Elevator anti-rolling control method and system

Abstract

The invention discloses an elevator anti-entrainment control method and system. The method comprises the steps that S1, an elevator anti-entrainment multi-mode data set is constructed; s2, multi-modal data preprocessing and feature extraction are carried out; s3, performing control strategy training based on deep reinforcement learning, performing strategy optimization by adopting a PPO algorithm, and adopting a two-stage training strategy of simulation pre-training and real system migration; s4, performing model optimization and robustness enhancement; and S5, actually deploying and using the model. Through a simulation-reality transfer learning mechanism and a hierarchical response strategy, an intelligent and self-adaptive integrated solution is provided for elevator safety protection, and the protection precision and the response speed are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an elevator anti-entanglement control method and system based on deep reinforcement learning. Background Art

[0002] Elevators, as an indispensable means of vertical transportation in modern buildings, carry hundreds of millions of passengers every day, and their safety is directly related to the safety of public life and property. However, elevator door systems (especially automatic doors) have certain safety hazards during operation, among which "involvement accidents" are one of the most common safety risks. This type of accident usually occurs when passengers' hands, clothes, luggage, etc. are caught by the elevator doors and move with the doors, resulting in squeezing, dragging, or even more serious injuries. According to data from the State Administration for Market Regulation, approximately 15% to 20% of elevator accidents are related to door system failures or improper operation, of which "involvement" accidents account for a large proportion. Therefore, the development of a systematic elevator anti-involvement control method and system is of great significance to improving the safety and reliability of elevators.

[0003] The development of a deep learning-based intelligent anti-entanglement control system significantly improves detection accuracy and response speed through multimodal sensor fusion and real-time dynamic decision-making, thereby effectively reducing the occurrence of accidents. This technology is not only a major breakthrough in elevator safety, but also an indispensable part of the smart building ecosystem. It has important implications for improving future urban safety and efficiency, and can bring multiple benefits to society and the economy. It has important practical significance and broad application prospects.

[0004] The existing elevator anti-entanglement control method is as follows:

[0005] Chinese invention patent application number CN202211119372.7 discloses an elevator anti-entanglement control method, system, elevator, and storage medium. The method requires the installation of a photoelectric sensor at the tangent point between the elevator's rotating components and the wire rope. This prevents human entanglement or foreign objects from entering the rotating components during operation and maintenance, potentially impacting the safety of operators and maintenance personnel, and improves elevator operation safety and stability.

[0006] Chinese invention patent application number CN201610098208.0 discloses an automatic elevator deceleration and braking system, comprising at least a car brake, a car opening deceleration device, an escalator anti-entrapment device, and a car-type elevator centralized monitoring and remote interaction system. This system achieves braking and monitoring functions without affecting the normal operation of the elevator, and features a simple structure, low cost, and excellent performance.

[0007] However, the above method has the following limitations:

[0008] 1) The detection method is single, relies on specific sensors, and is easily affected by environmental interference.

[0009] 2) Static threshold control, lack of intelligent dynamic decision-making capabilities.

[0010] 3) Mechanical braking solutions have slow response speeds and inherent delays.

[0011] 4) Insufficient intelligence and lack of independent learning ability.

[0012] 5) Remote monitoring functions are simple and lack in-depth analysis capabilities. Current technology only implements basic fault alarm and work dispatch functions, and is unable to conduct in-depth data analysis and predictive maintenance. It cannot provide more valuable decision support for elevator safety management, and the system lacks scalability and intelligence. Summary of the Invention

[0013] (1) Technical problems solved

[0014] In response to the shortcomings of the existing technology, the present invention provides an intelligent and adaptive integrated solution for elevator safety protection through a simulation-reality transfer learning mechanism and a hierarchical response strategy, which significantly improves the protection accuracy and response speed.

[0015] (2) Technical solution

[0016] To achieve the above object, the present invention provides the following technical solution: an elevator anti-entanglement control method, comprising the following steps:

[0017] Step S1: constructing a multimodal dataset for elevator anti-entanglement;

[0018] Step S2: multimodal data preprocessing and feature extraction;

[0019] Step S3: Control strategy training based on deep reinforcement learning, using the PPO algorithm for strategy optimization, and adopting a two-stage training strategy of simulation pre-training + real system migration;

[0020] Step S4: model optimization and robustness enhancement;

[0021] Step S5: Actual deployment and use of the model.

[0022] As a preferred solution, the step S1: constructing a multimodal dataset for elevator anti-entanglement, specifically includes the following steps:

[0023] S1-1: Multimodal sensor deployment and data acquisition: A collaborative deployment of depth cameras and high-precision thin-film pressure sensors enables the simultaneous acquisition of 3D visual information and mechanical parameters.

[0024] S1-2: The annotation work uses the professional CVAT annotation platform, and the implementation process is as follows:

[0025] First, we developed annotation specifications and defined three levels of safety status. The "safe" state requires the object to be more than 15cm from the door gap and the pressure to be less than 1N; the "warning" state requires the object to be 5-15cm away or the pressure to be 1-2N; and the "dangerous" state requires the object to be less than 5cm away (with a pressure greater than 2N). A detailed annotation example library was established for common object types (hands, clothing, luggage, etc.).

[0026]

[0027] where d obj is the shortest distance between the object and the door gap, and P is the peak pressure;

[0028] Collaborative labeling is performed simultaneously, with a three-person labeling team formed and cross-validated using the CVAT review model. Two labelers independently label each video segment, and any disagreements are arbitrated by experts. The labeling results must meet consistency standards.

[0029] S1-3: Multimodal dataset construction: The constructed dataset contains the following key features:

[0030] 1) Spatiotemporal alignment of data: Visual data and mechanical signals are precisely aligned by timestamp. Each sample contains a 30-frame RGB-D image sequence (approximately 1 second in length) and the corresponding pressure sensor array data.

[0031]

[0032] where ΔP t is the pressure difference feature, T is the time window size (30 frames by default), and D is the dimension;

[0033] 2) Hazard level labeling: In addition to the basic safety status label, a comprehensive hazard index is calculated. This index takes multiple dimensions into consideration and uses an expert scoring method to determine the weight of each dimension.

[0034] As a preferred solution, the step S2: multimodal data preprocessing and feature extraction specifically includes the following steps:

[0035] S2-1: Visual data processing flow: Visual data processing adopts a multi-stage pipeline architecture:

[0036] First, we use OpenCV's CLAHE algorithm to perform adaptive histogram equalization on the RGB image to effectively improve the image quality under low-light conditions. We also perform median filtering and hole filling on the depth map to eliminate the effects of sensor noise.

[0037] We use an improved MOG2 background modeling algorithm to accurately separate dynamic objects by analyzing the depth changes between consecutive frames. We have optimized the algorithm parameters to adapt to the scene changes when the elevator door opens and closes. The following is the background modeling idea of ​​the improved MOG2:

[0038] μ t =ρI t +(1-ρ)μ t-1

[0039] The dynamic learning rate ρ = 0.05

[0040] At the same time, the YOLOv8s model was deployed for real-time detection and fine-tuned for the elevator scenario: the input resolution was adjusted to 640×384 to balance accuracy and speed, and Focal Loss was used to address the class imbalance problem. While maintaining a mAP above 85%, the inference speed was improved. At the same time, the loss function of YOLOv8s was improved:

[0041]

[0042] where λ box =0.01,λ cls =0.05

[0043] S2-2: Mechanical signal processing: Key steps in pressure signal processing:

[0044] First, signal preprocessing is performed. A sliding window filter is designed with a window size of 1 second (100 sampling points). The mean and standard deviation of each channel signal are calculated in real time, and outliers are automatically identified and removed. To address the vibration interference during elevator operation, a frequency domain filtering method is used, using a low-pass filter to eliminate high-frequency noise. The frequency domain threshold of the adaptive filter is:

[0045] τ=μ w +3σ w

[0046] where w = 1s

[0047] Based on this, feature extraction is performed to construct a two-layer LSTM network to process time series pressure signals. The network hidden layer dimension is 128, and an additional activation function is used. A differential feature input is specially designed, and the deviation between the current pressure value and the historical mean is used as an additional input feature to enhance the ability to detect sudden anomalies. The LSTM state update steps are as follows:

[0048] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0049] The complete formula of the gating mechanism of the LSTM network is:

[0050] i t =σ(W ii x t +b ii +W hi h t-1 +b hi )

[0051] f t =σ(W if x t +b if +W hf h t-1 +b hf )

[0052] o t =σ(W io x t +b io +W ho h t-1 +b ho )

[0053]

[0054] h t =o t ⊙tanh(c t )

[0055] The features extracted in step S2 can be fed back to the GAN generator in step S1 to optimize the distribution of difficult examples in the synthetic data (such as increasing fast reaching samples).

[0056] As a preferred solution, step S3: control strategy training based on deep reinforcement learning specifically includes the following steps:

[0057] S3-1: State space definition: First, multimodal feature fusion is performed, that is, the visual features (object position, depth information) and pressure features (force magnitude, rate of change) extracted in step S2 are combined into a state vector. The specific process is as follows:

[0058] Visual feature processing: Perform maximum pooling on the 1024-dimensional features output by EfficientNet, retain key spatial information, and reduce the dimension to 256 dimensions;

[0059] Pressure feature processing: Calculate the mean (reflecting the persistence) and standard deviation (reflecting the mutation) of the pressure of the last 10 frames to obtain a 2D feature;

[0060] Temporal feature supplement: Use OpenCV's Farneback algorithm to calculate the optical flow of adjacent frames and obtain the object motion trend (2D vector);

[0061] Final state vector: The above features are concatenated into a 260-dimensional vector as the input state of DRL. The formula is simplified to:

[0062] s t =Concat(f v ,f p ,f o )

[0063] Among them, visual features Pressure characteristics Optical flow features State vector Concatenate visual, pressure, and motion features as the input state of DRL;

[0064] S3-2: Action Space Design: Control Instruction Mapping In order to map discrete actions to the actual control variables of the elevator door, four discrete actions corresponding to different control modes of the elevator door are designed:

[0065] 1) Continue closing the door (default speed 0.2m / s): for risk-free scenarios;

[0066] 2) Slow down door closing (speed 0.05m / s): triggered when an object is detected approaching the door gap (distance <10cm);

[0067] 3) Pause and hold: completely stop the door movement for 1 second, waiting for the risk to be eliminated (such as clothing flying away);

[0068] 4) Emergency rebound: 0.5m / s 2 Open the door in reverse acceleration mode to handle hard object obstruction (pressure > 5N):

[0069]

[0070] S3-3: Reward function design: A multi-objective balance strategy is introduced. The reward function consists of three parts:

[0071] Safety bonus: +10 points for successfully avoiding a pinch (according to the "Danger" label in S1-2), -100 points for a pinch.

[0072] Efficiency penalty: 1 point will be deducted for each unnecessary pause (no danger in the next 5 frames) to prevent excessive conservatism.

[0073] Smoothness bonus: If the difference between adjacent actions is too large (such as sudden stop and turn), 5 points will be deducted;

[0074] The formula is simplified to:

[0075] R=10·Safe-100·Danger-1·UnnecessaryStop-5·Jerk

[0076] S3-4: DRL algorithm training: Training process:

[0077] Offline stage: In the system simulation environment, load the S1 dataset and simulate 5 million steps of training with a batch size of 2048.

[0078] Online fine-tuning: Receive real elevator sensor data through ROS and prioritize replaying samples with large prediction errors:

[0079] The training process can be expressed as:

[0080]

[0081] Among them, the input state s t , action a t , the output θ is the policy parameter, and the training stability is improved by clipping the policy gradient.

[0082] As a preferred solution, step S4: model optimization and robustness enhancement, specifically includes the following steps:

[0083] S4-1: Prioritized Experience Replay: Sample Importance Weighting: Dynamically adjust sample priorities based on the following rules: Samples with large prediction errors (e.g., misjudged danger levels) have their priority increased by 50%. Experts label difficult examples with the highest priority. Using PyTorch's priority replay buffer, the priority is updated every 1000 steps. The following is the generated priority p t Method, where high error samples are trained first:

[0084] p t =|δ t ∣+0.5·KL(π θ ||π old )

[0085] S4-2: Dynamic Learning Rate Adjustment: To quickly adapt across scenarios, the MAML algorithm is used to enable the model to:

[0086] Inner layer update: fine-tuning on a single elevator data (10 steps of gradient descent);

[0087] Outer layer update: aggregate generalized parameters of 20 different elevator models (door width / speed differences);

[0088] S4-3: Risk Probability Control: Constrained optimization is used to ensure that when the predicted probability of danger exceeds 1%, the "emergency rebound" action is triggered. Lagrange multipliers are used to dynamically adjust the balance between safety and efficiency:

[0089]

[0090] This can control the hard limit hazard incidence to <1%;

[0091] S4-4: Online Adaptation: Use PyTorch's ReduceLROnPlateau scheduler to dynamically adjust the learning rate, adjusting the learning rate based on the average reward of the last 10 training runs:

[0092] When the reward is rising: the learning rate is increased by 10% (to speed up convergence), when the reward is falling: the learning rate is reduced by 50% (to avoid oscillation).

[0093] As a preferred solution, step S5 builds a low-latency inference system based on an edge computing platform. It designs a two-level response mechanism: Level 1 (warning deceleration) and Level 2 (emergency braking). Dynamic thresholds are set to implement risk-based management. The deployment solution includes a comprehensive anomaly monitoring and fault recovery mechanism, supporting hot updates and remote diagnosis. The system is deeply integrated with the elevator control system via the CAN bus, while retaining a hardware emergency stop circuit for safety redundancy.

[0094] An elevator anti-entanglement control system builds an all-round sensing system in the elevator door area, including:

[0095] The visual perception unit uses a depth camera installed at the top center of the elevator door frame, tilted downward at a 15-degree angle to ensure complete coverage of the door gap. The camera simultaneously outputs 1280×720 resolution RGB images and depth maps while maintaining a stable frame rate. Data acquisition is achieved through the driver, and a time synchronization service is used to ensure that the time alignment error of each sensor data is less than 5ms. The depth camera output data format is:

[0096]

[0097] The force sensing unit deploys ultra-thin pressure sensors symmetrically on both sides of the elevator door gap, evenly spaced 15 cm apart. The sensor signals are digitized using a high-precision acquisition card, with a sampling rate set to 100 Hz to capture rapid pressure changes. Each sensor channel is calibrated on-site to ensure a measurement error of less than 0.1 N. The output format is:

[0098]

[0099] The data synchronization system adopts a hardware synchronization solution based on the PTP protocol and allows all sensors to provide a unified time base. The collected data is stored in the ROS2 bag format with timestamp alignment for easy subsequent processing.

[0100] As a preferred solution, it also includes:

[0101] Edge computing module, the device needs to be equipped with multiple high-speed interfaces, including USB and GPIO, for connecting to the vision sensor and pressure sensor array;

[0102] The pixel calculation module pre-processes the original image to improve the accuracy of subsequent processing, and then uses an optimized object detection algorithm to analyze the pre-processed image to accurately identify various objects in the picture;

[0103] The hazard assessment module is responsible for performing real-time risk assessment of detected objects. This module needs to load a fully trained and optimized deep learning model and receive object information from the image processing module in real time during operation for time series analysis. The assessment results are expressed as a standardized hazard level with an accompanying confidence indicator.

[0104] The response module implements a hierarchical response strategy based on the output of the risk assessment module. The response mechanism is divided into multiple levels, each corresponding to a different processing strategy. Each response level has clear condition judgment criteria and corresponding execution actions. Response instructions are transmitted to the execution agency through a standardized interface to ensure that the instructions can be executed accurately and promptly.

[0105] The monitoring and debugging module collects the operating status information of each module in real time, including key indicators such as processing delay and resource utilization, and intuitively displays the system operating status and risk assessment results through a visual interface, making it easier for operators to understand the on-site situation.

[0106] (3) Beneficial effects

[0107] Compared with the prior art, the present invention provides an elevator anti-entanglement control method and system, which has the following beneficial effects:

[0108] 1. This invention achieves omnidirectional environmental perception through multimodal sensor fusion technology, and combines advanced target detection algorithms and time series modeling methods to accurately identify various types of involvement risks. It uses a deep reinforcement learning framework to build an adaptive decision-making system, and achieves an optimal balance between safety, operational efficiency, and control smoothness through a carefully designed multi-objective reward function. It innovatively applies optimization strategies such as dynamic priority and meta-learning to improve system robustness, and uses edge computing platforms to achieve rapid response. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] Figure 1 This is a schematic diagram of the technical process of the control method of the present invention;

[0110] Figure 2 This is a schematic diagram of the LSTM state of the present invention;

[0111] Figure 3This is a network structure diagram of deep reinforcement learning in the present invention;

[0112] Figure 4 Schematic diagram of the deep reinforcement learning strategy of the present invention;

[0113] Figure 5 Deploy an actual flow chart for the model of the present invention. DETAILED DESCRIPTION

[0114] In order to better understand the purpose, structure and function of the present invention, the elevator anti-entrapment control method and system of the present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0115] Example 1

[0116] refer to Figure 1-5 The present invention provides an elevator anti-entanglement control method, comprising the following steps:

[0117] Step S1: constructing a multimodal dataset for elevator anti-entanglement;

[0118] Step S2: multimodal data preprocessing and feature extraction;

[0119] Step S3: Control strategy training based on deep reinforcement learning, using the PPO algorithm for strategy optimization, and adopting a two-stage training strategy of simulation pre-training + real system migration;

[0120] Step S4: model optimization and robustness enhancement;

[0121] Step S5: Actual deployment and use of the model.

[0122] First, depth cameras and pressure sensors are deployed in the elevator door area to build a multidimensional perception network, collecting real-time visual information and pressure data from the elevator door area. Object detection algorithms and time series modeling methods are used to extract spatial-temporal features and construct a multimodal dataset. Next, a deep reinforcement learning framework is innovatively introduced to design a multi-objective reward function that integrates safety, efficiency, and smoothness requirements, and train an intelligent decision-making model. Dynamic prioritization and meta-learning optimization strategies are then employed to enhance model robustness, and simulation-to-reality transfer learning is used to ensure algorithm generalization. Finally, the system is deployed on an edge computing platform to implement a hierarchical control strategy with millisecond-level response.

[0123] Specifically, S1. Elevator anti-entanglement multimodal dataset construction

[0124] During this phase, a multimodal data system was constructed for elevator anti-entanglement tasks. A collaborative deployment of depth cameras and high-precision thin-film pressure sensors was adopted to achieve the simultaneous acquisition of three-dimensional visual information and mechanical parameters. Automatic labeling was achieved using the YOLOv8 target detection algorithm, and combined with manual verification, a labeling system with three levels of "safety / warning / danger" labels was constructed. The innovative introduction of generative adversarial network (GAN) technology to synthesize extreme case samples effectively addressed the scarcity of dangerous samples in real-world scenarios. The dataset covers different lighting conditions, object materials, and motion patterns, providing comprehensive data support for model training.

[0125] S2. Multimodal data preprocessing and feature extraction

[0126] A dual-modal parallel processing architecture was established: the visual channel uses the EfficientNet network to extract deep features from RGB-D images and combines it with optical flow to calculate motion vectors. The pressure signal channel uses an LSTM network to model temporal patterns and a dynamic threshold algorithm to detect anomalies. An innovative cross-modal feature alignment mechanism was designed to map and associate visual space coordinates with pressure distribution. Feature-level fusion forms a unified feature representation encompassing visual semantics, motion dynamics, and contact mechanics, providing multi-dimensional information input for decision-making and control.

[0127] S3. Control strategy training based on deep reinforcement learning

[0128] A discrete action space and three-dimensional reward function were designed, and the PPO algorithm was used for policy optimization. The state space integrates visual features (target position, motion trend) and pressure features to construct a multi-dimensional state vector. An innovative dynamic reward balancing mechanism was designed to address the trade-off between safety and operational efficiency through adaptive weight adjustment. A two-stage training strategy of "simulation pre-training + real system migration" was adopted, combined with human expert demonstration data to achieve policy fine-tuning, significantly improving the practicality and safety of the policy.

[0129] S4. Model Optimization and Robustness Enhancement Based on Dynamic Prioritization and Meta-Learning

[0130] The strategy of combining dynamic priority with meta-learning significantly improves the adaptability and robustness of the model. In terms of dynamic priority, the system constructs a dual weight mechanism based on the prediction error of the sample and the difficulty of expert labeling. Samples with large prediction errors and difficult samples that have been repeatedly corrected by humans will receive higher training weights. This mechanism is implemented through a weighted loss function, which enables the model to focus on learning key samples. In terms of meta-learning, the MAML framework is used for multi-task pre-training, so that the model can obtain the generalization ability to quickly adapt to new scenarios. At the same time, the Lagrange multiplier method is introduced to perform hard optimization of safety constraints, strictly control the hazard incidence rate to below 1%, and automatically balance the convergence speed and stability through a dynamic learning rate adjustment strategy. These optimization measures work together to make the final model more robust to real-world interference such as data distribution changes and sensor noise while maintaining high accuracy;

[0131] S5. Real-time deployment and system integration

[0132] A low-latency inference system is built on an edge computing platform. A two-level response mechanism, Level 1 (warning deceleration) and Level 2 (emergency braking), is designed, with dynamic thresholds implemented to implement risk-based management. The deployment solution includes a comprehensive anomaly monitoring and fault recovery mechanism, supporting hot updates and remote diagnostics. The system is deeply integrated with the elevator control system via the CAN bus, while retaining a hardware emergency stop circuit for safety redundancy.

[0133] Furthermore, in the above solution, step S1: constructing a multimodal dataset for elevator anti-entanglement, specifically includes the following steps:

[0134] S1-1: Multimodal sensor deployment and data collection to build a comprehensive perception system in the elevator door area, including:

[0135] 1) Visual Perception Unit: A depth camera is installed at the top center of the elevator door frame, tilted downward at a 15-degree angle to ensure complete coverage of the door gap. The camera simultaneously outputs a 1280×720 resolution RGB image and depth map while maintaining a stable frame rate. Data acquisition is achieved through a driver, and a time synchronization service is used to ensure that the time alignment error of each sensor data is less than 5ms. The depth camera output data format is:

[0136]

[0137] 2) Force Sensing Unit: Ultra-thin pressure sensors are symmetrically placed on both sides of the elevator door gap, evenly spaced 15 cm apart. The sensor signals are digitized using a high-precision acquisition card with a sampling rate of 100 Hz to capture rapid pressure changes. Each sensor channel is calibrated on-site to ensure measurement error of less than 0.1 N. The output format is:

[0138]

[0139] 3) Data Synchronization System: A hardware synchronization solution based on the PTP protocol is used, and all sensors provide a unified time base. The collected data is stored in the ROS2 bag format with timestamp alignment for easy subsequent processing.

[0140] S1-2: The annotation work uses the professional CVAT annotation platform, and the implementation process is as follows:

[0141] First, we developed annotation specifications and defined three levels of safety status. The "safe" state requires the object to be more than 15cm from the door gap and the pressure to be less than 1N; the "warning" state requires the object to be 5-15cm away or the pressure to be 1-2N; and the "dangerous" state requires the object to be less than 5cm away (with a pressure greater than 2N). A detailed annotation example library was established for common object types (hands, clothing, luggage, etc.).

[0142]

[0143] where d obj is the shortest distance between the object and the door gap, and P is the peak pressure;

[0144] Collaborative labeling is performed simultaneously, with a three-person labeling team formed and cross-validated using the CVAT review model. Two labelers independently label each video segment, and any disagreements are arbitrated by experts. The labeling results must meet consistency standards.

[0145] S1-3: Multimodal dataset construction: The constructed dataset contains the following key features:

[0146] 1) Spatiotemporal alignment of data: Visual data and mechanical signals are precisely aligned by timestamp. Each sample contains a 30-frame RGB-D image sequence (approximately 1 second in length) and the corresponding pressure sensor array data.

[0147]

[0148] where ΔP t is the pressure difference feature, T is the time window size (30 frames by default), and D is the dimension;

[0149] 2) Hazard level labeling: In addition to the basic safety status label, a comprehensive hazard index is calculated. This index takes multiple dimensions into consideration and uses an expert scoring method to determine the weight of each dimension.

[0150] Specifically, step S2: multimodal data preprocessing and feature extraction, specifically includes the following steps:

[0151] S2-1: Visual data processing flow: Visual data processing adopts a multi-stage pipeline architecture:

[0152] First, we use OpenCV's CLAHE algorithm to perform adaptive histogram equalization on the RGB image to effectively improve the image quality under low-light conditions. We also perform median filtering and hole filling on the depth map to eliminate the effects of sensor noise.

[0153] We use an improved MOG2 background modeling algorithm to accurately separate dynamic objects by analyzing the depth changes between consecutive frames. We have optimized the algorithm parameters to adapt to the scene changes when the elevator door opens and closes. The following is the background modeling idea of ​​the improved MOG2:

[0154] μ t =ρI t +(1-ρ)μ t-1

[0155] The dynamic learning rate ρ = 0.05

[0156] At the same time, the YOLOv8s model was deployed for real-time detection and fine-tuned for the elevator scenario: the input resolution was adjusted to 640×384 to balance accuracy and speed, and Focal Loss was used to address the class imbalance problem. While maintaining a mAP above 85%, the inference speed was improved. At the same time, the loss function of YOLOv8s was improved:

[0157]

[0158] where λ box =0.01,λ cls =0.05

[0159] S2-2: Mechanical signal processing: Key steps in pressure signal processing:

[0160] First, signal preprocessing is performed. A sliding window filter is designed with a window size of 1 second (100 sampling points). The mean and standard deviation of each channel signal are calculated in real time, and outliers are automatically identified and removed. To address the vibration interference during elevator operation, a frequency domain filtering method is used, using a low-pass filter to eliminate high-frequency noise. The frequency domain threshold of the adaptive filter is:

[0161] τ=μ w +3σ w

[0162] where w = 1s

[0163] Based on this, feature extraction is performed to construct a two-layer LSTM network to process time series pressure signals. The network hidden layer dimension is 128, and an additional activation function is used. A differential feature input is specially designed, and the deviation between the current pressure value and the historical mean is used as an additional input feature to enhance the ability to detect sudden anomalies. The LSTM state update steps are as follows:

[0164] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0165] The complete formula of the gating mechanism of the LSTM network is:

[0166] i t =σ(W ii x t +b ii +W hi h t-1 +b hi )

[0167] f t =σ(W if x t +b if +W hf h t-1 +b hf )

[0168] o t =σ(W io x t +b io +W ho h t-1 +b ho )

[0169]

[0170] h t =o t ⊙tanh(c t )

[0171] The features extracted in step S2 can be fed back to the GAN generator in step S1 to optimize the distribution of difficult examples in the synthetic data (such as increasing fast reaching samples).

[0172] Specifically, step S3: control strategy training based on deep reinforcement learning, specifically includes the following steps:

[0173] S3-1: State space definition: First, multimodal feature fusion is performed, that is, the visual features (object position, depth information) and pressure features (force magnitude, rate of change) extracted in step S2 are combined into a state vector. The specific process is as follows:

[0174] Visual feature processing: Perform maximum pooling on the 1024-dimensional features output by EfficientNet, retain key spatial information, and reduce the dimension to 256 dimensions;

[0175] Pressure feature processing: Calculate the mean (reflecting the persistence) and standard deviation (reflecting the mutation) of the pressure of the last 10 frames to obtain a 2D feature;

[0176] Temporal feature supplement: Use OpenCV's Farneback algorithm to calculate the optical flow of adjacent frames and obtain the object motion trend (2D vector);

[0177] Final state vector: The above features are concatenated into a 260-dimensional vector as the input state of DRL. The formula is simplified to:

[0178] s t =Concat(f v ,f p ,f o )

[0179] Among them, visual features Pressure characteristics Optical flow features State vector Concatenate visual, pressure, and motion features as the input state of DRL;

[0180] S3-2: Action Space Design: Control Instruction Mapping In order to map discrete actions to the actual control variables of the elevator door, four discrete actions corresponding to different control modes of the elevator door are designed:

[0181] 1) Continue closing the door (default speed 0.2m / s): for risk-free scenarios;

[0182] 2) Slow down door closing (speed 0.05m / s): triggered when an object is detected approaching the door gap (distance <10cm);

[0183] 3) Pause and hold: completely stop the door movement for 1 second, waiting for the risk to be eliminated (such as clothing flying away);

[0184] 4) Emergency rebound: 0.5m / s 2 Open the door in reverse acceleration mode to handle hard object obstruction (pressure > 5N):

[0185]

[0186] S3-3: Reward function design: A multi-objective balance strategy is introduced. The reward function consists of three parts:

[0187] Safety bonus: +10 points for successfully avoiding a pinch (according to the "Danger" label in S1-2), -100 points for a pinch.

[0188] Efficiency penalty: 1 point will be deducted for each unnecessary pause (no danger in the next 5 frames) to prevent excessive conservatism.

[0189] Smoothness bonus: If the difference between adjacent actions is too large (such as sudden stop and turn), 5 points will be deducted;

[0190] The formula is simplified to:

[0191] R=10·Safe-100·Danger-1·UnnecessaryStop-5·Jerk

[0192] S3-4: DRL algorithm training: Training process:

[0193] Offline stage: In the system simulation environment, load the S1 dataset and simulate 5 million steps of training with a batch size of 2048.

[0194] Online fine-tuning: Receive real elevator sensor data through ROS and prioritize replaying samples with large prediction errors:

[0195] The training process can be expressed as:

[0196]

[0197] Among them, the input state s t , action a t , the output θ is the policy parameter, and the training stability is improved by clipping the policy gradient.

[0198] Specifically, the step S4: model optimization and robustness enhancement, specifically includes the following steps:

[0199] S4-1: Prioritized Experience Replay: Sample Importance Weighting: Dynamically adjust sample priorities based on the following rules: Samples with large prediction errors (e.g., misjudged danger levels) have their priority increased by 50%. Experts label difficult examples with the highest priority. Using PyTorch's priority replay buffer, the priority is updated every 1000 steps. The following is the generated priority p t Method, where high error samples are trained first:

[0200] p t =|δ t ∣+0.5·KL(π θ ||π old )

[0201] S4-2: Dynamic Learning Rate Adjustment: To quickly adapt across scenarios, the MAML algorithm is used to enable the model to:

[0202] Inner layer update: fine-tuning on a single elevator data (10 steps of gradient descent);

[0203] Outer layer update: aggregate generalized parameters of 20 different elevator models (door width / speed differences);

[0204] S4-3: Risk Probability Control: Constrained optimization is used to ensure that when the predicted probability of danger exceeds 1%, the "emergency rebound" action is triggered. Lagrange multipliers are used to dynamically adjust the balance between safety and efficiency:

[0205]

[0206] This can control the hard limit hazard incidence to <1%;

[0207] S4-4: Online Adaptation: Use PyTorch's ReduceLROnPlateau scheduler to dynamically adjust the learning rate, and adjust the learning rate based on the average reward of the last 10 training runs:

[0208] When the reward is rising: the learning rate is increased by 10% (to speed up convergence), when the reward is falling: the learning rate is reduced by 50% (to avoid oscillation).

[0209] Specifically, step S5 builds a low-latency inference system based on the edge computing platform. Real-time deployment and system integration complete the entire technical roadmap, forming a complete closed loop with the previous four modules. S5 directly relies on the DRL control strategy trained in S3 and the model parameters optimized in S4. The risk assessment model generated in S3 and the decision algorithm adjusted in S4 are deployed to the edge device. The reward function designed in S3-3 directly determines the threshold setting of S5's graded response mechanism. Furthermore, S5's pixel calculation module is fully based on the multimodal feature extraction process of S2, inheriting S2's visual data processing methods and S2-2's pressure feature extraction scheme. The overall S5 architecture feeds back to optimize S1's data acquisition system. Edge cases discovered by the S5-5 monitoring module are fed back into the S1-2 data annotation phase to enhance the coverage of the original dataset, forming a complete technical chain from data to decision-making. The specific implementation steps include:

[0210] Example 2

[0211] This embodiment proposes an elevator anti-entrapment control system, which constructs an all-round sensing system in the elevator door area, including:

[0212] The visual perception unit uses a depth camera installed at the top center of the elevator door frame, tilted downward at a 15-degree angle to ensure complete coverage of the door gap. The camera simultaneously outputs 1280×720 resolution RGB images and depth maps while maintaining a stable frame rate. Data acquisition is achieved through the driver, and a time synchronization service is used to ensure that the time alignment error of each sensor data is less than 5ms. The depth camera output data format is:

[0213]

[0214] The force sensing unit deploys ultra-thin pressure sensors symmetrically on both sides of the elevator door gap, evenly spaced 15 cm apart. The sensor signals are digitized using a high-precision acquisition card, with a sampling rate set to 100 Hz to capture rapid pressure changes. Each sensor channel is calibrated on-site to ensure a measurement error of less than 0.1 N. The output format is:

[0215]

[0216] The data synchronization system adopts a hardware synchronization solution based on the PTP protocol and allows all sensors to provide a unified time base. The collected data is stored in the ROS2 bag format with timestamp alignment for easy subsequent processing.

[0217] The system of the present invention also includes: an edge computing module, a pixel computing module, a risk assessment module, a response module and a monitoring and debugging module. The specific implementation method is as follows:

[0218] Edge computing hardware deployment

[0219] During the system deployment phase, the primary task is to select appropriate edge computing devices to meet the requirements of real-time inference. Considering the balance between computing performance and power consumption, a high-performance edge computing module is selected. The device must be equipped with multiple high-speed interfaces, including USB and GPIO, for connecting to the visual sensor and pressure sensor array. The visual sensor must support high-resolution image acquisition and depth information acquisition, and the frame rate must meet real-time processing requirements. The pressure sensor must exhibit high sensitivity and fast response characteristics and be connected to the computing device via a dedicated analog-to-digital conversion module. Regarding system configuration, a dedicated embedded operating system and deep learning inference framework must be installed, and targeted performance optimization, including memory management strategies and power management modes, must be implemented to ensure continuous and stable system operation. A comprehensive cooling solution must also be established to prevent overheating and frequency throttling during prolonged high-load operation.

[0220] Pixel computing module deployment

[0221] The image processing module, serving as the front-end processing unit of the entire system, is responsible for real-time processing of the data stream from the visual sensor. This module first preprocesses the raw image to improve the accuracy of subsequent processing. It then uses an optimized object detection algorithm to analyze the preprocessed image and accurately identify objects within the image. The algorithm outputs precise pixel coordinate information for each detected object, including the position and size of its bounding box. To achieve efficient real-time processing, this module fully utilizes the hardware acceleration capabilities of the computing device, using a dedicated image processing library and deep learning inference engine to increase processing speed. Furthermore, a comprehensive error handling mechanism is implemented to ensure stable detection performance even in complex environments. The processing results are then transmitted to the subsequent hazard assessment module in a standardized data structure.

[0222] Hazard prediction model deployment

[0223] The hazard assessment module is the core decision-making unit of the entire system, responsible for real-time risk assessment of detected objects. This module requires a fully trained and optimized deep learning model that comprehensively analyzes multiple factors, including an object's motion trajectory and spatial position. Before deployment, the model undergoes specialized lightweight processing, including quantization and pruning, to accommodate the resource constraints of edge computing devices. During runtime, the module receives real-time object information from the image processing module for time series analysis. Assessment results are expressed as a standardized hazard level accompanied by a confidence indicator. To ensure real-time performance, the module requires an efficient memory management mechanism and data processing pipeline. Furthermore, a comprehensive exception handling process must be designed to maintain system stability in the event of input data anomalies or model inference issues.

[0224] Tiered response mechanism

[0225] Based on the output of the hazard assessment module, the system implements a hierarchical response strategy. The response mechanism is divided into multiple levels, each corresponding to a different handling strategy. Each response level has clear criteria for determining conditions and corresponding execution actions. Response instructions are transmitted to the actuator through a standardized interface, ensuring accurate and timely execution. The system also records all response events, including triggering conditions, response time, and execution results, for subsequent analysis and optimization. The response mechanism must undergo thorough testing and verification to ensure reliable operation under various operating conditions.

[0226] System monitoring and debugging

[0227] To ensure the long-term stability of the system, a comprehensive monitoring and debugging mechanism must be established. The monitoring system must collect real-time operational status information for each module, including key metrics such as processing latency and resource utilization. A visual interface should intuitively display system operational status and risk assessment results, allowing operators to easily understand the on-site situation. The system must have comprehensive logging capabilities to store all processing steps and decision-making basis, supporting post-event analysis and problem tracing. Debugging tools must support both individual module testing and overall joint debugging, simulating various input conditions to verify system responses.

[0228] like Figure 5 shown

[0229] This experiment, based on an actual elevator operating environment, continuously collected multimodal data over a six-month period, including RGB-D visual data (resolution 1280×720@30fps), pressure sensor signals (sampling rate 1kHz), and elevator door status data (door opening and closing speed, position, etc.). Data collection covered eight typical scenarios, including normal passage, carrying large luggage, children passing quickly, clothing entanglement, and hard object obstruction. A total of 25,000 valid samples were collected, divided into training and test sets in an 8:2 ratio:

[0230] 1. Hazard level prediction performance evaluation

[0231] Using precision, recall, and F1-score as evaluation indicators, the experimental results of the danger level prediction performance are shown in Table 1 below:

[0232] Hazard Level Precision(%) Recall (%) F1-score (%) Safety 98.7 99.1 98.2 Early Warning 95.4 96.3 95.1 Danger 97.2 98.3 97.9

[0233] Table 1

[0234] Experimental results show that the model's recognition recall rate for "dangerous" states reaches 98.3%, which can effectively avoid missed detections; at the same time, the accuracy rate for "safe" states is 98.7%, ensuring that there will be no excessive interference with normal operation.

[0235] 2. System response performance evaluation

[0236] The full-process delay from hazard detection to control command execution was measured, and its response time performance is shown in Table 2 below:

[0237] Scenario Type Average latency (ms) Maximum delay (ms) Standard deviation (ms) False positive rate 42 58 4.2 False negative rate 38 53 3.8 System availability 47 68 5.1

[0238] Table 2

[0239] In all test scenarios, the system response time was kept within 70ms, meeting elevator safety standards. In particular, in emergency braking scenarios, the average response time was only 38ms.

[0240] Experimental results demonstrate that this technical solution significantly outperforms traditional solutions in terms of hazard identification accuracy and response speed. In particular, the combined effects of the deep reinforcement learning decision model in step S3 and the meta-learning optimization model in step S4 demonstrate excellent generalization and adaptability, validating the effectiveness of the entire technical approach of this invention.

[0241] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. An elevator anti-entanglement control method, characterized in that: The steps include: Step S1: constructing a multimodal dataset for elevator anti-entanglement; Step S2: multimodal data preprocessing and feature extraction; Step S3: Control strategy training based on deep reinforcement learning, using the PPO algorithm for strategy optimization, and adopting a two-stage training strategy of simulation pre-training + real system migration; Step S4: model optimization and robustness enhancement; Step S5: Actual deployment and use of the model.

2. The elevator anti-entanglement control method according to claim 1, characterized in that: The step S1: constructing a multimodal dataset for elevator anti-entanglement, specifically includes the following steps: S1-1: Multimodal sensor deployment and data acquisition: A collaborative deployment of depth cameras and high-precision thin-film pressure sensors enables the simultaneous acquisition of 3D visual information and mechanical parameters. S1-2: The annotation work uses the professional CVAT annotation platform, and the implementation process is as follows: First, we developed annotation specifications and defined three levels of safety status. The "safe" state requires that the object be more than 15cm away from the door gap and the pressure be less than 1N; the "warning" state requires a distance of 5-15cm or a pressure of 1-2N; and the "dangerous" state requires a distance of less than 5cm. A detailed annotation example library was established for common object types. where d obj is the shortest distance between the object and the door gap, and P is the peak pressure; Collaborative labeling is performed simultaneously, with a three-person labeling team formed and cross-validated using the CVAT review model. Two labelers independently label each video segment, and any disagreements are arbitrated by experts. The labeling results must meet consistency standards. S1-3: Multimodal dataset construction: The constructed dataset contains the following key features: 1) Spatiotemporal alignment of data: Visual data and mechanical signals are precisely aligned by timestamp. Each sample contains a 30-frame RGB-D image sequence and corresponding pressure sensor array data. where ΔP t is the pressure difference feature, T is the time window size, and D is the dimension; 2) Hazard level labeling: In addition to the basic safety status label, a comprehensive hazard index is calculated. This index takes multiple dimensions into consideration and uses an expert scoring method to determine the weight of each dimension.

3. The elevator anti-entanglement control method according to claim 1, characterized in that: The step S2: multimodal data preprocessing and feature extraction specifically includes the following steps: S2-1: Visual data processing flow: Visual data processing adopts a multi-stage pipeline architecture: First, we use OpenCV's CLAHE algorithm to perform adaptive histogram equalization on the RGB image to effectively improve the image quality under low-light conditions. We also perform median filtering and hole filling on the depth map to eliminate the effects of sensor noise. We use an improved MOG2 background modeling algorithm to accurately separate dynamic objects by analyzing the depth changes between consecutive frames. We have optimized the algorithm parameters to adapt to the scene changes when the elevator door opens and closes. The following is the background modeling idea of ​​the improved MOG2: m t =ρI t +(1-p)m t-1 The dynamic learning rate ρ = 0.05 At the same time, the YOLOv8s model was deployed for real-time detection and fine-tuned for the elevator scenario: the input resolution was adjusted to 640×384 to balance accuracy and speed, and Focal Loss was used to address the class imbalance problem. While maintaining a mAP above 85%, the inference speed was improved. At the same time, the loss function of YOLOv8s was improved: among them box =0.01,λ cls =0.05 S2-2: Mechanical signal processing: Key steps in pressure signal processing: First, signal preprocessing is performed, and a sliding window filter is designed with a window size of 1 second. The mean and standard deviation of each channel signal are calculated in real time, and outliers are automatically identified and eliminated. To address the vibration interference during elevator operation, a frequency domain filtering method is used, using a low-pass filter to eliminate high-frequency noise. The frequency domain threshold of the adaptive filter is: τ=μ w +3s w where w = 1s Based on this, feature extraction is performed and a two-layer LSTM network is constructed to process time series pressure signals. The hidden layer dimension of the network is 128 and an additional activation function is used. A differential feature input is specially designed, and the deviation between the current pressure value and the historical mean is used as an additional input feature to enhance the detection capability of sudden anomalies. The LSTM state update steps are as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) The complete formula of the gating mechanism of the LSTM network is: i t =σ(W ii x t +b ii +W hi h t-1 +b hi ) f t =σ(W if x t +b if +W hf h t-1 +b hf ) o t =σ(W io x t +b io +W ho h t-1 +b ho ) h t =o t ⊙tanh(c t ) The features extracted in step S2 can be fed back to the GAN generator in step S1 to optimize the distribution of difficult examples in the synthesized data.

4. The elevator anti-entanglement control method according to claim 1, characterized in that: The step S3: control strategy training based on deep reinforcement learning specifically includes the following steps: S3-1: State space definition: First, multimodal feature fusion is performed, that is, the visual features and pressure features extracted in step S2 are combined into a state vector. The specific process is as follows: Visual feature processing: Perform maximum pooling on the 1024-dimensional features output by EfficientNet, retain key spatial information, and reduce the dimension to 256 dimensions; Pressure feature processing: Calculate the mean and standard deviation of the pressure of the last 10 frames to obtain a 2D feature; Temporal feature supplement: Use OpenCV's Farneback algorithm to calculate the optical flow of adjacent frames and obtain the object motion trend; Final state vector: The above features are concatenated into a 260-dimensional vector as the input state of DRL; the formula is simplified to: s t =Concat(f v ,f p ,f o ) Among them, visual features Pressure characteristics Optical flow features State vector Concatenate visual, pressure, and motion features as the input state of DRL; S3-2: Action Space Design: Control Instruction Mapping In order to map discrete actions to the actual control variables of the elevator door, four discrete actions corresponding to different control modes of the elevator door are designed: 1) Continue to close: for risk-free scenarios; 2) Slow down and close the door: triggered when an object is detected approaching the door gap; 3) Pause and hold: completely stop the door movement for 1 second and wait for the risk to be eliminated; 4) Emergency rebound: 0.5m / s 2 Open the door in reverse acceleration to deal with hard objects blocking it: S3-3: Reward function design: A multi-objective balance strategy is introduced. The reward function consists of three parts: Safety bonus: +10 points for successfully avoiding a pinch, -100 points for a pinch; Efficiency penalty: 1 point will be deducted for each unnecessary pause to prevent excessive conservatism; Smoothness bonus: If the difference between adjacent actions is too large, 5 points will be deducted; The formula is simplified to: R=10·Safe-100·Danger-1·UnnecessaryStop-5·Jerk S3-4: DRL algorithm training: Training process: Offline stage: In the system simulation environment, load the S1 dataset and simulate 5 million training steps with a batch size of 2048; Online fine-tuning: Receive real elevator sensor data through ROS and prioritize replaying samples with large prediction errors: The training process can be expressed as: Among them, the input state s t , action a t , the output θ is the policy parameter, and the training stability is improved by clipping the policy gradient.

5. The elevator anti-entanglement control method according to claim 1, characterized in that: The step S4: model optimization and robustness enhancement, specifically includes the following steps: S4-1: Priority Experience Replay: Sample Importance Weighting: Dynamically adjust sample priority according to the following rules: Samples with large prediction errors have their priority increased by 50%; Experts mark difficult examples with the highest priority; Use PyTorch's priority replay buffer to update the priority every 1000 steps. The following is the generated priority p t Method, where high error samples are trained first: p t =∣δ t ∣+0.5·KL(π θ ||π old ) S4-2: Dynamic Learning Rate Adjustment: To quickly adapt across scenarios, the MAML algorithm is used to enable the model to: Internal layer update: fine-tuning on single elevator data; Outer layer update: aggregate generalization parameters of 20 different elevator models; S4-3: Risk Probability Control: Constrained optimization is used to ensure that when the predicted probability of danger exceeds 1%, an "emergency rebound" action is triggered. Lagrange multipliers are used to dynamically adjust the balance between safety and efficiency. This can control the hard limit hazard incidence to <1%; S4-4: Online Adaptation: Use PyTorch's ReduceLROnPlateau scheduler to dynamically adjust the learning rate, and adjust the learning rate based on the average reward of the last 10 training runs: When the reward is rising: the learning rate is increased by 10%, and when the reward is falling: the learning rate is reduced by 50%.

6. The elevator anti-entanglement control method according to claim 1, characterized in that: In step S5, a low-latency inference system is built based on the edge computing platform, a Level 1 and Level 2 response mechanism is designed, and dynamic thresholds are set to implement risk classification and disposal. The deployment plan includes a complete abnormality monitoring and fault recovery mechanism, supports hot updates and remote diagnosis, and the system is deeply integrated with the elevator control system through the CAN bus, while retaining the hardware emergency stop circuit as a safety redundancy.

7. An elevator anti-entry control system, characterized in that: Build a comprehensive sensing system in the elevator door area, including: The visual perception unit uses a depth camera installed at the top center of the elevator door frame, tilted downward at a 15-degree angle to ensure complete coverage of the door gap. The camera simultaneously outputs 1280×720 resolution RGB images and depth maps while maintaining a stable frame rate. Data acquisition is achieved through the driver, and a time synchronization service is used to ensure that the time alignment error of each sensor data is less than 5ms. The depth camera output data format is: The force sensing unit deploys ultra-thin pressure sensors symmetrically on both sides of the elevator door gap, evenly spaced 15 cm apart. The sensor signals are digitized using a high-precision acquisition card with a sampling rate of 100 Hz to capture rapid pressure changes. Each sensor channel is calibrated on-site to ensure a measurement error of less than 0.1 N. The output format is: The data synchronization system adopts a hardware synchronization solution based on the PTP protocol and allows all sensors to provide a unified time base. The collected data is stored in the ROS2 bag format with timestamp alignment for easy subsequent processing.

8. An elevator anti-entrapment control system according to claim 7, characterized in that: Also includes: Edge computing module, the device needs to be equipped with multiple high-speed interfaces, including USB and GPIO, for connecting to the vision sensor and pressure sensor array; The pixel calculation module pre-processes the original image to improve the accuracy of subsequent processing, and then uses an optimized object detection algorithm to analyze the pre-processed image to accurately identify various objects in the picture; The hazard assessment module is responsible for performing real-time risk assessment of detected objects. This module needs to load a fully trained and optimized deep learning model and receive object information from the image processing module in real time during operation for time series analysis. The assessment results are expressed as a standardized hazard level with an accompanying confidence indicator. The response module implements a hierarchical response strategy based on the output of the risk assessment module. The response mechanism is divided into multiple levels, each corresponding to a different processing strategy. Each response level has clear condition judgment criteria and corresponding execution actions. The response instructions are transmitted to the execution agency through a standardized interface to ensure that the instructions can be executed accurately and timely. The monitoring and debugging module collects the operating status information of each module in real time, including key indicators such as processing delay and resource utilization, and intuitively displays the system operating status and risk assessment results through a visual interface, making it easier for operators to understand the on-site situation.

Images