Dangerous gas leakage tracing and positioning method and system based on artificial intelligence

Abstract

The invention relates to the technical field of gas leakage traceability positioning, in particular to a dangerous gas leakage traceability positioning method and system based on artificial intelligence, and the method comprises the steps: data collection: collecting environment parameters and gas concentration data from fixed and mobile sensors, and deploying the fixed sensors at key nodes to monitor gas diffusion; the mobile sensor is carried on the inspection robot or the unmanned aerial vehicle to dynamically fill the dead zone of the fixed sensor; preprocessing the data, and performing denoising, normalization and feature enhancement on the data; feature extraction and selection: extracting key features by using a machine learning algorithm; model training: training a deep neural network model to predict a leakage source position; monitoring and triggering: monitoring data and triggering traceability positioning when the data is abnormal; accurate positioning is carried out, and a leakage source is positioned by using the model and real-time data; intelligent decision making is achieved, and emergency response suggestions and optimization schemes are provided.

Description

Technical Field

[0001] The present invention relates to the technical field of gas leakage tracing and positioning, and in particular to an artificial intelligence-based method and system for tracing and positioning the source of hazardous gas leakage. Background Art

[0002] In modern industrial production, hazardous gas leaks are a significant safety concern. They not only threaten the lives of factory workers but can also cause severe damage to the surrounding environment. Therefore, effective detection and source location of hazardous gas leaks are crucial for ensuring safe industrial production. However, existing hazardous gas leak detection and source location technologies have numerous shortcomings. These deficiencies significantly limit the efficiency and effectiveness of monitoring systems and increase the risk of accidents.

[0003] First, traditional hazardous gas leak detection relies primarily on manual inspections and limited sensor networks. This approach has very limited coverage, making comprehensive monitoring difficult, especially in vast industrial areas or locations with complex terrain. Furthermore, manual inspections are inefficient and susceptible to human factors, such as blind spots and misjudgment, resulting in inaccurate detection. Furthermore, sensor networks can generate false alarms due to environmental interference or equipment failure, increasing the risk of false positives.

[0004] Secondly, existing monitoring systems have a slow response time. After detecting an anomaly, it takes a long time to locate the leak source, which can cause the leak to spread and increase the risk of an accident. Especially in emergency situations, quickly and accurately locating the leak source is crucial for timely response measures.

[0005] Thirdly, establishing and maintaining a comprehensive sensor network requires substantial capital investment, and costs increase dramatically as the monitoring scope expands. This, to a certain extent, limits the widespread adoption and application of monitoring systems. Furthermore, existing data analysis methods are often unable to effectively process large-scale, multi-source, and real-time data streams, resulting in a lack of timely and accurate analysis of leaks.

[0006] In addition, existing monitoring equipment performance degrades under extreme environmental conditions (such as high temperature, low temperature, and corrosive environments), and cannot provide reliable monitoring data. This limits the application of monitoring systems in various industrial scenarios.

[0007] Finally, the existing monitoring system lacks intelligent decision-making support. After a leak occurs, the lack of an effective intelligent decision-making support system to guide emergency response results in inaccurate and inefficient emergency measures, which, to a certain extent, affects the efficiency and effectiveness of incident handling.

[0008] Therefore, in view of the limitations of current gas leak tracing and positioning, it is particularly urgent and important to develop an artificial intelligence-based hazardous gas leak tracing and positioning method and system to promote the development of related technical fields. Summary of the Invention

[0009] The purpose of the present invention is to provide a method and system for tracing and locating the source of hazardous gas leaks based on artificial intelligence to solve the technical problems raised in the above background technology.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] A method for tracing and locating the source of a hazardous gas leak based on artificial intelligence, the method comprising the following steps:

[0012] S1. Data collection: collect data on environmental parameters and gas concentrations through fixed sensors and mobile sensors;

[0013] S2, data preprocessing: denoising, normalization and feature enhancement of the collected data;

[0014] S3. Feature extraction and selection: Use machine learning algorithms to extract and select key features from preprocessed data;

[0015] S4. Model training: training a deep neural network model that can predict the location of the leak source based on the extracted features;

[0016] S5, monitoring and triggering: monitoring sensor data and triggering traceability and positioning when anomalies are detected;

[0017] S6. Precise positioning: Use the trained deep learning model and real-time data to accurately locate the leak source;

[0018] S7. Intelligent decision-making: Provide emergency response recommendations and optimize emergency plans based on positioning results and environmental parameters.

[0019] Preferably, the fixed sensors in the data collection step are deployed at key nodes of industrial facilities to collect environmental parameters and gas concentrations in real time.

[0020] Preferably, the mobile sensors in the data collection step are mounted on inspection robots or drones to dynamically fill the blind spots of fixed sensors.

[0021] Preferably, the deep learning model includes multiple sub-networks, each sub-network is responsible for processing different types of sensor data.

[0022] Preferably, the training of the S4 deep learning model includes the following steps:

[0023] S401. Build annotated datasets covering various scenarios to support model training and validation.

[0024] S402, generating a three-dimensional map based on the factory geographic information system, marking the installation locations of the fixed sensors in the three-dimensional map and dividing the grid environment model;

[0025] S403, fusing the spatiotemporal data of the fixed sensor and the mobile sensor through a multi-source data fusion model, and outputting the three-dimensional coordinates of the leakage source;

[0026] S404, optimize the inspection path of mobile sensors to maximize blind spot coverage;

[0027] S405: When the mobile sensor detects a leak, it triggers local grid refinement to perform source location tracing;

[0028] S406. Adjust the parameters of the mesoscale turbulence diffusion model according to the real-time detection data to adapt to the real-time environmental changes.

[0029] Preferably, the multi-source data fusion model includes:

[0030] The spatiotemporal feature extraction module uses a graph convolutional network (GCN) to model the topological relationship of fixed sensors and extracts the time series features of mobile sensors through LSTM.

[0031] The fusion decision module aligns the feature vectors of fixed and mobile sensors through a cross-attention mechanism to generate a joint embedding vector.

[0032] Preferably, the number of layers of GCN is three to five.

[0033] Preferably, the mobile sensor plans the inspection path using a dynamic path planning module, and the dynamic path planning module uses a multi-agent deep deterministic policy gradient algorithm.

[0034] Preferably, the S3 feature extraction and selection step includes an adaptive feature fusion algorithm for fusing and optimizing features from different sensors.

[0035] Preferably, the S7 intelligent decision-making step includes a multi-objective optimization algorithm for selecting the optimal solution among multiple emergency response options, while considering response speed and resource consumption.

[0036] Technical effects and advantages of the present invention:

[0037] By combining multi-source data acquisition, deep learning models, adaptive feature extraction, and intelligent decision support, this method achieves a wider monitoring range, faster response, higher accuracy, and lower cost. Furthermore, the method exhibits excellent environmental adaptability and generalization capabilities, effectively operating in various industrial scenarios and providing optimal emergency response solutions. This significantly improves the efficiency and effectiveness of hazardous gas leak detection and source location, reducing the risk of accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the main process of the present invention;

[0039] Figure 2 Schematic diagram of the process of model training of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0041] Example 1

[0042] Reference Figures 1 to 2 As shown, the present invention proposes a method for tracing and locating the source of hazardous gas leaks based on artificial intelligence, and the method comprises the following steps:

[0043] S1. Data collection: collect data on environmental parameters and gas concentrations through fixed sensors and mobile sensors;

[0044] Fixed sensors are deployed at key nodes in industrial facilities to collect real-time data on environmental parameters and gas concentrations. Meanwhile, mobile sensors are mounted on inspection robots or drones to dynamically fill in the monitoring blind spots left by fixed sensors.

[0045] Fixed sensors monitor parameters including gas concentration, temperature, humidity, and pressure, as well as three-dimensional coordinate information based on the factory's Geographic Information System (GIS). The system utilizes a combination of multiple sensor types to comprehensively cover different gas types and concentration ranges, effectively avoiding monitoring errors caused by the performance limitations of a single sensor.

[0046] The mobile sensor system consists of two components: aerial and ground. The aerial component utilizes a multi-axis gas sampling probe mounted on a drone, flying at an altitude of 5 to 30 meters. This allows for dynamic coverage of areas difficult to reach by fixed sensors, ensuring comprehensive environmental monitoring. The ground component consists of an infrared spectrometer and lidar mounted on an inspection robot. The robot performs inspections along pre-set paths, enabling detailed monitoring of the ground environment.

[0047] Data collected by fixed and mobile sensors is transmitted in real time to a data processing unit via wired or wireless communication modules. Once data is collected, it is reliably transmitted to the gas monitoring system using IoT transmission technology.

[0048] S2, data preprocessing: denoising, normalization and feature enhancement of the collected data;

[0049] Data denoising: To address environmental interference (such as sudden changes in temperature and humidity, cross-gas interference) and sensor noise (drift, pulse anomalies) in gas concentration signals, wavelet decomposition is used to filter out high-frequency noise, and Kalman filtering is combined to dynamically correct sensor drift errors. Spatial clustering (DBSCAN) is used to eliminate outlier concentration points caused by sudden changes in wind direction or sensor failures, and a sliding window mean filter is used to smooth wind speed time series fluctuations. At the same time, a Gaussian plume diffusion model is introduced to physically constrain the concentration gradient, correct for non-physical mutations caused by turbulence, and retain the spatiotemporal distribution characteristics that are strongly correlated with the leakage source.

[0050] The Gaussian plume diffusion model is a classic physical mathematical model used to describe the diffusion law of gas in an open environment. Its mathematical model is:

[0051]

[0052] in:

[0053] Q is the leakage source strength (release amount per unit time);

[0054] σ y is the diffusion coefficient in the lateral direction;

[0055] σ z is the diffusion coefficient in the vertical direction;

[0056] u is the average wind speed;

[0057] H is the effective height of the leakage source.

[0058] Data normalization: To address the dimensional differences among multiple sensors (such as concentration in ppm, wind speed in m / s, and temperature in °C), Z-Score normalization is used to eliminate dimensional effects and enhance the model's sensitivity to global concentration gradients. Quantile normalization is used for sensor signals with long-tail distributions (such as instantaneous high-concentration pulses) to prevent outliers from dominating model training. For spatial data from multiple monitoring stations, Min-Max normalization is used to map concentrations to a local spatial grid (such as the 0-1 interval) to highlight the relative spatial patterns of leakage diffusion.

[0059] Based on the gas diffusion mechanism and spatiotemporal correlation, joint spatiotemporal features are constructed: lagged differences and window statistics (mean, variance) of concentration time series are extracted. Concentration gradients, diffusion direction consistency, and spatial interpolation concentration fields are calculated for adjacent sensors. The Gaussian plume model is used to infer leakage intensity and diffusion velocity as prior feature inputs. Small sample leakage scenario data is expanded by adding Gaussian noise or simulating diffusion trajectories at different wind speeds. Concentration-wind speed products (to simulate leakage flux) and temperature-humidity-concentration ratios (to correct for environmental interference) are constructed.

[0060] S3. Feature extraction and selection: Use machine learning algorithms to extract and select key features from preprocessed data;

[0061] Adaptive feature fusion algorithm is used to fuse and optimize features from different sensors according to the characteristics and importance of different sensor data.

[0062] S4. Model training: training a deep neural network model that can predict the location of the leak source based on the extracted features;

[0063] S5, monitoring and triggering: monitoring sensor data and triggering traceability and positioning when anomalies are detected;

[0064] By setting an abnormal threshold and monitoring the data transmitted by fixed sensors and mobile sensors in real time, when the data detected by the fixed sensor or mobile sensor exceeds the threshold, the traceability and positioning process is triggered to locate the gas leak point in real time.

[0065] S6. Precise positioning: Use the trained deep learning model and real-time data to accurately locate the leak source;

[0066] When the source location is activated, the data collected in real time by fixed and mobile sensors is input into the trained learning model, which then outputs the location of the leak source. Simultaneously, the leak source is precisely located on a map using the factory's Geographic Information System (GIS), displaying the danger zone (red overlay) and evacuation routes (green arrows) within a 20m radius of the leak source.

[0067] S7. Intelligent decision-making: Provide emergency response recommendations and optimize emergency plans based on positioning results and environmental parameters.

[0068] Using a multi-objective optimization algorithm, the optimal solution is selected from multiple emergency response options by considering factors such as response speed and resource consumption. For example, the optimal evacuation route and emergency rescue plan are selected based on the location of the leak source and the gas diffusion trend.

[0069] Reference Figure 2 As shown, in this application, the training of the S4 deep learning model includes the following steps:

[0070] S401. Build annotated datasets covering various scenarios to support model training and validation.

[0071] S402, generating a three-dimensional map based on the factory geographic information system, marking the installation locations of the fixed sensors in the three-dimensional map and dividing the grid environment model;

[0072] The gas leak source tracing and positioning learning model in this embodiment combines the multimodal data fusion technology of fixed sensors and mobile sensors to achieve precise positioning in three-dimensional space.

[0073] The data collection layer is mainly composed of fixed sensors and mobile sensors. Fixed sensors are deployed at key nodes such as pipeline connections and valves. Their three-dimensional coordinates are marked through the factory BIM / GIS system (accuracy ±0.1m), and the data sampling frequency is 10Hz.

[0074] The mobile sensor consists of two parts: a ground inspection robot and a drone. The ground inspection robot is equipped with an infrared spectrometer (sensitivity 0.1ppm), which inspects along a preset path at a speed of 2m / s. The drone is equipped with a multi-axis gas probe (response time <5s), flies at an altitude of 5 to 30m, and transmits three-dimensional position coordinates in real time.

[0075] 3D space modeling builds a 3D map based on the factory BIM / GIS system, marks the sensor location (X, Y, Z coordinates), and establishes a grid environment model (grid resolution 0.5m 3 ).

[0076] S403, fusing the spatiotemporal data of the fixed sensor and the mobile sensor through a multi-source data fusion model, and outputting the three-dimensional coordinates of the leakage source;

[0077] The multi-source data fusion model includes the following:

[0078] Input the coordinate data of the fixed sensor and the mobile sensor: the fixed sensor data is S fix ={C i ,T i ,H i ,Pi ,(x i ,y i ,z i )}, where C i is the gas concentration, T / H / P are environmental parameters. The mobile sensor data is S mobile ={C j (t),(x j (t),y j (t),z j (t))}, containing the dynamic coordinates of the time series.

[0079] The network structure includes a spatiotemporal feature extraction module and a fusion decision module. The spatiotemporal feature extraction module uses a graph convolutional network to model the topological relationship of fixed sensors and extracts the time series features of mobile sensors through LSTM. The fusion decision module aligns the feature vectors of fixed and mobile sensors through a cross-attention mechanism to generate a joint embedding vector.

[0080] The spatiotemporal feature extraction module consists of two parts: fixed sensors and mobile sensors:

[0081] Fixed sensors: Graph Convolutional Network (GCN) is used to model the sensor topology.

[0082] Among them, the number of layers of GCN is 3 to control the topological aggregation depth of the sensor network. The aggregation radius of each layer is 5m to ensure coverage of adjacent key nodes; node features include gas concentration (C_i), temperature (T_i), humidity (H_i), air pressure (P_i) and three-dimensional coordinates (x_i, y_i, z_i). The edge weight is calculated based on the three-dimensional Euclidean distance between sensors. d ij is the sensor spacing.

[0083] Mobile sensor: Use LSTM to extract time series features and combine it with the self-attention mechanism to capture concentration mutation points.

[0084] The LSTM hidden layer dimension (128 dimensions) is used to capture the time series concentration changes of mobile sensors (robots / drones), with the input being the dynamic coordinates (x_j(t), y_j(t), z_j(t)). The default sliding window length is 10 seconds, matching the drone inspection speed (approximately 2 m / s).

[0085] Fusion decision module:

[0086] Feature-level fusion: Align the spatiotemporal features of fixed and mobile sensors through the cross-attention mechanism to generate a joint embedding vector F fusion .

[0087] Number of attention heads (4): align the spatiotemporal features of fixed and mobile sensors to generate a joint embedding vector F fusion

[0088] Gating weight: Dynamically adjusted based on feature reliability, formula: Among them RI i is the data source reliability index.

[0089] Output layer: The fully connected layer is mapped to the leakage source coordinates (x, y, z), and the loss function uses Huber Loss (to balance the influence of outliers).

[0090] Huber Loss parameter: δ = 1.0 to balance the robustness of positioning error and reduce outlier interference. Adam optimizer is used, combined with gradient clipping (max_grad_norm = 5.0) to prevent gradient explosion.

[0091] S404, optimizing the inspection path of the mobile sensor through a dynamic path planning algorithm to maximize the blind spot coverage;

[0092] The dynamic path planning algorithm process includes:

[0093] Blind spot identification: Based on fixed sensor data, Gaussian process regression (GPR) is used to predict the concentration distribution of unmonitored areas and mark potential leakage areas. blind .

[0094] Path generation: Multi-agent Deep Deterministic Policy Gradient (MADDPG) is used to define the reward function:

[0095] R = α·Coverage + β·Concentration Gradient - γ·Transfer Energy Consumption

[0096] The reward function weight is:

[0097] α=0.6 (coverage): Based on the blind spot identification result Ω blind Grid coverage ratio;

[0098] β = 0.3 (concentration gradient): rate of change of concentration detected by the mobile sensor;

[0099] γ = 0.1 (movement energy consumption): The power consumption of the robot / drone is proportional to the moving distance and the square of the speed.

[0100] Collaborative obstacle avoidance: The Voronoi diagram is introduced to divide the inspection area. The inspection area is generated based on the density of sensor blind spots to ensure that there is no conflict between the trajectories of multiple devices.

[0101] S405: When the mobile sensor detects a leak, it triggers local grid refinement to perform source location tracing;

[0102] S406. Adjust the mesoscale turbulence diffusion model according to the real-time detection data to adapt to the real-time environmental changes.

[0103] A mesoscale turbulent diffusion model is introduced to simulate the gas diffusion path. The mesoscale turbulent diffusion model includes the turbulent diffusion coefficient and the dynamic calibration of environmental parameters.

[0104] The turbulent diffusion coefficient includes the eddy diffusion coefficient and the laminar diffusion coefficient:

[0105] Eddy diffusion coefficient:

[0106]

[0107] in:

[0108] k is the turbulent kinetic energy;

[0109] ε is the dissipation rate;

[0110] C μ The empirical constant is 0.09, which is applicable to laminar flow near the ground in industrial environments.

[0111] Laminar diffusion coefficient:

[0112]

[0113] in:

[0114] ν is the kinematic viscosity of air (1.5×10 -5 m 2 / s);

[0115] Sc is the Schmidt number (≈0.7).

[0116] Example 2

[0117] The present invention also provides an artificial intelligence-based hazardous gas leak tracing and positioning system, which is used to implement an artificial intelligence-based hazardous gas leak tracing and positioning method, including: a data acquisition module, a data preprocessing module, a feature extraction and selection module, a model training module, a monitoring and triggering module, a positioning module, a decision support module, and a model adaptive adjustment module.

[0118] The model adaptive adjustment module continuously collects data and effects of actual emergency responses, adaptively adjusts the model, automatically adjusts model parameters according to new data and environmental changes, optimizes emergency plans, and improves the model's generalization ability and adaptability.

[0119] By integrating multiple data streams, including sensors and emergency response records (e.g., leak resolution time and resource consumption), a dynamic data pool is constructed. Model performance is evaluated using pre-set quantitative metrics (e.g., positioning error rate and emergency response delay), generating feedback signals to drive parameter adjustments. If the model's predicted location deviates from the actual leak point by more than a set threshold (e.g., 5 meters), the system automatically triggers the parameter optimization process.

[0120] At the same time, based on local environmental changes (such as sudden changes in wind speed or new equipment), the contribution weights of sensor features are dynamically adjusted through the attention mechanism. When the wind speed suddenly changes, such as when the originally stable wind direction suddenly changes, or when the wind speed increases or decreases significantly in a short period of time, this change will significantly affect the diffusion path and range of the gas leak. At this time, the attention mechanism will keenly capture this change and automatically reduce the weights of sensor features that are less affected by wind speed or have low correlation, such as the weights of sensor data features that are fixed in a relatively closed space and less affected by external wind speed. At the same time, the weights of sensor features that are sensitive to wind speed changes and can better reflect new gas diffusion trends will be increased, such as the weights of sensor data features that are carried on drones and can flexibly perceive wind field changes and gas diffusion dynamics in large areas.

[0121] If large new equipment is added to a factory, it may alter local environmental factors such as airflow direction and temperature distribution, thereby affecting the propagation characteristics of gas leaks. The attention mechanism quickly analyzes these local environmental changes and reassesses the importance of each sensor feature for accurate traceability and positioning. Sensor features that are closer to the new equipment and more significantly affected by it are given higher weights, as these features can more promptly and accurately reflect the impact of environmental changes caused by the new equipment on gas leaks. Sensor features that are farther away from the new equipment and less significantly affected by it are appropriately weighted to avoid unnecessary interference from these features in model judgment. This dynamic adjustment of sensor feature contribution weights ensures that the model maintains high traceability and positioning accuracy and adaptability despite varying local environmental changes.

[0122] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for tracing and locating the source of hazardous gas leaks based on artificial intelligence, characterized in that: The traceability and positioning method comprises the following steps: S1. Data collection: collect data on environmental parameters and gas concentrations through fixed sensors and mobile sensors; S2, data preprocessing: denoising, normalization and feature enhancement of the collected data; S3. Feature extraction and selection: Use machine learning algorithms to extract and select key features from preprocessed data; S4. Model training: training a deep neural network model that can predict the location of the leak source based on the extracted features; S5, monitoring and triggering: monitoring sensor data and triggering traceability and positioning when anomalies are detected; S6. Precise positioning: Use the trained deep learning model and real-time data to accurately locate the leak source; S7. Intelligent decision-making: Provide emergency response recommendations and optimize emergency plans based on positioning results and environmental parameters.

2. The artificial intelligence-based method for tracing and locating the source of hazardous gas leaks according to claim 1 is characterized by: Fixed sensors in the data collection step are deployed at key nodes of industrial facilities to collect environmental parameters and gas concentrations in real time.

3. The artificial intelligence-based method for tracing and locating the source of hazardous gas leaks according to claim 2, characterized in that: The mobile sensors in the data collection step are mounted on inspection robots or drones to dynamically fill the blind spots of fixed sensors.

4. The method for tracing and locating the source of hazardous gas leaks based on artificial intelligence according to claim 2 is characterized in that: The deep learning model includes multiple sub-networks, each of which is responsible for processing different types of sensor data.

5. The method for tracing and locating the source of hazardous gas leakage based on artificial intelligence according to claim 4 is characterized in that: Training the S4 deep learning model consists of the following steps: S401. Build annotated datasets covering various scenarios to support model training and validation. S402, generating a three-dimensional map based on the factory geographic information system, marking the installation locations of the fixed sensors in the three-dimensional map and dividing the grid environment model; S403, fusing the spatiotemporal data of the fixed sensor and the mobile sensor through a multi-source data fusion model, and outputting the three-dimensional coordinates of the leakage source; S404, optimize the inspection path of mobile sensors to maximize blind spot coverage; S405: When the mobile sensor detects a leak, it triggers local grid refinement to perform source location tracing; S406. Adjust the parameters of the mesoscale turbulence diffusion model according to the real-time detection data to adapt to the real-time environmental changes.

6. The artificial intelligence-based method for tracing and locating the source of hazardous gas leaks according to claim 5 is characterized in that: Multi-source data fusion models include: The spatiotemporal feature extraction module uses a graph convolutional network (GCN) to model the topological relationship of fixed sensors and extracts the time series features of mobile sensors through LSTM. The fusion decision module aligns the feature vectors of fixed and mobile sensors through a cross-attention mechanism to generate a joint embedding vector.

7. The method for tracing and locating the source of hazardous gas leakage based on artificial intelligence according to claim 6 is characterized in that: The number of layers of GCN is three to five.

8. The method for tracing and locating the source of hazardous gas leaks based on artificial intelligence according to claim 4 is characterized in that: The mobile sensor plans the inspection path using a dynamic path planning module, which adopts a multi-agent deep deterministic policy gradient algorithm.

9. The method for tracing and locating the source of hazardous gas leaks based on artificial intelligence according to claim 1, characterized in that: The S3 feature extraction and selection step includes an adaptive feature fusion algorithm to fuse and optimize features from different sensors.

10. The method for tracing and locating the source of hazardous gas leakage based on artificial intelligence according to claim 1, characterized in that: The S7 intelligent decision-making step includes a multi-objective optimization algorithm to select the optimal solution among multiple emergency response options, taking into account response speed and resource consumption.

Images