Intelligent early warning method and system for elevator fire

Abstract

The invention discloses an elevator fire intelligent early warning method and system, and the method comprises the steps: generating a dynamic fire risk index according to smoke concentration, temperature gradient, gas components and infrared image data collected by a multi-mode sensor in an elevator shaft; constructing a multi-dimensional risk feature vector based on the dynamic fire risk index; according to the multi-dimensional risk feature vector, generating an escape path instruction including a stopping floor sequence and a speed curve; and based on infrared thermal imaging data and voice recognition data in the elevator car, the panic level and behavior mode of people are simulated through a confrontation generative network, an elevator door control strategy, a ventilation system regulation and control instruction and a voice guide instruction are generated in combination with the escape path instruction, and intelligent early warning control over elevator fire is achieved. By utilizing the embodiment of the invention, the fire risk can be evaluated in real time, the escape path is optimized, the elevator emergency system is accurately regulated and controlled, and the safe evacuation efficiency in a fire scene is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of elevator early warning, and particularly relates to an intelligent elevator fire early warning method and system. Background Art

[0002] In high-rise building fire accidents, as the core equipment for vertical transportation, the fire early warning and emergency control capabilities of elevators are directly related to the safety of people's lives. Traditional elevator fire early warning systems mainly rely on single smoke sensors or temperature sensors to trigger alarms, suffering from problems such as response lag, high false alarm rate, and lack of precise fire source location, making it difficult to cope with dynamic risks in complex fire scenarios. In the prior art, although the fire detection method based on multi-sensor fusion improves the detection accuracy, it fails to effectively solve the collaborative optimization problems of smoke diffusion path prediction, toxic gas concentration distribution, and elevator escape path planning. In addition, the traditional system is insufficient in identifying and coping with the panic behavior of people in the car, resulting in the disconnection between emergency guidance measures and actual needs. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent elevator fire early warning method and system to solve the deficiencies in the prior art, capable of real-time evaluating fire risks, optimizing escape paths, and precisely regulating the elevator emergency system, thereby improving the safety evacuation efficiency in fire scenarios.

[0004] An embodiment of the present application provides an intelligent elevator fire early warning method, which includes: According to the smoke concentration, temperature gradient, gas composition, and infrared image data collected by multi-modal sensors in the elevator shaft, using a spatio-temporal attention fusion model, by fusing the smoke diffusion rate and the thermal radiation distribution characteristics, a dynamic fire risk index is generated; Based on the dynamic fire risk index, a spatio-temporal convolutional neural network is used to decompose the fire source location characteristics, smoke diffusion path characteristics, and toxic gas concentration characteristics, and a multi-dimensional risk feature vector is constructed. Among them, the spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections; According to the multi-dimensional risk feature vector, an ant colony optimization algorithm is used to plan the dynamic escape path of the elevator. By introducing the fire spread prediction parameter to correct the path weight, an escape path instruction including the stop floor sequence and speed curve is generated; Based on the infrared thermal imaging data and voice recognition data in the elevator car, through an adversarial generation network, the panic level and behavior pattern of people are simulated, and combined with the escape path instruction, an elevator door control strategy, a ventilation system regulation instruction, and a voice guidance instruction are generated to realize intelligent elevator fire early warning control.

[0005] Optionally, based on the smoke concentration, temperature gradient, gas composition, and infrared image data collected by the multimodal sensors in the elevator shaft, a spatio-temporal attention fusion model is used to generate a dynamic fire risk index by fusing the smoke diffusion rate and the thermal radiation distribution characteristics, including: According to the multimodal sensor data, a time series alignment algorithm is used to denoise and synchronize the smoke concentration and temperature gradient, and a spatio-temporally aligned sensor data stream is generated; Based on the sensor data stream, the thermal radiation diffusion direction of the infrared image is calculated through the optical flow field, and the smoke diffusion rate is deduced by combining the gas composition data to generate a correlation matrix of thermal radiation-smoke diffusion; The correlation matrix is input into the spatio-temporal attention fusion model, the weights of the smoke concentration, temperature gradient, and thermal radiation are dynamically allocated, and a multimodal fusion feature vector is output; Based on the multimodal fusion feature vector, an exponential decay model is used to predict the fire growth trend, and a dynamic fire risk index map containing three-dimensional coordinates and hazard level labels is generated.

[0006] Optionally, based on the dynamic fire risk index, a spatio-temporal convolutional neural network is used to decompose the fire source location characteristics, smoke diffusion path characteristics, and toxic gas concentration characteristics, and a multi-dimensional risk feature vector is constructed. The spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections, including: The dynamic fire risk index map is segmented into spatio-temporal cubes, and the temperature mutation characteristics of the initial combustion point of the fire source are extracted through three-dimensional convolution to generate a spatio-temporal local feature map; Based on the spatio-temporal local feature map, residual connections are used to enhance the gradient transmission of small-scale fire source features, and a multi-scale fire source detection map is output; The multi-scale fire source detection map is processed by dilated convolution to capture the long-range dependence relationship of the smoke diffusion path, and a smoke diffusion probability distribution map is generated; The smoke diffusion probability distribution map is fused with the temporal differential characteristics of the toxic gas concentration, and weighted aggregation is performed through a channel attention mechanism to generate a multimodal fusion risk map; The multimodal fusion risk map is compressed and normalized in channels, and a multi-dimensional risk feature vector containing the fire source coordinates, smoke direction, and gas concentration gradient is output.

[0007] Optionally, according to the multi-dimensional risk feature vector, an ant colony optimization algorithm is used to plan the dynamic escape path of the elevator. By introducing fire spread prediction parameters to correct the path weights, an escape path instruction containing the sequence of stopping floors and the speed curve is generated, including: The multi-dimensional risk feature vector is mapped to the three-dimensional grid model of the elevator shaft, and the fire threat value and smoke density weight of each grid node are initialized; Based on the three-dimensional grid model, a dynamic heuristic function is used to incorporate the fire spread prediction parameters into the path weight matrix of the ant colony algorithm, generating a dynamically updated path cost map; On the path cost map, execute the pheromone update rule, and combine the time decay factor to simulate the impact of fire spread on historical paths, and output the pheromone concentration distribution map; Extract the optimal path according to the pheromone concentration distribution map, and combine the elevator mechanical parameters to generate an escape path instruction including the stop floor sequence and the acceleration curve.

[0008] Optionally, based on the infrared thermal imaging data and voice recognition data in the elevator car, through the generative adversarial network, simulate the panic level and behavior pattern of personnel, and combine the escape path instruction to generate an elevator door control strategy, a ventilation system regulation instruction and a voice guidance instruction, realizing intelligent early warning control of elevator fires, including: Perform human pose estimation on the infrared thermal imaging data in the car, and generate a personnel aggregation heat map through the heat radiation distribution density; Extract the speech rate jitter and frequency mutation parameters in the voiceprint features of the voice data, and quantitatively generate a panic level coefficient; Input the personnel aggregation heat map and the panic level coefficient into the generative adversarial network, simulate the spatio-temporal distribution of personnel pushing behavior, and output a personnel behavior prediction map; Fuse the escape path instruction and the personnel behavior prediction map, dynamically calculate the elevator door opening time and ventilation volume demand, and generate a door control strategy and a ventilation regulation curve; Adjust the intonation and volume of the voice guidance according to the panic level coefficient, generate a voice waveform instruction with decibel adaptation, and synchronously output the door control strategy, the ventilation instruction and the voice guidance instruction.

[0009] Another embodiment of the present application provides an intelligent early warning system for elevator fires, and the system includes: A fusion module, configured to generate a dynamic fire risk index by using a spatio-temporal attention fusion model through fusing the smoke diffusion rate and the heat radiation distribution characteristics according to the smoke concentration, temperature gradient, gas composition and infrared image data collected by multi-modal sensors in the elevator shaft; A decomposition module, configured to decompose the fire source position feature, the smoke diffusion path feature and the toxic gas concentration feature based on the dynamic fire risk index by using a spatio-temporal convolutional neural network to construct a multi-dimensional risk feature vector, wherein the spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections; A planning module, configured to plan an elevator dynamic escape path according to the multi-dimensional risk feature vector by using an ant colony optimization algorithm, and correct the path weight by introducing fire spread prediction parameters to generate an escape path instruction including a stop floor sequence and a speed curve; An early warning module, which is used to simulate the panic level and behavior patterns of people through a generative adversarial network based on the infrared thermal imaging data and voice recognition data inside the elevator car, and generate an elevator door control strategy, a ventilation system regulation instruction and a voice guidance instruction in combination with the escape path instruction, so as to realize intelligent early warning control of elevator fires.

[0010] Another embodiment of the present application provides a storage medium, in which a computer program is stored, and the computer program is set to execute the method described in any one of the above when running.

[0011] Another embodiment of the present application provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is set to run the computer program to execute the method described in any one of the above.

[0012] Compared with the prior art, an intelligent elevator fire early warning method provided by the present invention generates a dynamic fire risk index according to the smoke concentration, temperature gradient, gas composition and infrared image data collected by multi-modal sensors in the elevator shaft; constructs a multi-dimensional risk feature vector based on the dynamic fire risk index; generates an escape path instruction including a sequence of stop floors and a speed curve according to the multi-dimensional risk feature vector; and based on the infrared thermal imaging data and voice recognition data inside the elevator car, simulates the panic level and behavior patterns of people through a generative adversarial network, and generates an elevator door control strategy, a ventilation system regulation instruction and a voice guidance instruction in combination with the escape path instruction, so as to realize intelligent early warning control of elevator fires, thereby being able to evaluate the fire risk in real time, optimize the escape path and accurately regulate the elevator emergency system, and improve the safety evacuation efficiency in a fire scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a hardware structure block diagram of a computer terminal for an intelligent elevator fire early warning method provided by an embodiment of the present invention; Figure 2 It is a flow schematic diagram of an intelligent elevator fire early warning method provided by an embodiment of the present invention; Figure 3 It is a structure schematic diagram of an intelligent elevator fire early warning system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] The embodiments described below with reference to the drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] An embodiment of the present invention first provides an intelligent elevator fire early warning method, which can be applied to an electronic device, such as a computer terminal, specifically, an ordinary computer, etc.

[0016] The following takes running on a computer terminal as an example for a detailed description thereof. Figure 1 The following is a hardware structure block diagram of a computer terminal for an elevator fire intelligent early warning method provided by an embodiment of the present invention. As Figure 1 shown, the computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the memory may include a non-volatile storage medium and an internal memory.

[0017] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any elevator fire intelligent early warning method.

[0018] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0019] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any elevator fire intelligent early warning method.

[0020] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 the structure shown in is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0021] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0022] See Figure 2 , an embodiment of the present invention provides an elevator fire intelligent early warning method, which may include the following steps: S201. Based on the smoke concentration, temperature gradient, gas composition, and infrared image data collected by multimodal sensors in the elevator shaft, using a spatio-temporal attention fusion model, by fusing the smoke diffusion rate and the thermal radiation distribution characteristics, a dynamic fire risk index is generated; Specifically, according to the multimodal sensor data, a time series alignment algorithm can be used to denoise and synchronize the smoke concentration and temperature gradient, generating a spatio-temporally aligned sensor data stream; The multimodal sensors include four types of independent acquisition devices: Smoke concentration sensor (model XSA-2100): It collects the concentrations of PM2.5 and PM10 particles in the shaft every 500 milliseconds (unit: micrograms per cubic meter), and the data format is [timestamp, concentration value].

[0023] Temperature gradient sensor array (model TGA-8L): It consists of 8 thermocouples, vertically distributed at a spacing of 0.2 meters. It records the temperature values at different heights in the shaft every 300 milliseconds (unit: degrees Celsius), and the output data is [timestamp, temperature of height layer 1,..., temperature of height layer 8].

[0024] Gas composition analyzer (model GCA-400): It detects the concentrations of CO (carbon monoxide) and HCN (hydrogen cyanide) every 1 second (unit: ppm), and the data format is [timestamp, CO value, HCN value].

[0025] Infrared thermal imager (model IR-Thermal8): It outputs 10 thermal radiation images with a resolution of 640×480 per second, and the pixel values from 0 to 255 represent the temperature (accuracy ±2°C).

[0026] The core of the time series alignment algorithm is the clock synchronization protocol and the data interpolation technology: Clock synchronization: All sensors receive the PPS signal (Pulse Per Second) of the central processor through the CAN bus (Controller Area Network), and control the local clock error within ±20 milliseconds. For example, the infrared image timestamp 1700000000123 and the smoke data timestamp 1700000000125 are regarded as synchronized.

[0027] Denoising processing: The smoke concentration data uses a sliding window mean filter (Window Length = 0.5 seconds) to eliminate the instantaneous jitter caused by elevator vibration.

[0028] The temperature gradient data suppresses environmental electromagnetic interference through the Kalman filter algorithm (Kalman Filter Gain = 0.8) and retains the true temperature change trend.

[0029] Data resampling: Based on the frame rate of the infrared image (10 Hz), resample the data of smoke concentration (2 Hz), gas concentration (1 Hz), and temperature gradient (3.3 Hz) to 10 Hz using linear interpolation. For example, at the moment of 1700000000100 milliseconds, generate an interpolation (weight 0.5A + 0.5B) by weighting the adjacent smoke concentration values (value A 100 milliseconds ago and value B 200 milliseconds later) according to the distance.

[0030] Output format: Generate 10 sets of synchronous data packets per second, each containing [timestamp (milliseconds), smoke concentration (μg / m³), 8-layer temperature values (°C), CO concentration (ppm), HCN concentration (ppm), infrared image frame index].

[0031] Based on the sensor data stream, calculate the thermal radiation diffusion direction of the infrared image through the optical flow field, and deduce the smoke diffusion rate by combining the gas component data to generate a correlation matrix of thermal radiation - smoke diffusion; Optical Flow Field calculation is used to analyze the thermal radiation transfer direction between consecutive infrared image frames: Thermal radiation direction extraction: Select two adjacent infrared images (such as frame T and T + 100 milliseconds), and use the Farneback dense optical flow algorithm (parameters PyrScale = 0.5, Levels = 3, Winsize = 15) to calculate the motion vector (Δx, Δy) of each pixel.

[0032] Perform direction clustering statistics on the motion vectors: Divide the elevator shaft into a 20×20 grid, and statistically analyze the average direction angle (0° - 360°) of the vectors in each grid to generate a thermal radiation diffusion direction map (Direction Map).

[0033] Example: If most of the vectors in a certain grid point to the upper right (angle 45° ± 10°), mark that the thermal radiation in this area is diffusing right up.

[0034] Deduction of smoke diffusion rate: Concentration gradient calculation based on gas component data: Analyze the change rate of CO concentration along the height direction of the shaft. For example, at a height of 1 meter, CO = 50 ppm, and at a height of 2 meters, CO = 85 ppm, then the vertical diffusion rate = (85 - 50) / 0.5 seconds / 1 meter spacing = 70 ppm / second·meter.

[0035] Horizontal diffusion rate: Calculate the rate (unit: m / s) based on the distance between adjacent smoke concentration sensors (0.5 meters) and the concentration difference, combined with the time difference. For example, the concentration of the left sensor is 100 μg / m 3 and that of the right sensor is 150 μg / m 3, with a time difference of 0.5 seconds, then the horizontal rate = (150 - 100) / 0.5 seconds / 0.5 meters = 200 μg / second·meter.

[0036] ‌Associated matrix generation‌: Construct a three-dimensional tensor of a 20×20 grid (Dimensions: Grid serial number on the X-axis × Grid serial number on the Y-axis × Feature dimension): The feature dimensions include: the angle of thermal radiation direction (angle value), the intensity of thermal radiation (pixel mean), the horizontal smoke rate (μg / second·meter), the vertical smoke rate (μg / second·meter), the CO concentration gradient (ppm / second·meter), and the HCN concentration gradient (ppm / second·meter).

[0037] ‌Matrix filling logic‌: The direction of thermal radiation for each grid is taken from the direction map (Direction Map).

[0038] The smoke rate is obtained by interpolating the data of two adjacent smoke sensors.

[0039] The gas concentration gradient is deduced from the vertical distribution of eight-layer temperature gradient sensors.

[0040] ‌Output example‌: The data of grid (5, 8) is [45°, 180, 200, 70, 85, 30], indicating that the thermal radiation in this area spreads at 45°, the radiation intensity is 180, the horizontal smoke rate is 200 μg / second·meter, the vertical smoke rate is 70 μg / second·meter, the CO gradient is 85 ppm / second·meter, and the HCN gradient is 30 ppm / second·meter.

[0041] Input the associated matrix into the spatio-temporal attention fusion model, dynamically allocate the weights of smoke concentration, temperature gradient, and thermal radiation, and output a multi-modal fusion feature vector; ‌Spatio-Temporal Attention Fusion Model‌ is a three-channel neural network: ‌Model architecture design‌: Input layer‌: Receive a 20×20×6 associated matrix (6 feature dimensions).

[0042] ‌Spatio-Temporal Attention Module‌: Spatial Attention‌: Generate the importance weight (0 - 1) of each grid through a 1×1 convolutional layer. For example, the weight of the smoke concentration feature in the high-temperature area is increased to 0.9, and decreased to 0.3 in the low-temperature area.

[0043] ‌Temporal Attention‌: Analyze the feature change trend of 5 consecutive data packets (0.5 seconds). If a certain feature continues to increase (such as the continuous rise of CO concentration), its temporal weight is increased to 1.2 times.

[0044] Feature Fusion Mechanism: Perform 3D convolution (KernelSize 3×3×3) on the weighted smoke, temperature, and thermal radiation features respectively to extract local spatio-temporal features.

[0045] Automatically learn the fusion ratio of the three modalities through the channel attention mechanism (SENet module, compression ratio CR = 16). For example, at the initial stage of a fire, the weight of thermal radiation accounts for 60%, smoke accounts for 30%, and temperature accounts for 10%; after the fire spreads, the weight of smoke rises to 50%.

[0046] Dynamic Weight Allocation Example: When the temperature gradient in a certain grid is detected to be >10°C / second (mutation threshold), the attention weight of the temperature feature rises from the baseline value of 0.4 to 0.8.

[0047] If the angle between the thermal radiation diffusion direction and the smoke diffusion direction in a certain area is <30°, it is determined that the fire spread is consistent, and the fusion weight is increased by 20%.

[0048] Output Feature Vector: Flatten the fused feature map into a 2400-dimensional vector (20×20×6).

[0049] Compress it to a 256-dimensional multi-modal fusion feature vector through a fully connected layer (number of neurons 1024), and output it 10 times per second.

[0050] Vector Example: [0.76, -1.2, 0.33, ..., 2.1] (256 floating-point numbers), where positive values indicate an increase in risk and negative values indicate a decrease in risk.

[0051] Based on the multi-modal fusion feature vector, use the exponential decay model to predict the fire growth trend and generate a dynamic fire risk index map containing three-dimensional coordinates and danger level labels.

[0052] Exponential Decay Model: Trained with historical fire data to predict the fire evolution in the next 30 seconds: Fire Growth Trend Prediction: Input Feature: 256-dimensional multi-modal fusion feature vector.

[0053] Prediction Engine: Three-layer fully connected neural network (number of neurons 256 - 128 - 64), outputting three core parameters: Base Intensity: 0 - 100, the initial fire source energy value.

[0054] Spread Factor: 0.1 - 5.0, the larger the value, the faster the spread.

[0055] Decay Constant: 0.01 - 0.5, reflecting the flame retardancy of the material (the smaller the value, the longer the fire persists).

[0056] Prediction formula logic: Fire intensity in the next t seconds = Fire intensity base × e^(Spread rate coefficient × t) × e^(-Decay constant × t).

[0057] For example, when the base = 60, the spread coefficient = 0.8, and the decay constant = 0.1, the intensity after 10 seconds = 60×e 8 ×e -1 ≈60×2980×0.9≈160,920 (normalized to 0 - 100).

[0058] 3D risk map generation: Spatial mapping: Map the positions in the 20×20 grid correlation matrix to the three - dimensional coordinate system of the elevator shaft (X / Y - axis grid coordinates, and the Z - axis height is determined by the number of layers of the temperature gradient sensor).

[0059] Hazard level label: Level 1 (low risk): Predicted intensity < 30, marked blue.

[0060] Level 2 (medium risk): 30 - 60, marked yellow.

[0061] Level 3 (high risk): > 60, marked red.

[0062] Dynamic update mechanism: Refresh the map every 0.1 second, and mark the real - time risk index (0 - 100 integer) and predicted level for each 1 - cubic - meter grid.

[0063] Example of map output: The label for the grid (X = 3.2 m, Y = 1.5 m, Z = 5 m) is: { "risk_index": 73, "level": 3, "trend": "increasing", / / Growth trend "next_30s_peak": 89 / / Predicted peak in the next 30 seconds }.

[0064] A total of 1600 grid data (20×20×4 layers) are output for the entire shaft and transmitted to the control center via Gigabit Ethernet.

[0065] This step integrates the real-time monitoring data of multi-source sensors in the elevator shaft and uses a spatio-temporal attention mechanism to dynamically weight and fuse the smoke diffusion speed and heat radiation distribution. The spatio-temporal attention model can automatically identify the key feature regions in the fire development process, such as the rapidly spreading smoke front or the abnormally high-temperature region, and quantify these features into a computable fire risk index. This index not only includes the current fire status but also predicts the fire development trend through time series analysis, achieving the precise capture and quantitative evaluation of the early fire characteristics and breaking through the lag defect of traditional threshold alarms. By fusing multi-dimensional physical features, the recognition sensitivity to hidden fire sources and rapidly spreading fires is significantly improved, providing a scientific basis for subsequent emergency decision-making.

[0066] S202, based on the dynamic fire risk index, use a spatio-temporal convolutional neural network to decompose the fire source location features, smoke diffusion path features, and toxic gas concentration features, and construct a multi-dimensional risk feature vector, where the spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections; Specifically, the dynamic fire risk index map can be segmented into spatio-temporal cubes, and the temperature mutation features of the initial combustion point of the fire source are extracted through three-dimensional convolution to generate a spatio-temporal local feature map; The dynamic fire risk index map is a continuous data stream composed of a 20×20×4 three-dimensional grid (updated every 0.1 seconds). Each grid contains a risk index value (0 - 100), a danger level label (1 - 3 levels), and a prediction trend (such as "growing"). The spatio-temporal cube segmentation is to cut the map into overlapping small three-dimensional data blocks for refined analysis. Specific operations: Cube size definition: Each cube covers a grid area of 5×5×2 (physical range 1 meter × 1 meter × 1.2 meters), and adjacent cubes overlap by 2 grids (overlap rate 40%) to ensure that the fire source boundary is not truncated.

[0067] Temperature mutation feature extraction: Use a three-dimensional convolution operation (3D Convolution), with a convolution kernel size of 5×5×2 (length × width × time layer), and the stride is set to 2×2×1.

[0068] The convolution kernel design focuses on the temperature gradient threshold: When the temperature difference between adjacent time layers in the grid exceeds 15°C (defined as the mutation threshold), the feature extraction is activated. For example, if it is detected that the temperature in a certain grid rises from 45°C to 62°C within 0.2 seconds, the convolution kernel outputs a high response value (such as 0.9).

[0069] Feature map generation: The output tensor dimension is 10×10×2 (the length and width are each reduced by 50%, and the time layer is retained), and each channel stores a type of feature (such as channel 1: temperature mutation intensity; channel 2: risk index change rate).

[0070] Local feature optimization: Eliminate the sensor magnitude difference through Batch Normalization to accelerate the training convergence.

[0071] Use the ReLU activation function (Rectified Linear Unit) to filter negative responses and only retain significant features (e.g., set to zero when the output value is less than 0).

[0072] Output example: Mark the area with the number CUBE_12 in the spatio-temporal local feature map, where the temperature mutation intensity at the grid (3, 7) is 0.87 (a high-probability fire source point).

[0073] Based on the spatio-temporal local feature map, use residual connections to enhance the gradient transmission of small-scale fire source features and output a multi-scale fire source detection map; Small-scale fire sources (such as initial flames within 0.5 meters) are easily diluted in the deep network due to fewer covered grids. Residual connections retain the original feature details through cross-layer direct connection paths: Multi-scale feature extraction architecture: Main path: Three layers of three-dimensional convolution (the convolution kernel size decreases: 5×5×2 → 3×3×2 → 3×3×1), gradually expanding the receptive field to capture a large range of fire situations.

[0074] Residual bypass: Directly connect from the input layer to the output of the third layer and stack the original small-scale features (such as the fine temperature distribution of a 5×5 grid).

[0075] Feature fusion method: Element-wise addition of the main path output and bypass features. For example, the main path output value of 0.6 is stacked with the bypass value of 0.3, and after fusion, it is 0.9, strengthening the fire source signal.

[0076] Example of enhancing small-scale fire sources: When detecting a flame that only occupies a 3×3 grid (diameter about 0.6 meters), the main path may lose details due to multiple downsamplings (output response 0.2), but the residual bypass retains the original response of 0.7, and after stacking, it is increased to 0.9.

[0077] Output multi-scale fire source detection map: The final output is a four-dimensional tensor of 10×10×2×4: The first two dimensions: Spatial grid positions (10×10).

[0078] The third dimension: Time layer (2 layers, covering 0.2 seconds).

[0079] Fourth Dimension: Feature Channels (the 4 channels respectively represent fire source intensity, combustion stability, diffusion trend, and confidence score).

[0080] Visualization Example: The value of Channel 1 at grid (5,5) is 0.92, indicating a high-confidence fire source at this location.

[0081] Perform dilated convolution processing on the multi-scale fire source detection map to capture the long-range dependencies of the smoke diffusion path and generate a smoke diffusion probability distribution map; Smoke diffusion often spans several meters, and long-distance grid associations need to be modeled. Dilated Convolution expands the receptive field through spaced sampling: Dilated Convolution Parameter Configuration: Adopt three levels of dilation rates: 1, 3, 5 (covering 3×3, 7×7, and 11×11 grids respectively).

[0082] The convolution kernel has a fixed size of 3×3, but the actual coverage range doubles with the dilation rate: Dilation Rate 1: Adjacent grids (covering 1m×1m).

[0083] Dilation Rate 3: Sample with an interval of 2 grids (covering 3m×3m).

[0084] Dilation Rate 5: Sample with an interval of 4 grids (covering 5m×5m).

[0085] Long-range Dependence Modeling Logic: Calculate the smoke flow consistency in the area around the fire source point: If multiple dilated convolution layers detect smoke diffusing in the same direction (such as directly upward), it is determined as an effective path.

[0086] Probability Distribution Calculation: Each grid outputs the diffusion probabilities in four directions (up, down, left, right), the probability values range from 0 to 1, and the sum is 1. For example, the probability distribution of grid (8,9) is [up: 0.6, down: 0.1, left: 0.2, right: 0.1], indicating that the smoke mainly diffuses upward.

[0087] Anti-interference Mechanism: Elevator Airflow Correction: If the elevator is in the upward state, forcefully reduce the downward diffusion probability by 20% (airflow inhibition effect).

[0088] Structural Obstacle Shielding: Preset the coordinates of the shaft wall, and reset the probability to zero when the smoke hits the wall.

[0089] Output Example: The smoke diffusion probability distribution map marks grid (8,9) as a key diffusion node, with a dominant direction probability of 0.6.

[0090] Fuse the smoke diffusion probability distribution map with the temporal differential features of the toxic gas concentration, and through the channel attention mechanism, perform weighted aggregation to generate a multi-modal fusion risk map; The temporal differential feature quantifies the change speed of the gas concentration. Combining it with the smoke diffusion path can predict the threat of poisonous gas: Gas feature extraction: Calculate the first derivative (change rate) of the CO and HCN concentrations per second. For example, if the CO concentration rises from 50 ppm to 60 ppm, then the change rate = 10 ppm / second.

[0091] Extract the second derivative (acceleration). If the change rate increases from 5 ppm / second to 15 ppm / second, then the acceleration = 10 ppm / second 2 (Indicating the outbreak of poisonous gas).

[0092] Channel Attention: Feature splicing: Splice the smoke diffusion map (4 channels) with the gas features (CO change rate, acceleration, HCN change rate, acceleration, a total of 4 channels) into an 8-channel tensor.

[0093] Dynamic weight allocation: Compress the spatial information through Global Average Pooling to generate an 8-dimensional channel description vector.

[0094] Learn the weights through a two-layer fully connected network (the number of neurons is 8→4→8). For example, the weight of the channel in the dominant direction of the smoke = 0.7, and the weight of the CO acceleration channel = 0.9.

[0095] Weighted output: Multiply the features of each channel by the weights and then sum them. For example, the probability of the smoke spreading upward 0.6×weight 0.7 + CO acceleration 0.8×weight 0.9 = fusion value 1.14.

[0096] Generation of multi-modal risk map: Output a 2D grid map of 10×10×1 (collapse of the time layer), and each grid stores the comprehensive risk value (a floating point number from 0 to 10).

[0097] Risk value mapping rule: >8 is a fatal risk (red alert), 5-8 is a high risk (orange), <5 is a warning (yellow).

[0098] Example: Grid (6,2) has a CO acceleration of 12 ppm / second 2 and the smoke spreads upward, with a risk value = 9.3, triggering a red alert.

[0099] Perform channel compression and normalization on the multi-modal fusion risk map, and output a multi-dimensional risk feature vector containing the fire source coordinates, smoke direction, and gas concentration gradient.

[0100] Channel Compression converts the two-dimensional risk map into a structured vector for easy parsing by the path planning algorithm: Feature Dimensionality Reduction and Screening: Principal Component Analysis (PCA): Retaining 95% of the information, compressing the 10×10 = 100-dimensional space features to 20 dimensions.

[0101] Key Feature Extraction: Fire Source Coordinates: Take the three-dimensional coordinates of the center point of the grid with a risk value > 7 (e.g., [3.2m, 1.5m, 5.0m]).

[0102] Dominant Smoke Direction: Statistically, in the grids with a probability > 0.5, find the direction with the highest proportion (e.g., "upward" accounts for 70%).

[0103] Gas Gradient Extreme Value: Record the position of the maximum change rate of CO (e.g., 15 ppm / second at grid (4,7)).

[0104] Normalization: Divide the coordinate values by the shaft dimensions (length 10 m, width 3 m, height 25 m) and scale them to the [0,1] interval. For example, X = 3.2 → 0.32.

[0105] Divide the gas change rate by the upper limit of the range (CO upper limit 100 ppm / second, HCN upper limit 50 ppm / second). For example, 15 ppm / second → 0.15.

[0106] Structured Vector Output: Finally, generate a 48-dimensional feature vector, which in order contains: Fire source coordinates (3 dimensions: X, Y, Z) Main smoke direction (4 dimensions: probability values of up / down / left / right) Gas gradient (8 dimensions: maximum change rates of CO and HCN in four directions) Global statistics (33 dimensions: spatial distribution features after PCA compression) Update Frequency per Second: Generate a vector every 0.1 seconds and transmit it to the path planning module via the real-time bus.

[0107] Example Vector Fragment: [0.32, 0.15, 0.20, 0.61, 0.05, 0.22, 0.12,...] (The first 3 dimensions are the fire source coordinates, and the last 4 dimensions are the smoke direction probabilities).

[0108] This step uses a deep learning network with a residual structure to perform three-dimensional spatial analysis of fire risks. The network extracts fire features at different scales through hierarchical convolution, where the residual connections ensure the effective transmission of weak signals from early small fire sources. The spatio-temporal convolution module synchronously analyzes the fire source point, the direction of smoke flow, and the diffusion gradient of toxic gases, and finally outputs a structured feature vector containing spatial coordinates and hazard intensity, solving the pain points of traditional detection methods such as fuzzy fire source positioning and inaccurate prediction of smoke diffusion. The residual structure improves the detection accuracy for small-scale fire sources, and the multi-dimensional feature vector provides an accurate environmental threat model for path planning.

[0109] S203. According to the multi-dimensional risk feature vector, use the ant colony optimization algorithm to plan the dynamic escape path of the elevator, and correct the path weight by introducing the predicted parameters of fire spread to generate an escape path instruction containing the sequence of stopping floors and the speed curve. Specifically, the multi-dimensional risk feature vector can be mapped to the three-dimensional grid model of the elevator shaft, and the fire threat value and smoke density weight of each grid node are initialized. The three-dimensional grid model divides the elevator shaft into cube units of 0.5 m × 0.5 m × 0.5 m (a total of 10 m in length × 3 m in width × 25 m in height = 3000 grids). The multi-dimensional risk feature vector contains 48-dimensional data (fire source coordinates, smoke direction probability, gas gradient, etc.), and needs to be mapped to the grid nodes according to the spatial position: Calculation of fire threat value: Based on the fire source coordinates (such as [3.2m, 1.5m, 5.0m]), a three-dimensional Gaussian decay model is constructed. With the fire source as the center, the threat value decays exponentially with distance: the threat value at a distance of 1 meter = 80 (full score 100), 2 meters = 60, 3 meters = 40. For example, the grid (6, 3, 10) is 2.3 meters away from the fire source, and the threat value = 55.

[0110] Initialization of smoke density weight: According to the probability of the main direction of smoke (as above: 0.6), a direction-weighted diffusion model is constructed. If the grid is located in the main direction of smoke (such as the upwind direction), the weight is increased by 30%. For example, the grid directly above the fire source has an initial weight = 1.3 (base value 1.0 + 0.3) due to the smoke upward diffusion probability of 0.6.

[0111] Superposition of gas concentration: If the CO change rate in the grid > 10 ppm / second (high toxicity threshold), the threat value is increased by an additional 20 points. For example, the CO change rate of the grid (7, 2, 8) is 15 ppm / second, and the threat value is increased from 50 to 70.

[0112] Data structure of grid nodes: Each node stores [threat value (0 - 100), smoke weight (0.5 - 2.0), gas toxicity flag (0 / 1)], and is updated 10 times per second.

[0113] Based on the 3D grid model, a dynamic heuristic function is used to incorporate the fire spread prediction parameters into the path weight matrix of the ant colony algorithm, generating a dynamically updated path cost map; The ant colony optimization algorithm (ACO) simulates the swarm intelligence behavior of ants searching for food and needs to dynamically adjust the path cost in combination with the fire situation: Design of dynamic heuristic function: Base Cost: Euclidean distance between grids (e.g., the cost of 0.5 meters between adjacent grids = 5).

[0114] Fire threat correction: When the grid threat value > 60, the cost is multiplied by a coefficient of 1.5 (detour in high-risk areas); when the threat value < 30, the coefficient is 0.8 (priority in safe areas).

[0115] Time Decay Factor: If the fire spread prediction shows that the threat value in a certain area rises to 80 after 10 seconds, the current cost increases by 20% of the future threat (evade in advance).

[0116] Generation of path weight matrix: Construct a 3000×3000 fully connected weight matrix (Weight Matrix) to store the movement cost between grids.

[0117] Real-time update mechanism: Recalculate the weights every 0.1 second according to the latest threat value. For example, the cost from grid A (threat 70) to grid B (threat 30) = base cost 5 × threat correction 1.5 (risk in area A) × 0.8 (safety in area B) × 1.2 (future threat) = 7.2.

[0118] Output of path cost map: Visualize the cost heat map of all feasible paths in the shaft. Areas with high cost ( > 10) in red need to be avoided, and areas with low cost ( < 3) in green are recommended for passage.

[0119] Execute the pheromone update rule on the path cost map, simulate the impact of fire spread on historical paths in combination with the time decay factor, and output the pheromone concentration distribution map; Pheromone is a virtual chemical substance that marks the quality of paths in the ant colony algorithm: Pheromone update rule: Ant exploration mechanism: Release 1000 virtual ants, and each randomly wanders from the current position of the elevator to the target floor (such as the exit on the first floor).

[0120] Pheromone Deposition: When an ant passes through a low-cost path (cost < 4), it leaves pheromone on the grids along the way, and the deposition amount = 10 / path cost. For example, when passing through a grid with a cost of 2, the deposition amount = 5.

[0121] Pheromone Evaporation: The evaporation rate is 15% per second (Time Decay Factor = 0.85), simulating the spread of fire to cover old paths. For example, if a grid originally has 8 pheromones, it drops to 6.8 after 1 second.

[0122] Modeling the Impact of Fire Spread: If the fire spreads to a certain grid (threat value > 70), the pheromone in this grid is forced to zero (the path fails).

[0123] Chain Reaction Simulation: The pheromone evaporation rate within 3 meters around high-threat grids (> 60) is increased to 25% (accelerated disappearance).

[0124] Pheromone Concentration Distribution Map: Output the concentration values (0 - 10) of 3000 nodes. For example, the concentration of grid (5, 1, 15) = 7.2 (high-quality path), and the concentration of grid (8, 2, 20) = 0 (fire source restricted area).

[0125] The optimal path appears as a continuous high-concentration band (concentration > 5), extending from the elevator position to the safety exit.

[0126] Extract the optimal path according to the pheromone concentration distribution map, and generate an escape path instruction including the sequence of stopping floors and the acceleration curve in combination with the elevator mechanical parameters.

[0127] The extraction of the optimal path follows the "Maximum Pheromone Continuity Principle": Path Search Algorithm: Depth-First Search (DFS) starts from the current position and selects the adjacent grid with the highest concentration to progress until reaching the target floor (such as the 1st floor).

[0128] Dynamic Pruning Rule: If the concentration of three consecutive grids in the path < 3 (low-quality path), backtrack to the bifurcation point and search again.

[0129] Integration of Mechanical Parameter Constraints: Maximum Acceleration: The elevator rated value is 1.5m / s 2 , and the acceleration of the path curve shall not exceed the limit.

[0130] Decision on Stopping Floors: If the path passes through an intermediate floor (such as the 5th floor) and the concentration > 6, add a stopping point (for personnel evacuation).

[0131] If the concentration of the path to the target floor < 4 (high risk), automatically switch to an alternative floor (such as the B1 basement floor).

[0132] ‌Escape path instruction generation‌: Stop sequence‌: such as ["currently 8th floor → 5th floor → 1st floor"] (stop point concentration is > 6).

[0133] ‌Acceleration curve‌: The acceleration value is defined in segments with an interval of 0.1 seconds (unit: m / s 2 ).For example: { "8th floor to 5th floor": [0.8, 1.2, 1.5, ...], / / acceleration section "Stop on the 5th floor": [0, 0, 0, ...], / / Stop for 10 seconds "5th floor to 1st floor": [-1.0, -1.3, ...]}. / / deceleration section ‌Safety Verification‌: Absolute value of acceleration throughout the entire process ≤ 1.5m / s 2 , the stop should be >5 meters away from the fire source.

[0134] This step converts the fire environment characteristics into spatial constraints for path planning, and realizes dynamic path optimization by improving the ant colony algorithm. The algorithm integrates the spread speed data output by the fire prediction model in real time, and dynamically adjusts the travel cost weights of the path nodes. The output escape instructions not only include the target floor sequence, but also accurately specify the acceleration curve of the elevator in each section, breaking through the limitations of fixed escape plans and realizing minute-level dynamic updates of escape paths. The introduction of fire prediction parameters makes path planning forward-looking, and can gain an additional safe escape time window on average.

[0135] S204, based on the infrared thermal imaging data and voice recognition data in the elevator car, the panic level and behavior pattern of the personnel are simulated through the adversarial generative network, and the elevator door control strategy, ventilation system control instructions and voice guidance instructions are generated in combination with the escape path instructions to realize intelligent early warning control of elevator fire.

[0136] Specifically, the infrared thermal imaging data in the car can be used to estimate the posture of the human body, and the heat map of the gathering of people can be generated through the thermal radiation distribution density; The infrared thermal imager (model IR-Tracker5, resolution 320×240, frame rate 15fps) installed on the top of the car captures the temperature distribution of passengers in real time (accuracy ±0.5℃). Human posture estimation is divided into three layers of processing: ‌Joint key point positioning‌: Adopt the ‌OpenPose bone recognition algorithm‌ (18 key point model) to identify the head, shoulders, elbows and other parts through convolutional neural networks. For example, the coordinates of a passenger’s right elbow joint are detected to be (120,85) and the coordinates of the left knee are (95,150).

[0137] Anti-occlusion optimization: When multiple people overlap, individuals are separated by the difference in thermal radiation intensity. If the temperature difference between the torso areas of two people > 2°C, they are judged as independent individuals (e.g., passenger A has a body temperature of 37°C and passenger B has a body temperature of 39°C).

[0138] Aggregation density analysis: Divide the car floor into a 10×10 grid (each grid is 0.2m×0.2m), and count the number of key points in each grid.

[0139] Thermal radiation density weight: Higher weight is given to high-temperature areas (>38°C) (weight multiple W = 1.5) because high body temperature is usually accompanied by increased activity. For example, a certain grid contains 2 key points and the average temperature is 39°C, and the density value = 2×1.5 = 3.

[0140] Thermogram generation: Output a pseudo-color image: blue (density 0 - 1), green (1 - 3), red (>3).

[0141] Example: The grid near the car door (coordinate 3,7) is crowded with 5 people, and the density value = 8, marked as a dark red patch.

[0142] Extract the parameters of speech rate jitter and frequency mutation in the voiceprint features of voice data, and quantitatively generate a panic level coefficient; The car microphone array (sampling rate 16kHz) collects voice data and quantifies panic through three-layer feature extraction: Speech rate jitter analysis: Short-time energy segmentation: Detect speech paragraphs (SpeechSegment) with a 0.5-second window, and calculate the number of syllables per second (Syllables Per Second, SPS).

[0143] Jitter Index: The standard deviation of the difference in SPS between adjacent paragraphs. In the normal state, SPS = 4±0.5 (e.g., "Please calm down" contains 3 syllables / 0.8 seconds), and during panic, SPS rises to 6±1.2 (e.g., "Open the door quickly" contains 5 syllables / 0.7 seconds). When the jitter index exceeds the threshold of 0.8, an alarm is triggered.

[0144] Frequency mutation detection: Fundamental frequency extraction (Fundamental Frequency, F0): Use the YIN algorithm to calculate the vocal cord vibration frequency. For normal male voices, F0 = 120Hz±20, and for female voices, F0 = 220Hz±30.

[0145] Mutation parameter: When F0 changes > 50Hz within 0.2 seconds (e.g., 120Hz → 180Hz), it is judged as a scream feature.

[0146] ‌Panic Coefficient Synthesis‌: Formula Logic: The panic level coefficient P = 0.6×Normalized Jitter Index + 0.4×Normalized Mutation Frequency‌ ‌Classification Example‌: P = 0.3 (Calm): Jitter Index 0.4, Mutation Frequency 1 time / 10 seconds; P = 0.8 (Severe Panic): Jitter Index 1.0, Mutation Frequency 5 times / 10 seconds.

[0147] Input the personnel aggregation heat map and the panic level coefficient into the generative adversarial network to simulate the spatio-temporal distribution of personnel pushing behaviors and output a personnel behavior prediction map; ‌Generative Adversarial Network (GAN)‌ consists of a generator and a discriminator, and learns the pushing behavior pattern through a game: ‌Data Input Structuring‌: Generator Input: Heat map matrix (10×10 grid density values) Panic coefficient P (scalar) Current car position (e.g., "Ascending on the 8th floor") ‌Discriminator Input‌: Pushing frames annotated from real accident videos (Source: Fire drill database).

[0148] ‌Behavior Simulation Mechanism‌:‌Physical Engine Constraint‌: The generator simulates human collisions based on the Mass-Spring Model. When the grid density > 5 and P > 0.7, trigger the pushing dynamics calculation: Thrust F = 10N×P (panic coefficient) Direction of action: From the high-density area to the car door direction ‌Spatio-Temporal Distribution Output‌: Predict the personnel movement trajectory for the next 5 seconds every second (e.g., it is predicted that passenger A will move to the coordinates (5,2) after 3 seconds).

[0149] ‌Discrimination Optimization Loop‌: The discriminator compares the generated behavior with the real data, and feeds back the loss value to adjust the generator parameters. After 10,000 iterations, the generator can output a behavior prediction map with an error < 15%, marking the high-pushing areas (red warning areas) and safe areas (green).

[0150] Fuse the escape path instructions and the personnel behavior prediction map, dynamically calculate the elevator door opening timing and ventilation volume requirements, and generate a door control strategy and a ventilation regulation curve; The door control and ventilation strategies need to coordinate the path safety and the passenger state: ‌Door Opening Timing Decision‌: Safety Conditions‌: The threat value of the target floor grid < 30 (low fire risk) and the car has stopped steadily.

[0151] Behavior Constraint: If the behavior prediction map shows that the probability of shoving in the car door area > 60%: Delay the door opening by 0.5 seconds and play the voice "Please step back". Execute the door opening after the probability of shoving drops below 30%. Example Decision Chain: Path Instruction: Docking at the 5th floor → Checking the threat value on the 5th floor = 25 (safe) → Predicting the probability of car door shoving = 70% → Delaying the door opening → The probability drops to 25% after 0.5 seconds → Executing the door opening.

[0152] Dynamic Regulation of Ventilation Volume: Basic Air Volume: Set the minimum ventilation volume of 200 m 3 ) according to the car volume (5 m 3 / h.

[0153] Panic Increment: When the panic coefficient P > 0.5, an additional air volume of ΔV = 300×P m 3 / h (for example, when P = 0.8, the total air volume = 200 + 240 = 440 m 3 / h).

[0154] Toxic Gas Response: If toxic gas (CO > 50 ppm) is detected in the escape path, an additional HEPA filtration mode (wind speed increased by 40%) is added.

[0155] Regulation Curve Generation: Define the ventilation volume and door control state at intervals of 0.1 second: { "time_s": [0.0, 0.1, 0.2,...], / / Time series "vent_m3h": [200, 210, 440,...], / / Ventilation volume "door_state": [0, 0, 1,...]}. / / 0 closed, 1 open Adjust the intonation and volume of the voice guidance according to the panic level coefficient, generate a voice waveform command adapted to the decibels, and synchronously output the door control strategy, ventilation command, and voice guidance command.

[0156] The voice synthesis system (TTS engine) dynamically adjusts the output according to the panic coefficient: Intonation Mapping Rule: Calm State (P < 0.3): Use a flat intonation (fundamental frequency F0 = 180 Hz ± 10), speech rate 3 words per second ("Please move towards the exit").

[0157] Panic State (P > 0.6): Adopt a rising intonation for emphasis (F0 peak value 250 Hz), speech rate increased to 5 words per second ("Leave the elevator immediately!").

[0158] Volume dynamic adaptation: Basic volume: Set the voice to 70 dB when the background noise in the car is 60 dB.

[0159] Panic compensation: For every 0.1 increase in the panic coefficient, the volume is increased by 5 dB (e.g., when P = 0.8, the volume = 70 + 5×3 = 85 dB).

[0160] Multi-instruction synchronous output: Protocol encapsulation: Send a structured instruction packet through the RS485 bus, including: Door control instruction (switch state / delay time); Ventilation curve (air volume / filtration mode); Voice waveform (encoding format PCM 16bit).

[0161] Real-time guarantee: Instruction cycle 100 ms (updated every 0.1 second), transmission delay < 50 ms.

[0162] This step establishes a digital twin model of human behavior. By analyzing the personnel distribution density in the thermal imaging and the stress characteristics in the voice, it predicts possible crowded stampede behaviors. The generative adversarial network simulates the personnel movement patterns under different panic levels, and accordingly optimizes the opening and closing rhythm of the elevator door, the directional smoke exhaust strategy of the ventilation system, and the differentiated voice soothing content, realizing the closed-loop collaborative control of "environment - equipment - personnel", and reducing the risk of casualties. The dynamic voice guidance system can adjust the soothing strategy according to the psychological state of passengers, significantly improving the organization efficiency of emergency evacuation.

[0163] It can be seen that according to the smoke concentration, temperature gradient, gas composition and infrared image data collected by the multi-modal sensors in the elevator shaft, a dynamic fire risk index is generated; based on the dynamic fire risk index, a multi-dimensional risk feature vector is constructed; according to the multi-dimensional risk feature vector, an escape path instruction including a sequence of stopping floors and a speed curve is generated; based on the infrared thermal imaging data and voice recognition data in the elevator car, the generative adversarial network simulates the panic level and behavior patterns of personnel, and combines the escape path instruction to generate an elevator door control strategy, a ventilation system regulation instruction and a voice guidance instruction, realizing the intelligent early warning control of elevator fires, so as to be able to evaluate the fire risk in real time, optimize the escape path and accurately regulate the elevator emergency system, improving the safety evacuation efficiency in the fire scene.

[0164] Another embodiment of the present invention provides an elevator fire intelligent early warning system. Refer to Figure 3 , the system may include: A fusion module 301, configured to generate a dynamic fire risk index by using a spatio-temporal attention fusion model according to the smoke concentration, temperature gradient, gas composition and infrared image data collected by the multi-modal sensors in the elevator shaft, through fusing the smoke diffusion rate and the thermal radiation distribution characteristics; A decomposition module 302, configured to decompose the fire source location feature, the smoke diffusion path feature, and the toxic gas concentration feature based on the dynamic fire risk index by using a spatio-temporal convolutional neural network, and construct a multi-dimensional risk feature vector, wherein the spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections; A planning module 303, configured to plan a dynamic escape path of the elevator according to the multi-dimensional risk feature vector by using an ant colony optimization algorithm, and correct the path weight by introducing a fire spread prediction parameter to generate an escape path instruction including a stop floor sequence and a speed curve; An early warning module 304, configured to simulate the panic level and behavior pattern of personnel through an adversarial generation network based on the infrared thermal imaging data and the speech recognition data in the elevator car, and generate an elevator door control strategy, a ventilation system regulation instruction, and a voice guidance instruction in combination with the escape path instruction to implement intelligent early warning control of elevator fires.

[0165] It can be seen that according to the smoke concentration, temperature gradient, gas composition, and infrared image data collected by the multi-modal sensors in the elevator shaft, a dynamic fire risk index is generated; based on the dynamic fire risk index, a multi-dimensional risk feature vector is constructed; according to the multi-dimensional risk feature vector, an escape path instruction including a stop floor sequence and a speed curve is generated; based on the infrared thermal imaging data and the speech recognition data in the elevator car, the panic level and behavior pattern of personnel are simulated through an adversarial generation network, and an elevator door control strategy, a ventilation system regulation instruction, and a voice guidance instruction are generated in combination with the escape path instruction to implement intelligent early warning control of elevator fires, so as to be able to evaluate the fire risk in real time, optimize the escape path, and accurately control the elevator emergency system, and improve the safety evacuation efficiency in the fire scenario.

[0166] An embodiment of the present invention further provides a storage medium, in which a computer program is stored, and the computer program is configured to execute the steps in any one of the above method embodiments when running.

[0167] Specifically, in this embodiment, the above storage medium may be configured to store a computer program for executing the following steps: S201, according to the smoke concentration, temperature gradient, gas composition, and infrared image data collected by the multi-modal sensors in the elevator shaft, use a spatio-temporal attention fusion model to generate a dynamic fire risk index by fusing the smoke diffusion rate and the thermal radiation distribution feature; S202, based on the dynamic fire risk index, use a spatio-temporal convolutional neural network to decompose the fire source location feature, the smoke diffusion path feature, and the toxic gas concentration feature, and construct a multi-dimensional risk feature vector, wherein the spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections; S203. According to the multi-dimensional risk feature vector, use the ant colony optimization algorithm to plan the dynamic escape path of the elevator, correct the path weight by introducing the fire spread prediction parameter, and generate an escape path instruction including the sequence of stopping floors and the speed curve. S204. Based on the infrared thermal imaging data and voice recognition data in the elevator car, simulate the panic level and behavior pattern of people through the generative adversarial network, and combine the escape path instruction to generate an elevator door control strategy, a ventilation system regulation instruction, and a voice guidance instruction to achieve intelligent early warning control of elevator fires.

[0168] It can be seen that according to the smoke concentration, temperature gradient, gas composition, and infrared image data collected by the multi-modal sensors in the elevator shaft, a dynamic fire risk index is generated; based on the dynamic fire risk index, a multi-dimensional risk feature vector is constructed; according to the multi-dimensional risk feature vector, an escape path instruction including the sequence of stopping floors and the speed curve is generated; based on the infrared thermal imaging data and voice recognition data in the elevator car, simulate the panic level and behavior pattern of people through the generative adversarial network, and combine the escape path instruction to generate an elevator door control strategy, a ventilation system regulation instruction, and a voice guidance instruction to achieve intelligent early warning control of elevator fires, so as to be able to evaluate the fire risk in real time, optimize the escape path, and accurately regulate the elevator emergency system, and improve the safety evacuation efficiency in the fire scenario.

[0169] An embodiment of the present invention also provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0170] Specifically, the above electronic device may further include a transmission device and an input / output device. Among them, the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0171] Specifically, in this embodiment, the above processor may be configured to execute the following steps through a computer program: S201. According to the smoke concentration, temperature gradient, gas composition, and infrared image data collected by the multi-modal sensors in the elevator shaft, use the spatio-temporal attention fusion model to generate a dynamic fire risk index by fusing the smoke diffusion rate and the thermal radiation distribution characteristics. S202. Based on the dynamic fire risk index, use a spatio-temporal convolutional neural network to decompose the fire source position characteristics, smoke diffusion path characteristics, and toxic gas concentration characteristics, and construct a multi-dimensional risk feature vector, where the spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections. S203. According to the multi-dimensional risk feature vector, use the ant colony optimization algorithm to plan the dynamic escape path of the elevator. By introducing the fire spread prediction parameter to correct the path weight, generate an escape path instruction including the stop floor sequence and speed curve. S204. Based on the infrared thermal imaging data and voice recognition data in the elevator car, simulate the panic level and behavior pattern of people through the generative adversarial network, and combine the escape path instruction to generate the elevator door control strategy, ventilation system regulation instruction and voice guidance instruction, so as to realize the intelligent early warning control of elevator fire.

[0172] It can be seen that according to the smoke concentration, temperature gradient, gas composition and infrared image data collected by the multi-modal sensors in the elevator shaft, generate a dynamic fire risk index; based on the dynamic fire risk index, construct a multi-dimensional risk feature vector; according to the multi-dimensional risk feature vector, generate an escape path instruction including the stop floor sequence and speed curve; based on the infrared thermal imaging data and voice recognition data in the elevator car, simulate the panic level and behavior pattern of people through the generative adversarial network, and combine the escape path instruction to generate the elevator door control strategy, ventilation system regulation instruction and voice guidance instruction, so as to realize the intelligent early warning control of elevator fire, thereby being able to evaluate the fire risk in real time, optimize the escape path and accurately control the elevator emergency system, and improve the safety evacuation efficiency in the fire scene.

[0173] The structure, features and effects of the present invention have been described in detail based on the embodiments shown in the drawings. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the scope defined by the drawings. Any changes made according to the concept of the present invention, or modified into equivalent embodiments with equivalent changes, still within the spirit covered by the description and drawings, shall be within the protection scope of the present invention.

Claims

1. An intelligent early warning method for elevator fires, characterized in that, The method includes: According to the smoke concentration, temperature gradient, gas composition, and infrared image data collected by multimodal sensors in the elevator shaft, using a spatio-temporal attention fusion model, by fusing the smoke diffusion rate and the thermal radiation distribution characteristics, a dynamic fire risk index is generated; Based on the dynamic fire risk index, a spatio-temporal convolutional neural network is used to decompose the fire source location characteristics, smoke diffusion path characteristics, and toxic gas concentration characteristics, and a multi-dimensional risk feature vector is constructed. Among them, the spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections; According to the multi-dimensional risk feature vector, an ant colony optimization algorithm is used to plan the dynamic escape path of the elevator. By introducing the fire spread prediction parameter to correct the path weight, an escape path instruction including the stop floor sequence and the speed curve is generated; Based on the infrared thermal imaging data and speech recognition data in the elevator car, through a generative adversarial network, the panic level and behavior pattern of people are simulated, and combined with the escape path instruction, an elevator door control strategy, a ventilation system regulation instruction, and a voice guidance instruction are generated to realize the intelligent early warning control of elevator fires.

2. The method according to claim 1, wherein The step of generating a dynamic fire risk index according to the smoke concentration, temperature gradient, gas composition, and infrared image data collected by multimodal sensors in the elevator shaft, using a spatio-temporal attention fusion model, by fusing the smoke diffusion rate and the thermal radiation distribution characteristics, includes: According to the multimodal sensor data, a time series alignment algorithm is used to denoise and synchronize the smoke concentration and temperature gradient to generate a spatio-temporally aligned sensor data stream; Based on the sensor data stream, the thermal radiation diffusion direction of the infrared image is calculated through the optical flow field, and combined with the gas composition data, the smoke diffusion rate is deduced to generate a correlation matrix of thermal radiation-smoke diffusion; The correlation matrix is input into the spatio-temporal attention fusion model to dynamically allocate the weights of the smoke concentration, temperature gradient, and thermal radiation, and output a multi-modal fusion feature vector; Based on the multi-modal fusion feature vector, an exponential decay model is used to predict the fire growth trend, and a dynamic fire risk index map including three-dimensional coordinates and hazard level labels is generated.

3. The method according to claim 2, wherein The step of constructing a multi-dimensional risk feature vector by decomposing the fire source location characteristics, smoke diffusion path characteristics, and toxic gas concentration characteristics using a spatio-temporal convolutional neural network based on the dynamic fire risk index, where the spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections, includes: The dynamic fire risk index map is segmented into a spatio-temporal cube, and the temperature mutation characteristics of the initial combustion point of the fire source are extracted through three-dimensional convolution to generate a spatio-temporal local feature map; Based on the spatio-temporal local feature map, residual connections are used to enhance the gradient transfer of small-scale fire source features, and a multi-scale fire source detection map is output; The multi-scale fire source detection map is processed by dilated convolution to capture the long-range dependence relationship of the smoke diffusion path, and a smoke diffusion probability distribution map is generated; The smoke diffusion probability distribution map and the temporal differential features of the toxic gas concentration are fused, and weighted aggregation is performed through a channel attention mechanism to generate a multi-modal fusion risk map; Channel compression and normalization are performed on the multi-modal fusion risk map, and a multi-dimensional risk feature vector including the coordinates of the fire source, the direction of the smoke, and the gas concentration gradient is output.

4. The method according to claim 3, characterized in that, According to the multi-dimensional risk feature vector, an ant colony optimization algorithm is used to plan a dynamic escape path for the elevator. By introducing fire spread prediction parameters to correct the path weights, an escape path instruction including a sequence of stop floors and a speed curve is generated, including: The multi-dimensional risk feature vector is mapped to a three-dimensional grid model of the elevator shaft, and the fire threat value and the smoke density weight of each grid node are initialized. Based on the three-dimensional grid model, a dynamic heuristic function is used to incorporate the fire spread prediction parameters into the path weight matrix of the ant colony algorithm, and a dynamically updated path cost map is generated. The pheromone update rule is executed on the path cost map, and the influence of fire spread on the historical path is simulated by combining the time decay factor, and a pheromone concentration distribution map is output. The optimal path is extracted according to the pheromone concentration distribution map, and an escape path instruction including a sequence of stop floors and an acceleration curve is generated in combination with the elevator mechanical parameters.

5. The method according to claim 4, wherein Based on the infrared thermal imaging data and voice recognition data in the elevator car, the panic level and behavior patterns of people are simulated through a generative adversarial network, and a door control strategy, a ventilation system control instruction, and a voice guidance instruction are generated in combination with the escape path instruction to realize intelligent early warning control of elevator fires, including: Perform human pose estimation on the infrared thermal imaging data in the car, and generate a heat map of people gathering through the heat radiation distribution density. Extract the speech rate jitter and frequency mutation parameters in the voiceprint features of the voice data, and quantitatively generate a panic level coefficient. Input the heat map of people gathering and the panic level coefficient into a generative adversarial network to simulate the spatio-temporal distribution of people's pushing behavior, and output a human behavior prediction map. Fuse the escape path instruction and the human behavior prediction map, dynamically calculate the elevator door opening time and the ventilation volume requirement, and generate a door control strategy and a ventilation control curve. Adjust the tone and volume of the voice guidance according to the panic level coefficient, generate a voice waveform instruction with adapted decibels, and synchronously output the door control strategy, the ventilation instruction, and the voice guidance instruction.

6. An intelligent elevator fire warning system, characterized in that, The system includes: A fusion module for generating a dynamic fire risk index by using a spatio-temporal attention fusion model based on the smoke concentration, temperature gradient, gas composition, and infrared image data collected by multi-modal sensors in the elevator shaft, and by fusing the smoke diffusion rate and the heat radiation distribution characteristics. A decomposition module for decomposing the fire source position feature, the smoke diffusion path feature, and the toxic gas concentration feature based on the dynamic fire risk index by using a spatio-temporal convolutional neural network, and constructing a multi-dimensional risk feature vector, wherein the spatio-temporal convolutional neural network enhances the small-scale fire source detection ability through residual connections. A planning module for planning a dynamic escape path for the elevator according to the multi-dimensional risk feature vector by using an ant colony optimization algorithm, and generating an escape path instruction including a sequence of stop floors and a speed curve by introducing fire spread prediction parameters to correct the path weights. An early warning module, which is used to simulate the panic level and behavior patterns of people through a generative adversarial network based on the infrared thermal imaging data and voice recognition data in the elevator car, and generate an elevator door control strategy, a ventilation system regulation instruction and a voice guidance instruction in combination with the escape path instruction, so as to realize intelligent early warning control of elevator fires.

7. The system according to claim 6, characterized in that, The fusion module is specifically used for: According to the multi-modal sensor data, a temporal alignment algorithm is used to denoise and synchronize the smoke concentration and temperature gradient, and a sensor data stream with spatio-temporal alignment is generated; Based on the sensor data stream, the thermal radiation diffusion direction of the infrared image is calculated through the optical flow field, and the smoke diffusion rate is deduced in combination with the gas composition data, and a correlation matrix of thermal radiation-smoke diffusion is generated; The correlation matrix is input into a spatio-temporal attention fusion model to dynamically allocate the weights of the smoke concentration, temperature gradient and thermal radiation, and a multi-modal fusion feature vector is output; Based on the multi-modal fusion feature vector, an exponential decay model is used to predict the fire growth trend, and a dynamic fire risk index map containing three-dimensional coordinates and danger level labels is generated.

8. The system according to claim 7, characterized in that, The decomposition module is specifically used for: The dynamic fire risk index map is segmented into spatio-temporal cubes, and the temperature mutation characteristics of the initial combustion point of the fire source are extracted through three-dimensional convolution to generate a spatio-temporal local feature map; Based on the spatio-temporal local feature map, residual connections are used to enhance the gradient transfer of small-scale fire source features, and a multi-scale fire source detection map is output; The multi-scale fire source detection map is subjected to dilated convolution processing to capture the long-range dependence relationship of the smoke diffusion path, and a smoke diffusion probability distribution map is generated; The smoke diffusion probability distribution map and the temporal differential characteristics of the toxic gas concentration are fused, and weighted aggregation is performed through a channel attention mechanism to generate a multi-modal fusion risk map; The multi-modal fusion risk map is subjected to channel compression and normalization, and a multi-dimensional risk feature vector containing the fire source coordinates, smoke direction and gas concentration gradient is output.

9. A storage medium, characterized in that, A computer program is stored in the storage medium, wherein the computer program is set to execute the method according to any one of claims 1-5 when running.

10. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is set to run the computer program to execute the method according to any one of claims 1-5.

Images