An unattended intelligent detection method and system for leakage of hazardous chemical gas

Abstract

The application discloses a kind of dangerous chemical gas leakage unattended intelligent detection method and detection system, belong to artificial intelligence and safety monitoring technical field.The application is realized to the dynamic recurrence of the panoramic environment state of operating area by constructing the environment digital twin that fuses multi-modal sensing data and facility three-dimensional information;Leakage early anomaly is realized sensitive detection based on reconstruction error using spatiotemporal graph neural network model to learn and reconstruct twin data;Once confirming leakage, the leakage source position and intensity are inverted using Bayesian inference framework, and the computational fluid dynamics model coupled with real-time wind field is driven to predict gas diffusion situation;Finally, emergency response instructions such as equipment shutdown, personnel evacuation and regional isolation are automatically generated and executed in coordination.The application realizes the whole-process intelligent closed loop from perception, diagnosis, prediction to disposal, significantly improves the timeliness, accuracy of dangerous chemical gas leakage monitoring and early warning, and the initiative and scientificity of emergency disposal.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and safety monitoring technology, specifically relating to an unattended intelligent detection method and system for hazardous chemical gas leaks. Background Technology

[0002] In fields involving hazardous chemicals, such as chemical and energy industries, real-time monitoring and early warning of gas leaks are crucial for ensuring safety. Currently, mainstream technologies rely on fixed sensors based on concentration threshold alarms or manual inspections. These methods have significant shortcomings: fixed sensors are susceptible to environmental interference, leading to false alarms and missed alarms, and their fixed deployment makes it difficult to fully cover dynamic diffusion areas; manual inspections are inefficient, risky, and cannot provide continuous monitoring.

[0003] While existing improvement solutions attempt to integrate multiple sensors, they still have limitations: the systems mostly adopt static data fusion strategies, making it difficult to accurately capture early weak leakage signals; the algorithms often ignore the physical laws of gas diffusion, resulting in insufficient prediction reliability in complex scenarios; in addition, the monitoring, early warning, and response links are usually isolated from each other, failing to form an intelligent closed loop from detection to response.

[0004] Therefore, there is an urgent need for an intelligent system and method that can achieve early and accurate detection, leak source location, diffusion prediction, and automatically trigger targeted emergency response. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an unattended intelligent detection method and system for hazardous chemical gas leaks, which aims to overcome the shortcomings of the existing technology, such as single detection method, reliance on fixed thresholds, existence of monitoring blind spots, delayed response, and inability to perform situational analysis and risk prediction for leak events.

[0006] To address the aforementioned technical challenges, this invention provides an unattended intelligent detection method and system for hazardous chemical gas leaks. It constructs a high-fidelity digital twin of the environment, integrating multimodal sensor data and three-dimensional geometric information of the facility. It utilizes a spatiotemporal graph neural network to deeply mine the complex spatiotemporal correlations of environmental parameters under normal operating conditions and accurately captures weak early-stage anomaly signals based on reconstruction errors. Once an anomaly is detected, a Bayesian inference-based inverse calculation model is immediately activated to precisely locate the leak source and invert the leak intensity. A real-time computational fluid dynamics model is then driven to accurately predict the future diffusion trajectory and impact range of the leaked gas. Ultimately, dynamic and visualized risk assessment and emergency response commands are generated, thus achieving a fundamental shift from passive alarm to proactive early warning, from single-point monitoring to comprehensive perception, and from symptom alerts to root cause localization and situational analysis.

[0007] According to one aspect of the present invention, an unattended intelligent detection method for hazardous chemical gas leaks is provided, comprising the following steps: Construct an environmental digital twin covering the hazardous chemical operation area, synchronously collect multimodal sensor data within the operation area, and fuse it with the three-dimensional geometric information of the facilities to generate and dynamically update a continuous four-dimensional spatiotemporal data field characterizing the environmental status of the entire area. Based on the aforementioned environmental digital twin, leakage anomaly detection is performed. The trained spatiotemporal graph neural network model is used to reconstruct the real-time acquired four-dimensional spatiotemporal data field. By calculating the reconstruction error and comparing it with a dynamically adjusted judgment threshold, continuous judgment of early gas leakage anomalies is achieved. Based on the anomaly detection results, the leakage source is located by inversion. In response to the continuous determination of a leakage event, the spatial location and leakage intensity of the leakage source are solved by inverse calculation based on the concentration field and wind field data in the four-dimensional spatiotemporal data field at the current moment. Using the leakage source parameters obtained through inverse inversion and the real-time environmental wind field as input, the computational fluid dynamics model is driven to perform simulation calculations, predict the diffusion trajectory and concentration distribution of the leaked gas in the future, and generate a dynamic risk map. By integrating the dynamic risk map, static equipment asset information, and real-time personnel location information, and through spatial analysis and logical reasoning, control commands for process equipment, safety evacuation guidelines for personnel, and isolation and control commands for the physical environment are generated and issued.

[0008] Furthermore, constructing the environmental digital twin covering the hazardous chemical operation area specifically includes: A 3D spatial mesh covering the work area is generated based on the 3D model of the facility. Multimodal sensing data is mapped to corresponding nodes of the three-dimensional spatial grid, and the multimodal sensing data includes at least the target gas concentration, three-dimensional wind speed vector, ambient temperature, and humidity. For nodes in the three-dimensional spatial grid that are not directly covered by sensor data, estimate their environmental parameter values ​​and quantify the estimation uncertainty; The four-dimensional spatiotemporal data field is dynamically generated and updated by repeatedly performing data mapping and spatial interpolation on continuous time slices.

[0009] Furthermore, leakage anomaly detection based on the aforementioned environmental digital twin specifically includes: The spatiotemporal tensor is extracted from the four-dimensional spatiotemporal data field as input and reconstructed using a pre-trained spatiotemporal graph convolutional autoencoder model. The graph structure of the spatiotemporal graph convolutional autoencoder is constructed based on the geographical distance of spatial nodes and the historical correlation of environmental parameters. Calculate the reconstruction error between the input spatiotemporal tensor and the model reconstruction output; The baseline threshold is determined based on historical normal operating condition data, and the dynamic judgment threshold is calculated by combining real-time environmental background noise factor. When the reconstruction error continues to exceed the dynamic determination threshold for a preset duration, a gas leak event is determined to have occurred.

[0010] Furthermore, based on the anomaly detection results, the leak source is located by inversion, specifically including: Define an unknown parameter vector containing the coordinates of the leakage source and the leakage rate; A likelihood function is established that uses the Gaussian smoke model as the positive operator and integrates sensor measurement error and spatial interpolation uncertainty. The prior uniform distribution of the coordinates of the leakage source is set based on the spatial range of the potential leakage equipment, and the log-normal prior distribution of the leakage rate is set based on the characteristics of the process medium. The Markov chain Monte Carlo sampling algorithm is used to sample and solve the posterior distribution, and the optimal estimate of the leakage source parameters and their confidence intervals are calculated based on the sampling results.

[0011] Furthermore, emergency response strategies are generated and implemented based on the diffusion prediction results, specifically including: The predicted hazardous areas in the dynamic risk map are overlaid with the factory equipment asset geographic information system in three dimensions to identify the threatened key equipment, and an ordered sequence of equipment shutdown instructions is generated based on the process safety logic knowledge base. The predicted danger zone is integrated into the three-dimensional personnel navigation network as a dynamic obstacle, and the optimal evacuation route to avoid risks is planned and pushed for each person based on the real-time personnel location. Based on the properties of the leaked gas and the real-time wind field, instructions for isolating entrances and exits of hazardous areas and for directional control of the ventilation system are generated and issued.

[0012] Furthermore, the multimodal sensing data is collected through a heterogeneous sensor network deployed in the operating area, the heterogeneous sensor network comprising: A tunable diode laser absorption spectroscopy sensor for fixed-point concentration monitoring, a mobile inspection sensor mounted on an autonomous mobile platform, a three-dimensional ultrasonic anemometer array for constructing a refined wind field model, and a hyperspectral imaging device and an infrared thermal imaging device for identifying the morphology of gas clouds.

[0013] Furthermore, the training process of the spatiotemporal graph convolutional autoencoder model includes: Training samples were constructed using the four-dimensional spatiotemporal data field under historical normal operating conditions. The graph structure, constructed based on spatial distance and parameter correlation, is used as input; The latent vector is obtained by extracting and compressing spatiotemporal features through an encoder, and then reconstructed through a decoder; Unsupervised training is performed with the goal of minimizing the mean square error between the input and the reconstructed output.

[0014] According to another aspect of the present invention, a smart detection system for hazardous chemical gas leaks is provided, comprising: The multimodal sensor data acquisition module is used to synchronously acquire multimodal sensor data within the work area; The environmental digital twin construction module is used to integrate multimodal sensor data and facility three-dimensional geometric information to construct and dynamically update the environmental digital twin, namely a four-dimensional spatiotemporal data field. The leakage anomaly detection module is loaded with a trained spatiotemporal graph neural network model, which is used for real-time anomaly detection and leakage event determination based on the environmental digital twin. The leak source inversion and localization module integrates a Bayesian inference engine and a sampler, which is used to invert and locate the leak source after a leak event is determined. The diffusion prediction and risk assessment module integrates a computational fluid dynamics solver, which is used to perform gas diffusion simulation based on inversion results and generate dynamic risk maps. The emergency response strategy generation module is used to integrate dynamic risk maps, static asset information and dynamic personnel locations to generate and collaboratively issue emergency response instructions; The modules are connected through a communication interface to form a closed-loop intelligent detection and response system.

[0015] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention constructs an environmental digital twin that integrates multi-source data, reconstructing sparse sensor observations into continuous and complete physical states such as global concentration and wind field. By using a spatiotemporal graph neural network to learn complex patterns of normal operating conditions, it can keenly identify weak and early spatiotemporal anomaly patterns, thereby achieving early and highly sensitive leak detection and significantly reducing the risk of accidents escalating due to monitoring blind spots or response delays.

[0016] 2. Upon detecting an anomaly, this invention immediately initiates a Bayesian inference-based inversion model, using real-time concentration field and wind field data to inversely calculate the most probable leak source location and intensity. Then, using this precise source term to drive rapid computational fluid dynamics simulation coupled with the real wind field, it predicts the diffusion path and impact range of the leaked gas over a future period. This provides a core basis for accurate decision-making.

[0017] 3. This invention utilizes an intelligent strategy generation module to integrate and spatially analyze dynamic risk prediction maps, static asset information, and real-time personnel locations. This automatically generates and collaboratively issues a complete set of emergency commands, including process equipment shutdown, personalized personnel evacuation guidance, area isolation, and ventilation control. This forms a fully automated, rapid closed loop from perception, analysis, prediction to execution, greatly improving the speed, scientific rigor, and coordination of emergency response, and effectively controlling the consequences of accidents. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the leakage anomaly detection principle framework based on spatiotemporal graph convolutional autoencoder; Figure 2 This is a logical flowchart of the leakage source inversion and localization and gas diffusion prediction. Detailed Implementation

[0019] This invention provides an unattended intelligent detection system and method for hazardous chemical gas leaks. Its core lies in constructing a high-fidelity environmental digital twin that integrates multimodal sensor data and facility three-dimensional geometric information to achieve closed-loop control of the entire process, including early anomaly detection, precise location of the leak source, prediction of diffusion trends, and generation of emergency response strategies for gas leak events.

[0020] Reference Figures 1 to 2 As shown below, the specific implementation steps of the unattended intelligent detection method for hazardous chemical gas leaks will be elaborated in detail, as follows: The first step, S1, involves multimodal sensor data acquisition covering the hazardous materials operation area. Step S1 aims to achieve comprehensive and high-precision perception of the environmental conditions through the collaborative work of various heterogeneous sensors deployed within the operation area, providing a high-quality data source for the subsequent construction of a digital twin model. The specific implementation steps are as follows: S101: Deploy gas concentration sensors using tunable diode laser absorption spectroscopy technology in high-leakage risk areas such as key process nodes, the bottom of tank areas, and under pipe racks.

[0021] Gas concentration sensors are used to monitor the concentration of specific hazardous chemical gases such as ammonia, chlorine, or hydrogen sulfide, and their detection accuracy needs to reach the part per million (ppm) level.

[0022] The specific locations and number of sensors should be optimized based on the planar and three-dimensional layout of the work area and the distribution of potential leakage sources.

[0023] S102: To compensate for the spatial coverage blind spots of fixed sensors, autonomous navigation drones and ground robots are configured as mobile inspection platforms.

[0024] The mobile inspection platform is equipped with a photoionization detector and / or a non-dispersive infrared sensor. It can dynamically scan areas that are difficult to cover by fixed sensor networks, such as high altitudes, densely populated equipment areas, and corners, according to a preset or dynamically planned path, thereby supplementing the acquisition of comprehensive gas concentration data in three-dimensional space.

[0025] S103: At the boundaries and core areas of the work area, three-dimensional ultrasonic anemometer arrays are deployed in layers at different heights to collect wind speed vectors and wind direction information in three-dimensional space in real time, and to construct a refined wind field model covering the entire work area.

[0026] At the same time, temperature and humidity sensors and air pressure sensors are deployed simultaneously within the monitoring network to continuously record ambient temperature, relative humidity and atmospheric pressure data.

[0027] S104: Deploy hyperspectral imaging equipment in key areas or at high points. This equipment can identify and visualize the macroscopic morphology and spatial distribution contours of gas clouds by capturing image information of gas molecules in their characteristic absorption bands.

[0028] Meanwhile, infrared thermal imaging equipment was deployed to detect localized temperature anomalies that may be caused by throttling expansion or phase change processes due to high-pressure gas leaks, providing auxiliary thermal evidence for the determination of leak events.

[0029] S105: To ensure the accuracy of subsequent spatiotemporal analysis, the data acquisition clocks of all the aforementioned sensors must be strictly synchronized.

[0030] Specifically, sub-millisecond time synchronization among all sensor nodes is achieved by employing a Network Time Protocol (NTP) or its enhanced version. The acquired multimodal data is encapsulated according to a predefined, unified frame format and transmitted in real time to the backend data processing center via a highly reliable industrial wireless network.

[0031] In summary, step S1 not only covers gas concentration, but also simultaneously acquires key wind field and environmental parameters that drive gas diffusion, and effectively eliminates monitoring blind spots through mobile inspection and imaging technology.

[0032] The strictly synchronized multimodal data streams collected form the basis for subsequent analysis. Based on this perception layer data, the system can proceed to the next stage—step S2—to build a dynamic environmental digital twin model.

[0033] For step S2: Constructing an environmental digital twin model, step S2 follows the acquisition of strictly synchronized multimodal sensor data covering the work area in step S1. This discrete sensor data is fused with the three-dimensional geometric information of the facilities to construct a continuous four-dimensional spatiotemporal data field capable of representing the real-time environmental state of the entire area—that is, a high-fidelity environmental digital twin model. The specific implementation process is broken down as follows: S201: First, import the building information model or computer-aided design model of the work area into the system.

[0034] Subsequently, the 3D model is processed into a high-resolution mesh. Specifically, the system generates a 3D Cartesian coordinate mesh covering the entire operating area based on preset monitoring accuracy requirements, such as for areas requiring fine prediction of gas diffusion.

[0035] The grid cell size can be set and adjusted between 0.5 meters and 2 meters to ensure a balance between computational resources and model accuracy.

[0036] S202: Next, all the sensor data collected in real time in step S1 are mapped to the corresponding nodes or cells of the three-dimensional mesh generated in S201 based on their physical installation location or the real-time trajectory coordinates of the mobile platform and with timestamp alignment.

[0037] For example, the readings of a stationary gas sensor are assigned to the grid node corresponding to its location; The data measured by the UAV at a given moment is then assigned to the grid cell containing its GPS / RTK positioning coordinates at that time. This process correlates discrete observation data with a continuous spatial frame.

[0038] S203: For nodes or cells in a 3D mesh that are not directly covered by sensors, their environmental parameters, such as gas concentration, wind speed, and temperature, need to be estimated through spatial interpolation. This embodiment uses a Gaussian process regression algorithm to accomplish this task, as detailed below: The algorithm uses the grid nodes of existing observation data in S202 as training points, and their measured values ​​as observation values. The core of the Gaussian process is to define a covariance function, or kernel function, to describe the similarity of environmental parameters between any two points in space. Here, the quadratic exponential kernel function is used as the basis, and its formula is:

[0039] in , For spatial location, For signal variance, For length scales, these two hyperparameters can be learned from historical observation data by maximizing the marginal likelihood function.

[0040] To achieve physical consistency, the interpolation process is not performed independently for each environmental parameter. The algorithm constructs a multi-output Gaussian process model, considering the coupling relationships between different variables, such as concentration and wind speed, and temperature and humidity, which are reflected in the structure of the covariance matrix, and simultaneously estimates the complete set of environmental parameters for all missing points.

[0041] In addition, the algorithm not only outputs the estimated values ​​of the parameters, but also provides the uncertainty of the estimates.

[0042] Specific calculations: For any unobserved point The posterior distribution of its environmental parameters can be analytically calculated using the standard inference formula of a Gaussian process based on data from all observation points. This completes the reconstruction from discrete observations to a continuous spatial field.

[0043] Steps S201 and S202 establish a digital spatial skeleton for the entire region and link it with observation data. Step S203 then uses Gaussian process regression, a probabilistic machine learning method, to fill in the data in the observation gaps based on the physical prior of spatial correlation and quantify the uncertainty. These three steps together transform sparse, heterogeneous sensor data into a physically meaningful and spatially continuous environmental state estimation field.

[0044] S204: Repeat the mapping step S202 and interpolation step S203 in each consecutive time slice, such as at 1-second intervals. The system continuously receives new sensor data and dynamically updates the environmental parameter values ​​and uncertainties of each node on the entire 3D mesh.

[0045] This generates a high-resolution four-dimensional spatiotemporal data field that evolves over time and covers the entire work area.

[0046] This dynamic data field, which includes multi-physics information such as concentration field, wind field, temperature and humidity field, is the environmental digital twin model constructed by this system, which can reproduce the environmental state of the work area at any time in real time.

[0047] In summary, step S2 constructs a high-fidelity digital twin model of the environment through a process of geometric meshing, data mapping, intelligent interpolation, and dynamic updating.

[0048] This model addresses the problem that traditional point-based monitoring cannot obtain full-field state information. It integrates limited sensor observations into a continuous and consistent virtual environment image that includes uncertainty assessment. This is not only the foundation for visual monitoring, but more importantly, it provides a unified, reliable, and complete data base and computational framework for all subsequent analyses. This enables subsequent algorithms to perform reasoning and computation across the entire spatial dimension, which is the key to realizing the transition from point monitoring to field perception.

[0049] For step S3, train the spatiotemporal graph convolutional autoencoder anomaly detection model: Step S3 builds upon the digital twin model, i.e., the four-dimensional spatiotemporal data field, constructed in step S2 that can characterize the panoramic environmental state of the work area. The goal of step S3 is to use data from this model under historical normal working conditions to train an intelligent model capable of learning complex spatiotemporal correlation paradigms between environmental parameters.

[0050] This model will serve as a benchmark for accurately identifying abnormal states that deviate from normal patterns during real-time monitoring, thereby enabling early detection of gas leaks. The specific implementation process is broken down as follows: S301: Data Preparation: Extract a four-dimensional spatiotemporal data field of the work area under verified normal operating conditions over a continuous period, such as 30 consecutive days, from the historical database. Preprocess the raw data, including linear interpolation imputation of missing values, and perform Z-score standardization on each environmental parameter sequence, such as concentration, wind speed component, and temperature, so that its mean is 0 and its standard deviation is 1.

[0051] Constructing spatiotemporal samples: Training samples are constructed from the preprocessed time series using a sliding window approach. Each sample is a spatiotemporal tensor. X is a real tensor, R represents the real set, N is the number of spatial nodes, i.e. the key grid nodes or sensor locations of the digital twin model, T is the time step, for example, T=60 represents 1 minute of data with a time resolution of 1 second, and F is the feature dimension, i.e. the number of environmental parameters, such as 6 dimensions: target gas concentration, U / V / W wind speed, temperature, and humidity.

[0052] Define the graph structure: Abstract the monitored area as an undirected graph G=(V,E,A). Each node... This corresponds to a spatial location. Constructing the adjacency matrix A is crucial: First, calculate the Euclidean distance between nodes. .

[0053] Secondly, using all historical normal data, calculate the overall correlation of the multidimensional time series of each pair of nodes across all features. Specifically, the maximum information coefficient or the average absolute value of the Pearson correlation coefficients for each feature can be used.

[0054] Then, adjacency relationships are constructed based on a threshold: if two nodes satisfy... (Distance threshold, such as 50 meters) and correlation coefficient If the correlation threshold is 0.3, then an edge is considered to exist between the two nodes. .

[0055] Finally, the elements of adjacency matrix A It can be defined as binarized (0 or 1), or assigned a value. As a weighting value, where γ is the distance scaling factor.

[0056] S302: Define the spatiotemporal graph convolutional autoencoder network architecture. Define a spatiotemporal graph convolutional autoencoder for unsupervised learning, which consists of an encoder and a decoder. The specific architecture configuration is as follows: The encoder consists of two spatiotemporal graph convolutional blocks connected in series, followed by a global pooling layer and a fully connected layer to generate latent vectors.

[0057] Spatiotemporal graph convolutional blocks: Spatial Graph Convolutional Layer: Employs a Chebyshev polynomial first-order approximation, i.e., a Geometric Graph Convolutional Layer (GCN). Its forward propagation formula is as follows: .in, It is an adjacency matrix with self-loops. It is its degree matrix. σ is the trainable weight matrix for this layer, and σ is the ReLU activation function.

[0058] Temporal convolutional layer: After spatial convolution, a one-dimensional causal convolutional layer is applied to the feature sequence of each node. The kernel size is 3, the stride is 1, and the padding is 1 to maintain the length of the temporal dimension. This layer is followed by a ReLU activation.

[0059] Latent Vector Generation: After two ST blocks, a global average pooling layer is used to aggregate the features of all nodes into a global feature vector, which is then mapped to a fixed-dimensional latent space vector through a fully connected layer. Among them, the latent dimensions It can be set to 16 or 32.

[0060] Decoder: The decoder is an approximate mirror image of the encoder. First, z is expanded through a fully connected layer, and then upsampling and reconstruction are performed through two spatiotemporal graph convolutional transpose blocks.

[0061] Each block contains a temporal deconvolutional layer with a kernel size of 3 and a stride of 1, and a spatial graph convolutional layer, with the same structure as the encoder. The final layer uses a linear activation function to output the reconstructed spatiotemporal tensor. Its dimension is the same as the input X.

[0062] S303: Perform model training and optimization.

[0063] Training setup: Randomly divide the dataset into training, validation, and test sets at a ratio of 70% / 15% / 15%. Use the Adam optimizer with an initial learning rate of [missing value]. The ReduceLROnPlateau strategy was adopted, and the learning rate was multiplied by 0.5 when the validation loss did not decrease within 5 epochs, and the batch size was set to 32.

[0064] Loss function and training process: The loss function is the mean squared error (MSE). Where |B| is the batch size, It is the Frobenius norm.

[0065] During training, the loss is evaluated on the validation set after each epoch, and an early stopping strategy is employed: training stops when the validation loss fails to decrease for 10 consecutive epochs to prevent overfitting. The total number of training epochs is typically between 50 and 150.

[0066] Regularization: During training, Dropout regularization is applied after each fully connected layer and graph convolutional layer of the encoder and decoder, with a dropout rate of 0.1.

[0067] S304: Complete model validation, evaluation, and storage.

[0068] Performance evaluation: Evaluate the trained model on an independent test set. Calculate the average reconstruction error across all samples on the test set. and its standard deviation .

[0069] Threshold determination: Based on the reconstruction error distribution of the test set, a dynamic anomaly detection threshold is set. .make ,in This is the average error of the test set. This threshold will be used for subsequent real-time detection.

[0070] Model saving: The weights of the trained encoder, the graph structure A, and the mean and standard deviation vectors required for standardization are serialized and saved together, for example, using PyTorch's torch.save or TensorFlow's SavedModel format. The saved model package is the leak anomaly detection model, which can be loaded during deployment.

[0071] In summary, step S3, through unsupervised learning, deeply understands the normal patterns of the interaction and evolution of environmental parameters in the work area under safe conditions.

[0072] The core principle is that the model has a small reconstruction error for data conforming to the pattern, but the reconstruction error increases significantly for data containing anomalies such as leaks. This provides a crucial basis for judgment in the next step. Step S4: Real-time leakage determination based on the anomaly detection model. In step S3, the training of the spatiotemporal graph convolutional autoencoder model is completed. This model can reconstruct environmental data under normal working conditions with high accuracy.

[0073] Step S4 involves deploying the trained model into the real-time monitoring pipeline. By calculating and analyzing the reconstruction error of real-time data, the system achieves automatic, sensitive, and reliable detection of gas leak events. The specific implementation process is as follows: S401: Real-time data preprocessing and model inference. The system extracts real-time spatiotemporal data blocks from the latest environmental digital twin model constructed in step S2 at fixed intervals. Where N is the number of spatial nodes, T is the time step (consistent with training, e.g., T=60), and F is the dimension of environmental features (e.g., F=6).

[0074] Subsequently, the real-time data block undergoes the same standardization process as during the training phase: that is, for each feature dimension... ,calculate ,in and These are the mean and standard deviation of the k-th feature, calculated on the training set and saved in step S304.

[0075] Processed standardized data blocks The data is input into the loaded leak anomaly detection model, i.e., the encoder-decoder network, and the model outputs the corresponding reconstructed data block. .

[0076] S402: Calculate the instantaneous reconstruction error.

[0077] Define the instantaneous reconstruction error at the current time t. This is used to quantify the overall deviation between real-time data and the model reconstruction results. Its calculation formula is the root mean square error: ; in, and These represent the values ​​of the standardized input data and the model reconstructed data at node i, time step τ, and feature k, respectively. This error comprehensively evaluates the reconstruction accuracy across the entire field, all time periods, and all parameters.

[0078] S403: Calculate the dynamic anomaly detection threshold.

[0079] Determination threshold It is not a fixed value, but rather dynamically adjusted based on ambient background noise to achieve a balance between sensitivity and false alarm rate. The calculation involves two steps: Determine the baseline statistical threshold ( In the model evaluation phase (S304), the reconstruction error sequence is calculated using test set data under pure normal operating conditions. sample mean and sample standard deviation The baseline threshold is set as follows: .

[0080] This threshold is based on the normal distribution assumption and is designed to cover the vast majority of normal fluctuations.

[0081] Evaluation of real-time background noise factor ( And generate a dynamic threshold: calculate an adjustment factor in real time. This is to reflect the current environmental background noise level. A comprehensive evaluation is conducted using the following formula:

[0082] in: This represents the current average wind speed in the area. For reference wind speed, such as historical average wind speed.

[0083] For recent data, such as the standard deviation of the temperature series over the past 1 minute, Use this as a reference value.

[0084] This is the combined variance of all sensor readings (each feature) at the current moment. Use this as a reference value.

[0085] α, β, γ are preset weighting coefficients that satisfy α + β + γ = 1. For example, α = 0.4, β = 0.3, and γ = 0.3 can be chosen.

[0086] After calculation, the value will be constrained to the range of [0.8, 1.2]. Finally, the dynamic anomaly detection threshold for the current moment is:

[0087] S404: Perform a persistent leak event determination. To effectively distinguish between persistent leaks and transient disturbances, the system employs time-duration-based determination logic: Initialization and instantaneous comparison: Set an alarm duration threshold K (e.g., K=10). The system calculates the alarm duration every second. and .

[0088] Continuous state monitoring: The system maintains a state counter c. At each time t, if... > If the counter is positive, the counter increments by 1 (c = c + 1); otherwise, the counter is reset to zero (c = 0).

[0089] Alarm decision: When the counter c ≥ K, a continuous gas leak event is determined. The system immediately triggers the highest level alarm and automatically initiates the subsequent leak source inversion and emergency response process, specifically step S5 and beyond.

[0090] False alarm suppression and reset: If a false alarm occurs at any time before the cumulative time reaches K seconds... ≤ If the fluctuation is determined to be a benign environmental fluctuation or transient interference, the system will reset the counter c to zero, not trigger an alarm, and continue routine monitoring.

[0091] In summary, step S4 obtains anomaly quantification indicators by calculating the reconstruction deviation between real-time data and the "digital benchmark" model; it dynamically generates reasonable alarm thresholds by combining historical statistical patterns with real-time environmental noise assessment; and finally, by requiring abnormal signals to persist for a certain period of time, it effectively filters out instantaneous interference, achieving a balance between high sensitivity and low false alarm rate.

[0092] Step S5: Leak source inversion and localization based on Bayesian inference. After confirming the occurrence of a continuous gas leak event in step S4, the system starts the leak source inversion and localization module.

[0093] Step S5 utilizes real-time 3D concentration field and 3D wind field data provided by the environmental digital twin model to inversely solve for the most probable leak source location and intensity based on a Bayesian inference framework, thus upgrading from "detecting an anomaly" to "locating the root cause." Specifically: S501: Define the inversion problem and input data.

[0094] When the leak alarm is triggered, the system immediately locks the alarm time and extracts two key data fields from the environmental digital twin model: 1) The three-dimensional gas concentration distribution across the entire field at this moment ; 2) The three-dimensional wind speed vector field U at this moment.

[0095] At the same time, define the unknown parameter vector θ to be inverted.

[0096] For common point source leaks ,in Let Q be the three-dimensional spatial coordinates of the leakage source, and let Q be the mass leakage rate per unit time.

[0097] S502: Construct a Bayesian inversion probability model.

[0098] This step constructs a probabilistic framework that links unknown parameters to observed data.

[0099] Forward physical model: A computationally fast Gaussian smoke plume model is used as the forward operator. For a given candidate parameter θ, the system first extracts the candidate source location from the real-time wind field U. Calculate the average wind speed vector for a nearby area, such as within a 50-meter radius. and turbulence intensity, the latter used to estimate diffusion parameters. .

[0100] Based on these inputs, the Gaussian smoke model calculates the predicted concentration vector at all observation locations (i.e., digital twin grid nodes) across the entire field. .

[0101] Define the likelihood function: assuming the observed concentration With model prediction The difference (error) between them follows a multivariate Gaussian distribution, and its covariance matrix Σ represents the uncertainty. Then the likelihood function is defined as:

[0102] The key point lies in the construction of the covariance matrix Σ. This invention sets it as a diagonal matrix, with the i-th diagonal element... It is the sum of two items: the variance of sensor measurement error. (Given by the sensor calibration certificate) and the prediction variance generated at that location by the Gaussian process spatial interpolation in step S2. .Right now, This objectively integrates the uncertainties in perception and modeling.

[0103] Define the prior distribution: Introduce prior knowledge about the parameters of the leakage source.

[0104] Coordinate Prior P Based on the factory's BIM / design drawings, the three-dimensional spatial extent of all potential leaking equipment, such as valves, flanges, and tank interfaces, is merged to form one or more three-dimensional bounding boxes. The prior coordinates follow a uniform distribution within this bounding box, and the probability is zero outside this bounding box.

[0105] Leakage Prior P(Q): The prior setting of the mass leakage rate Q is a log-normal distribution, i.e. Logarithmic mean It can be estimated based on the typical leakage rate of the process medium; for example, for medium-pressure gas phase pipelines, it can be set as follows: (corresponding to a magnitude of 0.1 kg / s); logarithmic standard deviation It can be set to 0.5 to 1 to cover a possible range of half to one order of magnitude.

[0106] Forming the posterior distribution: According to Bayes' theorem, the posterior probability distribution of the parameters is: ; The posterior distribution reflects both the evidence from the observational data and prior knowledge.

[0107] S503: Perform adaptive MCMC sampling solution.

[0108] To obtain from the posterior distribution Samples were obtained from the data using the adaptive Metropolis-Hastings Markov chain Monte Carlo algorithm.

[0109] Initialization: Set algorithm parameters. Total number of iterations. Pre-firing period From the prior distribution Randomly select a point as the initial state of the chain. Initialize the proposal distribution (suggestion distribution) as a set of... Centered on, with covariance as The multivariate Gaussian distribution, It can be set as a diagonal matrix, whose elements are proportional to the prior range of each parameter.

[0110] Iterative sampling: for i=0 to -1: a. Proposal: Based on the current distribution of proposals Generate a candidate parameter .

[0111] b. Calculate the acceptance rate: Calculate the probability of acceptance. The posterior probability is obtained by multiplying the likelihood function on its right side by the prior distribution.

[0112] c. Accept / Reject: Generate a uniformly random number If u < α, then accept the candidate point and let Otherwise, reject the candidate point, and let .

[0113] Adaptive adjustment: during the pre-burning period Each completed After each iteration, an adaptive update is performed. Specifically, based on a past window period (e.g., the most recent 2...),... Calculate the empirical covariance matrix of the samples that have been accepted within (times). Then follow the formula. Update the covariance matrix of the proposal distribution, where The scaling factor (usually taken as...) (where d is the parameter dimension), and ϵ is a local minimum (e.g., ...). This ensures the matrix is ​​positive definite. This adaptive mechanism can significantly improve sampling efficiency.

[0114] S504: Analyze the sampling results and output a location report.

[0115] After sampling, post-processing is performed to obtain the final inversion result.

[0116] Chain diagnosis and pre-burn period removal: First, visually inspect the trajectory of the sampled chains for each parameter to confirm that it has stabilized. Then, discard the chains before burning. Of the samples from the pre-firing period, only the subsequent 80,000 samples were retained for analysis; these samples can be considered as coming from the posterior distribution. The independent and identically distributed approximation.

[0117] Calculate point estimates and uncertainties: Optimal estimate: for the coordinates of the leakage source The mode of spatial locations in the posterior sample is taken (which can be obtained by estimating the kernel density and then finding its maximum). For the leakage rate Q, the median of the posterior sample is taken because it is more robust to long-tailed distributions.

[0118] Confidence Interval: For each parameter, calculate its 95% highest posterior density interval. That is, find an interval that maximizes the posterior probability density and includes 95% of the samples. This gives the range of uncertainty for the parameter estimate.

[0119] Results Output and Transmission: The system generates a structured inversion report, including: the most probable three-dimensional coordinates of the leakage source. The estimated mass leakage rate Q, and their respective 95% HPDI.

[0120] The report will be automatically passed to the diffusion prediction module in step S6 as input for accurate simulation.

[0121] Step S5 involves constructing a complete Bayesian inversion model to uniformly encode the gas diffusion patterns, multi-source observation data and their uncertainties, as well as prior knowledge about the leakage source.

[0122] By utilizing efficient MCMC sampling technology, we can not only calculate the most likely location and intensity of the leak source, but also quantitatively provide the confidence interval of the estimation results.

[0123] This step marks a leap in the system's capabilities from "perceiving anomalies" to "diagnosing the root cause," providing a basis for subsequent decision-making on diffusion prediction based on precise source terms (step S6) and the development of targeted emergency response strategies (step S7), thus forming a complete intelligent closed loop of detection-location-prediction-disposal.

[0124] Step S6 uses the precise leak source parameters output from step S5 as input to drive a high-fidelity computational fluid dynamics model coupled with the real environment in real time. This model simulates and predicts the diffusion process of the leaked gas over a future period, generating a visualized risk map. The specific process is as follows: S601: Prepare simulation inputs and initialize the computational domain. This step prepares all the necessary input conditions for the CFD simulation.

[0125] Geometry and Mesh: The 3D computational mesh corresponding to the environmental digital twin model constructed in step S2 is used as the computational domain for CFD simulation. This mesh already includes details of key geometric structures such as factory buildings and equipment.

[0126] Leakage source setting: The optimal leakage source coordinates obtained from step S5 are used for inversion. Mapping to the computational grid. Specifically, search the 3D computational grid for a grid cell containing the coordinate point. If the point happens to be on a grid node, assign it to the nearest cell. Apply a mass source term to this cell; its source strength (mass released per unit time) is the mass leakage rate Q retrieved in step S5. The composition of the released gas is determined based on the type of leaked substance.

[0127] Boundary and initial conditions: Inlet boundary: Real-time 3D wind field data is provided by the environmental digital twin model. The system extracts wind speed and turbulence parameters (such as turbulent kinetic energy k and dissipation rate ε) at various locations at the inlet boundary of the computational domain from the digital twin, and sets them as transient inlet boundary conditions for CFD simulation through trilinear interpolation.

[0128] Wall conditions: Apply no-slip wall boundary conditions to all solid surfaces such as buildings and equipment, and use standard wall functions to handle near-wall flow.

[0129] Initial field: The initial flow rate, pressure, temperature and gas concentration of all cells in the computational domain are set to the current global state provided by the environmental digital twin model for interpolation initialization.

[0130] S602: This step configures the solver and executes the calculation.

[0131] Governing equations and physical model: The Reynolds-averaged Navier-Stokes equations are solved using a standard k-ε turbulence model. Simultaneously, the component transport equations are solved to describe the convection and diffusion processes of the leaked gas.

[0132] Numerical methods and solution settings: Spatial discretization: The governing equations are discretized on a grid using the finite volume method, and the convection terms are discretized using a second-order upwind scheme.

[0133] Pressure-velocity coupling: The SIMPLE algorithm is used to solve the coupling problem of pressure and velocity fields.

[0134] Time progression: A second-order implicit time format is used, with a time step set to 0.5 seconds.

[0135] Computation Execution: The simulation task is submitted to a high-performance computing cluster for parallel computation. The total simulation time is set to 1800 seconds. By optimizing the parallel algorithm, the system can complete the computation in approximately 30 seconds.

[0136] S603: Extract and generate risk prediction maps.

[0137] During the simulation calculation, the system outputs full-field data at fixed intervals.

[0138] Concentration field extraction: Save the mass fraction data of leaked gas in the entire computational domain at each output time.

[0139] Risk zone delineation: The system pre-sets hazard concentration thresholds for different hazardous chemicals, such as the immediately life-threatening or health-threatening concentration (IDLH). The predicted concentration field at each moment is compared with the threshold, automatically identifying and marking all grid areas where the concentration exceeds the threshold, forming a dynamic hazard zone cloud map.

[0140] Trajectory and situation analysis: Calculate the centroid position of the leaking gas cloud at each moment, and connect them to obtain the predicted movement trajectory.

[0141] S604: Integrate the data obtained from S603 into a standardized forecast report, including: Predicted concentration distribution data for time series.

[0142] Dynamic danger zone evolution map for the next 30 minutes.

[0143] The predicted trajectory of the gas cloud's center of mass movement.

[0144] The results are output in both structured data format and visual graphics.

[0145] In summary, step S6 involves simultaneously embedding the precisely located leakage source and the real-time environmental wind field into a high-fidelity CFD model, performing rapid transient simulation, and generating an intuitive and quantitative risk impact area map.

[0146] Step S6 extends the system's capabilities to pre-event prediction, and the generated high-precision risk prediction map provides a key decision-making basis for the subsequent step S7 to conduct scientific and dynamic emergency response.

[0147] In step S7, this predicted future risk is intelligently integrated and spatially analyzed with the factory's current real-time status to automatically generate and execute a complete set of specific, executable, and coordinated emergency response instructions, thus forming a closed-loop decision-making process from intelligent perception and prediction to automatic handling. The specific process is as follows: S701: After determining a leakage event, the system immediately synchronizes and integrates key information from three independent data sources, providing a unified situational awareness foundation for subsequent intelligent decision-making. Dynamic risk prediction data: Receive the real-time output from step S6. This data exists in the form of a three-dimensional grid field with time series, and each time slice defines the predicted concentration distribution of leaked gas throughout the plant.

[0148] Static equipment asset information: This database accesses a pre-built factory asset geographic information system database. This database is stored in a layered vector map format. The core layer contains the precise three-dimensional spatial coordinates, unique equipment codes, functional attributes, and control point addresses in the industrial control system for all critical safety equipment in the plant, such as process shut-off valves, emergency isolation valves, pumps, tank breather valves, and emergency ventilation fans.

[0149] Dynamic personnel location information: obtained through a real-time location system (RTLS, such as one based on UWB ultra-wideband technology) deployed in the factory area.

[0150] The system pushes the three-dimensional coordinates (accuracy better than 1 meter) and identity ID of each person wearing a positioning tag to the system in real time at a frequency of no less than 1 Hz.

[0151] S702: Based on integrated data, the system first executes protective control logic for process equipment.

[0152] Spatial overlay and equipment identification: Perform 3D spatial intersection analysis between the 3D boundaries of high-concentration hazardous areas predicted at the current moment and in the near future (such as the next 1 minute) and the equipment asset GIS layer.

[0153] It automatically identifies all critical equipment located within or at a preset buffer distance (e.g., 3 meters) of the edge of a hazardous area and generates a "List of Threatened Critical Equipment".

[0154] Logical reasoning and priority ranking: The system accesses a pre-built process safety logic knowledge base. This knowledge base stores the topological connections and shutdown logic rules of equipment within key process units in a graph structure.

[0155] For each device in the list, the system automatically infers its upstream and downstream impact relationships based on the knowledge base. The generation of shutdown instructions strictly follows the principle of "prioritizing cutting off the leakage source and protecting critical assets," and is prioritized according to the order of process isolation.

[0156] For example, for a pipeline involved in a leak, the system will first generate a command to close the upstream root valve, and then the downstream isolation valve.

[0157] Command sequence generation: Based on the sorting results, each shutdown action is converted into a specific, standardized industrial control command.

[0158] Each instruction contains at least: the target device's unique address in the control system, a predefined control command word, and optional control parameters. These instructions are encapsulated into an ordered sequence of instructions.

[0159] S703: The system synchronously plans personalized safe evacuation routes for every person in the factory area.

[0160] Constructing a basic 3D navigation network: Based on the factory's building information model, all accessible corridors, staircases, platforms, and safety exits within the factory area are abstracted into a 3D graph model using automatic or semi-automatic algorithms. Nodes in the graph represent key points in the passageway, edges represent connecting paths, and the initial weight of each edge is equal to the geometric length of the path.

[0161] Incorporate dynamic risk barriers: Dynamically map the predicted danger areas on all time slices within a future warning time window, as output in step S6, onto this three-dimensional navigation map.

[0162] For any navigation edge, if the space it traverses at some point in the future is predicted to be a dangerous area, the system temporarily sets the weight of that edge to an extremely large value, or directly removes the edge from the graph temporarily, thereby ensuring that the path planning algorithm can proactively avoid risk areas that change over time.

[0163] Real-time personalized path calculation: For each person who needs to be evacuated, the system uses the real-time coordinates provided by their RTLS as the starting point and the predefined coordinates of the nearest safe exit as the ending point. On a navigation map that incorporates dynamic risk obstacles, the system uses Dijkstra's shortest path algorithm or A* heuristic search algorithm to calculate a safe path from the starting point to the ending point with the minimum cumulative weight in real time.

[0164] Generate and push navigation guidance: Convert the calculated path into a series of step-by-step navigation instructions, and push them in real time to the person's smart mobile terminal in the form of encrypted messages via the factory's wireless data network, providing graphic and voice prompts.

[0165] S704: The system also generates direct control commands for the physical space and environment.

[0166] Physical area isolation command: Based on the core danger zone delineated by the risk prediction map, the system automatically associates with the access control systems around that area. A command is generated to activate the corresponding electronic access control system and trigger the audible and visual alarm system for that area, achieving automatic physical isolation and visual and auditory warnings for the dangerous area.

[0167] Emergency ventilation control instructions: Based on the physical properties of the leaked gas and the current real-time wind field, the system generates a ventilation system control scheme according to the preset strategy model.

[0168] For example, for gases denser than air, the system automatically generates commands to activate the exhaust fan downwind of the leak point and the supply fan upwind of the leak point to create a directional dilution airflow; simultaneously, it shuts down conventional ventilation circuits that could cause gas backflow or accumulation. The commands explicitly specify the start / stop status and speed settings of the specific fans.

[0169] S705: Integrates, securely reinforces, and distributes the various commands generated above to each execution terminal.

[0170] Command encapsulation and security hardening: All commands issued to the industrial control system are encapsulated and transmitted through an industrial security gateway using a secure communication protocol compliant with IEC 62443, preventing command tampering or eavesdropping. Navigation and warning information issued to personnel mobile terminals is also protected through application-layer encryption.

[0171] Multi-channel collaborative distribution: Control command channel: The sequence of equipment control commands is sent in real time through the industrial ring network to the corresponding programmable logic controller or distributed control system for execution.

[0172] Information and command channels: Personnel evacuation routes, regional alarm information, etc., are pushed to personal mobile terminals and situational awareness screens in the emergency command center via enterprise-level wireless networks.

[0173] Execution status monitoring and feedback: The system continuously monitors the execution feedback signals of key instructions.

[0174] If no confirmation is received within the preset time, the system will determine that the instruction has failed, automatically trigger an alarm, and attempt to execute a backup plan (such as notifying nearby personnel to operate manually or activating redundant equipment).

[0175] Step S7 realizes a fundamental shift in emergency response from "human experience-driven, multi-stage transmission" to "data intelligence-driven, millisecond-level automatic triggering".

[0176] By deeply spatially integrating and logically reasoning dynamic risk prediction, static asset network and real-time personnel location, a collaborative response strategy covering the three dimensions of "equipment-environment-personnel" is generated.

[0177] This step not only transforms the perception, positioning, and predictive intelligence accumulated in all previous stages (S1-S6) into substantial protective actions to safeguard life and property, but also ensures the accurate execution of instructions through a safe and reliable control and communication architecture, thus forming a complete unattended, autonomous, and rapidly closed-loop intelligent emergency response system for hazardous chemical gas leaks.

[0178] The above-described steps constitute a complete closed-loop intelligent detection and response process. To support the implementation of this method, this invention also provides an unattended intelligent detection system for hazardous chemical gas leaks. This system includes a multimodal sensor data acquisition module, an environmental digital twin construction module, a leak anomaly detection module, a leak source inversion and localization module, a diffusion prediction and risk assessment module, and an emergency response strategy generation module.

[0179] The multimodal sensing data acquisition module consists of the aforementioned sensors and their communication units, and is responsible for acquiring and transmitting raw data. The environmental digital twin construction module runs on the data processing center server, executes grid generation, data mapping, and Gaussian interpolation algorithms, and outputs a four-dimensional spatiotemporal data field.

[0180] The leakage anomaly detection module is deployed on a dedicated inference server, loading a pre-trained spatiotemporal graph convolutional autoencoder model to perform real-time anomaly detection. The leakage source inversion and localization module integrates a Bayesian inference engine and a Markov chain Monte Carlo sampler, running on a high-performance computing node.

[0181] The diffusion prediction and risk assessment module calls upon a computational fluid dynamics solver to perform rapid simulations on a GPU-accelerated cluster. The emergency response strategy generation module, acting as the central control logic unit, integrates risk prediction, asset information, and personnel locations to generate and distribute control commands.

[0182] Each module communicates loosely with the application programming interface through message queues, ensuring the scalability and stability of the system.

[0183] In summary, this invention, through the synergistic operation of the aforementioned methods and systems, achieves early warning, accurate source tracing, situation prediction, and intelligent response to hazardous chemical gas leaks, fundamentally solving the core problems of existing technologies such as monitoring blind spots, delayed response, and lack of predictive capabilities.

[0184] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0185] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for unattended intelligent detection of hazardous chemical gas leaks, characterized in that, Includes the following steps: Construct an environmental digital twin covering the hazardous chemical operation area, synchronously collect multimodal sensor data within the operation area, and fuse it with the three-dimensional geometric information of the facilities to generate and dynamically update a continuous four-dimensional spatiotemporal data field characterizing the environmental status of the entire area. Based on the aforementioned environmental digital twin, leakage anomaly detection is performed. The trained spatiotemporal graph neural network model is used to reconstruct the real-time acquired four-dimensional spatiotemporal data field. By calculating the reconstruction error and comparing it with a dynamically adjusted judgment threshold, continuous judgment of early gas leakage anomalies is achieved. Based on the anomaly detection results, the leakage source is located by inversion. In response to the continuous determination of a leakage event, the spatial location and leakage intensity of the leakage source are solved by inverse calculation based on the concentration field and wind field data in the four-dimensional spatiotemporal data field at the current moment. Using the leakage source parameters obtained through inverse inversion and the real-time environmental wind field as input, the computational fluid dynamics model is driven to perform simulation calculations, predict the diffusion trajectory and concentration distribution of the leaked gas in the future, and generate a dynamic risk map. By integrating the dynamic risk map, static equipment asset information, and real-time personnel location information, and through spatial analysis and logical reasoning, control instructions for process equipment, safety evacuation guidelines for personnel, and isolation and control instructions for the physical environment are generated and issued. Among them, the prediction of gas diffusion based on the leakage source inversion results specifically includes: Using the three-dimensional computational grid corresponding to the digital twin of the environment as the computational domain, the inverted leakage source coordinates and leakage intensity are set as simulation mass source terms; The real-time three-dimensional wind field data provided by the four-dimensional spatiotemporal data field is used as the transient entry boundary condition for the simulation, and the current state of the entire field environment is used as the initial field for the simulation. Numerical simulation of the convective diffusion process of leaked gas under the action of a real-time wind field was performed. Extract the temporal concentration field from the simulation output, and combine it with the preset hazardous concentration threshold to generate a dynamic hazardous area evolution map and gas cloud movement trajectory over a future period.

2. The unattended intelligent detection method for hazardous chemical gas leaks according to claim 1, characterized in that, Constructing the aforementioned environmental digital twin covering the hazardous chemical operation area specifically includes: A 3D spatial mesh covering the work area is generated based on the 3D model of the facility. Multimodal sensing data is mapped to corresponding nodes of the three-dimensional spatial grid, and the multimodal sensing data includes at least the target gas concentration, three-dimensional wind speed vector, ambient temperature, and humidity. For nodes in the three-dimensional spatial grid that are not directly covered by sensor data, estimate their environmental parameter values ​​and quantify the estimation uncertainty; The four-dimensional spatiotemporal data field is dynamically generated and updated by repeatedly performing data mapping and spatial interpolation on continuous time slices.

3. The unattended intelligent detection method for hazardous chemical gas leaks according to claim 1 or 2, characterized in that, Leakage anomaly detection based on the aforementioned environmental digital twin specifically includes: The spatiotemporal tensor is extracted from the four-dimensional spatiotemporal data field as input and reconstructed using a pre-trained spatiotemporal graph convolutional autoencoder model. The graph structure of the spatiotemporal graph convolutional autoencoder is constructed based on the geographical distance of spatial nodes and the historical correlation of environmental parameters. Calculate the reconstruction error between the input spatiotemporal tensor and the model reconstruction output; The baseline threshold is determined based on historical normal operating condition data, and the dynamic judgment threshold is calculated by combining real-time environmental background noise factor. When the reconstruction error continues to exceed the dynamic determination threshold for a preset duration, a gas leak event is determined to have occurred.

4. The unattended intelligent detection method for hazardous chemical gas leaks according to claim 1, characterized in that, Based on the anomaly detection results, the leakage source is located by inversion, specifically including: Define an unknown parameter vector containing the coordinates of the leakage source and the leakage rate; A likelihood function is established that uses the Gaussian smoke model as the positive operator and integrates sensor measurement error and spatial interpolation uncertainty. The prior uniform distribution of the coordinates of the leakage source is set based on the spatial range of the potential leakage equipment, and the log-normal prior distribution of the leakage rate is set based on the characteristics of the process medium. The Markov chain Monte Carlo sampling algorithm is used to sample and solve from the posterior distribution, and the optimal estimate of the leakage source parameters and their confidence intervals are calculated based on the sampling results.

5. The unattended intelligent detection method for hazardous chemical gas leaks according to claim 1, characterized in that, Emergency response strategies are generated and implemented based on diffusion prediction results, specifically including: The predicted hazardous areas in the dynamic risk map are overlaid with the factory equipment asset geographic information system in three dimensions to identify the threatened key equipment, and an ordered sequence of equipment shutdown instructions is generated based on the process safety logic knowledge base. The predicted danger zone is integrated into the three-dimensional personnel navigation network as a dynamic obstacle, and the optimal evacuation route to avoid risks is planned and pushed for each person based on the real-time personnel location. Based on the properties of the leaked gas and the real-time wind field, instructions for isolating entrances and exits of hazardous areas and for directional control of the ventilation system are generated and issued.

6. The unattended intelligent detection method for hazardous chemical gas leaks according to claim 2, characterized in that, The multimodal sensing data is collected through a heterogeneous sensor network deployed in the work area, the heterogeneous sensor network comprising: A tunable diode laser absorption spectroscopy sensor for fixed-point concentration monitoring, a mobile inspection sensor mounted on an autonomous mobile platform, a three-dimensional ultrasonic anemometer array for constructing a refined wind field model, and a hyperspectral imaging device and an infrared thermal imaging device for identifying the morphology of gas clouds.

7. The unattended intelligent detection method for hazardous chemical gas leaks according to claim 3, characterized in that, The training process of the spatiotemporal graph convolutional autoencoder model includes: Training samples were constructed using the four-dimensional spatiotemporal data field under historical normal operating conditions. The graph structure, constructed based on spatial distance and parameter correlation, is used as input; The latent vector is obtained by extracting and compressing spatiotemporal features through an encoder, and then reconstructed through a decoder; Unsupervised training is performed with the goal of minimizing the mean square error between the input and the reconstructed output.

8. A hazardous chemical gas leak unattended intelligent detection system, used to implement the hazardous chemical gas leak unattended intelligent detection method according to any one of claims 1 to 7, characterized in that, include: The multimodal sensor data acquisition module is used to synchronously acquire multimodal sensor data within the work area; The environmental digital twin construction module is used to integrate multimodal sensor data and facility three-dimensional geometric information to construct and dynamically update the environmental digital twin, namely a four-dimensional spatiotemporal data field. The leakage anomaly detection module is loaded with a trained spatiotemporal graph neural network model, which is used for real-time anomaly detection and leakage event determination based on the environmental digital twin. The leak source inversion and localization module integrates a Bayesian inference engine and a sampler, which is used to invert and locate the leak source after a leak event is determined. The diffusion prediction and risk assessment module integrates a computational fluid dynamics solver, which is used to perform gas diffusion simulation based on inversion results and generate dynamic risk maps. The emergency response strategy generation module is used to integrate dynamic risk maps, static asset information and dynamic personnel locations to generate and collaboratively issue emergency response instructions; the modules are connected through communication interfaces to form a closed-loop intelligent detection and response system.

Images