentration H2 = 120 ppm consistently exceeds the dynamic weight trigger threshold of 52 ppm (center value 50 ppm, ±2 ppm hysteresis). Therefore, the system automatically switches to a high-risk dynamic weighting ratio (0.3, 0.1, 0.6) and recalculates the dynamic comprehensive score. Score_dynamic = 0.3×0 + 0.1×(8.2 / 35) +0.6×max(120 / 160,45 / 200)=0+0.0234+0.45=0.4734≈0.47. Even after dynamic weight optimization, the score is still lower than the conventional fire judgment threshold of 0.5. The system then triggers the hydrogen single-item high-risk feature mandatory alarm mechanism: H2=120ppm is detected, which is greater than the special alarm threshold of 100ppm. The conventional scoring and judgment logic is skipped, and it is directly judged as a real fire of lithium battery thermal runaway. The main control unit 104 immediately triggers the sound and light alarm, camera fixed-point verification and pre-fire extinguishing linkage, and completes fire intervention in the ultra-early critical stage, effectively avoiding the risk of lithium battery thermal runaway spread, fire and explosion, and filling the technical gap that traditional fire-fighting equipment cannot detect early hidden dangers of lithium batteries.

[0045] This invention constructs an integrated edge node on the hardware side, integrating aspirating smoke detectors, PoE network switching, edge AI inference, and multi-gas sensing and acquisition functions. Through a physically isolated architecture between the SPI management channel and the RMII data channel, it realizes hardware-level network configuration and data forwarding. By using a factory-fixed hardware mapping table, it establishes a unique binding relationship between physical ports and actual fire-fighting locations, eliminating location offsets at the hardware level.

[0046] The method of this invention is based on a hardware architecture to realize the whole process of edge local linkage, including threshold triggering, hardware addressing, fast wake-up, multi-mode acquisition, dynamic weight fusion judgment, high-risk mandatory alarm, and closed-loop handling. It does not require cloud or host computer intervention, which greatly improves response speed and judgment accuracy, and is suitable for both conventional fires and special fire scenarios such as lithium battery thermal runaway.

[0047] This invention integrates an aspirating smoke detection unit, a multi-port PoE switching unit, a main control unit with a built-in NPU, and a multi-gas sensor acquisition interface on the hardware side. It adopts a dual-path architecture with physically separated SPI management channel and RMII data channel to realize hardware-level data forwarding and device configuration. Based on a hardware mapping table of physical ports, device MAC, and CAD points, this invention constructs a "port as point" physical anti-misalignment mechanism. It triggers hardware-level MAC layer addressing and rapid wake-up of multiple probes through smoke concentration thresholds, integrates visible light / thermal infrared AI inference features, temperature rise slope, and multi-gas concentration parameters, and relies on a dynamically weighted normalized weighting algorithm and a high-risk gas single-item forced alarm mechanism to achieve accurate fire judgment.

Claims

1. A fire detection edge node and multi-probe linkage control system, characterized in that, include: The inhalation-type smoke detection unit (101) has a built-in auxiliary control MCU, which is used to collect the ambient smoke concentration signal and transmit the smoke warning signal to the main control unit (104) via the SPI bus. The multi-gas sensor acquisition interface (102) is compatible with both I2C and SPI buses and is connected to the gas sensor data common bus (105) to acquire ambient gas concentration data and transmit it to the main control unit (104). The AI ​​inference unit (103) is integrated inside the main control unit (104). It runs on the NPU computing power of the main control unit (104) and executes the improved YOLOv8n dual-stream neural network model with embedded ESCFM-RS attention module to identify early fires. The main control unit (104) is connected to the aspirating smoke detection unit (101) via a gas sensing data common bus (105) to receive smoke warning signals and perform fire judgment and linkage control; The multi-port PoE switching unit (106) is provided with a physically isolated SPI management channel and RMII data channel between itself and the main control unit (104). The SPI management channel is used for configuring the static address forwarding table, and the RMII data channel is used for hardware forwarding of service data frames. The downlink PoE port of the multi-port PoE switching unit (106) is connected to the probe terminal. The fire alarm control unit (107) is connected to the multi-port PoE exchange unit (106) to receive alarm information and location data.

2. The fire detection edge node and multi-probe linkage control system as described in claim 1, characterized in that, The inhalation-type smoke detection unit (101) has a built-in 375nm ultraviolet laser light source and an auxiliary control MCU, the MCU model being STM32H743.

3. The fire detection edge node and multi-probe linkage control system as described in claim 1, characterized in that, The built-in switching chip of the multi-port PoE switching unit (106) is Microchip KSZ8863RLL series, and the PoE PSE controller is TPS23861.

4. A method for controlling fire detection edge nodes and multi-probe linkage, characterized in that, Includes the following steps: Step 401: Smoke threshold warning triggered. When the ambient smoke concentration rises and exceeds the preset warning threshold, the auxiliary control MCU triggers an SPI interrupt, wakes up the main control unit (104) and enters the linkage handling process; Step 402: Hardware addressing and probe wake-up. The main control unit (104) retrieves the hardware mapping table pre-stamped in the non-volatile memory, determines the target physical port and the corresponding MAC address, writes the static address forwarding table entry to the multi-port PoE switching unit (106) through the SPI management channel, constructs and sends the Magic Packet wake-up frame through the RMII data channel, and the multi-port PoE switching unit (106) performs Layer 2 hardware forwarding to the target port according to the static address forwarding table to wake up the corresponding camera. Step 403: Multimodal data synchronous acquisition. The awakened camera completes the preset position scan and transmits back the visible light and thermal infrared dual-channel video streams. The AI ​​inference unit outputs the smoke confidence C and the temperature rise slope T. The gas sensor synchronously outputs the hydrogen concentration and CO concentration. The value of C is in the range of [0, 1], and the unit of T is ℃ / min. Step 404: Determine the weight of normal operating conditions according to the normalization formula. Calculate the overall score, where W1, W2, and W3 are all conventional weighting coefficients, satisfying the following conditions: Under normal operating conditions High-risk working conditions T_sat is the upper limit of temperature rise saturation, in °C / min; H2_base and CO_base are the normalized bases for hydrogen and carbon monoxide, respectively. It is a comprehensive fire situation score, with a value range of [0, +∞); Step 405: Determine whether the concentration of high-risk gas exceeds the dynamic weight trigger threshold, and determine whether the current hydrogen concentration exceeds the dynamic weight trigger threshold. If H2 is less than or equal to the dynamic weight trigger threshold, the system maintains the normal operating weight (0.3, 0.3, 0.4) and calculates the normal comprehensive score according to the normalization formula described in step 404, and then proceeds to step 408 to perform the normal fire threshold determination; if H2 is greater than the dynamic weight trigger threshold, the system immediately switches to the high-risk operating dynamic weight (0.3, 0.1, 0.6), recalculates the dynamic comprehensive score according to the normalization formula described in step 404, and then proceeds to step 406 to further determine whether the hydrogen-specific alarm threshold is exceeded. Step 406: Further determine whether the hydrogen-specific alarm threshold has been exceeded? If yes, proceed to step 407 to perform a lithium battery-specific mandatory alarm determination; if no, proceed to step 408 to perform a regular fire threshold determination. Step 407: A lithium battery-specific mandatory alarm is triggered, directly identifying it as a lithium battery thermal runaway fire, and proceeding to step 409 to execute coordinated response. Step 408: Standard fire threshold determination. Compare the comprehensive score calculated in step 404 according to the current effective weight with the preset fire judgment threshold. If the comprehensive score is greater than or equal to the preset fire judgment threshold, it is determined to be a real fire, and proceed to step 409 to execute the joint response. If the overall score is less than the preset fire judgment threshold, return to step 401 to continue monitoring; Step 409: The system activates the audible and visual alarms and fire extinguishing devices. Step 410: Upload the location information to the host computer.

5. The fire detection edge node and multi-probe linkage control method as described in claim 4, characterized in that, The preset warning threshold range is 0.001~0.8%obs / m.

6. The fire detection edge node and multi-probe linkage control method as described in claim 4, characterized in that, The center value of the dynamic weight trigger threshold is 50ppm, and a hysteresis anti-jitter range of ±2ppm is configured, that is, the rising edge trigger point is 52ppm and the falling edge backoff point is 48ppm; the hydrogen-specific alarm threshold is 100ppm.

7. The fire detection edge node and multi-probe linkage control method as described in claim 4, characterized in that, The normalized baseline for hydrogen is 160 ppm, and the normalized baseline for carbon monoxide is 200 ppm.

Images