A smart alarm and fire extinguishing device based on AI visual image analysis for fire location

Abstract

This invention discloses an intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location, belonging to the field of AI image recognition. The device includes a perception layer, a decision layer, and an execution layer. The perception layer includes an AI smoke and fire recognition camera; the decision layer includes an edge computing unit and a fire controller; the edge computing unit is used to fuse and analyze multi-source data collected by the perception layer, fuse temperature field data to construct a real-time fire thermal imaging map, identify the location of lithium battery thermal runaway flames, and complete fire confirmation; the fire controller is used to realize active fire reporting, power-off control of charging power supply, and start / stop control of fire pump group; the execution layer includes zoned sprinkler heads, electric control valves, fire pump group, audible and visual alarm devices, and a remote communication module; the execution layer adopts a zoned control strategy. This invention can significantly improve the efficiency of extinguishing lithium battery fires and reduce fire losses.

Description

Technical Field

[0001] This invention relates to the field of AI image recognition, and in particular to an intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location. Background Technology

[0002] With the continuous growth of electric bicycle ownership, the number of electric bicycles in China exceeded 350 million by 2025, with an annual growth rate of 20%. Fire safety issues at centralized parking and charging sites for electric bicycles are becoming increasingly prominent. Lithium battery fires are characterized by rapid fire spread, high risk of explosion, high reignition rate, and toxic release, classifying them as Class A, the highest fire protection level, which traditional firefighting techniques struggle to address effectively.

[0003] Currently, fire protection systems for electric bicycle parking sheds mainly employ traditional detection methods such as smoke detectors, heat detectors, and infrared / ultraviolet flame detectors, combined with fixed sprinkler systems to extinguish fires. While some existing systems have incorporated AI cameras for fire identification, they generally suffer from insufficient fire location accuracy, limited fire suppression targeting, and slow response times, making them ineffective in addressing the unique characteristics of lithium battery fires.

[0004] The defects and shortcomings of the existing technology are as follows:

[0005] Low fire location accuracy: Traditional detection methods can only determine the zone where the fire is located, but cannot accurately locate the specific ignition point. This leads to the large-scale spraying of extinguishing agents, causing unnecessary water damage and agent waste, and it is difficult to target the burning battery for cooling and extinguishing.

[0006] Insufficient response speed: Existing systems typically require more than 15 seconds to respond from detecting to confirming a fire, while the thermal runaway of lithium batteries spreads extremely quickly, making it very easy to miss the best opportunity to extinguish the fire.

[0007] The fire suppression system lacks specificity: most existing sprinkler systems use fixed-angle nozzles to spray evenly, which cannot adjust the spray angle and agent flow rate according to the location of the fire. This results in limited cooling effect on lithium battery fires and makes them prone to reignition.

[0008] Insufficient fusion of multi-source data: The existing system's detection devices operate independently, lacking fusion analysis of multi-source data such as images, temperature, and smoke, resulting in a high false alarm rate and an inability to dynamically assess the development of fires.

[0009] Lack of dynamic feedback on fire extinguishing effectiveness: Once the existing system is started, it sprays the agent according to the preset program, and cannot adjust the fire extinguishing strategy according to the real-time changes in the fire situation, which may lead to insufficient fire extinguishing or excessive use of agents.

[0010] Therefore, there is an urgent need for a low-cost, easy-to-implement, and adaptable lithium battery fire combustion model that can be combined with intelligent decision-making algorithms to achieve precise fire suppression control. Summary of the Invention

[0011] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes an intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location. It achieves precise location of the fire point through AI image recognition, and combined with an adjustable-angle intelligent sprinkler system, achieves accurate fire extinguishing, significantly improving the efficiency of fighting lithium battery fires and reducing fire losses.

[0012] This invention provides an intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location, comprising: a perception layer, a decision layer, and an execution layer. The perception layer includes an AI smoke and fire recognition camera, used to acquire visible light and infrared thermal images of the charging shed, and generate temperature field features and smoke features through AI image analysis. The decision layer includes an edge computing unit and a fire controller. The edge computing unit incorporates an AI smoke and fire recognition algorithm to fuse and analyze the dual-light images and generated features, constructing a real-time fire thermal imaging map, identifying the location of the lithium battery thermal runaway flame, and confirming the fire. The fire controller is used to actively report fires, control the power supply to the charging station, control the start and stop of the fire pump, and control audible and visual alarms. The execution layer includes zoned sprinkler heads, electrically controlled valves, fire pumps, audible and visual alarm devices, and a remote communication module. The sprinkler heads are compatible with lithium battery-specific fire extinguishing inhibitors, which can reduce the temperature of the ignition point to below 40°C within 20 seconds. The execution layer adopts a zoned control strategy, extinguishing the fire only by activating the sprinkler heads and electrically controlled valves in the zone corresponding to the fire.

[0013] According to some embodiments of the present invention, the sensing layer is configured with environmental sensors for collecting ventilation parameters and ambient temperature and humidity data of the charging shed.

[0014] According to some embodiments of the present invention, the AI ​​smoke and fire recognition algorithm built into the edge computing unit is based on a convolutional neural network (CNN), constructed using visible light and infrared thermal imaging dual-light images as input, and sequentially includes a feature extraction layer, a target detection layer, and a temporal classification layer; the feature extraction layer adopts a ResNet-18 network, and performs transfer learning based on a dataset of early-stage pale white smoke, weak flames, and thermal imaging temperature field scenes from lithium battery thermal runaway in a charging shed, simultaneously extracting visible light smoke and fire features and infrared temperature field features, and suppressing interference from strong light, shadows, and reflective noise; The target detection layer employs an improved YOLOv8 algorithm, integrates dual-light features, adds a branch for detecting small targets with occlusion, and optimizes the anchor frame size to fit the narrow space of the charging shed, thus identifying early fire areas obscured by electric vehicles or pillars. The temporal classification layer uses an improved LSTM long short-term memory network, adds temperature rise time series and smoke diffusion time series input channels, extracts temperature rise time series features, smoke diffusion time series features, and visible light smoke and fire time series features, and distinguishes between lithium battery thermal runaway white smoke and dust, water vapor, and light flickering interference sources through the dynamic change features of smoke and fire.

[0015] According to some embodiments of the present invention, the AI ​​smoke and fire recognition algorithm further includes a fusion verification layer, which jointly verifies the temperature rise time series features, smoke diffusion time series features, and visible light smoke and fire time series features. Only when each of the three types of time series features meets a preset threshold is it determined to be a real fire.

[0016] According to some embodiments of the present invention, the zoning control strategy is as follows: the charging shed is divided into several independent fire extinguishing zones, each zone is equipped with an independent sprinkler head group and an electric control valve, and fire extinguishing operations are only initiated in the zone where the fire occurs.

[0017] According to some embodiments of the present invention, the remote communication module adopts 4G / 5G communication mode to report fire information, equipment operating status and fire extinguishing progress data to the superior management platform in real time.

[0018] According to some embodiments of the present invention, the device automatically terminates the fire extinguishing process after the fire is completely extinguished, continues to monitor, and restarts the fire extinguishing device again if the fire reignites, until there is no risk of reignition.

[0019] According to some embodiments of the present invention, the fire controller is linked to the power supply circuit of the charging shed. After a fire is confirmed, the charging power supply is immediately cut off, and the fire pump group and the audible and visual alarm device are started simultaneously.

[0020] According to some embodiments of the present invention, the edge computing unit of the decision layer only loads and activates the dynamic fire extinguishing decision optimization algorithm when the fire is confirmed and enters the automatic fire extinguishing execution stage. The algorithm, based on publicly available industry standards and measured data from lithium battery fires, constructs a fire combustion model that conforms to the thermal runaway characteristics of lithium batteries. Using temperature field characteristics, smoke characteristics, fire data, and ventilation parameters as inputs, and combining the fire location, scale, and development trend, it dynamically adjusts the number of sprinkler heads opened, the spray pressure, and the agent ratio. The dynamic fire extinguishing decision optimization algorithm employs a gradient descent-based multi-objective linear optimization algorithm to solve for the optimal combination of the number of sprinkler heads opened, the spray pressure, and the agent ratio. The optimization formula and iterative formula are as follows: Let the input vector X = [T, S, A, V, K], where T is the temperature field characteristic, S is the smoke characteristic, A is the fire scale, V is the ventilation rate, and K is the scene correction coefficient; let the optimization variable vector U = [N, P, R], where N is the number of sprinkler heads opened, P is the spray pressure, and R is the agent ratio; the optimization objective function is: The gradient descent iterative formula is: Where F(U,X) is the optimization objective function under the action of input vector X, and η(X), γ(X), and C(X) are the extinguishing efficiency, reignition rate, and agent consumption per unit time corresponding to input vector X, respectively. U is the weight coefficient, and U is the control parameter vector for the k-th iteration. k+1 Let F(U) be the control parameter vector for the (k+1)th iteration, α be the iteration step size, and ▽F(U) k Let X be the objective function in U. k The gradient under the action of X.

[0021] The embodiments of this invention achieve at least the following beneficial effects: By achieving centimeter-level positioning of the ignition point, combined with intelligently openable sprinkler heads, precise spraying is achieved, increasing agent utilization by more than 70%, and significantly reducing water damage and agent consumption. Fast response speed effectively curbs fire spread. The system response time is ≤2 seconds, more than 7 times faster than traditional systems, allowing for timely intervention in the early stages of lithium battery thermal runaway, extinguishing the fire in its nascent stage and significantly reducing fire losses. High fire extinguishing efficiency and low risk of reignition. Using a dedicated lithium battery fire suppressant, combined with precise spraying technology, it can directly and deeply cool the burning battery. The highest temperature on site does not exceed 50°C within 24 hours after extinguishing the fire, fundamentally eliminating the possibility of reignition. Low false alarm rate and high system reliability. The multi-source data fusion confirmation mechanism reduces the false alarm rate to below 0.1%, avoiding unnecessary losses caused by accidental spraying. High degree of intelligence and low operation and maintenance costs. The system has functions such as automatic inspection, fault self-diagnosis, and remote monitoring, which can significantly reduce manual operation and maintenance costs.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0024] Figure 1 This is a schematic block diagram of the device according to an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the four-level fire response process according to an embodiment of the present invention.

[0026] Figure label:

[0027] Perception layer 100, decision-making layer 200, execution layer 300. Detailed Implementation

[0028] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0029] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0030] Reference Figure 1This invention proposes an intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location, comprising: a perception layer, a decision layer, and an execution layer; the perception layer includes an AI smoke and fire recognition camera, used to collect visible light images and infrared thermal imaging dual-light images inside the charging shed, and generate temperature field features and smoke features through AI image analysis; the decision layer includes an edge computing unit and a fire controller; the edge computing unit has a built-in AI smoke and fire recognition algorithm, which fuses and analyzes the dual-light images and generated features to construct a real-time fire thermal imaging map, identify the location of the lithium battery thermal runaway flame, and complete fire confirmation; the fire controller is used to realize active fire reporting, power-off control of charging power supply, start and stop control of fire pump group, and audible and visual alarm control; the execution layer includes zoned sprinkler heads, electric control valves, fire pump groups, audible and visual alarm devices, and a remote communication module; the sprinkler heads are adapted to lithium battery-specific fire extinguishing inhibitors, which can reduce the temperature of the ignition point to below 40°C within 20 seconds; the execution layer adopts a zoned control strategy, only activating the sprinkler heads and electric control valves of the zone corresponding to the fire to extinguish the fire.

[0031] In some embodiments, the sensing layer is configured with environmental sensors to collect ventilation parameters and ambient temperature and humidity data of the charging shed.

[0032] In some embodiments, the AI ​​smoke and fire recognition algorithm built into the edge computing unit is based on a convolutional neural network (CNN), constructed using visible light and infrared thermal imaging dual-light images as input. It sequentially includes a feature extraction layer, a target detection layer, and a temporal classification layer. The feature extraction layer uses a ResNet-18 network and performs transfer learning based on a dataset of early-stage pale white smoke, weak flames, and thermal imaging temperature field scenes from lithium battery thermal runaway in charging sheds. It simultaneously extracts visible light smoke and fire features and infrared temperature field features, while suppressing interference from strong light, shadows, and reflective noise. The target detection layer uses an improved YOLOv8 algorithm, fusing dual-light features, adding a branch for detecting small, obstructed targets, and optimizing the anchor frame size to fit the narrow space of the charging shed. This is used to identify early fire areas obscured by electric vehicles or pillars. The temporal classification layer uses an improved LSTM long short-term memory network, adding temperature rise time series and smoke diffusion time series input channels. It extracts temperature rise time series features, smoke diffusion time series features, and visible light smoke and fire time series features, and distinguishes between lithium battery thermal runaway white smoke and interference sources such as dust, water vapor, and flashing lights through dynamic smoke and fire characteristics.

[0033] In some embodiments, the AI ​​smoke and fire recognition algorithm further includes a fusion verification layer, which jointly verifies the temperature rise time series features, smoke diffusion time series features, and visible light smoke and fire time series features. Only when each of the three types of time series features meets a preset threshold is it determined to be a real fire.

[0034] This embodiment, based on the general CNN architecture for fireworks detection, performs layered customized optimizations to address the specific pain points of the charging shed scenario. The algorithm takes visible light and infrared thermal imaging dual-light images as input, and the layered feature extraction and temporal verification logic is as follows:

[0035] The bottom feature extraction layer adopts a lightweight ResNet-18 network and completes transfer learning through a dedicated scene dataset of charging carports. It extracts static features of temperature field from infrared images and static features of smoke and smoke from visible light images. It focuses on enhancing the fine-grained feature extraction capabilities of early light white smoke from lithium batteries and weak flames that are blocked, while filtering out environmental noise interference such as strong light, shadows, and roof reflections.

[0036] Mid-level target detection layer: Based on the improved YOLOv8 algorithm, it integrates dual-light image features, optimizes the anchor frame size for narrow carports, and adds a small target detection branch to specifically identify small early fire areas obscured by electric vehicles and pillars, thus solving the problem of missed detection in traditional single-light detection.

[0037] Upper-level temporal classification layer: By improving the LSTM long short-term memory network, dedicated input channels for temperature rise time series and smoke diffusion time series are added to dynamically capture the rate of change of fire temperature rise, the movement law of smoke diffusion, and the frequency of flame flickering. The temperature rise time series features, smoke diffusion time series features, and visible light smoke and fire time series features are extracted to accurately distinguish the white smoke from lithium battery thermal runaway from interference sources such as dust, water vapor, light flickering, and static obstruction from the dynamic time series dimension.

[0038] The final fusion verification layer performs joint verification on the three types of time-series features output by the LSTM. Only when the temperature rise time-series feature, smoke diffusion time-series feature, and visible light smoke and fire time-series feature simultaneously meet the preset threshold can it be determined as a real lithium battery fire, effectively eliminating the problem of false alarms due to single features.

[0039] The training dataset used in this algorithm is a dedicated smoke and fire dataset for charging carports. The samples cover scenarios such as early-stage light white smoke from lithium batteries, small open flames, strong backlight, roof shadows, electric vehicle obstruction, dust, water vapor, headlight reflections, and personnel movement. The sample annotation strictly distinguishes between real fires and environmental interference, and focuses on strengthening the proportion of early weak fire samples for transfer learning of the ResNet-18 model.

[0040] In some embodiments, the zoning control strategy is as follows: the charging shed is divided into several independent fire extinguishing zones, each zone is equipped with an independent sprinkler head group and an electric control valve, and fire extinguishing operations are only initiated in the zone where the fire occurs.

[0041] In some embodiments, the remote communication module uses 4G / 5G communication to report fire information, equipment operating status, and fire extinguishing progress data to the upper-level management platform in real time.

[0042] In some embodiments, the device automatically terminates the fire extinguishing process after the fire is completely extinguished, continues to monitor, and restarts the fire extinguishing device if the fire reignites, until there is no risk of reignition.

[0043] In some embodiments, the fire controller is linked to the power supply circuit of the charging shed. Upon confirmation of a fire, the charging power supply is immediately cut off, and the fire pump set and the audible and visual alarm device are started simultaneously.

[0044] In some embodiments, the edge computing unit of the decision layer only loads and activates the dynamic fire suppression decision optimization algorithm when the fire is confirmed and enters the automatic fire suppression execution stage. The algorithm relies on publicly available industry standards and measured data from lithium battery fires to construct a fire combustion model that conforms to the thermal runaway characteristics of lithium batteries. It takes temperature field characteristics, smoke characteristics, fire data, and ventilation parameters as inputs, and dynamically adjusts the number of sprinkler heads opened, the spray pressure, and the agent ratio based on the fire location, scale, and development trend. The dynamic fire suppression decision optimization algorithm employs a gradient descent-based multi-objective linear optimization algorithm to solve for the optimal combination of the number of sprinkler heads opened, the spray pressure, and the agent ratio. The optimization formula and iterative formula are as follows: Let the input vector X = [T, S, A, V, K], where T is the temperature field characteristic, S is the smoke characteristic, A is the fire scale, V is the ventilation rate, and K is the scene correction coefficient; the optimization variable vector U = [N, P, R], where N is the number of sprinkler heads opened, P is the spray pressure, and R is the agent ratio; the optimization objective function is: The gradient descent iterative formula is: Where F(U,X) is the optimization objective function under the action of input vector X, and η(X), γ(X), and C(X) are the extinguishing efficiency, reignition rate, and agent consumption per unit time corresponding to input vector X, respectively. U is the weight coefficient, and U is the control parameter vector for the k-th iteration. k+1 Let F(U) be the control parameter vector for the (k+1)th iteration, α be the iteration step size, and ▽F(U) k Let X be the objective function in U. k The gradient under the action of X.

[0045] The existing technology foundation for constructing the lithium battery fire combustion model in this embodiment is derived from publicly available industry-standard conclusions and authoritative materials, specifically including:

[0046] (1) The lithium battery thermal runaway triggering conditions (heating to 130℃ at 5℃ / min and holding for 30 minutes) and the safety threshold after thermal runaway as specified in the national standard GB / T 31484-2015 "Safety Requirements for Power Batteries for Electric Vehicles";

[0047] (2) The thermal runaway warning temperature threshold (85℃) and the requirement that thermal runaway of a single cell does not spread to the module as specified in the national standard GB 43854-2024 "Safety Technical Specification for Lithium-ion Batteries for Electric Bicycles";

[0048] (3) The measured data of heat release rate (HRR) and flame temperature of lithium iron phosphate and ternary lithium batteries published in the August 2024 issue of "Research on Thermal Runaway of Lithium-ion Batteries of Electric Bicycles in Semi-enclosed Spaces" in Fire Science and Technology;

[0049] (4) The average time from thermal runaway to open flame (4.2 minutes) and the fire spread rate (0.3-0.8 m / min) disclosed in the "White Paper on Fire Prevention and Control of Electric Bicycles" published by the China Fire and Rescue Academy.

[0050] The construction process of the lithium battery fire combustion model in this embodiment is as follows: This embodiment adopts a lightweight engineering approach of "graded threshold division + empirical parameter fitting + scenario correction" to construct a simplified lithium battery fire combustion model. The specific steps are as follows:

[0051] Step 1: Classify the fire level. Based on the three-stage theory of lithium battery thermal runaway (diaphragm decomposition, electrolyte decomposition, and total thermal runaway) in publicly available information, and combined with the actual evolution of charging shed fires, lithium battery fires are classified into four quantifiable levels: Level I (early warning), Level II (initial stage), Level III (development), and Level IV (spread). Complex combustion dynamics calculations are abandoned, and the level is determined solely through core characteristic parameters.

[0052] Step 2: Set the fire severity threshold. Based on the publicly available data, fit the core characteristic parameter thresholds corresponding to each fire severity level, including battery surface temperature, temperature rise rate, smoke concentration, flame area, and heat release rate (HRR), to form a threshold comparison table. If any parameter meets the corresponding level threshold, the fire can be determined as a fire of that level. The specific thresholds are as follows:

[0053] Level I (Warning): Battery surface temperature > 85℃, or temperature rise rate > 2℃ / min, or smoke concentration > 0.1mg / L;

[0054] Level II (Initial): Flame area <0.1m², or temperature <300℃, or HRR <50kW;

[0055] Level III (Development): Flame area 0.1-0.5m², or temperature 300-600℃, or HRR 50-200kW;

[0056] Level IV (Spread): Flame area > 0.5 m², or temperature > 600 °C, or HRR > 200 kW.

[0057] Step 3: Scene Adaptation and Correction. Based on the space dimensions and ventilation conditions of the charging shed, a scene correction coefficient K is set to fine-tune the fire suppression control parameters and adapt to the fire prevention and control needs of different shed scenarios. The specific correction coefficients are as follows:

[0058] Small enclosed carport (<20㎡, no natural ventilation): K=1.2;

[0059] Medium-sized semi-enclosed carport (20-50㎡, single-sided ventilation): K=1.0;

[0060] Large open carport (>50㎡, double-sided ventilation): K=0.8.

[0061] Step 4: Model Encapsulation. Integrate the above-mentioned level classification, threshold lookup table, and scene correction coefficients into a simple model consisting of a parameter range lookup table and linear interpolation. No complex code or simulation tools are required. It can be directly embedded into a PLC or microcontroller for rapid calling and calculation, resulting in extremely low engineering implementation costs.

[0062] In some embodiments, the system of this invention embeds an intelligent fire extinguishing decision optimization system based on a lithium battery fire combustion model, including:

[0063] (1) Data acquisition: Real-time data such as battery surface temperature, smoke concentration, flame area, and temperature rise rate are collected through image acquisition equipment in the charging shed. At the same time, scene parameters such as shed space size and ventilation conditions are collected and transmitted to the model.

[0064] (2) Level determination: The model receives the collected multi-source data, compares it with the preset level threshold comparison table, automatically determines the current fire level, and determines the correction coefficient K based on the scene parameters;

[0065] (3) Decision output: The model sends the fire level and correction coefficient K to the dynamic fire extinguishing decision optimization algorithm. Based on this input, the algorithm outputs the optimal fire extinguishing control parameters, including the number of sprinkler heads opened, spray pressure, and agent ratio, in combination with the objectives of "highest fire extinguishing efficiency, lowest reignition rate, and least agent consumption".

[0066] (4) Iterative optimization: The actuator starts the sprinkler system according to the control parameters, and at the same time collects fire extinguishing feedback data (temperature and smoke concentration changes) in real time, transmits it back to the model, and fine-tunes the thresholds and correction coefficients of each level to realize the dynamic iterative optimization of the model.

[0067] In some embodiments, the system of this invention combines the fire level L(X) and development trend ΔL(X) output by the fire combustion model, and the constraint formula and iteration step size adjustment formula for the optimization variable U are as follows: When L(X) = Level I, N∈[1,1], P∈[0.8,1.0], R∈[100:0,100:0]; when L(X) = Level II, N∈[2,3], P∈[1.0,1.2], R∈[95:5,95:5]; when L(X) = Level III, N∈[4,6], P∈[1.2,1.5], R∈[90:10,90:10]; when L(X) = Level IV, N∈[Nmax,Nmax], P∈[1.5,2.0], R∈[85:15,85:15]; the iteration step size adjustment formula is: Where β is the step size adjustment coefficient, with a value range of 0.1~0.3, which can be flexibly adjusted according to the fire prevention and control accuracy requirements of the charging shed; Nmax is the total number of sprinkler heads in the charging shed; ΔL(X) is the fire development trend corresponding to the input vector X, with a value of +1 indicating escalation, 0 indicating maintenance, and -1 indicating de-escalation.

[0068] In some embodiments, the core formula for dynamic optimization is as follows: The iterative solution formula is: The scene correction formula is: Where U is the optimal control parameter vector obtained through iteration, Ufinal is the final control parameter issued, and K is the scene correction coefficient in the input vector X. The iteration is performed until F(U,X) reaches its minimum value.

[0069] In some embodiments, the present invention optimizes the fire extinguishing decision-making algorithm based on on-site operational feedback data, including real-time temperature changes ΔT, smoke concentration changes ΔS, flame area changes ΔA, total agent consumption Ctotal, and monitoring data on reignition after fire extinguishing. The iterative correction formula is as follows: The deviation calculation formula is: The threshold judgment condition is: , The set value is used; if it is not met, the weighting coefficient is adjusted. And the initial value of α, and substitute it back into step S200 to solve.

[0070] Reference Figure 2 In a specific embodiment, fire handling is divided into four levels: Level II potential risk, Level II suspected fire, Level II confirmed fire, and Level IV fire extinguishing execution. The system assigns the trigger judgment, response action, and execution command of each level to the data collection of the perception layer, the hierarchical analysis of the decision-making layer, and the hierarchical linkage of the execution layer, respectively.

[0071] When a potential risk is identified as Level II (battery abnormality, wire softening): The perception layer is triggered by capturing visual anomalies such as battery bulging, wire deformation, and insulation aging through an AI smoke and fire recognition camera, and extracting weak anomaly features of early non-open flames based on a customized ResNet-18 network; The decision layer responds by outputting response action instructions when the edge computing unit determines that there is a potential risk of thermal runaway, without triggering an alarm, only marking the abnormal location; The execution layer does not activate sprinklers or alarms, but only continuously monitors the temperature and image data of the abnormal location, and synchronously links with the charging circuit monitoring to provide a data basis for subsequent fire prediction.

[0072] When a suspected fire of level II (trace smoke, small flame) is identified: the perception layer is triggered: the improved YOLOv8 small target detection algorithm identifies trace white smoke and small flames in the early stage of lithium battery fire, and the LSTM time series network distinguishes smoke from dust and water vapor interference; the decision layer responds: if a suspected fire is identified, a preliminary assessment is initiated, but no remote reporting is performed; the execution layer takes action: a local alarm is triggered, the audible and visual alarm device is activated to remind on-site personnel, the sprinkler system is kept off, and the fire situation is continuously monitored.

[0073] When a Level II confirmed fire is identified (image + sensor verification): The perception layer is triggered to perform multi-source fusion verification of the image and sensor (AI smoke and fire recognition results), eliminating interference from lights, reflections, dust, etc., and confirming the authenticity of the fire. The decision layer responds by determining that the fire controller is a real fire, executing a remote alarm, and proactively reporting the fire location, level, and time to the superior management platform, while simultaneously issuing an instruction to cut off the charging power. The execution layer takes action by pushing alarm information through the remote communication module, cutting off power to the charging circuit, and putting the fire pump unit into standby mode.

[0074] When the fire is classified as Level 4 (fire confirmation → manual intervention / automatic response → sprinkler activation): This level has two response paths to adapt to the on-site management needs of the charging shed. Path 1: Manual intervention: The system pushes full fire information to the upper-level platform. After confirming the fire through remote monitoring or on-site inspection, the management personnel manually trigger the fire extinguishing command and start the sprinkler system. Path 2: Automatic response: Decision layer: The edge computing unit calls the lightweight lithium battery fire combustion model + dynamic fire extinguishing decision optimization algorithm, and dynamically calculates the number of sprinkler heads to be opened, the spray pressure, and the agent ratio based on the fire location, scale, and development trend. Execution layer: A zoned control strategy is adopted, only opening the sprinkler heads and electric control valves in the zone where the fire is located, starting the fire pump group, and spraying special fire extinguishing inhibitors to reduce the temperature of the ignition point to below 40°C within 20 seconds. Closed-loop management: The entire fire extinguishing process is monitored and reported in real time. After the fire is extinguished, continuous monitoring is carried out for 24 hours. The process ends after confirming that there is no risk of reignition.

[0075] This invention combines a customized AI recognition algorithm for charging sheds with a dynamic fire extinguishing optimization algorithm to solve the shortcomings of traditional fire protection systems, such as missed early fire reports, false alarms, fixed fire extinguishing parameters, and inability to adapt to the thermal runaway characteristics of lithium batteries.

[0076] In a specific embodiment, taking a standard electric vehicle charging shed that can accommodate 20 charging spaces as an example, the system deployment scheme of this embodiment of the invention is as follows:

[0077] Install one AI smoke and fire recognition camera to cover the entire charging area and achieve comprehensive monitoring.

[0078] The charging shed is divided into two zones, each equipped with four fine water mist nozzles, controlled by independent solenoid valves.

[0079] Equipped with a 200L capacity fire extinguishing agent storage tank, filled with water-based lithium battery-specific inhibitors, it can meet the simultaneous fire extinguishing needs of two fire zones.

[0080] Deploy one cabinet-mounted fire controller to achieve local logic control and equipment linkage, and support 4G communication access to the cloud platform.

[0081] During operation, the system runs continuously 24 / 7, collecting and analyzing various monitoring data in real time. When a suspected fire is detected, a multi-source data confirmation process is initiated. Once the fire is confirmed, the fire extinguishing procedure is immediately executed: Within 1 second, the coordinates of the ignition point are calculated with a positioning accuracy of ±10cm; within 2 seconds, an audible and visual alarm is activated, and alarm information is pushed to the monitoring center and management personnel's mobile phones; within 3 seconds, the corresponding area's sprinkler heads are activated, electric valves and pump units are opened, and fire extinguishing inhibitors are sprayed. The system continuously monitors changes in the fire situation and dynamically adjusts spray parameters until the fire is extinguished. After the fire is extinguished, the fire extinguishing process automatically ends, and continuous monitoring continues. If reignition occurs, the fire extinguishing device will be activated again until there is no risk of reignition.

[0082] During the maintenance phase, the system automatically performs daily inspections to check the equipment's operating status and agent levels; conducts a simulated fire extinguishing test every 6 months; and performs a comprehensive calibration and maintenance annually.

[0083] The method applied to embodiments of the present invention includes the following steps:

[0084] S1, the perception layer collects multi-source data such as images, temperature and smoke in the charging shed in real time through an AI smoke and fire recognition camera;

[0085] S2, the edge computing unit of the decision layer integrates multi-source data and temperature field information to construct a real-time fire thermal imaging map, locate the ignition point of lithium battery thermal runaway and complete fire confirmation.

[0086] S3. The fire controller actively reports the fire to the upper-level platform, simultaneously cuts off the charging power, starts the fire pump set, and triggers the audible and visual alarms.

[0087] S4. The execution layer, based on the zoning control strategy, opens the sprinkler heads and electric control valves of the corresponding fire zone to spray lithium battery-specific fire extinguishing inhibitors for precise fire extinguishing; when the automatic fire extinguishing mode is adopted, the edge computing unit starts the dynamic fire extinguishing decision optimization algorithm to dynamically adjust the sprinkler parameters.

[0088] S5. The entire fire extinguishing process is monitored and reported in real time. After the fire is extinguished, the fire extinguishing process is automatically terminated. The system continues to monitor the fire and will restart the fire extinguishing device if the fire reignites, until there is no risk of reignition.

[0089] Although specific embodiments are described herein, those skilled in the art will recognize that many other modifications or alternative embodiments are also within the scope of this disclosure. For example, any of the functions and / or processing capabilities described in connection with a particular device or component can be performed by any other device or component. Furthermore, while various exemplary embodiments and architectures have been described according to embodiments of this disclosure, those skilled in the art will recognize that many other modifications to the exemplary embodiments and architectures described herein are also within the scope of this disclosure.

[0090] The foregoing description, with reference to block diagrams and flowcharts of systems, methods, systems, and / or computer program products according to exemplary embodiments, has described certain aspects of this disclosure. It should be understood that one or more blocks in the block diagrams and flowcharts, as well as combinations of blocks in the block diagrams and flowcharts, can be implemented by executing computer-executable program instructions, respectively. Similarly, according to some embodiments, some blocks in the block diagrams and flowcharts may not need to be executed in the order shown, or may not all need to be executed. Furthermore, additional components and / or operations beyond those shown in the blocks in the block diagrams and flowcharts may exist in some embodiments.

[0091] Therefore, blocks in block diagrams and flowcharts support combinations of means for performing a specified function, combinations of elements or steps for performing a specified function, and program instruction means for performing a specified function. It should also be understood that each block in a block diagram and flowchart, and combinations of blocks in block diagrams and flowcharts, can be implemented by a dedicated hardware computer system or a combination of dedicated hardware and computer instructions that performs a specific function, element, or step.

[0092] The program modules, applications, etc., described herein may include one or more software components, including, for example, software objects, methods, data structures, etc. Each such software component may include computer-executable instructions that, in response to execution, cause at least a portion of the functionality described herein (e.g., one or more operations of the exemplary methods described herein) to be performed.

[0093] Software components can be coded using any of a variety of programming languages. An exemplary programming language could be a low-level programming language, such as assembly language associated with a specific hardware architecture and / or operating system platform. Software components including assembly language instructions may need to be converted into executable machine code by an assembler before being executed by the hardware architecture and / or platform. Another exemplary programming language could be a higher-level programming language that is portable across multiple architectures. Software components including higher-level programming languages ​​may need to be converted into an intermediate representation by an interpreter or compiler before execution. Other examples of programming languages ​​include, but are not limited to, macro languages, shell or command languages, job control languages, scripting languages, database query or search languages, or report writing languages. In one or more exemplary embodiments, a software component containing instructions from one of the above-described programming language examples can be executed directly by the operating system or other software components without first being converted into another form.

[0094] Software components can be stored as files or other data storage structures. Software components of similar type or related function can be stored together in a specific directory, folder, or library. Software components can be static (e.g., pre-defined or fixed) or dynamic (e.g., created or modified at runtime).

[0095] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location, characterized in that, include: The layers are: perception layer, decision-making layer, and execution layer. The perception layer includes an AI smoke and fire recognition camera, which is used to collect visible light images and infrared thermal imaging dual-light images inside the charging shed, and generate temperature field features and smoke features through AI image analysis. The decision-making layer includes an edge computing unit and a fire controller; the edge computing unit has a built-in AI smoke and fire recognition algorithm, which fuses and analyzes dual-light images and generated features to construct a real-time fire thermal imaging map, identify the location of lithium battery thermal runaway flames and complete fire confirmation; the fire controller is used to realize active fire reporting, power outage control of charging power supply, start and stop control of fire pump group and audible and visual alarm control. The execution layer includes zoned sprinkler heads, electrically controlled valves, fire pump sets, audible and visual alarm devices, and a remote communication module; the sprinkler heads are equipped with lithium battery-specific fire extinguishing inhibitors, which can reduce the temperature of the ignition point to below 40°C within 20 seconds; the execution layer adopts a zoned control strategy, extinguishing the fire by only activating the sprinkler heads and electrically controlled valves in the zone corresponding to the fire.

2. The intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location as described in claim 1, characterized in that, The sensing layer is equipped with environmental sensors to collect ventilation parameters and ambient temperature and humidity data of the charging shed.

3. The intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location as described in claim 1, characterized in that, The AI ​​fireworks recognition algorithm built into the edge computing unit is based on a convolutional neural network (CNN), constructed using visible light and infrared thermal imaging dual-light images as input, and includes a feature extraction layer, a target detection layer, and a temporal classification layer in sequence. The feature extraction layer uses a ResNet-18 network and performs transfer learning based on a dataset of early-stage light white smoke, weak flames, and thermal imaging temperature field scenes from lithium battery thermal runaway in a charging carport. It simultaneously extracts visible light smoke and infrared temperature field features and suppresses interference from strong light, shadows, and reflective noise. The target detection layer adopts the improved YOLOv8 algorithm, integrates dual-light features, adds a small target detection branch for occlusion, and optimizes the anchor frame size to adapt to the narrow space of the charging shed, so as to identify early fire areas that are blocked by electric vehicles or pillars. The time series classification layer adopts an improved LSTM long short-term memory network, adds temperature rise time series and smoke diffusion time series input channels, extracts temperature rise time series features, smoke diffusion time series features and visible light smoke and fire time series features, and distinguishes lithium battery thermal runaway white smoke from dust, water vapor and light flickering interference sources through the dynamic change features of smoke and fire.

4. The intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location as described in claim 3, characterized in that, The AI ​​smoke and fire recognition algorithm also includes a fusion verification layer, which jointly verifies the temperature rise time series features, smoke diffusion time series features, and visible light smoke and fire time series features. Only when each of the three types of time series features meets a preset threshold is it determined to be a real fire.

5. The intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location as described in claim 1, characterized in that, The zoning control strategy is as follows: the charging shed is divided into several independent fire extinguishing zones, each zone is equipped with an independent sprinkler head group and an electric control valve, and fire extinguishing operations are only initiated in the zone where the fire occurs.

6. The intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location as described in claim 1, characterized in that, The remote communication module uses 4G / 5G communication to report fire information, equipment operating status, and fire extinguishing progress data to the upper management platform in real time.

7. The intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location as described in claim 1, characterized in that, After the fire is completely extinguished, the device automatically ends the fire extinguishing process, continues to monitor, and if the fire reignites, it will restart the fire extinguishing device until there is no risk of reignition.

8. The intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location as described in claim 1, characterized in that, The fire controller is linked to the power supply circuit of the charging shed. Once a fire is confirmed, the charging power is immediately cut off, and the fire pump set and the audible and visual alarm device are started simultaneously.

9. The intelligent alarm and fire extinguishing device based on AI visual image analysis for fire location as described in claim 1, characterized in that, The edge computing unit of the decision layer only loads and starts the dynamic fire extinguishing decision optimization algorithm when the fire is confirmed and the automatic fire extinguishing execution phase is entered. The algorithm relies on publicly available industry standards and measured data of lithium battery fires to construct a fire combustion model that fits the thermal runaway characteristics of lithium batteries. It takes temperature field characteristics, smoke characteristics, fire data and ventilation parameters as inputs, and dynamically adjusts the number of sprinkler heads opened, the spray pressure and the agent ratio in combination with the fire location, scale and development trend. The dynamic fire extinguishing decision optimization algorithm employs a gradient descent-based multi-objective linear optimization algorithm to solve for the optimal combination of the number of sprinkler heads opened, the spray pressure, and the agent ratio. The optimization formula and iterative formula are as follows: Let the input vector X=[T,S,A,V,K], where T is the temperature field feature, S is the smoke feature, A is the fire scale, V is the ventilation rate, and K is the scene correction coefficient, and the optimization variable vector U=[N,P,R], where N is the number of sprinkler heads opened, P is the spray pressure, and R is the agent ratio; The objective function to be optimized is: ; The gradient descent iterative formula is: ; Where F(U,X) is the optimization objective function under the action of input vector X, and η(X), γ(X), and C(X) are the extinguishing efficiency, reignition rate, and agent consumption per unit time corresponding to input vector X, respectively. U is the weight coefficient, and U is the control parameter vector for the k-th iteration. k+1 Let F(U) be the control parameter vector for the (k+1)th iteration, α be the iteration step size, and ▽F(U) k Let X be the objective function in U. k The gradient under the action of X.

Images