Elevator environment abnormality monitoring method and system

Abstract

The present invention discloses a method and system for monitoring elevator environmental anomalies. The method comprises: using a multimodal sensor array to collect environmental data within the elevator car in real time and generate a dynamic environmental characteristic map; outputting a high-confidence environmental health index based on the dynamic environmental characteristic map; generating a partitioned abnormality probability heat map based on the environmental health index and passenger density detection data; superimposing real-time environmental parameters, the partitioned abnormality probability heat map, and risk warning prompts on the elevator car display screen based on the partitioned abnormality probability heat map; and generating an environmental deterioration warning signal using a spatiotemporal prediction model based on the duration and spatial diffusion trend of the partitioned abnormality probability heat map. When the environmental deterioration warning signal level reaches a preset signal level, a gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy. The embodiments of the present invention can improve the accuracy and response efficiency of elevator environmental safety management.

Description

Technical Field

[0001] The present invention belongs to the technical field of elevator monitoring, and in particular to a method and system for monitoring abnormal elevator environments. Background Art

[0002] With the prevalence of high-rise buildings and the increasing frequency of elevator use, environmental safety issues within elevator cabins are receiving increasing attention. Traditional elevator environmental monitoring primarily relies on single temperature or gas sensors, which suffer from insufficient monitoring dimensions and isolated data analysis, making it difficult to comprehensively assess the overall environmental quality within the cabin. Existing IoT-based monitoring systems, while capable of multi-parameter acquisition, still suffer from significant deficiencies in data processing. Firstly, the spatiotemporal correlations of multi-source sensor data are not fully utilized, resulting in limited anomaly detection accuracy. Secondly, environmental anomaly warning mechanisms lack dynamic adaptability and are unable to intelligently adjust monitoring strategies based on real-time risk levels. Furthermore, current systems often employ fixed threshold alarms, making it difficult to accurately predict environmental deterioration trends, and their coordinated response with building management systems is inefficient. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for monitoring elevator environmental anomalies to address the deficiencies in the existing technology, and to improve the accuracy and response efficiency of elevator environmental safety management through multimodal data fusion, dynamic risk assessment and predictive warning.

[0004] An embodiment of the present application provides a method for monitoring abnormal elevator environment, the method comprising:

[0005] The system collects environmental data in the cabin in real time through a multimodal sensor array, and uses a spatiotemporal convolutional network to fuse the spatial distribution characteristics of the sensor data with the time series fluctuation characteristics to generate a dynamic environmental feature map. The environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds, and voiceprint data.

[0006] Based on the dynamic environmental feature map, a denoised environmental state model is reconstructed through a generative adversarial network. The generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomalies based on a historical normal environment database and outputs a high-confidence environmental health index.

[0007] Based on the environmental health index and passenger density detection data, a multimodal anomaly scoring model is constructed, and a graph neural network is used to correlate the environmental parameter correlations between cabin areas to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the level of anomaly;

[0008] Based on the partition abnormality probability heat map, monitoring resources are dynamically allocated through an adaptive attention mechanism, sensor sampling frequency in high-risk areas is prioritized, and real-time environmental parameters, partition abnormality probability heat map and risk warning prompts are superimposed on the car display screen;

[0009] Based on the duration and spatial diffusion trend of the partitioned anomaly probability heat map, an environmental deterioration warning signal is generated through a spatiotemporal prediction model. When the environmental deterioration warning signal level reaches the preset signal level, the gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, forming a closed-loop monitoring response.

[0010] Optionally, the environmental data in the car is collected in real time through a multimodal sensor array, and the spatial distribution characteristics and time series fluctuation characteristics of the sensor data are fused using a spatiotemporal convolutional network to generate a dynamic environmental feature map, wherein the environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds and voiceprint data, including:

[0011] Based on the original environmental data stream collected by the multimodal sensor array, a sliding time window is used to segment the temperature, humidity, CO2 concentration, volatile organic compounds and voiceprint signals to generate time-aligned multimodal data slices;

[0012] The multimodal data slices are spatially encoded, the car is divided into gridded areas, and the topological relationships of the sensor nodes are annotated to construct a spatiotemporal correlation tensor.

[0013] The spatiotemporal correlation tensor is input into a 3D convolution-long short-term memory hybrid network. The local region features are extracted through the spatial convolution kernel, and the fluctuation pattern across the time window is captured by combining the time axis LSTM. The feature vector set that integrates the spatiotemporal features is output.

[0014] Regional features are aggregated based on the feature vector set, and a graph pooling algorithm is used to generate a dynamic environmental feature map covering the entire area of ​​the car. The graph nodes encode the regional environmental status, and the edge weights represent the correlation between environmental parameters between regions.

[0015] Optionally, based on the dynamic environmental feature map, a denoised environmental state model is reconstructed through a generative adversarial network, wherein the generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomaly in combination with a historical normal environment database and outputs a high-confidence environmental health index, including:

[0016] The dynamic environment feature map is input into the generator network, and the denoised environment state map is reconstructed through the multi-scale convolutional layer with residual connections. The generator output contains the estimated values ​​of the environment parameters after noise suppression.

[0017] The estimated environmental parameters and the standard state spectrum in the historical normal environment database are input into the discriminator network, and the distribution difference between the two is calculated through spectral normalization processing, and the probability score of environmental abnormality is output;

[0018] During the adversarial training phase, the generator reversely optimizes parameters based on the anomaly probability fed back by the discriminator, uses the Wasserstein distance constraint to ensure the authenticity of the generated data, and iteratively improves the denoising capability.

[0019] The parameter deviation of each node in the denoised environmental state map is integrated, and the comprehensive impact weights of temperature, humidity, CO2 concentration, and volatile organic compounds are calculated using the entropy weight method to generate a multidimensional environmental health index.

[0020] The multidimensional environmental health index is normalized and calibrated, and a high-confidence environmental health index in the range of 0-1 is generated by combining the abnormal probability output by the discriminator.

[0021] Optionally, a multimodal anomaly scoring model is constructed based on the environmental health index and passenger density detection data, and a graph neural network is used to correlate the environmental parameter correlations between cabin areas to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the anomaly level, including:

[0022] The environmental health index is spatiotemporally aligned with camera-based crowd density detection data. The environmental health weights and passenger density weights of different areas are dynamically assigned through an attention mechanism to generate a fusion weight matrix.

[0023] A graph neural network model is constructed, mapping the car area into graph nodes, using the correlation of environmental parameters as edge attributes, and modeling the abnormal propagation path between regions through the graph attention layer;

[0024] Update the graph node features according to the fusion weight matrix, use the graph convolutional network to iterate the anomaly propagation signal, calculate the anomaly accumulation value of each node, and generate the initial anomaly score distribution;

[0025] The initial anomaly score is spatially interpolated, and a heat diffusion algorithm is used to simulate the propagation trend of anomalies in the car. A partitioned anomaly probability heat map is output to calibrate the location of high-risk areas and anomaly levels.

[0026] Optionally, based on the partition abnormality probability heat map, monitoring resources are dynamically allocated through an adaptive attention mechanism, the sensor sampling frequency in high-risk areas is preferentially increased, and real-time environmental parameters, the partition abnormality probability heat map and risk warning prompts are superimposed on the car display screen, including:

[0027] The monitoring priority is divided according to the abnormality level in the partition abnormality probability heat map, and the resource allocation coefficient of the high-risk area is dynamically calculated using the attention weight function. The resource allocation coefficient value is exponentially related to the abnormality level.

[0028] Adjust the sampling strategy of the sensor array, enable super-resolution sampling mode for high-risk areas, and increase the sampling frequency of temperature and humidity sensors to three times the baseline value;

[0029] A heat map overlay module is embedded in the car display rendering engine, and real-time environmental parameters, partition abnormality probability heat maps, and camera video streams are integrated at the pixel level using a transparency gradient algorithm.

[0030] Build a risk warning dynamic labeling system. When the abnormality level in a certain area exceeds the preset level, a flashing warning box and voice prompt will be automatically superimposed at the corresponding location.

[0031] Optionally, based on the duration and spatial diffusion trend of the partitioned abnormality probability heat map, an environmental deterioration warning signal is generated through a spatiotemporal prediction model. When the environmental deterioration warning signal level reaches a preset signal level, a gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, forming a closed-loop monitoring response, including:

[0032] Extract the spatiotemporal variation characteristics of the partitioned anomaly probability heat map, model the duration, diffusion speed, and correlation strength of the abnormal area through the spatiotemporal graph convolutional network, and output the spatiotemporal feature vector;

[0033] The spatiotemporal feature vectors are input into a pre-trained LSTM-Transformer hybrid prediction model to predict environmental deterioration trends within a preset number of minutes and generate probabilistic early warning signals.

[0034] A three-level gradient alarm mechanism is designed. When the probabilistic warning signal reaches level one, an audible and visual alarm is triggered in the elevator car. When it reaches level two, the ventilation system is activated to enhance ventilation. When it reaches level three, an emergency elevator dispatch request is sent to the building management system.

[0035] Establish a closed-loop verification mechanism to compare actual environmental change data with prediction results, dynamically update the hybrid prediction model parameters through reinforcement learning, continuously optimize the accuracy of early warnings, and form a closed-loop monitoring response.

[0036] Another embodiment of the present application provides an elevator environment abnormality monitoring system, the system comprising:

[0037] A fusion module is used to collect environmental data in the car in real time through a multimodal sensor array, and use a spatiotemporal convolutional network to fuse the spatial distribution characteristics of the sensor data with the time series fluctuation characteristics to generate a dynamic environmental feature map, where the environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds, and voiceprint data;

[0038] A reconstruction module is used to reconstruct a denoised environmental state model based on the dynamic environmental feature map through a generative adversarial network, wherein the generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomaly based on a historical normal environment database and outputs a high-confidence environmental health index;

[0039] a correlation module, configured to construct a multimodal anomaly scoring model based on the environmental health index and passenger density detection data, and utilize a graph neural network to correlate the environmental parameter correlations between cabin zones to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the level of anomaly;

[0040] A display module is used to dynamically allocate monitoring resources based on the partition abnormality probability heat map through an adaptive attention mechanism, prioritize increasing the sensor sampling frequency in high-risk areas, and superimpose real-time environmental parameters, the partition abnormality probability heat map, and risk warning prompts on the car display screen;

[0041] The early warning module is used to generate an environmental deterioration early warning signal through a spatiotemporal prediction model based on the duration and spatial diffusion trend of the partitioned anomaly probability heat map. When the environmental deterioration early warning signal level reaches the preset signal level, the gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, forming a closed-loop monitoring response.

[0042] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0043] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0044] Compared with the existing technology, the present invention provides a method for monitoring elevator environmental anomalies, which collects environmental data in the car in real time through a multimodal sensor array to generate a dynamic environmental feature map; based on the dynamic environmental feature map, it outputs a high-confidence environmental health index; based on the environmental health index and passenger density detection data, it generates a partitioned abnormality probability heat map; based on the partitioned abnormality probability heat map, it superimposes and displays real-time environmental parameters, partitioned abnormality probability heat map and risk warning prompts on the car display screen; based on the duration and spatial diffusion trend of the partitioned abnormality probability heat map, it generates an environmental deterioration warning signal through a spatiotemporal prediction model; when the environmental deterioration warning signal level reaches a preset signal level, it triggers a gradient alarm mechanism and links the building management system to adjust the elevator scheduling strategy, thereby improving the accuracy and response efficiency of elevator environmental safety management through multimodal data fusion, dynamic risk assessment and predictive warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A hardware structure block diagram of a computer terminal for an elevator environment abnormality monitoring method provided by an embodiment of the present invention;

[0046] Figure 2 A flowchart of a method for monitoring abnormal elevator environment provided by an embodiment of the present invention;

[0047] Figure 3 A schematic structural diagram of an elevator environment abnormality monitoring system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0049] The embodiment of the present invention first provides a method for monitoring abnormal elevator environment. The method can be applied to electronic equipment, such as a computer terminal, specifically a common computer.

[0050] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal of an elevator environment abnormality monitoring method provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0051] 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 one of the elevator environment abnormality monitoring methods.

[0052] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0053] 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 environment abnormality monitoring method.

[0054] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure 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 shown in the figure, or combine certain components, or have a different component arrangement.

[0055] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0056] See also Figure 2 , an embodiment of the present invention provides a method for monitoring abnormal elevator environment, which may include the following steps:

[0057] S201, collecting environmental data in the car in real time through a multimodal sensor array, and fusing the spatial distribution characteristics of the sensor data with the time series fluctuation characteristics using a spatiotemporal convolutional network to generate a dynamic environmental feature map, wherein the environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds, and voiceprint data;

[0058] Specifically, based on the original environmental data stream collected by the multimodal sensor array, a sliding time window can be used to segment temperature, humidity, CO2 concentration, volatile organic compounds and voiceprint signals to generate time-aligned multimodal data slices;

[0059] A multimodal sensor array (model: EnvSense Pro) is installed in the elevator car, consisting of 16 nodes evenly distributed on the ceiling and side walls. The sensor collects temperature (unit: Celsius), humidity (percentage RH), CO2 concentration (unit: ppm), volatile organic compound (VOC) (unit: μg / m 3 ) and voiceprint (16kHz sampling rate). The original data stream needs to solve the problem of time asynchrony:

[0060] Sliding time window design: The window length is fixed at 5 seconds (covering 500 sets of data points), and the step size is 0.5 seconds (i.e., a new window is generated every 0.5 seconds).

[0061] Alignment rules: The CO2 concentration collection time is used as the base timestamp (because it has the slowest response time), and the remaining data is aligned using linear interpolation. For example, if a temperature sensor has a delay of 0.1 seconds, its equivalent value at the base time is calculated through interpolation.

[0062] Multimodal Slice Generation: Each time window outputs a structured data packet containing:

[0063] Temperature: instantaneous values ​​of 16 nodes (e.g., [23.5, 24.1, ..., 22.8]);

[0064] Voiceprint: A 5-second audio waveform is compressed into a 20-dimensional feature vector using Mel-Frequency Cepstral Coefficients (MFCC).

[0065] Example slice: In the slice with timestamp T = 12.5 seconds, the CO2 concentration at the car door is 850ppm (the normal threshold is 1000ppm), and the VOC on the rear wall is 120μg / m 3 (Exceeding the threshold of 80 μg / m 3 ).

[0066] The multimodal data slices are spatially encoded, the car is divided into gridded areas, and the topological relationships of the sensor nodes are annotated to construct a spatiotemporal correlation tensor.

[0067] Spatial location encoding converts physical locations into machine-readable topological structures:

[0068] ‌Car grid partition‌: The car bottom is divided into a 5×3 grid (each grid is 0.4m×0.6m), and divided into 2 layers in the vertical direction (height 1.2m / 2.4m), forming 30 three-dimensional units (5×3×2).

[0069] Sensor Mapping: Each sensor is labeled with the grid number it belongs to (e.g., node 7 is located in the upper layer of grid B3).

[0070] Topological relationship construction: Adjacency Matrix marks the spatial relationship between grids:

[0071] Horizontally adjacent: weight 1.0 (e.g., grids A1 and A2);

[0072] Vertically adjacent: weight 0.8 (e.g., upper and lower layers of A1);

[0073] Diagonally adjacent: weight 0.5 (such as A1 and B2);

[0074] Dynamic Correlation: If the environmental parameters of two grids change synchronously for 3 consecutive seconds (e.g., the temperature difference is less than 0.5°C), the weight is increased by an additional 0.2.

[0075] ‌Spatiotemporal correlation tensor generation‌:

[0076] Output 4D tensor (dimensions: 30 grids × 5 seconds × 5 modes × 3 topology weights):

[0077] Modal dimensions: temperature, humidity, CO2, VOC, and voiceprint features;

[0078] Topological dimension: horizontal / vertical / diagonal relationship weights.

[0079] Example: The temperature of the upper layer of mesh C2 is 24.3°C at T=12 seconds, and its topology weight with the adjacent mesh B2 is 1.0.

[0080] The spatiotemporal correlation tensor is input into a 3D convolution-long short-term memory hybrid network. The local region features are extracted through the spatial convolution kernel, and the fluctuation pattern across the time window is captured by combining the time axis LSTM. The feature vector set that integrates the spatiotemporal features is output.

[0081] The ‌Three-Dimensional Convolution-LSTM Hybrid Network‌ (3D CNN-LSTM) processes in two stages:

[0082] Spatial Feature Extraction (3D CNN): Convolution kernel size 3×3×2 (length×width×time layer), stride 1×1×1, covering adjacent grids and temporal segments.

[0083] Feature Mapping Example: Focus on local high-temperature areas (e.g., a temperature difference > 2°C within a 3×3 grid) and output a feature map to mark abnormal hotspots (e.g., a response value of 0.85 in grid B2).

[0084] Use the Leaky ReLU activation function: negative inputs are multiplied by a factor of 0.01 to prevent vanishing gradients.

[0085] ‌Long Short-Term Memory‌ (LSTM): Long Short-Term Memory processes the time dimension, with each unit having a memory capacity of 100 states.

[0086] Fluctuation pattern modeling:

[0087] Input: Convolutional features of 5 consecutive time windows (2 windows per second, covering 2.5 seconds of history); Output: Identify periodic changes (such as CO2 concentration rising by 50ppm every 10 seconds) or sudden changes (such as VOC soaring by 40μg / m in 0.2 seconds). 3 ).

[0088] Feature vector set generation: Each grid outputs a 128-dimensional feature vector, including: spatial features (64 dimensions): aggregated values ​​of local area parameters; temporal features (64 dimensions): encoding of fluctuation trends (such as "constant rise" and "violent oscillation").

[0089] Example: In the grid A1 vector, the spatial dimension contains a temperature gradient of 0.7 (high gradient), and the temporal dimension contains a CO2 acceleration flag of 1 (continuous growth).

[0090] Regional features are aggregated based on the feature vector set, and a graph pooling algorithm is used to generate a dynamic environmental feature map covering the entire area of ​​the car. The graph nodes encode the regional environmental status, and the edge weights represent the correlation between environmental parameters between regions.

[0091] Graph Pooling aggregates scattered features into a global graph:

[0092] Graph structure construction: Node: 30 grid units, attributes are 128-dimensional feature vectors.

[0093] Edge: Topological weight (adjacency matrix value in step 2), reflecting the influence strength between regions.

[0094] Feature aggregation mechanism: Node feature update: The weighted influence of adjacent nodes is calculated through the Graph Attention Network (GAT): the weight of grid A1 influenced by B1 is 0.6, and the weight of grid A1 influenced by C1 is 0.3.

[0095] Update formula: A1 new feature = original feature × 0.7 + B1 feature × 0.6 × 0.3 + C1 feature × 0.3 × 0.3.

[0096] Hierarchical pooling: Cluster 30 nodes into 6 super nodes (such as front / back / left / right / upper / lower), and use the node degeneration algorithm to retain key features.

[0097] Dynamic environment feature map output:

[0098] Node data: A summary of the environmental status of each supernode (e.g. "high temperature and high humidity at the rear").

[0099] Edge weight: correlation coefficient of parameters between regions (0-1), such as the correlation between front and back temperatures is 0.92 (strong correlation), and the correlation between left and right VOC is 0.35 (weak correlation).

[0100] ‌Visualization map‌: Red highlighted super nodes indicate abnormal areas (such as the upper temperature node abnormal value > 0.8), and the thickness of the connecting line indicates the strength of the correlation.

[0101] This method uses multiple types of sensors distributed across the car to synchronously collect environmental parameters such as temperature, humidity, and gas concentration, as well as sound feature data. This data is deeply processed using a spatiotemporal convolutional network, capturing the spatial correlations between sensor nodes and analyzing how each parameter changes over time, ultimately forming a feature map reflecting the overall environmental state of the car. This fusion process overcomes the limitations of single sensors or single-moment data, enabling full-dimensional perception and fusion of environmental data, providing high-precision input features for subsequent anomaly detection. Joint analysis of spatiotemporal features can uncover potential environmental anomaly patterns. For example, localized high temperatures accompanied by specific sound patterns may indicate equipment failure, creating conditions for early warning.

[0102] S202, based on the dynamic environmental feature map, reconstructing a denoised environmental state model through a generative adversarial network, wherein the generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomaly based on a historical normal environment database and outputs a high-confidence environmental health index;

[0103] Specifically, the dynamic environment feature map can be input into the generator network, and the denoised environment state map can be reconstructed through the residual connected multi-scale convolution layer. The generator output contains the estimated values ​​of the environment parameters after noise suppression.

[0104] The generator network uses a deep convolutional architecture (12 layers deep) specifically designed to eliminate sensor noise. Its core is to process input data through multi-scale convolutional layers with residual connections:

[0105] Multi-scale convolution design: Small-scale convolution kernel (3×3 grid coverage area): Captures local mutations (such as instantaneous temperature spikes in a 0.4m×0.6m grid).

[0106] Medium-scale convolution kernel (5×5 grid area): detects regional correlation anomalies (such as a synchronous increase in humidity across five adjacent grids).

[0107] Large-scale convolution kernel (covering 1 / 2 of the car): identifies global trends (such as overall CO2 accumulation at the rear of the car).

[0108] ‌Residual Connection Mechanism‌:

[0109] The output of each layer is superimposed with the original input (residual connection), preserving the true data background. For example, if a small-scale convolution mistakenly identifies a grid temperature noise (such as a 0.5°C fluctuation caused by air conditioning) as an anomaly, the residual connection can reduce its weight to below 0.2.

[0110] Denoising and reconstruction process:

[0111] Input dynamic environment feature map (30 super node features), after being refined step by step by convolutional layer:

[0112] Suppress transient interference: Eliminate sudden high-frequency noise (such as the mechanical sound of an elevator) that can cause false triggering of the voiceprint sensor.

[0113] Correct drift error: calibrate the humidity sensor to eliminate long-term reading deviations due to condensation (e.g., a continuous 3% RH increase).

[0114] The output is a denoised environmental state map, where each node contains the optimized parameter value (e.g. the temperature of grid B2 is corrected from 24.8°C to 24.1°C).

[0115] The estimated environmental parameters and the standard state spectrum in the historical normal environment database are input into the discriminator network, and the distribution difference between the two is calculated through spectral normalization processing, and the probability score of environmental abnormality is output;

[0116] The discriminator network acts as a "quality inspector" and compares the generator output with historical normal data:

[0117] Historical database construction: Stores 10,000 hours of normal operation data (temperature 18-26°C, humidity 40-60%RH, CO2<1000ppm, VOC<80μg / m 3 ).

[0118] Standard state map: Modeling by season / time period (e.g., standard CO2 during the morning peak in summer = 900±50ppm).

[0119] Spectral Normalization:

[0120] Core function: Limit the singular value of the discriminator weight matrix to prevent adversarial training from diverging.

[0121] Implementation process: Calculate the spectral norm of the weight matrix W (Spectral Norm) σ(W) = maximum singular value.

[0122] Normalization operation: Replace W with W / σ(W) to ensure that the discriminator satisfies Lipschitz continuity.

[0123] ‌Distribution difference calculation‌: Use ‌Jensen-Shannon Divergence‌ (JSD) to quantify similarity:

[0124] Input: The estimated environmental parameter value output by the generator (such as the temperature of grid C3, 25.2°C) vs. the historical average value of the same location in the database (24.5°C).

[0125] Output: Difference score (0-1), such as 0.7 indicates high anomaly.

[0126] ‌Anomaly Probability Score‌: The difference score is mapped to a probability value using the Sigmoid function (e.g., JSD = 0.7 → anomaly probability 0.85).

[0127] During the adversarial training phase, the generator reversely optimizes parameters based on the anomaly probability fed back by the discriminator, uses the Wasserstein distance constraint to ensure the authenticity of the generated data, and iteratively improves the denoising capability.

[0128] Adversarial training continuously optimizes the game between the generator and the discriminator:

[0129] ‌Reverse optimization mechanism‌: The generator receives the abnormal probability output by the discriminator (such as 0.85) and updates the convolutional layer weights through the ‌backpropagation algorithm‌ (Backpropagation):

[0130] If a convolution kernel causes multiple grid anomalies with a probability greater than 0.8, its weight is reduced by 30%.

[0131] If a residual connection effectively suppresses false positives (abnormal probability < 0.2), its weight is increased by 15%.

[0132] Wasserstein distance constraint:

[0133] Core function: Replace the cross entropy loss of traditional GAN ​​and solve the problem of training instability.

[0134] Implementation principle: Calculate the EarthMover's distance between the generated data distribution P_g and the real data distribution P_r.

[0135] Constraint objective: Minimize the Wasserstein distance (target value < 0.05) to ensure that the generated data is close to the true distribution.

[0136] For example, when the generator misclassifies high-temperature noise as a real fire, the Wasserstein distance rises to 0.12, triggering a weight penalty.

[0137] ‌Iterative improvement process‌: Each round of training consists of 100 generator-discriminator games.

[0138] Termination condition: Wasserstein distance < 0.05 and abnormal probability error < 5% for 10 consecutive rounds.

[0139] Training results: The correlation coefficient between the denoised data and the real data reached 0.98.

[0140] The parameter deviation of each node in the denoised environmental state map is integrated, and the comprehensive impact weights of temperature, humidity, CO2 concentration, and volatile organic compounds are calculated using the entropy weight method to generate a multidimensional environmental health index.

[0141] The Entropy Weight Method objectively quantifies the influence of different environmental parameters:

[0142] Parameter deviation calculation:

[0143] Definition: The absolute deviation between the current value and the historical baseline value (e.g., grid A1 temperature 24.8°C vs. baseline 23.5°C → deviation 1.3°C).

[0144] Normalization: Divide the deviation by the maximum historical deviation (e.g., a maximum temperature deviation of 5°C → a normalized value of 0.26).

[0145] ‌Entropy weight method weight distribution‌:

[0146] Information entropy calculation: Temperature entropy value E_temp = -Σ(p_i × ln p_i), where p_i is the temperature deviation ratio of each grid.

[0147] The smaller the entropy value (e.g. 0.2 vs 0.6), the greater the parameter variability and the higher the weight.

[0148] Weight formula: Temperature weight W_temp = (1 - E_temp) / Σ(1 - E_k) (k covers all 4 parameters).

[0149] Example: If E_temp=0.2, E_hum=0.5, E_co2=0.3, E_voc=0.4, then: W_temp = (1-0.2) / [(1-0.2)+(1-0.5)+(1-0.3)+(1-0.4)] = 0.8 / 2.6 ≈ 0.31.

[0150] Multi-dimensional health index synthesis: Node health index = Σ(parameter deviation × weight).

[0151] Example: A grid has a temperature deviation of 0.3, humidity of 0.1, CO2 of 0.4, and VOC of 0.2. The health index is 0.3 × 0.31 + 0.1 × 0.19 + 0.4 × 0.27 + 0.2 × 0.23 = 0.28.

[0152] The multidimensional environmental health index is normalized and calibrated, and a high-confidence environmental health index in the range of 0-1 is generated by combining the abnormal probability output by the discriminator.

[0153] The final health index needs to integrate objective data and subjective evaluation of the discriminator:

[0154] ‌Normalization Calibration‌:

[0155] Linear scaling: Project the multidimensional health index to the range of 0-1 (e.g., the original value of 0.28 corresponds to the calibrated value of 0.35).

[0156] Nonlinear correction: Use the segmented S-shaped curve to enhance the discrimination of extreme values:

[0157] Health index < 0.2 → output close to 0 (safe);

[0158] Health Index > 0.8 → Output close to 1 (high risk).

[0159] ‌Anomaly Probability Fusion‌:

[0160] The abnormal probability output by the discriminator (such as 0.85) is used as the confidence coefficient:

[0161] Weighted formula: Final health index = calibrated index × 0.7 + abnormality probability × 0.3.

[0162] Example: Calibration value 0.35 × 0.7 + 0.85 × 0.3 = 0.245 + 0.255 = 0.50.

[0163] ‌High confidence output‌:

[0164] Threshold classification: 0-0.3 (green safety); 0.3-0.6 (yellow warning); 0.6-1.0 (red alert).

[0165] Application Example: When the VOC level at the rear of an elevator exceeds the standard, the final index = 0.73 (red alert), triggering the generation of a partitioned heat map.

[0166] Leveraging the adversarial learning mechanism of a generative adversarial network, the generator eliminates sensor noise and outliers to reconstruct the true environmental state. The discriminator identifies the degree of deviation from the current environment by comparing it to historical normal data distributions. Through continuous adversarial optimization, the two ultimately output a quantitative environmental health score that accurately reflects the cabin's environmental status, effectively addressing misjudgments caused by sensor noise and occasional interference, and improving the reliability of anomaly detection. This health index provides a standardized basis for subsequent decision-making, transforming raw data into actionable indicators.

[0167] S203: Based on the environmental health index and passenger density detection data, a multimodal anomaly scoring model is constructed, and a graph neural network is used to correlate the environmental parameter correlations between cabin areas to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the anomaly level;

[0168] Specifically, the environmental health index can be spatiotemporally aligned with camera-based crowd density detection data, and the environmental health weights and passenger density weights of different areas can be dynamically allocated through the attention mechanism to generate a fusion weight matrix.

[0169] Spatiotemporal alignment is the foundation for fusing multi-source data. A wide-angle camera (1920×1080 resolution, 30fps) mounted on the car roof captures passenger distribution in real time, while the environmental health index is derived from the high-confidence output of the previous steps (updated every 0.5 seconds). The alignment process consists of three steps:

[0170] Timestamp Synchronization: Each camera frame is stamped with a millisecond timestamp (e.g., 13:05:27.845). The time difference between the environmental health index generation time and the time interval must be less than 50 milliseconds. If the environmental health index is updated at T=27.890 seconds, the closest camera frame within the T=27.845-27.945 interval (e.g., T=27.892 frame) is selected.

[0171] Spatial Grid Mapping: The camera image is divided into a 5×3 grid (corresponding to the physical grid in the previous step). The YOLOv5 object detection algorithm (You Only Look Once, a real-time object recognition model) is used to count the number of passengers within each grid. For example, if grid B2 detects three people, the passenger density is 3 (the maximum density threshold is set to 5 people per grid).

[0172] ‌Attention weight distribution‌:

[0173] Dynamic weight function: The environmental health weight (W_e) is determined by the grid health index: when the index is greater than 0.6, W_e=0.8; when the index is less than 0.3, W_e=0.2.

[0174] The passenger density weight (W_p) is nonlinearly related to the number of people: 1 person → 0.3, 3 people → 0.7, 5 people → 1.0.

[0175] Fusion rule: High-risk areas (e.g., densely populated areas with unusual environments) are weighted. For example, a grid with a C3 health index of 0.65 (highly unusual) and a density of 4 people → W_e = 0.8, W_p = 0.9 → fusion weight = √(0.8 × 0.9) = 0.85.

[0176] Finally, a weight matrix (5×3×2) of 30 grids is output, storing the environmental weight and passenger density weight of each grid.

[0177] A graph neural network model is constructed, mapping the car area into graph nodes, using the correlation of environmental parameters as edge attributes, and modeling the abnormal propagation path between regions through the graph attention layer;

[0178] The core of Graph Neural Network (GNN) is to transform the car into a topological network:

[0179] ‌Graph structure initialization‌:

[0180] Node: 30 grid cells. Node features include environmental health index, passenger density weight, and historical abnormal frequency.

[0181] ‌Edge‌: Based on the topological weights (adjacent mesh correlation strengths) from the previous step, dynamic parameter dependencies are superimposed:

[0182] If the temperature difference between the two grids is less than 1°C for 5 consecutive seconds, the correlation weight is increased by 0.1.

[0183] If passengers frequently move, resulting in interactions between areas (e.g., from A1 to B1), the movement path weight is increased by 0.2.

[0184] ‌Graph Attention Layer (GAT)‌ Working Mechanism:

[0185] Each node calculates the influence of its neighbors:

[0186] Grid A1 focuses on the adjacent B1, A2, and A1 layers below:

[0187] B1's environmental weight is 0.7 → its impact score = 0.7 × its relevance weight is 0.9 = 0.63;

[0188] A2's environmental weight is 0.5 → score = 0.5 × 0.6 = 0.3.

[0189] Weighted aggregation: A1 new feature = (original feature × 0.4 + B1 feature × 0.63 + A2 feature × 0.3).

[0190] Abnormal propagation modeling: The high-temperature area (grid C3) transfers heat parameters to adjacent grids, and the propagation intensity is controlled by the edge weight (for example, the weight from C3 to B3 is 0.8, and the propagation efficiency is 80%).

[0191] Dynamic parameter updates: Edge attributes are recalculated every 5 seconds to adapt to topological changes caused by passenger movement.

[0192] Update the graph node features according to the fusion weight matrix, use the graph convolutional network to iterate the anomaly propagation signal, calculate the anomaly accumulation value of each node, and generate the initial anomaly score distribution;

[0193] ‌Graph Convolutional Network (GCN)‌ captures anomaly propagation through multiple layers of iteration:

[0194] Node feature updates:

[0195] Input the fusion weight matrix (environment weight + density weight) from step 1 and concatenate it with the original node features:

[0196] Grid B2 features = [Health index 0.65, Density weight 0.9, Historical anomaly 0.3] → After concatenation, it becomes [0.65, 0.9, 0.3, W_e=0.8, W_p=0.9].

[0197] Updated through two layers of GCN: the first layer outputs a 128-dimensional vector (extracting local abnormal features);

[0198] The second layer outputs a 64-dimensional vector (fused with the global propagation signal).

[0199] Exception propagation iteration:

[0200] Iteration rule: Each node absorbs the outliers of adjacent nodes and transfers them according to the edge weight decay:

[0201] Grid C3 has an initial anomaly score of 0.8, which is passed to B3 (weight 0.8): B3 receives a score of 0.8 × 0.8 = 0.64, and B3's own score is 0.6 → updated to 0.6 + 0.64 = 1.24 (normalization required).

[0202] ‌Cumulative mechanism‌: If the grid score increases by >0.1 for three consecutive iterations, a positive feedback loop is triggered (the score increases at an accelerated rate).

[0203] Initial anomaly score generation: Final score = node health index × 0.6 + neighbor propagation contribution × 0.4.

[0204] Example distribution: The posterior high-temperature area (C2, C3) has a score of 0.75-0.92; the anterior low-risk area (A1, A2) has a score of 0.15-0.30; output is a 30-grid score matrix (5×3), with a value range of 0-1.

[0205] The initial anomaly score is spatially interpolated, and a heat diffusion algorithm is used to simulate the propagation trend of anomalies in the car. A partitioned anomaly probability heat map is output to calibrate the location of high-risk areas and anomaly levels.

[0206] Heat map generation requires solving the mapping from discrete grid to continuous space:

[0207] Spatial Interpolation

[0208] ‌Bicubic Interpolation‌: Expands the 5×3 grid into a 50×30 pixel image, smoothing the transition between score values.

[0209] Grid B2 score 0.8 → the corresponding pixel area (x10-20, y20-30) is assigned a value of 80 (0.8×100);

[0210] The A2 score of the adjacent grid is 0.3 → the pixel gradient at the intersection (for example, the value of the coordinate (19,21) = (0.8×0.7+0.3×0.3)=0.65).

[0211] Thermal Diffusion Algorithm:

[0212] Physics simulation anomaly propagation:

[0213] Diffusion source: Meshes with anomaly scores > 0.7 are considered "heat sources" (e.g., C3 score 0.92).

[0214] Diffusion equation: The score value diffuses to the surrounding pixels every 0.1 seconds, and the attenuation coefficient = 0.05 / pixel distance.

[0215] Example: C3 (coordinates 25, 15) spreads to B3 (coordinates 20, 15):

[0216] Pixel distance = 5 → Score increment after attenuation = 0.92 × (0.05 / 5) = 0.0092; New score of B3 = original value 0.75 + 0.0092 = 0.7592.

[0217] Dynamic trend visualization: Red areas (score > 0.7) expand over time, indicating the spread of the anomaly; arrows indicate the direction of spread (e.g., from back to front).

[0218] ‌Partition anomaly probability heatmap output‌:

[0219] Color coding: Blue (0-0.3): safe; Yellow (0.3-0.6): warning; Red (0.6-1.0): high risk.

[0220] High-risk area demarcation: Continuous red areas greater than 0.4 m2 are marked as high-risk areas (e.g., the 1.2 m2 area at the rear). Abnormality level = average score × area (e.g., 0.82 × 1.2 = 0.984 → Level 1 high-risk).

[0221] Application Example: Crowded passengers cause CO2 accumulation at the rear of an elevator, resulting in a red patch in the rear third of the generated heat map, marked "LV2."

[0222] By integrating environmental health data with visually recognized passenger distribution information, graph neural networks are used to analyze the interplay between environmental parameters in different areas. For example, elevated CO2 levels in high-density passenger areas can impact adjacent areas. This ultimately generates a heat map that intuitively displays risk distribution, clearly marking the location and severity of abnormal areas. This allows for precise location and grading of environmental anomalies, providing a basis for targeted response. The heat map visualizes risk distribution, facilitating rapid identification of problem areas and improving response efficiency.

[0223] S204: Based on the partitioned abnormality probability heat map, monitoring resources are dynamically allocated through an adaptive attention mechanism, the sensor sampling frequency in high-risk areas is prioritized, and real-time environmental parameters, the partitioned abnormality probability heat map, and risk warning prompts are superimposed on the car display screen;

[0224] Specifically, the monitoring priority can be divided according to the anomaly level in the partition anomaly probability heat map, and the resource allocation coefficient of the high-risk area can be dynamically calculated using the attention weight function. The resource allocation coefficient value is exponentially related to the anomaly level.

[0225] The system analyzes the color codes and numerical labels of the heat map to divide the car area into four levels of monitoring priority:

[0226] ‌Prioritization Rules‌:

[0227] ‌Blue Area (Anomaly Level 0-0.3): Marked as P4 level (lowest priority), the resource allocation coefficient is fixed at 1.0 (i.e., the baseline sampling frequency).

[0228] Yellow area (0.3-0.6): Marked as P3 level (medium concern), the coefficient increases linearly (e.g., anomaly level 0.4 → coefficient 1.4).

[0229] Orange area (0.6-0.8): Marked as P2 (high priority), triggering index calculation (e.g., level 0.7 → coefficient = 1.5^2 = 2.25).

[0230] Red area (>0.8): Marked as P1 level (highest risk), coefficient = 1.5^(level × 10). For example, anomaly level 0.85 → coefficient = 1.5^8.5≈25.6 (but constrained by the hardware upper limit of 3.0).

[0231] ‌Example of how the attention weight function works‌:

[0232] When the heat map shows that grid No. 2 at the rear of the car is red (abnormality level 0.88) and grid No. 4 on the side wall is orange (0.72):

[0233] Back-end resource allocation coefficient = 1.5^(0.88×10) ≈ 1.5^8.8 ≈ 28.7 → 3.0 after hardware limiting;

[0234] Sidewall coefficient = 1.5^(0.72×10) = 1.5^7.2 ≈ 17.8 → after clipping, take 3.0;

[0235] The coefficients are sent to the sensor control module via the CAN bus (Controller Area Network, an in-vehicle communication protocol).

[0236] Adjust the sampling strategy of the sensor array, enable super-resolution sampling mode for high-risk areas, and increase the sampling frequency of temperature and humidity sensors to three times the baseline value;

[0237] Dynamic sampling strategy adjustment is divided into two stages: hardware reconfiguration and data flow optimization:

[0238] ‌Hardware Reconfiguration‌:

[0239] The default sampling frequency of the temperature / humidity sensor (model SHT40) is 1Hz (1 time per second). When the resource allocation coefficient is ≥ 2.0:

[0240] Switch to ‌Turbo Mode‌: By rewriting I 2C register (Inter-Integrated Circuit, two-wire communication protocol), increasing the sampling rate to 3Hz.

[0241] Power consumption management: Boost the power supply voltage (from 3.3V to 5V) only for sensors in high-risk grids to avoid a surge in energy consumption for the entire array of sensors.

[0242] Example: When the rear grid temperature sensor No. 2 is in P1 level, the sampling interval is compressed from 1000 milliseconds to 333 milliseconds, and the single sampling time is shortened from 20 milliseconds to 5 milliseconds (high-speed mode).

[0243] Data flow optimization: High-frequency sampling generates three times the amount of data (for example, the temperature value increases from 16 points per second to 48 points per second), requiring a dedicated buffer for temporary storage. An independent memory partition (8KB capacity) is created, and a ring buffer structure is used to store temporary data.

[0244] Data compression: Delta encoding is performed on consecutive sample values, storing only the changes (e.g., a temperature change of 0.1°C is recorded as +01).

[0245] The Edge Computing Unit (ECU) filters valid data in real time, eliminating transient interference (such as a 0.2°C temperature fluctuation caused by a passing passenger) and retaining persistent abnormal signals.

[0246] A heat map overlay module is embedded in the car display rendering engine, and real-time environmental parameters, partition abnormality probability heat maps, and camera video streams are integrated at the pixel level using a transparency gradient algorithm.

[0247] Multi-source information fusion display relies on a 10-inch LCD display (resolution 1280×720) on the top of the car. The rendering engine adopts a three-layer superposition architecture:

[0248] ‌ Bottom-layer video stream‌: The camera (2-megapixel, fisheye lens) video stream is transmitted at 30 fps (Frames Per Second) and mapped to the bottom layer of the display after distortion correction.

[0249] ‌Mid-level heatmap overlay‌:

[0250] Alpha Blending: Each pixel in the heatmap is assigned a transparency value (Alpha value): Alpha = 0.3 (semi-transparent) in the blue area and Alpha = 0.8 (high coverage) in the red area.

[0251] Fusion formula: Final pixel color = video pixel color × (1-Alpha) + heat map pixel color × Alpha.

[0252] Example: The rear seat pixel (RGB=120,80,60) in the video is superimposed on the red heatmap pixel (RGB=255,0,0), Alpha=0.8 → After blending, RGB = (120×0.2 + 255×0.8, 80×0.2 + 0×0.8, 60×0.2 +0×0.8) = (222, 16, 12).

[0253] Dynamic Update: The heat map is refreshed every 0.5 seconds, using GPU (Graphics Processing Unit) for accelerated rendering.

[0254] ‌Top-level environment parameters‌:

[0255] Key values ​​are displayed floating in the four corners of the screen: upper left corner: average cabin temperature (such as 24.3°C) and humidity (52%RH); lower right corner: highest abnormal area level (such as "rear LV2"); font color changes dynamically: normal value green (temperature <26°C), exceeding threshold value flashing red (CO2>1000ppm).

[0256] Build a risk warning dynamic labeling system. When the abnormality level in a certain area exceeds the preset level, a flashing warning box and voice prompt will be automatically superimposed at the corresponding location.

[0257] The multimodal warning system consists of visual annotation and voice announcement:

[0258] ‌Visual Annotation‌:

[0259] Flashing warning box: Using SVG vector graphics (Scalable Vector Graphics), it is drawn in real time based on the coordinates of the high-risk areas on the heat map:

[0260] LV1 warning (0.6-0.8): yellow square, flashing at 1Hz (on and off for 0.5 seconds each); LV2 and above (>0.8): red octagonal frame, flashing at 2Hz (on for 0.25 seconds / off for 0.25 seconds).

[0261] ‌Intelligent Obstacle Avoidance Algorithm‌: When the selected area overlaps with a passenger's face (via face detection), it is automatically offset by 50 pixels to avoid occlusion.

[0262] Voice prompts:

[0263] Level 1 triggers pre-recorded voice: "Attention, local environment abnormalities" (volume 60dB); Level 2 triggers TTS synthesis (Text-to-Speech): "Warning! High temperature and high risk in the rear" (volume 75dB).

[0264] Sound source directional technology: Through the four-microphone array on the top of the car, the voice sound image is localized to the direction of the abnormal area (for example, the sound source is located behind the display screen during rear alarm).

[0265] Emergency Brake Interaction: When the same area remains in LV2 for 10 seconds, a CAN signal is automatically sent to the elevator control system, triggering the following responses: the car fan speed is increased to 3000 RPM (Revolutions Per Minute); and an "Environmental Emergency" event code (Event Code 0xE3) is sent to the building management system.

[0266] Monitoring strategies are automatically adjusted based on the risk level of the heat map: sampling frequency is increased in high-risk areas to ensure real-time data. Environmental data and the risk heat map are superimposed on the elevator display screen, with high-risk areas highlighted through color changes and flashing prompts. This allows for intuitive communication of information, optimizes resource allocation, and concentrates monitoring efforts on addressing key risks. Real-time visualization enhances risk awareness among passengers and maintenance personnel, facilitating rapid response.

[0267] S205: Based on the duration and spatial diffusion trend of the partitioned anomaly probability heat map, an environmental deterioration warning signal is generated through a spatiotemporal prediction model. When the environmental deterioration warning signal level reaches a preset signal level, a gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, forming a closed-loop monitoring response.

[0268] Specifically, the spatiotemporal variation characteristics of the partitioned anomaly probability heat map can be extracted, and the duration, diffusion speed, and correlation strength of the abnormal area and adjacent areas can be modeled through the spatiotemporal graph convolutional network to output the spatiotemporal feature vector.

[0269] The system collects a 10-minute continuous partitioned anomaly probability heat map (resolution 50×30 pixels) at a rate of 1 frame per second and extracts three key features using the Spatio-Temporal Graph Convolutional Network (ST-GCN):

[0270] Duration Quantification: For each high-risk pixel block (e.g., the red area between coordinates (35,20) and (45,25)), the duration of its abnormal state is calculated (in seconds). For example, if a high-temperature area persists for 120 seconds without being eliminated, it is marked as a "long-term anomaly" and the characteristic value is normalized to 0.85 (120 seconds / maximum set threshold of 141 seconds).

[0271] Dynamic decay mechanism: If the abnormal area is interrupted during the diffusion process (for example, the red area disappears due to ventilation and cooling), the duration characteristic value decays by 5% per second (for example, the original value of 0.8 drops to 0.8×(0.95^3)≈0.69 after a 3-second interruption).

[0272] Diffusion Speed ​​Modeling: Detects abnormal spread trends through differences between adjacent heatmap frames: Diffusion Speed ​​(Speed) = Newly Added Abnormal Pixel Area ÷ Time Interval (Seconds).

[0273] Example: A VOC pollution source spreads from 5 pixels to 50 pixels in 5 seconds → Speed ​​= (50-5) / 5 = 9 pixels / second.

[0274] Direction Encoding: Combined with the car's spatial coordinate system, the diffusion direction is quantified as an angle value (0°-360°). For example, if the anomaly spreads from the back to the front (along the positive Y-axis), it is recorded as 90°, and if it spreads to the left and front, it is recorded as 45°.

[0275] Regional correlation strength analysis:

[0276] The graph attention mechanism (GAT) is used to calculate the coupling coefficient (CouplingCoefficient) between the abnormal source and the affected area:

[0277] The correlation strength from the high-temperature source area (A) to the adjacent area (B) = A anomaly value × heat transfer coefficient × spatial distance attenuation factor;

[0278] Example: The rear high temperature area (outlier value 0.9) is 2 meters away from the front area (attenuation factor 0.5), and the transmission coefficient is 0.8 → the correlation strength = 0.9×0.8×0.5=0.36;

[0279] Output 256-dimensional spatiotemporal feature vectors (including sub-features such as duration, diffusion speed, direction angle, and correlation strength) for use in subsequent prediction models.

[0280] The spatiotemporal feature vectors are input into a pre-trained LSTM-Transformer hybrid prediction model to predict environmental deterioration trends within a preset number of minutes and generate probabilistic early warning signals.

[0281] The ‌LSTM-Transformer hybrid model‌ consists of a Long Short-Term Memory (LSTM) network and a Transformer encoder, processing spatiotemporal features in stages:

[0282] ‌LSTM time series modeling stage‌:

[0283] Input 100 consecutive spatiotemporal feature vectors (covering the past 100 seconds) and extract the temporal dependency pattern through two layers of LSTM units (128 neurons per layer):

[0284] The first layer of LSTM outputs a hidden state to capture short-term fluctuations (such as the change in abnormal area every 10 seconds);

[0285] The second layer of LSTM outputs the cell state (Cell State), which memorizes long-term trends (such as continuous temperature rise for more than 3 minutes).

[0286] Key parameter example: When the hidden state value exceeds the threshold of 0.7, it is judged as "accelerated deterioration" (for example, the CO2 concentration in a certain area increases by 200 ppm per minute).

[0287] ‌Transformer spatial dependency modeling phase‌:

[0288] Input the temporal features output by LSTM into the Transformer encoder (4-layer structure, 8 attention heads per layer):

[0289] The self-attention mechanism calculates the mutual influence of features in each region: the attention weight of the features in the rear high-temperature area and the central area is 0.85 (strong correlation); the weight of the low-temperature area on the side wall is only 0.12 (weak correlation); positional encoding preserves spatial topology information (for example, the coordinate encoding value of the rear of the car is higher than that of the front).

[0290] Output a comprehensive spatiotemporal feature matrix (128×100) to characterize the future risk evolution pattern.

[0291] Probabilistic early warning signal generation:

[0292] The fully connected layer maps the feature matrix into three types of warning probabilities: probability of deterioration within the next 2 minutes (e.g., 82%); probability of deterioration within the next 5 minutes (e.g., 64%); and probability of deterioration within the next 10 minutes (e.g., 28%).

[0293] Decision rule: If any probability value is ≥ 70%, a graded warning signal is generated (e.g., "Level 1 warning: 82% probability of deterioration of rear high temperature within 2 minutes").

[0294] A three-level gradient alarm mechanism is designed. When the probabilistic warning signal reaches level one, an audible and visual alarm is triggered in the elevator car. When it reaches level two, the ventilation system is activated to enhance ventilation. When it reaches level three, an emergency elevator dispatch request is sent to the building management system.

[0295] The three-level alarm mechanism adopts a cumulative triggering strategy, and the response intensity increases with the warning level:

[0296] Level 1 alarm (warning probability ≥ 70%):

[0297] Sound and light warning: The LED light strip on the top of the car switches to red breathing flashing mode (frequency 1Hz, brightness 1000 lumens).

[0298] Play a pre-recorded voice: "Abnormal environment detected, please pay attention to safety" (volume 65 decibels, repeated 3 times).

[0299] ‌Physical Positioning‌: The warning signal source points to the direction of the high-risk area on the heat map (for example, when there is an abnormality in the rear, the sound is emitted from the rear speaker).

[0300] Level 2 alarm (warning probability ≥ 85%):

[0301] ‌Ventilation Enhancement‌: Send a command to the fan controller via the Modbus protocol (a communication standard for industrial equipment) to increase the speed from 1500 RPM to 3000 RPM (Revolutions Per Minute).

[0302] Start emergency ventilation mode: the fresh air valve opening increases from 30% to 100%, and the ventilation volume increases by 3 times.

[0303] Local purification: If VOC levels exceed the limit, the nanofilter above the high-risk area (such as the rear grids B3-C3) will be activated.

[0304] Level 3 alarm (warning probability ≥ 95%):

[0305] Elevator dispatch linkage: Generates event code 0xE7 (environmental emergency evacuation command) and sends it to the building management system via the OPC UA protocol (industrial automation communication protocol).

[0306] Building system execution strategy: dispatch the target elevator to the nearest floor (for example, if it was originally planned to go up to the 20th floor, it will open the door at the 15th floor instead); neighboring elevators will be shut down (to prevent passengers from entering dangerous elevators).

[0307] ‌Escape Guidance‌: The car display switches to green arrows indicating the exit path (such as "Please move toward the front door").

[0308] Establish a closed-loop verification mechanism to compare actual environmental change data with prediction results, dynamically update the hybrid prediction model parameters through reinforcement learning, continuously optimize the accuracy of early warnings, and form a closed-loop monitoring response.

[0309] The closed-loop verification system performs model optimization every 24 hours. The process is as follows:

[0310] Data comparison and analysis:

[0311] Extract historical forecast records (such as 10 "Level 1 warnings") and actual sensor data (such as the temperature rise):

[0312] Correct prediction: The actual deterioration occurs within the warning time window (for example, deterioration is predicted within 2 minutes, but it is actually triggered at the 110th second). Failed prediction: There is no actual deterioration or the deterioration times out (for example, deterioration is predicted within 5 minutes, but it actually occurs after 10 minutes).

[0313] Quantization error metric:

[0314] Time error = |predicted deterioration time - actual deterioration time| (e.g., an error of 120 seconds is recorded as -2 minutes);

[0315] Spatial error = the distance between the center point of the predicted position and the actual position (e.g., a deviation of 0.5 meters).

[0316] ‌Reinforcement Learning Parameter Update‌:

[0317] Use Proximal Policy Optimization (PPO algorithm) to adjust model weights:

[0318] Reward Function Design: Successful prediction with temporal error < 30 seconds: reward +1.0; failed prediction with spatial error < 0.3 meters: reward -0.5; complete false alarm (no actual deterioration): reward -1.5.

[0319] Policy Network Update: If the prediction time error is >60 seconds for three consecutive times, reduce the weight of the LSTM forget gate by 10%. If the spatial error is persistently high, increase the weight of the Transformer position encoding layer by 15%. After the model is updated, it is deployed to the Edge Computing Unit, replacing the old version.

[0320] Closed-loop response verification: The new model's performance was tested by simulating sudden environmental change scenarios (such as the injection of high-temperature gas). Before optimization, the average error in predicted deterioration time was 45 seconds. After optimization, the error was reduced to 12 seconds, and the warning accuracy rate increased to 93%. Real-time linkage verification: After the third-level alarm command was issued to the building system, the measured elevator dispatch delay was shortened from 20 seconds to 8 seconds.

[0321] By analyzing the duration and diffusion rate of abnormal areas, we predict future environmental trends. Based on these predictions, we trigger a tiered response: a primary alert alerts passengers, a secondary one activates the ventilation system, and a tertiary one adjusts elevator operation strategies, such as prioritizing the elevator to the nearest floor for ventilation. This completes a closed loop from detection to action, enabling an upgrade from passive monitoring to active prediction. This tiered response mechanism ensures that the intensity of the action aligns with the severity of the risk. This systemic linkage enables intelligent management, effectively preventing environmental incidents from escalating.

[0322] It can be seen that the environmental data in the cabin is collected in real time through the multimodal sensor array to generate a dynamic environmental feature map; based on the dynamic environmental feature map, a high-confidence environmental health index is output; based on the environmental health index and passenger density detection data, a partition abnormality probability heat map is generated; based on the partition abnormality probability heat map, real-time environmental parameters, partition abnormality probability heat map and risk warning prompts are superimposed on the cabin display screen; based on the duration and spatial diffusion trend of the partition abnormality probability heat map, an environmental deterioration warning signal is generated through the spatiotemporal prediction model. When the environmental deterioration warning signal level reaches the preset signal level, the gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, thereby improving the accuracy and response efficiency of elevator environmental safety management through multimodal data fusion, dynamic risk assessment and predictive warning.

[0323] Another embodiment of the present invention provides an elevator environment abnormality monitoring system, see Figure 3 , the system may include:

[0324] Fusion module 301 is used to collect environmental data in the car in real time through a multimodal sensor array, and fuse the spatial distribution characteristics and time series fluctuation characteristics of the sensor data using a spatiotemporal convolutional network to generate a dynamic environmental feature map, where the environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds, and voiceprint data;

[0325] A reconstruction module 302 is configured to reconstruct a denoised environmental state model based on the dynamic environmental feature map using a generative adversarial network, wherein the generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomalies based on a historical normal environment database and outputs a high-confidence environmental health index;

[0326] Correlation module 303 is configured to construct a multimodal anomaly scoring model based on the environmental health index and passenger density detection data, and utilize a graph neural network to correlate the environmental parameter correlations between cabin zones to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the level of anomaly;

[0327] The display module 304 is used to dynamically allocate monitoring resources based on the partition abnormality probability heat map through an adaptive attention mechanism, prioritize the increase in sensor sampling frequency in high-risk areas, and overlay real-time environmental parameters, the partition abnormality probability heat map, and risk warning prompts on the car display screen;

[0328] The early warning module 305 is used to generate an environmental deterioration early warning signal through a spatiotemporal prediction model based on the duration and spatial diffusion trend of the partition abnormality probability heat map. When the environmental deterioration early warning signal level reaches a preset signal level, the gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy to form a closed-loop monitoring response.

[0329] It can be seen that the environmental data in the cabin is collected in real time through the multimodal sensor array to generate a dynamic environmental feature map; based on the dynamic environmental feature map, a high-confidence environmental health index is output; based on the environmental health index and passenger density detection data, a partition abnormality probability heat map is generated; based on the partition abnormality probability heat map, real-time environmental parameters, partition abnormality probability heat map and risk warning prompts are superimposed on the cabin display screen; based on the duration and spatial diffusion trend of the partition abnormality probability heat map, an environmental deterioration warning signal is generated through the spatiotemporal prediction model. When the environmental deterioration warning signal level reaches the preset signal level, the gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, thereby improving the accuracy and response efficiency of elevator environmental safety management through multimodal data fusion, dynamic risk assessment and predictive warning.

[0330] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in the above-mentioned embodiment of the elevator environment abnormality monitoring method when running.

[0331] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0332] S201, collecting environmental data in the car in real time through a multimodal sensor array, and fusing the spatial distribution characteristics of the sensor data with the time series fluctuation characteristics using a spatiotemporal convolutional network to generate a dynamic environmental feature map, wherein the environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds, and voiceprint data;

[0333] S202, based on the dynamic environmental feature map, reconstructing a denoised environmental state model through a generative adversarial network, wherein the generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomaly based on a historical normal environment database and outputs a high-confidence environmental health index;

[0334] S203: Based on the environmental health index and passenger density detection data, a multimodal anomaly scoring model is constructed, and a graph neural network is used to correlate the environmental parameter correlations between cabin areas to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the anomaly level;

[0335] S204: Based on the partitioned abnormality probability heat map, monitoring resources are dynamically allocated through an adaptive attention mechanism, the sensor sampling frequency in high-risk areas is prioritized, and real-time environmental parameters, the partitioned abnormality probability heat map, and risk warning prompts are superimposed on the car display screen;

[0336] S205: Based on the duration and spatial diffusion trend of the partitioned anomaly probability heat map, an environmental deterioration warning signal is generated through a spatiotemporal prediction model. When the environmental deterioration warning signal level reaches a preset signal level, a gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, forming a closed-loop monitoring response.

[0337] It can be seen that the environmental data in the cabin is collected in real time through the multimodal sensor array to generate a dynamic environmental feature map; based on the dynamic environmental feature map, a high-confidence environmental health index is output; based on the environmental health index and passenger density detection data, a partition abnormality probability heat map is generated; based on the partition abnormality probability heat map, real-time environmental parameters, partition abnormality probability heat map and risk warning prompts are superimposed on the cabin display screen; based on the duration and spatial diffusion trend of the partition abnormality probability heat map, an environmental deterioration warning signal is generated through the spatiotemporal prediction model. When the environmental deterioration warning signal level reaches the preset signal level, the gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, thereby improving the accuracy and response efficiency of elevator environmental safety management through multimodal data fusion, dynamic risk assessment and predictive warning.

[0338] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps of the embodiment of the elevator environment abnormality monitoring method.

[0339] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0340] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0341] S201, collecting environmental data in the car in real time through a multimodal sensor array, and fusing the spatial distribution characteristics of the sensor data with the time series fluctuation characteristics using a spatiotemporal convolutional network to generate a dynamic environmental feature map, wherein the environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds, and voiceprint data;

[0342] S202, based on the dynamic environmental feature map, reconstructing a denoised environmental state model through a generative adversarial network, wherein the generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomaly based on a historical normal environment database and outputs a high-confidence environmental health index;

[0343] S203: Based on the environmental health index and passenger density detection data, a multimodal anomaly scoring model is constructed, and a graph neural network is used to correlate the environmental parameter correlations between cabin areas to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the anomaly level;

[0344] S204: Based on the partitioned abnormality probability heat map, monitoring resources are dynamically allocated through an adaptive attention mechanism, the sensor sampling frequency in high-risk areas is prioritized, and real-time environmental parameters, the partitioned abnormality probability heat map, and risk warning prompts are superimposed on the car display screen;

[0345] S205: Based on the duration and spatial diffusion trend of the partitioned anomaly probability heat map, an environmental deterioration warning signal is generated through a spatiotemporal prediction model. When the environmental deterioration warning signal level reaches a preset signal level, a gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, forming a closed-loop monitoring response.

[0346] It can be seen that the environmental data in the cabin is collected in real time through the multimodal sensor array to generate a dynamic environmental feature map; based on the dynamic environmental feature map, a high-confidence environmental health index is output; based on the environmental health index and passenger density detection data, a partition abnormality probability heat map is generated; based on the partition abnormality probability heat map, real-time environmental parameters, partition abnormality probability heat map and risk warning prompts are superimposed on the cabin display screen; based on the duration and spatial diffusion trend of the partition abnormality probability heat map, an environmental deterioration warning signal is generated through the spatiotemporal prediction model. When the environmental deterioration warning signal level reaches the preset signal level, the gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, thereby improving the accuracy and response efficiency of elevator environmental safety management through multimodal data fusion, dynamic risk assessment and predictive warning.

[0347] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for monitoring abnormal elevator environment, characterized in that: The method comprises: The system collects environmental data in the cabin in real time through a multimodal sensor array, and uses a spatiotemporal convolutional network to fuse the spatial distribution characteristics of the sensor data with the time series fluctuation characteristics to generate a dynamic environmental feature map. The environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds, and voiceprint data. Based on the dynamic environmental feature map, a denoised environmental state model is reconstructed through a generative adversarial network. The generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomalies based on a historical normal environment database and outputs a high-confidence environmental health index. Based on the environmental health index and passenger density detection data, a multimodal anomaly scoring model is constructed, and a graph neural network is used to correlate the environmental parameter correlations between cabin areas to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the level of anomaly; Based on the partition abnormality probability heat map, monitoring resources are dynamically allocated through an adaptive attention mechanism, sensor sampling frequency in high-risk areas is prioritized, and real-time environmental parameters, partition abnormality probability heat map and risk warning prompts are superimposed on the car display screen; Based on the duration and spatial diffusion trend of the partitioned anomaly probability heat map, an environmental deterioration warning signal is generated through a spatiotemporal prediction model. When the environmental deterioration warning signal level reaches the preset signal level, the gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, forming a closed-loop monitoring response.

2. The method according to claim 1, characterized in that The multimodal sensor array is used to collect environmental data in the car in real time, and the spatial distribution characteristics and time series fluctuation characteristics of the sensor data are integrated using a spatiotemporal convolutional network to generate a dynamic environmental feature map, wherein the environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds and voiceprint data, including: Based on the original environmental data stream collected by the multimodal sensor array, a sliding time window is used to segment the temperature, humidity, CO2 concentration, volatile organic compounds and voiceprint signals to generate time-aligned multimodal data slices; The multimodal data slices are spatially encoded, the car is divided into gridded areas, and the topological relationships of the sensor nodes are annotated to construct a spatiotemporal correlation tensor. The spatiotemporal correlation tensor is input into a 3D convolution-long short-term memory hybrid network. The local region features are extracted through the spatial convolution kernel, and the fluctuation pattern across the time window is captured by combining the time axis LSTM. The feature vector set that integrates the spatiotemporal features is output. Regional features are aggregated based on the feature vector set, and a graph pooling algorithm is used to generate a dynamic environmental feature map covering the entire area of ​​the car. The graph nodes encode the regional environmental status, and the edge weights represent the correlation between environmental parameters between regions.

3. The method according to claim 2, characterized in that Based on the dynamic environmental feature map, a denoised environmental state model is reconstructed through a generative adversarial network, wherein the generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomaly based on a historical normal environment database and outputs a high-confidence environmental health index, including: The dynamic environment feature map is input into the generator network, and the denoised environment state map is reconstructed through the multi-scale convolutional layer with residual connections. The generator output contains the estimated values ​​of the environment parameters after noise suppression. The estimated environmental parameters and the standard state spectrum in the historical normal environment database are input into the discriminator network, and the distribution difference between the two is calculated through spectral normalization processing, and the probability score of environmental abnormality is output; During the adversarial training phase, the generator reversely optimizes parameters based on the anomaly probability fed back by the discriminator, uses the Wasserstein distance constraint to ensure the authenticity of the generated data, and iteratively improves the denoising capability. The parameter deviation of each node in the denoised environmental state map is integrated, and the comprehensive impact weights of temperature, humidity, CO2 concentration, and volatile organic compounds are calculated using the entropy weight method to generate a multidimensional environmental health index. The multidimensional environmental health index is normalized and calibrated, and a high-confidence environmental health index in the range of 0-1 is generated by combining the abnormal probability output by the discriminator.

4. The method according to claim 3, characterized in that The multimodal anomaly scoring model is constructed based on the environmental health index and passenger density detection data, and the graph neural network is used to associate the environmental parameter correlations between the cabin areas to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the anomaly level, including: The environmental health index is spatiotemporally aligned with camera-based crowd density detection data. The environmental health weights and passenger density weights of different areas are dynamically assigned through an attention mechanism to generate a fusion weight matrix. A graph neural network model is constructed, mapping the car area into graph nodes, using the correlation of environmental parameters as edge attributes, and modeling the abnormal propagation path between regions through the graph attention layer; Update the graph node features according to the fusion weight matrix, use the graph convolutional network to iterate the anomaly propagation signal, calculate the anomaly accumulation value of each node, and generate the initial anomaly score distribution; The initial anomaly score is spatially interpolated, and a heat diffusion algorithm is used to simulate the propagation trend of anomalies in the car. A partitioned anomaly probability heat map is output to calibrate the location of high-risk areas and anomaly levels.

5. The method according to claim 4, characterized in that Based on the partition abnormality probability heat map, monitoring resources are dynamically allocated through an adaptive attention mechanism, the sensor sampling frequency in high-risk areas is prioritized, and real-time environmental parameters, partition abnormality probability heat map and risk warning prompts are superimposed on the car display screen, including: The monitoring priority is divided according to the abnormality level in the partition abnormality probability heat map, and the resource allocation coefficient of the high-risk area is dynamically calculated using the attention weight function. The resource allocation coefficient value is exponentially related to the abnormality level. Adjust the sampling strategy of the sensor array, enable super-resolution sampling mode for high-risk areas, and increase the sampling frequency of temperature and humidity sensors to three times the baseline value; A heat map overlay module is embedded in the cabin display rendering engine, which uses a transparency gradient algorithm to fuse real-time environmental parameters, partition abnormality probability heat maps, and camera video streams at the pixel level. Build a risk warning dynamic labeling system. When the abnormality level in a certain area exceeds the preset level, a flashing warning box and voice prompt will be automatically superimposed at the corresponding location.

6. The method according to claim 5, characterized in that According to the duration and spatial diffusion trend of the partition abnormality probability heat map, an environmental deterioration warning signal is generated through a spatiotemporal prediction model. When the environmental deterioration warning signal level reaches a preset signal level, a gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, forming a closed-loop monitoring response, including: Extract the spatiotemporal variation characteristics of the partitioned anomaly probability heat map, model the duration, diffusion speed, and correlation strength of the abnormal area through the spatiotemporal graph convolutional network, and output the spatiotemporal feature vector; The spatiotemporal feature vectors are input into a pre-trained LSTM-Transformer hybrid prediction model to predict environmental deterioration trends within a preset number of minutes and generate probabilistic early warning signals. A three-level gradient alarm mechanism is designed. When the probabilistic warning signal reaches level one, an audible and visual alarm is triggered in the elevator car. When it reaches level two, the ventilation system is activated to enhance ventilation. When it reaches level three, an emergency elevator dispatch request is sent to the building management system. Establish a closed-loop verification mechanism to compare actual environmental change data with prediction results, dynamically update the hybrid prediction model parameters through reinforcement learning, continuously optimize the accuracy of early warnings, and form a closed-loop monitoring response.

7. An elevator environment abnormality monitoring system, characterized in that: The system comprises: A fusion module is used to collect environmental data in the car in real time through a multimodal sensor array, and use a spatiotemporal convolutional network to fuse the spatial distribution characteristics of the sensor data with the time series fluctuation characteristics to generate a dynamic environmental feature map, where the environmental data includes temperature, humidity, CO2 concentration, volatile organic compounds, and voiceprint data; A reconstruction module is used to reconstruct a denoised environmental state model based on the dynamic environmental feature map through a generative adversarial network, wherein the generator eliminates sensor noise interference through adversarial training, and the discriminator calculates the probability of environmental anomaly based on a historical normal environment database and outputs a high-confidence environmental health index; a correlation module, configured to construct a multimodal anomaly scoring model based on the environmental health index and passenger density detection data, and utilize a graph neural network to correlate the environmental parameter correlations between cabin zones to generate a partitioned anomaly probability heat map, wherein the partitioned anomaly probability heat map identifies the location of high-risk areas and the level of anomaly; A display module is used to dynamically allocate monitoring resources based on the partition abnormality probability heat map through an adaptive attention mechanism, prioritize increasing the sensor sampling frequency in high-risk areas, and superimpose real-time environmental parameters, the partition abnormality probability heat map, and risk warning prompts on the car display screen; The early warning module is used to generate an environmental deterioration early warning signal through a spatiotemporal prediction model based on the duration and spatial diffusion trend of the partitioned anomaly probability heat map. When the environmental deterioration early warning signal level reaches the preset signal level, the gradient alarm mechanism is triggered and the building management system is linked to adjust the elevator scheduling strategy, forming a closed-loop monitoring response.

8. The system according to claim 7, characterized in that The fusion module is specifically used to: Based on the original environmental data stream collected by the multimodal sensor array, a sliding time window is used to segment the temperature, humidity, CO2 concentration, volatile organic compounds and voiceprint signals to generate time-aligned multimodal data slices; The multimodal data slices are spatially encoded, the car is divided into gridded areas, and the topological relationships of the sensor nodes are annotated to construct a spatiotemporal correlation tensor. The spatiotemporal correlation tensor is input into a 3D convolution-long short-term memory hybrid network. The local region features are extracted through the spatial convolution kernel, and the fluctuation pattern across the time window is captured by combining the time axis LSTM. The feature vector set that integrates the spatiotemporal features is output. Regional features are aggregated based on the feature vector set, and a graph pooling algorithm is used to generate a dynamic environmental feature map covering the entire area of ​​the car. The graph nodes encode the regional environmental status, and the edge weights represent the correlation between environmental parameters between regions.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

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 configured to run the computer program to perform the method according to any one of claims 1 to 6.

Images