A multi-phase system lithium battery special intelligent directional jet fire extinguishing method and system

Abstract

The application discloses a kind of multi-phase system lithium battery special intelligent directional jet fire extinguishing system and method, system is by mobile cabinet, liquid storage unit, pumping unit, jet water cannon, double-shaft holder, control cabinet and multi-source sensor module composition.Method is completed fire preliminary judgment and coarse positioning by solar blind ultraviolet sensor, combined with infrared thermal imaging, visible light vision and laser ranging data, and the three-dimensional space coordinates of fire source are solved by multi-source feature fusion;initial aiming angle is solved by built-in air resistance jet trajectory prediction model, and spraying posture is corrected in real time relying on visual landing point feedback and incremental PID control algorithm, and lithium battery thermal runaway reignition discrimination mechanism is matched to guarantee the completeness of fire extinguishing.The system can be independently transported and deployed, and is suitable for lithium battery protection scenes such as charging station and energy storage warehouse.The fire identification is reliable, the spraying precision is high, and the early disposal efficiency of lithium battery fire can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of lithium battery fire extinguishing technology, specifically to a smart directional jet fire extinguishing method and system for multiphase lithium batteries. Background Technology

[0002] The risk of thermal runaway fires is prominent in scenarios such as lithium battery energy storage sites and electric vehicle charging stations. These sites burn rapidly, are prone to reignition, and produce toxic fumes. Traditional fixed fire protection networks are insufficient to cover temporarily deployed or mobile fire protection scenarios. Existing mobile fire extinguishing devices generally use a single ultraviolet or infrared detection scheme, which is susceptible to interference from high-temperature power distribution cabinets, lighting fixtures, sunlight reflection, and welding arcs on site, resulting in high false alarm and false negative rates. Furthermore, they do not take into account actual operating conditions such as lateral wind disturbance and the viscosity of multiphase extinguishing agents, leading to large deviations in open-loop aiming. At the same time, most devices lack a specific assessment mechanism for reignition in cases of lithium battery thermal runaway, making secondary deflagration possible after the open flame is extinguished, resulting in low overall protection reliability. Summary of the Invention

[0003] The purpose of this application is to provide a smart directional jet fire extinguishing method and system specifically for multiphase lithium batteries, so as to improve the reliability of the fire extinguishing system.

[0004] In a first aspect, embodiments of this application provide a smart directional jet fire extinguishing method specifically for multiphase lithium batteries, comprising the following steps:

[0005] The S1 fire extinguishing device is powered on and initialized and enters inspection mode, continuously collecting ultraviolet radiation intensity signals of the detection area through ultraviolet sensors;

[0006] When the ultraviolet radiation intensity signal exceeds the set trigger threshold, S2 drives the pan-tilt unit to scan the detection area, simultaneously acquiring thermal imaging data output by the infrared thermal imager, distance data output by the laser rangefinder, and visual image data output by the high-definition camera. Based on the orientation of the ultraviolet radiation intensity signal, multi-source signal fusion processing is performed on the thermal imaging data and visual image data to determine the pixel coordinates of the flame area and calculate the three-dimensional spatial coordinates of the fire source by combining the distance data.

[0007] S3 inputs the three-dimensional spatial coordinates of the fire source as boundary conditions into the air resistance jet trajectory prediction model. The air resistance jet trajectory prediction model outputs the horizontal angle and pitch angle of the water cannon target. Based on the horizontal angle and pitch angle of the target, the gimbal is driven to rotate to achieve water cannon alignment.

[0008] S4 starts the pump unit to deliver fire extinguishing agent to the water cannon and opens the water cannon valve. During the spraying process, it continuously collects feedback visual image data, extracts the pixel coordinates of the jet landing point from the feedback visual image data, calculates the deviation value between the pixel coordinates of the jet landing point and the pixel coordinates of the flame area, and calls the incremental PID control algorithm to calculate the target horizontal angle compensation and target pitch angle compensation in real time to dynamically correct the water cannon attitude. When the deviation value is continuously less than the set error threshold, the water cannon valve is closed and the pump unit is stopped, and the control unit is reset to the inspection state.

[0009] In some embodiments, step S2 involves multi-source signal fusion processing of thermal imaging data and visual image data based on the orientation of the ultraviolet radiation intensity signal, including the following sub-steps:

[0010] S21, calculate the coarse positioning horizontal azimuth and coarse positioning elevation azimuth based on the peak intensity position of the ultraviolet radiation pulse signal output by the ultraviolet sensor within the detection area;

[0011] S22, based on the coarse positioning horizontal azimuth and coarse positioning elevation azimuth, extract the set region of interest centered on the coarse positioning coordinates in the output image of the infrared thermal imager, and extract the temperature gradient distribution features within the set region of interest to lock the suspected fire source area.

[0012] S23, acquire the local image region corresponding to the suspected fire source area in the output image of the high-definition camera, perform threshold segmentation on the pixel features in the local image region to extract the dynamic flickering edge produced by open flame combustion, and determine the centroid coordinates of the dynamic flickering edge as the pixel coordinates of the flame region.

[0013] In some embodiments, calculating the three-dimensional spatial coordinates of the fire source in step S2 includes the following sub-steps:

[0014] S31. Based on the gimbal scanning action, obtain the gimbal horizontal absolute encoder angle and gimbal pitch absolute encoder angle at the current moment, and establish a geodetic rectangular coordinate system with the gimbal rotation center as the origin.

[0015] S32, calculate the horizontal and vertical observation angles of the flame relative to the rotation center of the pan-tilt unit based on the pixel coordinates of the flame area and the pre-calibration intrinsic parameter matrix of the high-definition camera.

[0016] S33: Based on the distance data output by the laser rangefinder, the straight-line distance of the flame relative to the rotation center of the gimbal is used as the distance. The straight-line distance, horizontal observation angle and vertical observation angle are substituted into the spatial geometric projection formula to calculate the three-dimensional Cartesian coordinates of the flame area in the current coordinate system, which are used as the three-dimensional spatial coordinates of the fire source.

[0017] In some embodiments, step S3, constructing an air resistance jet trajectory prediction model, includes the following sub-steps:

[0018] S41, construct the differential equation of jet motion in a two-dimensional plane, where the differential term of the jet velocity in the horizontal direction is determined based on air density, jet air resistance coefficient, nozzle cross-sectional area and jet resultant velocity, and the differential term of the jet velocity in the vertical direction is determined based on air density, jet air resistance coefficient, nozzle cross-sectional area, jet resultant velocity and gravitational acceleration.

[0019] S42 uses the three-dimensional spatial coordinates of the fire source as the expected landing point of the jet, and substitutes the current attitude of the water cannon as the initial condition into the differential equation of jet motion. A numerical integration algorithm of a set order is used to perform time-series integration iteration on the differential equation of jet motion to solve the initial spray angle when the jet trajectory coincides with the expected landing point of the fire source, which is used as the target horizontal angle and the target pitch angle.

[0020] In some embodiments, the incremental PID control algorithm is invoked based on the deviation value in step S4, including the following sub-steps:

[0021] S51, calculate the pixel difference between the pixel coordinates of the jet landing point and the pixel coordinates of the flame area in the horizontal and vertical directions respectively;

[0022] S52, multiply the pixel difference in the horizontal direction by the horizontal spatial scale mapping coefficient to obtain the horizontal spatial deviation, and multiply the pixel difference in the vertical direction by the vertical spatial scale mapping coefficient to obtain the vertical spatial deviation.

[0023] S53 takes the horizontal and vertical spatial deviations as inputs, calls the incremental PID control algorithm to obtain the horizontal angle correction increment and vertical angle correction increment of the current control cycle, and superimposes the horizontal angle correction increment and vertical angle correction increment to the target horizontal angle and target pitch angle of the previous control cycle, respectively, to obtain the corrected target horizontal angle and corrected target pitch angle of the current control cycle.

[0024] In some embodiments, in step S2, multi-source signal fusion processing is performed on thermal imaging data and visual image data based on the orientation of the ultraviolet radiation intensity signal, including the following steps:

[0025] The radiation intensity in the ultraviolet band is detected by a solar-blind ultraviolet sensor to eliminate interference from sunlight and ambient stray light. The area of ​​the temperature hot spot in the thermal imaging data is compared with the area of ​​the dynamic flicker edge in the local area of ​​the image to determine the degree of overlap.

[0026] When the distance between the geometric center of the hot spot and the geometric center of the dynamic flashing edge is less than the set overlap threshold, the multi-source signal fusion is determined to be successful; otherwise, it is determined to be environmental interference and the process re-enters the inspection state in step S1.

[0027] In some embodiments, step S3, using the three-dimensional spatial coordinates of the fire source as boundary conditions to input the air resistance jet trajectory prediction model, includes the following steps:

[0028] Obtain the liquid pressure and liquid density measurements from the pump output pipeline, and substitute the liquid density measurements into the extinguishing medium property parameters in the air resistance jet trajectory prediction model to correct the dynamic head attenuation coefficient in the jet motion differential equation.

[0029] The target jet velocity in the air resistance jet trajectory prediction model is adjusted according to the flame combustion intensity in the current high-definition camera visual image, so as to dynamically adapt to the coverage flow required for fire sources of different intensities.

[0030] In some embodiments, calculating the deviation between the pixel coordinates of the jet impact point and the pixel coordinates of the flame region in step S4 includes the following steps:

[0031] After the pump unit starts and the water cannon valve is opened, the feedback visual image data from the high-definition camera is captured in real time at a preset sampling frequency.

[0032] The system performs region segmentation based on the color features of the water column on the feedback visual image data to identify the center coordinates of the pixel region where the jet lands; it also calculates the difference between the pixel coordinates of the flame region and the center coordinates of the pixel region where the jet lands to obtain the horizontal and vertical pixel errors at the current moment.

[0033] In some embodiments, step S4, when the deviation value is continuously less than a set error threshold, closes the water cannon valve and stops the pump unit, including the following steps:

[0034] If, in a set number of consecutive feedback visual image frames, the horizontal and vertical pixel errors between the center coordinates of the jet landing pixel area and the flame area pixel coordinates are both less than the set error threshold, it is determined that the fire has been suppressed and a stop command is issued.

[0035] After the stop command is issued, the dynamic correction result of the current control cycle is kept for a set time to observe whether there are signs of reignition. If the pixel coordinates of the flame area reach the set size area again within the set time, the drive pan-tilt unit in step S2 is reset to scan the detection area.

[0036] Secondly, this application provides a multiphase lithium battery-specific intelligent directional jet fire extinguishing system for performing the above-mentioned method, including a mobile cabinet, a liquid storage unit, a pumping unit, a jet water cannon, a dual-axis gimbal, a control cabinet, and a multi-source sensor module.

[0037] The control cabinet is electrically connected to the pumping unit, the jet water cannon, the dual-axis gimbal, and the multi-source sensor module; the multi-source sensor module includes an ultraviolet flame sensor, an infrared thermal imager, a high-definition camera, and a laser rangefinder.

[0038] The control cabinet drives the gimbal to rotate based on the radiation signal from the ultraviolet flame sensor. It combines data from the infrared thermal imager, high-definition camera, and laser rangefinder to perform multi-source signal fusion and calculate the three-dimensional coordinates of the fire source. The fire source coordinates are input into the air resistance jet trajectory prediction model to obtain the target angle and drive the gimbal to align. The jet water cannon obtains the extinguishing agent in the storage unit through the pumping unit and executes the spraying. Based on the jet landing point feedback from the high-definition camera, the control cabinet corrects the water cannon spraying attitude in real time through an incremental PID control algorithm.

[0039] The beneficial effects of this application are as follows: This solution, through the combination of a multimodal sensing collaborative architecture and a fully open intelligent algorithm system, comprehensively improves the fire situation identification capability and directional spraying accuracy of mobile lithium battery fire extinguishing equipment. The multi-level identification link integrates ultraviolet initial screening, dynamic temperature segmentation, lightweight convolutional neural network recognition, and frequency domain flicker verification, coupled with a channel attention adaptive fusion mechanism, which can automatically adapt to complex environments such as dense smoke and strong light, significantly reducing the probability of false alarms and missed alarms. Coupled with a mechanical eccentricity and environmental temperature drift error compensation model, combined with three-dimensional wind disturbance jet dynamics equations and incremental PID closed-loop correction, it effectively improves the aiming accuracy of long-distance impact points. A built-in lithium battery-specific reignition risk identification model adapts to the characteristics of thermal runaway fires, reducing the risk of secondary deflagration. All AI models fully disclose their structural hierarchy, training process, and core parameters, completely overcoming the black-box defect and facilitating engineering reproduction and iterative optimization. The integrated mobile cabinet and manual / automatic pipeline design cater to both unattended operation and manual emergency response needs, resulting in stronger scenario adaptability. Attached Figure Description

[0040] Figure 1 This is a flowchart of the fire extinguishing method in this application;

[0041] Figure 2 This is a flowchart of step S2;

[0042] Figure 3 This is a flowchart of step S3;

[0043] Figure 4 This is a flowchart of step S4;

[0044] Figure 5 A model diagram for intelligent flame recognition;

[0045] Figure 6 This is a block diagram of the fire extinguishing system in this application. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0047] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0048] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0049] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0050] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] Example 1

[0052] The first aspect of this application discloses a smart directional jet fire extinguishing method specifically for multiphase lithium batteries, referring to... Figure 1 This includes the following steps.

[0053] Step 1: Device initialization and ultraviolet inspection. The fire extinguishing device is powered on and initialized, and enters the inspection state. The ultraviolet sensor continuously collects the ultraviolet radiation intensity signal of the detection area.

[0054] After the fire extinguishing device is connected to the power supply, it first enters the power-on initialization process. The control unit in the control cabinet sequentially completes the self-test and loading of operating parameters for each hardware module. After confirming that all functional units are normal, it enters the routine inspection state. The ultraviolet sensor continuously collects the ultraviolet radiation signal of the detection area to complete the first-level fire pre-screening. The device adopts a box-type cabinet structure with a set of moving wheels and a fixed support structure at the bottom, which can realize the switching between transportation and fixed deployment. The side of the cabinet is equipped with an operation panel and a push handle. The control box is integrated below the panel for installing control circuit boards and electronic components. The two sides of the cabinet are equipped with maintenance structures, and the front adopts a double door design for easy maintenance and replacement of internal components. The top of the cabinet is equipped with a liquid filling port and a liquid level observation structure, and also has a main pipeline and water cannon installation interface. The internal components include a liquid storage unit, a pumping unit, pipeline valves, and a manual spray assembly, forming a complete agent delivery and spraying path.

[0055] Step 1.1 Power on and perform initial self-test of the entire machine.

[0056] After the fire extinguishing device is powered on, it enters the power-on initialization process, sequentially completing self-tests of the power circuit, peripheral interfaces, mechanism transmission, sensors, pipeline valves, and human-machine interface module. Pre-set operating parameters are loaded, and after all self-tests pass, it enters a routine inspection state. It continuously collects ultraviolet radiation signals from the detection area using a solar-blind ultraviolet sensor to complete the first-level fire pre-screening. The device adopts a box-type cabinet structure with wheels and a fixed support structure at the bottom, enabling switching between transport and fixed-point deployment.

[0057] Step 1.2 Normal signal acquisition in the ultraviolet channel

[0058] During inspection, the control unit collects ultraviolet pulse signals at a fixed sampling frequency and performs a first-level coarse judgment through sliding window mean statistics and an adaptive trigger threshold model. The adaptive threshold is dynamically adjusted based on the on-site smoke concentration and ambient temperature. When the smoke concentration increases, the ultraviolet signal attenuates, and the threshold is raised accordingly to avoid missed detections. Under high temperature conditions, the dark current of the sensor increases, and the threshold is adjusted accordingly to suppress false alarms. It can also filter interference signals generated by instantaneous arcs and static electricity. If the trigger threshold is not reached, the inspection continues. Once the threshold is reached, multi-sensor synchronous acquisition is immediately initiated, and the fire identification and location process begins.

[0059] Step 2: Multi-source fusion positioning and error compensation: When the ultraviolet radiation intensity signal exceeds the set trigger threshold, the device initiates the precise fire positioning process, driving the pan-tilt unit to scan within the detection area, simultaneously acquiring thermal imaging data output by the infrared thermal imager, distance data output by the laser ranging unit, and visual image data output by the high-definition imaging unit. The entire positioning process is guided by the directional direction of the ultraviolet radiation intensity signal, sequentially completing multiple stages including temperature feature extraction, AI visual flame recognition, frequency domain feature verification, attention network multi-source fusion, system error compensation, 3D coordinate calculation, and overlap verification. This gradually narrows the target range and eliminates interference, ultimately calculating the 3D spatial coordinates of the fire source.

[0060] The device's sensing and detection module integrates multiple types of detection devices. Among them, the ultraviolet flame sensing unit is responsible for rapid fire early warning and coarse positioning; the infrared positioning signal unit is responsible for high-precision azimuth and elevation angle measurement; the laser pointer and ranging unit work together to directly measure the target distance; the infrared temperature sensing unit is used for ambient background temperature filtering to eliminate interference sources such as lights and high-temperature objects; and the infrared temperature measurement array unit is used for continuous temperature tracking in the flame area. Finally, it forms a multi-level detection and positioning system.

[0061] Step 2.1 Coarse UV positioning azimuth calculation

[0062] Reference Figure 2The gimbal performs a step-by-step scan in the horizontal and pitch directions according to a preset scanning trajectory, recording the ultraviolet radiation intensity signal value at each position in real time during the scan. The control unit traverses the signal intensity data throughout the scan, finds the gimbal attitude corresponding to the peak signal intensity, and calculates the coarse horizontal azimuth and coarse pitch azimuth of the fire source. The horizontal azimuth corresponds to the rotation angle of the gimbal's horizontal axis, using the gimbal's initial zero position as a reference, describing the fire source's orientation relative to the device in the horizontal plane; the pitch azimuth corresponds to the rotation angle of the gimbal's pitch axis, describing the fire source's elevation angle relative to the device in the vertical plane.

[0063] In this embodiment, the coarse positioning result is used to narrow the processing range of subsequent infrared and visual detection. The corresponding area is extracted as the region of interest and input into the AI ​​recognition model. There is no need to perform full-image calculation on the entire image, which can reduce the amount of data processing and improve the positioning response speed. The scanning process simultaneously triggers the full-speed acquisition of the infrared dot array and thermal imaging unit to prepare data for subsequent precise positioning.

[0064] Step 2.2 Extraction of infrared region of interest temperature features

[0065] Based on the horizontal and vertical azimuth angles obtained from coarse positioning, a region centered on the coarse positioning coordinates is selected as the region of interest from the entire thermal image output by the infrared thermal imaging unit. The processing of this localized region within the entire image is performed only on the pixels within that region. In this embodiment, the infrared thermal imaging unit operates based on the principle of infrared radiation detection. It senses the infrared thermal radiation emitted by objects through a detection array, converts the radiation intensity into corresponding temperature values, and ultimately generates a thermal image reflecting the temperature distribution of the scene, enabling the identification of high-temperature targets in dense smoke environments.

[0066] Within the defined region of interest, suspected fire source areas are extracted using an infrared adaptive temperature segmentation model. The complete set of formulas is as follows:

[0067] Single pixel temperature rise difference:

[0068]

[0069] Dynamic temperature segmentation threshold:

[0070]

[0071] Binary discrimination of high-temperature flame region:

[0072]

[0073] Total pixel area of ​​the high-temperature region:

[0074]

[0075] Secondary temperature criterion:

[0076]

[0077] In the formula, The temperature of the k-th thermal image corresponding to the coordinate pixel; The long-term ambient background temperature of this pixel is updated in real time via a sliding average. For pixels, the temperature rise relative to the environment is relatively small; for objects in a steady state at high temperatures, the temperature rise is close to 0. It is the base temperature rise threshold, which is the minimum temperature rise reference constant for lithium battery ignition; λ is the instantaneous rate of change of the highest temperature in the current frame; λ is the weighting coefficient for the temperature rise rate. A frame-by-frame dynamic temperature segmentation threshold; For the binary mask, the value 1 represents a pixel that is suspected to be a high-temperature flame; This represents the total pixel area of ​​the high-temperature region in a single frame. The minimum effective flame area threshold is used to filter noise from single-point high-temperature components.

[0078] By combining relative temperature rise with temperature rise rate, only rapidly heating dynamic high-temperature areas are marked as suspected fire points, thereby automatically filtering out steady-state high-temperature equipment. This adapts to the rapid temperature rise characteristics of lithium battery thermal runaway, distinguishing between static heat sources and dynamic combustion sources. Simultaneously, ambient background temperature data collected by a single-point infrared temperature sensor is used to compensate for the overall ambient temperature rise, further eliminating misjudgments caused by overall ambient temperature increases. The extracted high-temperature mask area is used as prior information and input into the subsequent AI flame recognition model, thereby narrowing the model's computational scope.

[0079] Step 2.3 Multi-scale lightweight convolutional neural network for intelligent flame recognition: Obtain the local image region corresponding to the suspected fire source area in the output image of the high-definition camera. Perform threshold segmentation on the pixel features within the local image region to extract the dynamic flickering edge generated by open flame combustion. Determine the centroid coordinates of the dynamic flickering edge as the pixel coordinates of the flame region. Specifically, this includes:

[0080] Reference Figure 5 The model adopts a lightweight dual-input dual-branch architecture. The inputs are visible light ROI images and infrared temperature pseudo-color images, and the outputs are flame confidence, flame centroid pixel coordinates, and flame normalized area. The model is divided into four modules from input to output: input preprocessing layer, backbone feature extraction layer, multi-scale feature enhancement layer, and dual-task output layer. The modules are connected in series, backbone features are shared, and classification and localization tasks are output in parallel.

[0081] The input preprocessing layer receives the visible light ROI image captured by coarse ultraviolet localization and the infrared temperature matrix of the corresponding region. First, the infrared temperature matrix is ​​converted into a single-channel grayscale image through pseudo-color mapping, and then stitched with the three-channel visible light image to form a four-channel input. All inputs are uniformly scaled to a size of 256×256 pixels, and pixel value normalization is performed, that is, the pixel value is divided by 255 and then the mean of the dataset is subtracted to complete the input standardization process, thereby improving the recognition effect in dense smoke scenes.

[0082] The backbone feature extraction layer employs a three-level depthwise separable convolutional architecture, downsampling and extracting features at each level to adapt to embedded deployments. In this embodiment, the three-level modules are connected sequentially, with the output of the previous level serving as the input of the next level.

[0083] Level 1: Depthwise separable convolutional units, kernel size 3×3, output channels 16, stride 2, activation function ReLU6, followed by a 2×2 max pooling layer; input size 256×256×4, output size 128×128×16, extracting bottom edge and texture features.

[0084] Second level: Depthwise separable convolutional units, kernel size 3×3, output channels 32, stride 2, activation function is ReLU6, followed by 2×2 max pooling layer; input size 128×128×16, output size 64×64×32, extracting mid-layer flame shape and contour features.

[0085] Level 3: Depthwise separable convolutional units with a kernel size of 3×3, 64 output channels, a stride of 2, and a ReLU6 activation function; input size 64×64×32, output size 32×32×64, extracting high-level semantic and global structural features.

[0086] The multi-scale feature enhancement layer employs three parallel convolutional branches to extract features from different receptive fields, which are then concatenated and fused to address the adaptation issue of flame size varying with distance.

[0087] Branch 1: 1×1 standard convolution, 32 output channels, stride 1, extracts fine-grained local features.

[0088] Branch 2: 3×3 dilated convolution with an inflation rate of 2, 32 output channels, and a stride of 1 to expand the receptive field and extract mesoscale features.

[0089] Branch 3: After 5×5 average pooling, a 1×1 convolution is performed, resulting in 32 output channels and a stride of 1, to extract global large-scale features.

[0090] The three outputs are spliced ​​and fused along the channel dimension, and the output feature map size is 32×32×96, realizing complementary enhancement of multi-scale features.

[0091] The dual-task output layer consists of a classification branch and a localization branch, which are connected in parallel and share the features output by the multi-scale enhancement layer.

[0092] Classification branch: Outputs a single-value flame confidence score, ranging from 0 to 1. The higher the value, the greater the probability that it is a real flame.

[0093] Location branch: Outputs a three-dimensional vector, which consists of the normalized x-coordinate of the flame centroid, the normalized y-coordinate, and the normalized area of ​​the flame region.

[0094] The model uses a labeled fire dataset as training samples. The dataset is constructed with positive samples including lithium battery thermal runaway combustion frames, electric vehicle fire frames, and standard oil pan fire frames, covering different distances, light intensities, and smoke concentrations. The negative samples include easily misjudged scenarios such as high-temperature power distribution cabinets, lighting fixtures, sunlight reflections, electric welding arcs, heating pipes, and cigarette butt interference. All samples are simultaneously labeled with real labels, flame centroid coordinates, and flame area. The dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0095] The data augmentation strategy performs data augmentation online during the training process to improve the model's generalization ability: Geometric augmentation includes random horizontal flipping, random scaling from 0.5 to 1.5 times, random rotation from -30° to 30°, and random cropping; Luminous augmentation includes ±20% random brightness adjustment, ±15% random contrast adjustment, adding Gaussian noise, and random simulated smoke occlusion.

[0096] The model is specifically optimized for lithium battery fire extinguishing scenarios. The training set contains a large number of lithium battery thermal runaway combustion samples, which can accurately identify the special shape and temperature characteristics of lithium battery flames. The inference process only processes the ROI region extracted by coarse ultraviolet localization. The flame confidence score output by the model will be used as the input of multi-source fusion visual features, the output flame centroid coordinates will be directly used as the pixel coordinates of the flame region, and the output flame area will be used to evaluate the combustion intensity.

[0097] Step 2.4 Auxiliary Verification of Flame Flicker Frequency Domain Features

[0098] Real flames exhibit periodic brightness flickering within a fixed range, while high-temperature equipment and static light sources show no periodic fluctuations. The temporal spectrum of pixels in candidate flame regions is extracted, and the amplitude of the dominant frequency is used to determine whether it is a real open flame, serving as an auxiliary verification for AI recognition results. A high-definition imaging unit continuously acquires multiple frames of images at a fixed sampling frame rate, caching the brightness temporal data of the candidate fire source center. Spectral features are extracted using Fast Fourier Transform. If a significant peak exists within the inherent frequency range of the flame, it is determined to possess dynamic flame flickering characteristics. This method can distinguish between random light and shadow jitter and periodic flame flickering, filtering out non-flame dynamic interference, serving as a supplementary verification after AI visual recognition. Both conditions must be met for the visual feature to be considered valid.

[0099] Step 2.5 Multimodal Channel Attention Adaptive Fusion Network Decision

[0100] After extracting four independent features from ultraviolet, infrared temperature, AI vision, and scintillation frequency domain, a channel attention fusion network is introduced to achieve intelligent fusion of multi-source features:

[0101] Step 2.5.1 Integrating Network Modules and Hierarchical Structure

[0102] The model is a fully connected feedforward network structure, consisting of an input normalization layer, a channel attention generation layer, a weight normalization layer, and a weighted fusion output layer. The layers are connected in series. The input normalization layer receives four single feature confidence scores, eliminates the differences in the dimensions of different features, and outputs a 4-dimensional normalized feature vector.

[0103] The channel attention generation layer consists of two fully connected network layers. It generates attention weights for each channel through feature compression and activation. The weight normalization layer performs normalization processing on the original attention weights to ensure that the sum of all weights is 1. The weighted fusion output layer performs weighted summation of the four features and the corresponding normalized weights to obtain the comprehensive fire confidence score, ranging from 0 to 1, which serves as the basis for the final fire determination.

[0104] Step 2.5.2 Network Training and Weight Generation Logic

[0105] The model employs supervised training, using multi-sensor synchronous data from real fire and interference scenarios as training samples. The optimization objective is to achieve the classification accuracy after fusion. After training, the network can automatically generate weights adapted to the current environment based on the four input features. In dense smoke scenarios, it automatically reduces the weight of the ultraviolet channel and increases the weights of the infrared and visual channels. In strong light scenarios, it automatically increases the weight of the ultraviolet channel and decreases the weight of the visual channel, thus achieving intelligent fusion that adapts to the environment.

[0106] After the network outputs the overall fire confidence score, it compares it with the set confidence threshold. If the overall confidence score is greater than or equal to the judgment threshold, a valid fire source is confirmed, and the subsequent coordinate calculation process begins; if it is lower than the threshold, it is determined to be environmental interference, the positioning process is terminated, and the system returns to the inspection state, including:

[0107] Single-feature normalized confidence, with numerical constraints between 0 and 1:

[0108]

[0109] Dynamic calculation of adaptive weights for the environment:

[0110]

[0111] Normalized weight constraint correction:

[0112]

[0113] Overall fire confidence level:

[0114]

[0115] Comprehensive criteria:

[0116]

[0117] In the formula, Do not use single-feature confidence levels for ultraviolet, infrared temperature, fractal morphology, and scintillation frequency domain; This is a limiting function that constrains the value to the interval between 0 and 1; These are the original, unnormalized weights; The value represents the real-time smoke concentration; the higher the smoke concentration, the more exponentially the ultraviolet weight decreases. For ambient light intensity, the weight of infrared light decreases linearly in strong light scenarios; For fixed environmental correction factors; These are the normalized effective weights; The overall fire situation confidence level after fusion; The confidence threshold for determining an effective ignition source.

[0118] Step 2.6 Real-time compensation for coupled system errors

[0119] The gimbal camera, laser ranging module, and rotation center have a fixed installation eccentricity. Deformation of the metal frame and encoder temperature drift under high and low temperature environments can introduce systematic positioning errors. To improve positioning angle accuracy, a multi-coupled error correction equation system is constructed to compensate for the horizontal and pitch angles of the gimbal in real time. The built-in temperature sensor unit of the gimbal collects the frame temperature in real time and automatically substitutes it into the equation system to complete the compensation after each encoder angle reading. No manual on-site secondary calibration is required. It can simultaneously compensate for two types of systematic errors: mechanical fixed eccentricity and dynamic temperature drift, eliminate positioning drift caused by long-term use of the equipment under high and low temperature environments, and stabilize and control the angle positioning error.

[0120] Step 2.7 Solving the three-dimensional spatial coordinates of the fire source

[0121] After completing the angle error compensation, a geodetic rectangular coordinate system is constructed with the gimbal rotation center as the origin, based on the straight-line distance output by the laser ranging unit. The horizontal axis is the X-axis, the horizontal axis is the Y-axis, and the vertical axis is the Z-axis. The three coordinate axes are perpendicular to each other and are used to describe the absolute position of the fire source in space. The three-dimensional Cartesian coordinates of the fire source are calculated through spatial geometric projection relationship. After multi-source fusion to determine the effective fire source, the gimbal accurately turns to the candidate area, and the laser ranging unit outputs the distance value. The three-dimensional coordinates are then calculated in one go by substituting them into the geometric equation system. These coordinates will serve as the target landing point boundary conditions for the subsequent jet prediction model and provide input for the jet angle calculation.

[0122] Step 2.8 Multi-source signal overlap verification and interference elimination

[0123] In the multi-source signal fusion processing, the basic interference from sunlight and ambient stray light is first eliminated by utilizing the band characteristics of the solar-blind ultraviolet sensor. Then, the overlap between the temperature hotspot obtained from infrared thermal imaging and the dynamic flicker edge obtained from the visible light image is verified. Specifically, the pixel distance between the geometric center of the temperature hotspot and the geometric center of the dynamic flicker edge in the image is calculated separately, and then converted into the actual spatial distance by combining the spatial scale coefficient.

[0124] When the distance is less than the set overlap threshold, it is determined that the two types of sensors detect the same target, the multi-source signal fusion is successful, and the target is confirmed as a real fire source; when the distance exceeds the overlap threshold, it is determined that the two types of signals correspond to different targets, the abnormal signal is environmental interference, and the subsequent fire extinguishing process is not triggered. The device terminates the positioning operation and returns to the inspection state to continue monitoring.

[0125] After completing the positioning, the control unit logs information such as positioning angle, fire level, and detection time, forming a full-state operation record for easy subsequent operation and maintenance and traceability.

[0126] Step 3: Jet trajectory prediction and initial alignment

[0127] After obtaining the three-dimensional spatial coordinates of the fire source, these are used as boundary conditions for the expected landing point of the jet. An air resistance jet trajectory prediction model with lateral wind disturbance is input, and the model calculates and outputs the target horizontal and vertical angles that the water cannon needs to achieve. The gimbal is then driven to rotate to the corresponding angles, completing the initial alignment of the water cannon. This stage is an open-loop feedforward control system, capable of quickly completing large-range angle adjustments, providing a foundation for subsequent closed-loop fine correction. The device's piping system adopts a low-flow-resistance design; the pipe section connecting the top to the water cannon uses a smooth bend structure to reduce flow resistance losses caused by right-angle pipes. The pumping unit uses a positive displacement screw pump structure, which can stably deliver multiphase extinguishing agents with a certain viscosity, providing stable output pressure and small flow pulsation, suitable for the pressure-stabilized spray requirements of jet extinguishing.

[0128] Step 3.1 Construction of the three-dimensional air resistance jet differential equation

[0129] Reference Figure 3 A three-dimensional differential equation for jet motion is constructed to describe the motion of the jet under the combined effects of air resistance and crosswind. By introducing a pre-defined square air resistance model and adding a three-dimensional crosswind velocity component, the velocity changes in the three axes are modeled respectively: in the horizontal and longitudinal directions, the jet is only affected by air resistance and wind disturbance, and the velocity continuously decreases; in the vertical direction, the jet is simultaneously affected by gravity, air resistance, and wind disturbance, and the velocity change is the superposition of gravitational acceleration and acceleration generated by resistance and wind disturbance.

[0130] This step uses a set of three-dimensional jet dynamics differential equations with lateral wind disturbance as the trajectory prediction model:

[0131] The three-dimensional velocity components of the jet element are The three-dimensional velocity of the environmental lateral wind is The relative air velocity component of the infinitesimal element is ;

[0132] Air resistance acceleration is proportional to the square of the relative velocity, and gravity acts only in the negative Z-axis direction:

[0133]

[0134] Displacement integral relationship:

[0135]

[0136] Iterative criteria for solving the target angle:

[0137]

[0138] In the formula, Ambient air density; denoted as ; A is the nozzle exit cross-sectional area; m is the mass of a unit jet element; g is the acceleration due to gravity. These are the three-dimensional lateral wind velocity components; Let be the coordinates of the jet landing point in the k-th step of the numerical iteration; The coordinates of the three-dimensional target of the fire source; This is the threshold for coordinate convergence error.

[0139] In this embodiment, the three-dimensional differential equations form the basis of the jet trajectory prediction model. Compared with the two-dimensional model, the added coupling effect of lateral wind makes it suitable for jet calculations in outdoor and ventilated scenarios, and more closely reflects the jet motion law under actual working conditions. Especially at longer ranges, it can significantly improve the accuracy of trajectory prediction, providing a reliable calculation basis for open-loop feedforward aiming. The modeling process also considers the density and viscosity properties of the extinguishing agent, corrects the drag coefficient and dynamic pressure attenuation characteristics, and adapts to the spraying conditions of specialized extinguishing agents, rather than being applicable only to water media.

[0140] Step 3.2 Numerical integration to solve for the target jet angle

[0141] The three-dimensional spatial coordinates of the fire source are converted into the expected landing point coordinates of the jet. Using the initial position and initial velocity direction of the water cannon nozzle outlet as initial conditions, these are substituted into the aforementioned differential equation of jet motion. The differential equation is then iterated over time to progressively solve for the jet's trajectory. The solution process employs an iterative approximation method: first, a set of initial injection angle trial values ​​are given, and the landing point coordinates of the corresponding jet trajectory are calculated through numerical integration. The calculated landing point is compared with the target landing point to obtain the deviation. Based on the deviation, the initial injection angle is adjusted, and the integration calculation is performed again. This process is repeated iteratively until the deviation between the calculated landing point and the target landing point is less than a set calculation accuracy threshold. The corresponding initial horizontal angle and initial pitch angle at this point are the solved target horizontal angle and target pitch angle.

[0142] After the solution is completed, the control unit sends the target angle command to the gimbal servo driver, driving the horizontal and pitch axes of the gimbal to rotate synchronously, moving the water cannon to the target angle, and completing the initial aiming and alignment. This open-loop feedforward calculation can quickly complete a large range of angle adjustments, making the jet impact point close to the target area, significantly reducing the number of iterations for subsequent closed-loop corrections, and improving the overall response speed.

[0143] Step 3.3 Dynamic adaptation and adjustment of medium properties and combustion intensity

[0144] Before inputting the values ​​into the jet trajectory prediction model for calculation, the pressure sensing unit and the medium parameter detection module in the pipeline are used to obtain the liquid pressure measurement value and the density measurement value of the extinguishing medium at the pump output pipeline. The medium density parameter is then substituted into the extinguishing medium physical property parameter term in the model to correct the correlation coefficient in the jet motion differential equation, so as to adapt to the physical characteristics of different types of extinguishing agents and avoid trajectory calculation deviations caused by the difference between the agent density and that of water.

[0145] Simultaneously, the control unit determines the flame intensity based on the flame area output by the AI ​​visual model and the highest temperature obtained from infrared thermal imaging, and adjusts the target jet velocity and corresponding flow rate set in the model accordingly. When the flame intensity is high, the jet outlet velocity and flow rate are appropriately increased to ensure sufficient extinguishing agent coverage and impact strength; when the flame intensity is low, the velocity and flow rate can be appropriately reduced to reduce agent consumption while ensuring extinguishing effect and extending the continuous operation time of the device. An overflow pressure stabilizing structure is installed in the pipeline to stabilize the working pressure in the pipeline and prevent pressure fluctuations from affecting the stability of the jet trajectory.

[0146] Step 4: Closed-loop injection correction and reset shutdown

[0147] After initial alignment of the water cannon, the device starts the pump unit and opens the corresponding pipeline valves. Under the pressure of the pump unit, the extinguishing agent in the storage tank is transported to the water cannon through the pipeline and sprayed outwards. The device's pipeline system has three operating modes: automatic jet extinguishing mode, manual hose spraying mode, and agent refilling mode. In automatic extinguishing mode, the control unit automatically controls the valves and pump unit to complete the spraying operation. In manual mode, the operator can control the valves and pan / tilt head through the panel buttons and pull out the hose reel to complete close-range manual spraying. In refilling mode, the agent in the storage tank can be replenished through the external refill interface.

[0148] During the spraying process, the device continuously collects visual feedback images, identifies the actual landing point of the jet, calculates the positional deviation between the actual landing point and the flame area, and corrects the spray angle of the water cannon in real time through an incremental PID control algorithm until the landing point deviation stabilizes within the allowable range. After completing the reignition risk assessment based on the lithium battery ignition characteristics, the valves and pumps are shut off, and the gimbal resets to the inspection state.

[0149] Step 4.1 Visual recognition of jet landing point and calculation of pixel deviation

[0150] After the pump unit starts and the water cannon valve is opened, the high-definition imaging unit acquires real-time images of the scene at a preset sampling frequency, obtaining feedback visual image data. For each frame of feedback image, jet impact point recognition processing is performed. The processing flow is as follows: First, color thresholding is used to segment and filter out pixel regions that match the color characteristics of the jet medium; then, morphological filtering is used to remove small noise points, and adjacent jet pixels are connected to obtain a complete jet region; subsequently, the boundary position between the jet region and the protective plane is extracted to determine the pixel region of the jet impact point, and the center coordinates of the impact point region are calculated.

[0151] After completing the landing point recognition, the difference between the pixel coordinates of the flame area in the current frame and the pixel coordinates of the center of the jet landing point is calculated to obtain the pixel error in the horizontal direction and the pixel error in the vertical direction.

[0152] Step 4.2 Landing Point-Ignition Source Coupling Error Compensation

[0153] After obtaining the pixel deviation, the inherent deviation of the model is first corrected by a coupled iterative error compensation model, and then input into the subsequent control loop. The complete set of formulas is as follows:

[0154] Original value of 3D landing point deviation:

[0155]

[0156] Deviation after coupling compensation:

[0157]

[0158] The convergence criterion is that the deviation of a set number of consecutive frames is less than a threshold.

[0159]

[0160] In the formula, The three-dimensional coordinates of the fire source; The three-dimensional coordinates of the actual impact point of the jet; This represents the original landing point deviation. These are the coupling compensation coefficients, used to compensate for inherent model errors caused by jet atomization and agent viscosity; The compensated effective deviation is used as the input to the subsequent PID controller; This is the convergence error threshold.

[0161] Step 4.3 Spatial Scale Mapping and Bias Conversion

[0162] Reference Figure 4 Pixel-level deviations cannot be directly used for angle correction and need to be converted into actual spatial deviations. The actual spatial deviations are obtained by multiplying the pixel differences in the horizontal and vertical directions by their corresponding spatial scale mapping coefficients. The spatial scale mapping coefficient represents the actual spatial distance corresponding to a single pixel. Multiplying the horizontal pixel difference by the horizontal spatial scale mapping coefficient yields the horizontal spatial deviation; multiplying the vertical pixel difference by the vertical spatial scale mapping coefficient yields the vertical spatial deviation.

[0163] Step 4.4 Fuzzy PID Intelligent Closed-Loop Angle Correction

[0164] The horizontal and vertical spatial deviations are used as inputs to two independent controllers. An incremental PID control algorithm is used to correct the angle, and the PID parameters are automatically tuned online based on the deviation status.

[0165] The horizontal and pitch axes are each equipped with independent fuzzy PID controllers, each calculating its corresponding angle correction increment. After the calculation is completed, the horizontal angle correction increment is added to the target horizontal angle of the previous control cycle, and the vertical angle correction increment is added to the target pitch angle of the previous control cycle, resulting in the corrected target horizontal and pitch angles for the current control cycle. The control unit sends the corrected target angles to the gimbal driver, driving the gimbal to perform the corresponding angle adjustment, thus achieving dynamic correction of the jet attitude. Through continuous closed-loop iterative calculations, the jet impact point gradually converges towards the fire source position, ultimately stabilizing the impact point deviation within the allowable error range.

[0166] Step 4.5 Lithium-ion battery-specific reignition risk assessment

[0167] Lithium-ion battery thermal runaway fires have the characteristic of internal residual heat accumulation. Even after the open flame is extinguished, the internal reaction of the cell may continue, making secondary reignition likely. Therefore, it is necessary to assess the reignition risk through a quantitative model. When the horizontal and vertical pixel errors between the jet impact point center and the flame area are both less than the set error threshold in a set number of consecutive feedback image frames, and the area and temperature of the flame area continue to decrease, the lithium-ion battery-specific reignition risk assessment process is activated, but the system is not immediately shut down.

[0168] This step uses a lithium battery thermal runaway reignition probability discrimination model to complete the risk assessment. The complete formula set is as follows:

[0169] Temperature rise decay rate:

[0170]

[0171] Overall risk rating for fire hazards:

[0172]

[0173] Reignition detection logic:

[0174]

[0175] In the formula, These are the highest temperatures in the fire source area for the current and previous cycles, respectively. The area of ​​the pixel with residual high temperature; a, b, and c are the weighting coefficients for lithium battery thermal runaway; A comprehensive score for the risk of reignition; This is the shutdown safety threshold.

[0176] In this embodiment, the model quantifies the probability of reignition by parameters such as temperature decay rate and area of ​​residual high-temperature region, maintains injection suppression under high-risk conditions, adapts to the special characteristics of lithium battery fire, and reduces the risk of secondary deflagration.

[0177] Step 4.6 Shutdown Reset and Multi-mode Extension Instructions

[0178] Once the reignition risk score falls below the safety threshold, the fire is deemed effectively contained. The device issues a stop command, sequentially closing the water cannon pipeline valves and the main pipeline valve, and then stopping the pumping unit. After shutdown, the device does not immediately reset but maintains its current monitoring posture for a set duration to monitor for reignition in the target area. If, within this duration, a pixel area matching flame characteristics reappears in the target area, and its area reaches the set reignition determination threshold, the fire source location process is retried, and a complete fire extinguishing operation is performed again. If no reignition signal is detected within the set duration, the pan-tilt unit is reset to its initial zero position, all sensors return to the inspection and sampling mode, the audible and visual alarms are deactivated, and the device returns to normal inspection status.

[0179] In addition to the automatic fire extinguishing mode, the device also supports manual operation and refilling modes. In manual mode, operators can switch to manual control via the manual / automatic switch button on the control panel, manually adjust the pan-tilt angle using the directional control unit, and control the main valve, water cannon valve, manual valve, and pump unit start / stop using the buttons. Pulling out the hose reel on the side of the cabinet allows for close-range manual spraying, adapting to manual handling needs in special working conditions. In refilling mode, connecting the external pipeline to the refill interface, opening the corresponding refill and return valves, and starting the pump unit will deliver external agents to the storage tank, completing the agent replenishment. The liquid level can be monitored during the operation using the liquid level observation structure.

[0180] The device's control unit has a linkage interface, which can be linked with external fire alarm systems, smoke exhaust systems and other equipment. It can receive external fire alarm signals to trigger the start, and can also upload fire and equipment status information to external systems. It also has full status feedback and log recording functions, which can store operating data such as positioning angle, pipeline pressure, flow rate, fault status, and spraying duration. It supports historical data query and export, which is convenient for equipment operation and maintenance and fault diagnosis.

[0181] Example 2

[0182] The second aspect of this embodiment discloses a smart directional jet fire extinguishing system for multiphase lithium batteries, used to perform the above-mentioned fire extinguishing method. The overall module structure is shown in Figure 6.

[0183] The system includes a mobile cabinet, a liquid storage unit, a pumping unit, pipeline valve components, a jet water cannon, a dual-axis gimbal, a control cabinet, and a multi-source sensor module.

[0184] The control cabinet adopts a hierarchical control architecture, which is electrically connected to the pumping unit, jet water cannon, dual-axis gimbal, and multi-source sensor module. It has five built-in algorithm modules for fire identification, spatial positioning, jet prediction, closed-loop control, and status management, and is responsible for receiving sensor data, running control logic, and outputting drive commands.

[0185] The multi-source sensor module is divided into two categories: flame detection and environmental perception. The flame detection unit includes a solar-blind ultraviolet flame sensor, a dot matrix infrared thermal imager, a high-definition imaging camera, and a laser rangefinder, which respectively realize fire early warning, temperature feature extraction, visual image acquisition, and distance measurement. The environmental perception unit includes smoke, temperature and humidity, light, and wind speed sensors, which provide parameter inputs for environmental adaptive correction of various algorithms.

[0186] The liquid storage unit is fixed inside the mobile cabinet and is used to store multiphase lithium battery-specific fire extinguishing agents. It is equipped with liquid level monitoring and liquid addition and return interfaces. The pumping unit adopts a positive displacement screw pump structure. The liquid inlet is connected to the liquid storage unit, and the liquid outlet is connected to the jet water cannon through pipeline valves. It is equipped with a pressure sensor and overflow pressure stabilizing component to provide stable pressure and flow power for agent delivery.

[0187] The water jet cannon is fixed to the support end of the dual-axis gimbal, with a dedicated atomizing nozzle at the front end. It adjusts its horizontal and vertical angles synchronously with the gimbal. The dual-axis gimbal adopts a dual-servo drive structure with built-in absolute encoders and temperature sensors, driving the water jet and sensor module to rotate synchronously, achieving full-range detection and aiming.

[0188] The pipeline valve assembly includes the main pipeline, water cannon branch, manual spray branch and liquid filling branch. Each branch is equipped with independent electric and manual valves, and the valve group logic switches to realize three working modes: automatic jet fire extinguishing, manual hose spraying and agent filling.

[0189] The mobile cabinet has an integrated box-type structure with universal wheels and a lockable support structure at the bottom, which can flexibly switch between transfer deployment and fixed-point operation. The cabinet is divided into functional zones and units to form an integrated independent fire extinguishing system that does not require external fixed pipelines.

[0190] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0191] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A smart directional jet fire extinguishing method specifically for multiphase lithium batteries, characterized in that, Includes the following steps: The S1 fire extinguishing device is powered on and initialized and enters inspection mode, continuously collecting ultraviolet radiation intensity signals of the detection area through ultraviolet sensors; When the ultraviolet radiation intensity signal exceeds the set trigger threshold, S2 drives the pan-tilt unit to scan the detection area, simultaneously acquiring thermal imaging data output by the infrared thermal imager, distance data output by the laser rangefinder, and visual image data output by the high-definition camera. Based on the orientation of the ultraviolet radiation intensity signal, multi-source signal fusion processing is performed on the thermal imaging data and visual image data to determine the pixel coordinates of the flame area and calculate the three-dimensional spatial coordinates of the fire source by combining the distance data. S3 inputs the three-dimensional spatial coordinates of the fire source as boundary conditions into the air resistance jet trajectory prediction model. The air resistance jet trajectory prediction model outputs the horizontal angle and pitch angle of the water cannon target. Based on the horizontal angle and pitch angle of the target, the gimbal is driven to rotate to achieve water cannon alignment. S4 starts the pump unit to deliver fire extinguishing agent to the water cannon and opens the water cannon valve. During the spraying process, it continuously collects feedback visual image data, extracts the pixel coordinates of the jet landing point from the feedback visual image data, calculates the deviation value between the pixel coordinates of the jet landing point and the pixel coordinates of the flame area, and calls the incremental PID control algorithm to calculate the target horizontal angle compensation and target pitch angle compensation in real time to dynamically correct the water cannon attitude. When the deviation value is continuously less than the set error threshold, the water cannon valve is closed and the pump unit is stopped, and the control unit is reset to the inspection state.

2. The method according to claim 1, characterized in that, In step S2, multi-source signal fusion processing is performed on thermal imaging data and visual image data based on the orientation of the ultraviolet radiation intensity signal, including the following sub-steps: S21, calculate the coarse positioning horizontal azimuth and coarse positioning elevation azimuth based on the peak intensity position of the ultraviolet radiation pulse signal output by the ultraviolet sensor within the detection area; S22, based on the coarse positioning horizontal azimuth and coarse positioning elevation azimuth, extract the set region of interest centered on the coarse positioning coordinates in the output image of the infrared thermal imager, and extract the temperature gradient distribution features within the set region of interest to lock the suspected fire source area. S23, acquire the local image region corresponding to the suspected fire source area in the output image of the high-definition camera, perform threshold segmentation on the pixel features in the local image region to extract the dynamic flickering edge produced by open flame combustion, and determine the centroid coordinates of the dynamic flickering edge as the pixel coordinates of the flame region.

3. The method according to claim 1, characterized in that, The calculation of the three-dimensional spatial coordinates of the fire source in step S2 includes the following sub-steps: S31. Based on the gimbal scanning action, obtain the gimbal horizontal absolute encoder angle and gimbal pitch absolute encoder angle at the current moment, and establish a geodetic rectangular coordinate system with the gimbal rotation center as the origin. S32, calculate the horizontal and vertical observation angles of the flame relative to the rotation center of the pan-tilt unit based on the pixel coordinates of the flame area and the pre-calibration intrinsic parameter matrix of the high-definition camera. S33: Based on the distance data output by the laser rangefinder, the straight-line distance of the flame relative to the rotation center of the gimbal is used as the distance. The straight-line distance, horizontal observation angle and vertical observation angle are substituted into the spatial geometric projection formula to calculate the three-dimensional Cartesian coordinates of the flame area in the current coordinate system, which are used as the three-dimensional spatial coordinates of the fire source.

4. The method according to claim 1, characterized in that, The air resistance jet trajectory prediction model in step S3 includes the following sub-steps: S41, construct the differential equation of jet motion in a two-dimensional plane, where the differential term of the jet velocity in the horizontal direction is determined based on air density, jet air resistance coefficient, nozzle cross-sectional area and jet resultant velocity, and the differential term of the jet velocity in the vertical direction is determined based on air density, jet air resistance coefficient, nozzle cross-sectional area, jet resultant velocity and gravitational acceleration. S42 uses the three-dimensional spatial coordinates of the fire source as the expected landing point of the jet, and substitutes the current attitude of the water cannon as the initial condition into the differential equation of jet motion. A numerical integration algorithm of a set order is used to perform time-series integration iteration on the differential equation of jet motion to solve the initial spray angle when the jet trajectory coincides with the expected landing point of the fire source, which is used as the target horizontal angle and the target pitch angle.

5. The method according to claim 1, characterized in that, The incremental PID control algorithm invoked based on the deviation value in step S4 includes the following sub-steps: S51, calculate the pixel difference between the pixel coordinates of the jet landing point and the pixel coordinates of the flame area in the horizontal and vertical directions respectively; S52, multiply the pixel difference in the horizontal direction by the horizontal spatial scale mapping coefficient to obtain the horizontal spatial deviation, and multiply the pixel difference in the vertical direction by the vertical spatial scale mapping coefficient to obtain the vertical spatial deviation. S53 takes the horizontal and vertical spatial deviations as inputs, calls the incremental PID control algorithm to obtain the horizontal angle correction increment and vertical angle correction increment of the current control cycle, and superimposes the horizontal angle correction increment and vertical angle correction increment to the target horizontal angle and target pitch angle of the previous control cycle, respectively, to obtain the corrected target horizontal angle and corrected target pitch angle of the current control cycle.

6. The method according to claim 1, characterized in that, In step S2, multi-source signal fusion processing is performed on thermal imaging data and visual image data based on the orientation of the ultraviolet radiation intensity signal, including the following steps: The radiation intensity in the ultraviolet band is detected by a solar-blind ultraviolet sensor to eliminate interference from sunlight and ambient stray light. The area of ​​the temperature hot spot in the thermal imaging data is compared with the area of ​​the dynamic flicker edge in the local area of ​​the image to determine the degree of overlap. When the distance between the geometric center of the hot spot and the geometric center of the dynamic flashing edge is less than the set overlap threshold, the multi-source signal fusion is determined to be successful; otherwise, it is determined to be environmental interference and the process re-enters the inspection state in step S1.

7. The method according to claim 1, characterized in that, In step S3, the three-dimensional spatial coordinates of the fire source are used as boundary conditions to input the air resistance jet trajectory prediction model, including the following steps: Obtain the liquid pressure and liquid density measurements from the pump output pipeline, and substitute the liquid density measurements into the extinguishing medium property parameters in the air resistance jet trajectory prediction model to correct the dynamic head attenuation coefficient in the jet motion differential equation. The target jet velocity in the air resistance jet trajectory prediction model is adjusted according to the flame combustion intensity in the current high-definition camera visual image, so as to dynamically adapt to the coverage flow required for fire sources of different intensities.

8. The method according to claim 1, characterized in that, In step S4, the deviation between the pixel coordinates of the jet impact point and the pixel coordinates of the flame area is calculated, including the following steps: After the pump unit starts and the water cannon valve is opened, the feedback visual image data from the high-definition camera is captured in real time at a preset sampling frequency. The system performs region segmentation based on the color features of the water column on the feedback visual image data to identify the center coordinates of the pixel region where the jet lands; it also calculates the difference between the pixel coordinates of the flame region and the center coordinates of the pixel region where the jet lands to obtain the horizontal and vertical pixel errors at the current moment.

9. The method according to claim 1, characterized in that, In step S4, when the deviation value remains below the set error threshold, the water cannon valve is closed and the pump unit is stopped, including the following steps: If, in a set number of consecutive feedback visual image frames, the horizontal and vertical pixel errors between the center coordinates of the jet landing pixel area and the flame area pixel coordinates are both less than the set error threshold, it is determined that the fire has been suppressed and a stop command is issued. After the stop command is issued, the dynamic correction result of the current control cycle is kept for a set time to observe whether there are signs of reignition. If the pixel coordinates of the flame area reach the set size area again within the set time, the drive pan-tilt unit in step S2 is reset to scan the detection area.

10. A smart directional jet fire extinguishing system specifically for multiphase lithium batteries, used to execute the method as described in any one of claims 1-9, characterized in that, It includes a mobile cabinet, a liquid storage unit, a pumping unit, a jet water cannon, a dual-axis gimbal, a control cabinet, and a multi-source sensor module; The control cabinet is electrically connected to the pumping unit, the jet water cannon, the dual-axis gimbal, and the multi-source sensor module; the multi-source sensor module includes an ultraviolet flame sensor, an infrared thermal imager, a high-definition camera, and a laser rangefinder. The control cabinet drives the gimbal to rotate based on the radiation signal from the ultraviolet flame sensor. It combines data from the infrared thermal imager, high-definition camera, and laser rangefinder to perform multi-source signal fusion and calculate the three-dimensional coordinates of the fire source. The fire source coordinates are then input into the air resistance jet trajectory prediction model to obtain the target angle, which drives the gimbal to align. The jet water cannon obtains the extinguishing agent from the storage unit through the pumping unit and executes the spraying. The control cabinet uses an incremental PID control algorithm to correct the water cannon's spraying attitude in real time based on the jet landing point feedback from the high-definition camera.

Images