A method, apparatus, equipment, storage medium, and product for sensing risks during chemical transportation.

Abstract

This invention discloses a method, apparatus, equipment, storage medium, and product for sensing risks in chemical transportation. It acquires monitoring data related to the transportation of chemicals, including at least two of vehicle operation monitoring data, driver monitoring data, and environmental monitoring data. Based on the monitoring data, it generates node features and an initial single risk index. On a pre-constructed causal graph structure, a pre-trained graph neural network is used for message passing and node state updates to correct the initial single risk index through causal reasoning, outputting the corrected single risk index corresponding to each monitoring data point. Finally, all corrected single risk indices are combined to obtain a comprehensive risk index. Therefore, this invention, by acquiring multi-source monitoring data related to the chemicals transported by vehicles and performing feature fusion and causal reasoning correction based on a causal graph structure and graph neural network, effectively improves the accuracy of risk assessment in the chemical transportation process, contributing to risk prevention and control.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a method, apparatus, equipment, storage medium, and product for sensing risks during chemical transportation. Background Technology

[0002] Chemicals are easily affected by various factors such as vehicle condition, driving behavior, and environment during transportation, posing significant safety hazards such as leaks and explosions. Therefore, monitoring the risks of chemicals during transportation is crucial.

[0003] Currently, risk monitoring of chemicals during transportation mainly relies on monitoring vehicle operating status and identifying driver behavior. For example, sensors installed on the vehicle monitor data such as vehicle speed, and in-vehicle cameras capture driver fatigue, improper operation, and other behaviors. However, in existing technologies, various monitoring data are usually processed separately and used for their own independent risk assessments. This fragmented assessment method has low information utilization and poor accuracy in overall risk assessment. Summary of the Invention

[0004] Based on this, the present invention provides a method, device, equipment, storage medium and product for sensing risks in chemical transportation, which can effectively improve the accuracy of risk assessment in the chemical transportation process by acquiring multi-source monitoring data and performing multi-dimensional feature fusion and causal reasoning correction.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for perceiving risks during chemical transportation, comprising: Acquire monitoring data related to the chemicals transported by the vehicle; wherein the monitoring data includes at least two of the following: vehicle operation monitoring data, driver monitoring data, and environmental monitoring data; Based on the monitoring data, multi-dimensional node features and an initial single risk index are extracted and generated; Based on a causal graph structure pre-constructed according to domain knowledge, a pre-trained graph neural network is used to perform message passing and node state updates on the node features, and causal reasoning is used to correct the initial single risk index, outputting the corrected single risk index corresponding to each of the monitoring data. The composite risk index is obtained by combining all the individual risk indices after correction.

[0006] To achieve the above objectives, embodiments of the present invention also provide a chemical transportation risk sensing device, comprising: The data acquisition module is used to acquire monitoring data related to the chemicals transported by the vehicle; wherein the monitoring data includes at least two of the following: vehicle operation monitoring data, driver monitoring data, and environmental monitoring data. The initial processing module is used to extract and generate multi-dimensional node features and an initial single risk index based on the monitoring data. The risk correction module is used to perform message passing and node state updates on the node features using a pre-trained graph neural network on a causal graph structure pre-constructed based on domain knowledge, to perform causal reasoning correction on the initial single risk index, and to output the corrected single risk index corresponding to each of the monitoring data. The risk fusion module is used to synthesize all the individual risk indices after correction to obtain a comprehensive risk index.

[0007] To achieve the above objectives, embodiments of the present invention also provide a chemical transportation risk perception device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the chemical transportation risk perception method as described in any of the above embodiments.

[0008] To achieve the above objectives, embodiments of the present invention also provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the chemical transportation risk perception method as described in any of the above embodiments.

[0009] To achieve the above objectives, embodiments of the present invention also provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the chemical transportation risk perception method as described in any of the above embodiments.

[0010] Compared with existing technologies, the chemical transportation risk perception method, apparatus, equipment, storage medium, and product disclosed in this invention first acquire monitoring data related to the chemicals transported by the vehicle; wherein, the monitoring data includes at least two of vehicle operation monitoring data, driver monitoring data, and environmental monitoring data; then, based on the monitoring data, multi-dimensional node features and an initial single risk index are extracted and generated; next, on a causal graph structure pre-constructed based on domain knowledge, a pre-trained graph neural network is used to perform message passing and node state updates on the node features, and causal reasoning is used to correct the initial single risk index, outputting the corrected single risk index corresponding to each monitoring data; finally, all corrected single risk indices are combined to obtain a comprehensive risk index. Therefore, this invention, by acquiring multi-source monitoring data related to the chemicals transported by the vehicle and performing multi-dimensional feature fusion and causal reasoning correction based on a causal graph structure and graph neural network, effectively improves the accuracy of risk assessment in the chemical transportation process, contributing to risk prevention and control. Attached Figure Description

[0011] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic flowchart of a chemical transportation risk perception method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a chemical transportation risk sensing device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a chemical transportation risk sensing device provided in an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] See Figure 1 This is a flowchart illustrating a method for perceiving risks in chemical transportation according to an embodiment of the present invention. Specifically, the method for perceiving risks in chemical transportation includes steps S1 to S4: S1. Acquire monitoring data related to the chemicals transported by the vehicle; wherein the monitoring data includes at least two of the following: vehicle operation monitoring data, driver monitoring data, and environmental monitoring data; S2. Based on the monitoring data, extract and generate multi-dimensional node features and an initial single risk index; S3. On the causal graph structure pre-constructed based on domain knowledge, a pre-trained graph neural network is used to perform message passing and node state updates on the node features, and causal reasoning is performed to correct the initial single risk index, outputting the corrected single risk index corresponding to each of the monitoring data. S4. Combine all the individual risk indices after adjustment to obtain the comprehensive risk index.

[0015] Specifically, this embodiment is applicable to risk assessment of hazardous chemicals during transportation, especially in chemical industrial parks. For example, the method is applied to a risk assessment system, which includes a vehicle detection module, a data acquisition module, a data processing and feature fusion module, an edge computing auxiliary module, a smart management platform integration module, and a dynamic risk assessment model. The functions of each part are illustrated below: 1. Vehicle detection module When a vehicle enters the chemical industrial park, this module is responsible for acquiring basic vehicle information (e.g., tanker truck, license plate number XXX) and information about the chemicals it is transporting (e.g., highly toxic chemicals, such as liquid chlorine, with a median lethal dose (LD50) of 30 mg / kg, no flash point, and strong corrosiveness). Subsequently, this module sends this information to the intelligent management platform integration module, enabling the platform to connect with the sensors installed on the transport vehicle.

[0016] 2. Data Acquisition Module This module includes: sensors installed on the transport vehicle, vehicle-mounted cameras, driver's cab cameras, and park environment sensors.

[0017] Status data is collected through sensors on the vehicle: vehicle position (Global Positioning System, outputting latitude and longitude coordinates), speed (vehicle speed sensor, unit: km / h), acceleration (accelerometer, unit: m / s²), vibration (vibration sensor, unit: g), and braking frequency (calculated by wheel speed sensor in conjunction with brake switch signal, unit: times / hour). This data is sampled at a frequency of 10Hz and stored as a time-series file.

[0018] The system uses an in-vehicle camera (1080p resolution, 30fps) to monitor driver actions and obtain driver behavior data.

[0019] The current weather conditions, road congestion index, and hazardous gas concentration in the chemical industrial park are obtained through environmental sensors.

[0020] 3. Edge computing auxiliary module Edge computing gateways (such as NVIDIA Jetson Xavier NX) are deployed on transport vehicles to perform preliminary processing on data collected by sensors on the vehicle and driver behavior data. Lightweight neural network models (such as object detection algorithms using MobileNetV3 combined with Single Shot MultiBox Detector (SSD)) are then used to filter out potential risk events based on vehicle location and driver behavior data.

[0021] For example, lane departure can be determined by the relative position of the vehicle to the lane line: the lane departure threshold set by the model is half the width of the lane line (assuming the lane line width is 3m, then the threshold is 1.5m). If the deviation distance between the vehicle and the lane line reaches 1.6m, exceeding the threshold, it is determined as a lane departure event, i.e., there is a potential risk.

[0022] For example, driver behavior data can be analyzed to determine blink frequency and eyelid closure. The preset safe threshold for blink frequency is 5 times per minute, and the threshold for eyelid closure is 80%. If the analysis results show that the driver's blink frequency is 4 times per minute and the eyelid closure is 85%, both exceeding the threshold, it is judged as a sign of fatigue and a potential risk.

[0023] When a potential risk event is identified, the edge computing module controls the data acquisition module to transmit the relevant data to the data processing and feature fusion module.

[0024] 4. Data Processing and Feature Fusion Module This module preprocesses the collected data (monitoring data related to the chemicals transported by the vehicle), extracts features, and generates an initial single risk index.

[0025] Data preprocessing: Digital filtering techniques (such as Kalman filtering) are used to remove high-frequency noise (e.g., 50 Hz power frequency interference), and data such as vehicle position, speed, acceleration, vibration, and braking frequency are smoothed.

[0026] 5. Dynamic Risk Assessment Model: This model is used to further process the features and initial single risk indices output by the data processing and feature fusion module, and finally output a comprehensive risk index.

[0027] Specifically, this implementation method collects at least two types of real-time data related to chemical transportation, such as vehicle operation monitoring data, driver monitoring data, and environmental monitoring data, capturing risk information from three dimensions: vehicle, personnel, and environment. From the aforementioned monitoring data, multi-dimensional node features are extracted and generated, and an initial single risk index corresponding to each type of monitoring data is preliminarily calculated. On a causal graph structure pre-constructed based on chemical industry expert knowledge (i.e., domain knowledge) (this structure represents the causal relationships between various risk factors), a pre-trained graph neural network (GNN) is used to perform message passing and node state updates on this causal graph, allowing each risk node to "see" information from other related nodes, outputting a single risk index corrected by causal reasoning. All corrected single risk indices (vehicle risk, driver risk, environmental risk, etc.) are fused and calculated to obtain a comprehensive risk index, which represents the overall risk level of the current chemical transportation task.

[0028] Compared with existing technologies, this implementation method effectively improves the accuracy of risk assessment in the chemical transportation process by acquiring multi-source monitoring data related to the chemicals transported by the vehicle, and performs multi-dimensional feature fusion and causal reasoning correction based on causal graph structure and graph neural network. It can also capture complex risks caused by the coupling of multiple factors in a timely manner, which is helpful for risk prevention and control.

[0029] In a preferred embodiment, the cause-effect graph structure is constructed in the following manner: Acquire domain knowledge; Based on the domain knowledge, nodes and directed edges are predefined to obtain a causal graph structure; The nodes in the cause-effect graph structure include root cause nodes, intermediate nodes, and result nodes; the root cause nodes represent the root cause of the risk, the intermediate nodes represent the manifestation of the risk, and the result nodes represent the consequences of the risk.

[0030] Specifically, causal relationships between risk factors are extracted from expert experience or historical accident reports in areas such as chemical safety transportation management, and this information is used as domain knowledge. Based on this domain knowledge, nodes and directed edges are defined to obtain a causal graph structure. Nodes represent various risk-related factors, and directed edges represent the direction of causal relationships. Node types include root cause nodes, intermediate nodes, and result nodes. Root cause nodes are the fundamental causes of risks, with only outgoing edges and no incoming edges; intermediate nodes represent the manifestations or intermediate states of risks; and result nodes are the final consequences of risks, with only incoming edges and no outgoing edges. Based on the constructed causal graph structure, the extracted node features and initial single risk indices are used as inputs. Graph neural networks are then used to analyze the causal relationships between risk factors. For example, it is identified that "high environmental risk index (slippery road)" is the root cause of "increased vehicle operation risk index (intensified vibration, frequent braking)" and "increased driver behavior risk index (mental stress)," rather than vehicle malfunction causing "increased vehicle operation risk index." The identified information is then used to adjust each risk index.

[0031] For example, the following is a simple illustration of how to construct a cause-effect graph structure: 1. Node Definition (Define three types of core nodes) (1) Root cause nodes represent the possible root causes of risks, mainly including: N_env: Environmental factors (road conditions, weather, gas concentration); N_vehicle: Vehicle mechanical condition; N_driver: Driver's physiological state; N_chemical: Chemical properties.

[0032] (2) Intermediate nodes represent observable risk manifestations, mainly including: N_vibration: Vehicle vibration; N_braking: Braking frequency; N_fatigue: Fatigue characteristics; N_distraction: Distraction feature.

[0033] (3) Result node, representing the final risk consequences, mainly includes: N_leak_risk: Leakage risk.

[0034] 2. Edge Definition: Directed edges in the causal graph are initialized based on domain knowledge, representing possible causal relationships, for example: N_env → N_vibration, N_env → N_braking, N_env → N_fatigue; N_vehicle → N_vibration, N_vehicle → N_braking; N_driver → N_fatigue, N_driver → N_distraction, N_driver → N_braking; N_chemical → N_leak_risk; N_vibration → N_leak_risk, N_braking → N_leak_risk.

[0035] Construct a causal graph structure based on the nodes and edges defined above.

[0036] In a preferred embodiment, the graph neural network is trained in the following manner: Obtain historical samples corresponding to the monitoring data; The features in the historical samples are assigned to the corresponding nodes of the causal graph structure; The graph neural network is invoked to perform multiple rounds of message passing on the causal graph structure, learns the weights of the directed edges through a causal attention mechanism, and adjusts the parameters of the graph neural network.

[0037] Specifically, when using Graph Neural Networks (GNNs) for causal reasoning, the core objective is to learn the weights (i.e., causal strength) of edges from real-world data and identify the most probable causal path. This approach primarily utilizes message-passing neural networks to perform information propagation and reasoning on a causal graph structure. The specific steps are as follows: (a) Node initialization: Each node of the cause-effect graph structure initial feature vector It is obtained from its corresponding input feature mapping: ; in, and These are nodes The input projection weight matrix and the input projection bias vector, It is a node The corresponding feature subset (e.g., feature vectors of environment-related nodes). The feature subset is obtained by splitting the node features.

[0038] (b) Message passing (multiple iterations): In each layer Message passing occurs along the directed edges of the causal graph structure.

[0039] Message function: ; in, From node to neighboring nodes The message; It is the edge → The features are initialized as prior causal strength based on domain knowledge; It is the first Layer message functions; It is a node In the Hidden representation of layers, It is a node In the Hidden representation of layers.

[0040] Aggregate functions: ; in, It is a node Aggregate all incoming messages, It is a node The set of incoming neighbors. The aggregation function (AGG) can be a summation, an average, or an attention-weighted summation.

[0041] Update function: ; Typically implemented as follows: ; in, It updates the node representation based on aggregated messages, where || represents vector concatenation. It is an activation function; It is the first The layer update function, and They are the first The weight matrix and bias vector in the layer update function.

[0042] (c) Causal attention mechanism: Introducing attention weights in message passing Used to learn the importance of edges (i.e., causal strength): ; in, It is an attention vector. In all pointing nodes Neighbors Above calculation, and These are for nodes and The node representation obtained by performing a linear transformation on the representation; (d) Perform hypothesis analysis using the "do-operation" to verify the identified causal relationships and quantify the causal effects. For the suspected root cause node N_cause, perform the intervention do(N_cause=value). Then, rerun the GNN inference on the graph after the intervention and observe the change in the result node N_leak_risk: Causal_Effect=P(N_leak_risk|do(N_cause=value))-P(N_leak_risk|do(N_cause=baseline)); Here, Causal_Effect is the quantified causal effect value; the probability P is obtained by applying softmax to the node representation of the GNN output.

[0043] The model's final output includes the identified root cause nodes and their causal strength (providing a basis for precise intervention), as well as the causal path, the causal strength of each edge, and the confidence level based on counterfactual reasoning results, and the revised individual risk indices, etc.

[0044] Model reasoning process: GNNs perform multi-round message passing on a causal graph structure, and the learned attention weights are as follows: N_env→N_vibration: weight 0.8; N_env→N_braking: weight 0.7; N_env→N_fatigue: weight 0.6; N_vehicle→N_vibration: weight 0.2; N_driver→N_fatigue: weight 0.3.

[0045] Counterfactual verification: Assuming intervention in N_env (e.g., improving road conditions): P(N_leak_risk|do(road_condition=good))=0.2; where P is the probability; do(road_condition=good) is the intervention operation that forces the road condition to be set to "good".

[0046] Baseline case: P(N_leak_risk|do(road_condition=poor))=0.8; where do(road_condition=poor) is the intervention operation that forces the road condition to be set to "poor".

[0047] Causal effect: 0.8-0.2=0.6, the result is significant.

[0048] The output example is as follows: Root cause node: The most likely root cause node identified, such as N_env.

[0049] Causal path: A complete causal chain from the root cause node to the result node, such as: N_env (bad road conditions) → N_vibration (increased vibration) → N_leak_risk (increased risk of leakage).

[0050] Causal strength: The quantified causal effect value of each edge, with a value range of [0,1].

[0051] Confidence score: a confidence score based on the results of counterfactual reasoning.

[0052] Individual risk indices derived from causal reasoning: vehicle operation risk index Rv, driver behavior risk index Rd, and environmental risk index Re.

[0053] In a preferred embodiment, the method further includes: Obtain characteristic data of the chemicals transported by the vehicle; The comprehensive risk index is obtained by combining all the individual risk indices after comprehensive adjustment, including: Obtain a predefined correlation matrix; wherein the correlation matrix records the predefined weights of each chemical characteristic in each risk dimension; Based on the predefined weights, calculate the basic weights for each of the risk dimensions according to the characteristic data; Based on the aforementioned basic weights, the weighted sum of all the modified individual risk indices is used to obtain the comprehensive risk index.

[0054] Specifically, the weights for each individual risk index are determined based on the characteristic data of the chemicals transported by the vehicle, and all risk indices are then combined based on these weights. The basic technical logic is as follows: 1. Input: (1) Individual risk indices derived from causal reasoning: vehicle operation risk index Rv, driver behavior risk index Rd, and environmental risk index Re.

[0055] (2) Dynamic characteristic vector of chemical transportation (i.e. characteristic data of the chemicals transported by the vehicle): C = [LD50, flash point, corrosivity, interaction terms].

[0056] 2. Processing Objectives: Based on the characteristics of the chemicals, the fusion weights for each risk dimension are dynamically allocated. Weight adjustments are made as needed, taking into account real-time contextual factors.

[0057] The specific processing steps are as follows: (a) Define the characteristic-risk correlation matrix: Define a learnable weight matrix to map chemical properties to three risk dimensions:

[0058] The table records the contribution weight of each chemical characteristic to each risk dimension (i.e., predefined weights).

[0059] (b) Dynamic weight calculation: For each risk dimension k∈{v,d,e}, its basic weights Calculated using matrix multiplication: ; in: It is the first of the chemical property vectors One component; These are the weight parameters at the corresponding positions in the correlation matrix; It is the sigmoid function, used to compress the result to the range (0,1).

[0060] (c) Furthermore, we introduce context-enhancing weight adjustments: Fine-tune the weights by considering real-time contextual factors (such as weather and time of day): ; in, It is the first Each situational factor (e.g., rainy day = 0.3, nighttime = 0.2). It is the situation sensitivity coefficient.

[0061] To maintain the comparability of the weights across all dimensions, all adjusted weights are ultimately normalized so that their sum equals 1.

[0062] Based on the normalized weights, all modified individual risk indices are weighted and fused. The fusion method can be linear weighted fusion or nonlinear fusion considering the risk superposition effect. Specifically, for high-risk scenarios, geometric mean or maximum operator is used to amplify the risk superposition effect, resulting in a comprehensive risk index. ; in, It is the weight of the k-th risk dimension after normalization. It is a single risk index of the k-th risk dimension after causal adjustment.

[0063] After performing the above steps, the final output is: a dynamic weight vector. (This vector represents the final weight allocation for each risk dimension), the comprehensive risk index R_fused∈[0,1] and the basis for weight allocation (the main basis for this weight allocation, such as "based on high toxicity characteristics" or "based on flammability characteristics").

[0064] Example explanation: Scenario 1: When transporting liquid chlorine (highly toxic and corrosive), the system will significantly increase the weight of the environmental risk index (especially the gas concentration) and the vehicle operation risk index (related to the probability of leakage).

[0065] Scenario 2: When transporting highly flammable materials, the system will increase the weight of temperature factors in the driver behavior risk index (misoperation may cause open flame) and the environmental risk index accordingly.

[0066] Furthermore, based on the aforementioned basic weights, a weighted sum of all the modified individual risk indices is performed to obtain a comprehensive risk index, including: Real-time contextual factors are introduced to fine-tune the basic weights; Based on all the adjusted basic weights, the weighted sum of all the corrected individual risk indices is used to obtain the comprehensive risk index.

[0067] In a preferred embodiment, it further includes: The calculation process of the comprehensive risk index incorporates Bayesian uncertainty quantification, resulting in an output comprehensive risk index with an uncertainty range.

[0068] Specifically, the method is applied to a risk assessment system, which includes a vehicle detection module, a data acquisition module, a data processing and feature fusion module, an edge computing auxiliary module, a smart management platform integration module, and a dynamic risk assessment model. Descriptions of each component can be found in the above implementation method and will not be repeated here. In particular, Bayesian uncertainty quantification is introduced into the dynamic risk assessment model, giving the output comprehensive risk index an uncertainty range. The technical logic of this is explained below: 1. Input: Intermediate calculation results of each layer within the dynamic risk assessment model.

[0069] 2. Processing Objective: To introduce randomness into key parts of the model and use Bayesian neural networks for probabilistic reasoning, so that the output includes not only the risk value, but also the uncertainty range of that risk value (e.g., a certain risk index = 0.75 ± 0.05), thereby providing a reliable basis for subsequent decision-making.

[0070] 3. Specific implementation method: (a) Weight randomness - cognitive uncertainty Using a Bayesian neural network, the deterministic weights (i.e., the network weight parameters) are replaced with a probability distribution: Where D represents the training data. Specifically, variational inference is employed, assuming that each weight parameter follows a Gaussian distribution: That is, each weight parameter is represented by two variational parameters: the mean and the standard deviation.

[0071] (b) Activation of randomness - random uncertainty Introduce random dropout in the output of the critical layer to maintain activation randomness even during testing: ; in, ; where, after dropout processing, the randomized activation vector, ⊙ represents element-wise multiplication, It is a mask vector. Represents the mask vector Each element is sampled from a Bernoulli distribution.

[0072] This method, known as Monte Carlo dropout, can be considered an effective approximation of Bayesian inference.

[0073] (c) Output randomness The final comprehensive risk index is modeled as a probability distribution, and since the risk index ranges from [0,1], it is assumed to follow a beta distribution: ; in, Given shape parameters and Under the condition of random variable The probability density function, It's a gamma function, a shape parameter. and It is generated by the output layer of a Bayesian neural network.

[0074] 4. Bayesian inference process: Based on Bayesian neural networks, uncertainty estimates for risk prediction are obtained through probabilistic reasoning, specifically including the following steps.

[0075] (a) Variational reasoning framework Define variational posterior To approximate the true posterior Training is performed by optimizing the lower bound of evidence: ; in, It is the lower bound of evidence; These are variational parameters. It conforms to a Gaussian distribution; It is the mathematical expectation; It is a prior distribution; It means Divergence; It represents how well the model fits the training data D given a specific set of weight parameters W.

[0076] (b) Reparameterization techniques To ensure the model is differentiable and supports gradient descent training, a reparameterization technique is used to sample the weights: ; in, , It is the mean. It is the standard deviation.

[0077] (c) Monte Carlo sampling inference During the prediction phase, T forward propagation samplings are performed to approximate Bayesian inference: For t=1 to T, sample weights from the variational posterior: Calculate the risk index for this sampling: ;in, It is a set of specific weight parameter values ​​sampled during the t-th forward propagation; This is the predicted risk index value under the t-th sampling. Is using Parameterized neural network models; This is the input feature for this part, and its value is the risk index after dynamic weight fusion.

[0078] (d) Uncertainty calculation Based on the results of T samplings, the following statistics are calculated: Predicted mean: ; Prediction variance (total uncertainty): ; Cognitive uncertainty (model uncertainty): ; Random uncertainty (data noise): ; in, It is the output variance of the t-th sample.

[0079] (e) Confidence Interval Calculation The confidence interval for the risk index is calculated based on the sampling results: ; ; ; in, It is an estimate of the mean. It is an estimate of the standard deviation. It is a 95% confidence interval.

[0080] (f) Uncertainty classification Based on the size of the uncertainty interval, the current prediction results are classified and corresponding operational suggestions are given, as follows: Low uncertainty (e.g., 0.72 ± 0.03): The model is confident in its predictions and can make decisions automatically; Moderate uncertainty (e.g., 0.65±0.08): It is recommended to increase the monitoring frequency and prepare for manual review; High uncertainty (e.g., 0.55 ± 0.15): Immediately triggers manual review, which may require additional sensor data.

[0081] 5. The output data mainly includes: (1) the comprehensive risk index after multi-dimensional fusion and uncertainty calibration, with a value range of [0,1]; (2) analysis of the main risk types and root causes: such as "the dominant risk type: environmental risk; the root cause: road congestion leads to frequent vehicle start-stop, which aggravates the fatigue stress of the corrosive chemical tank"; (3) uncertainty estimation: a quantitative representation of the confidence level of the comprehensive risk index.

[0082] In addition, the output data may also include: tiered early warning suggestions generated based on the identified root causes. For example, when the comprehensive risk index is greater than 0.7, a "red alert" is triggered, suggesting that the platform automatically issue an instruction to "force the vehicle to slow down to a safe zone" and notify the emergency response team to prepare for intervention.

[0083] In a preferred embodiment, the step of extracting and generating multi-dimensional node features and an initial single risk index based on the monitoring data includes: Feature extraction is performed on each of the monitoring data to obtain an initial feature vector corresponding to each of the monitoring data; Based on the initial feature vectors, determine the initial single risk index corresponding to each monitoring data; All the initial feature vectors are concatenated to obtain a concatenated vector; and local short-term feature extraction and long-term dependency feature capture are performed on the concatenated vector to generate enhanced temporal features; The enhanced time-series features are concatenated with all the initial single risk indices to obtain multi-dimensional node features.

[0084] Specifically, the method is applied to a risk assessment system, which includes a vehicle detection module, a data acquisition module, a data processing and feature fusion module, an edge computing auxiliary module, a smart management platform integration module, and a dynamic risk assessment model. Descriptions of each component can be found in the above implementation method. The data processing and feature fusion module includes a vehicle status analysis model, a driver behavior analysis model, an environmental risk analysis model, and a chemical property quantification model. The dynamic risk assessment model further processes the feature data and single risk indices extracted by the data processing and feature fusion module, finally outputting a comprehensive risk index. The working principles of some models are described below.

[0085] 1. Vehicle State Analysis Model Function: Analyzes risks related to the vehicle's operating condition, such as mechanical failure and aggressive driving.

[0086] Implementation steps: (1) Input: Preprocessed vehicle sensor time series data (i.e. vehicle operation monitoring data).

[0087] (2) Feature engineering: First, calculate the statistical features in the time domain. For example: Location characteristics: mean longitude, mean latitude, variance of longitude and latitude; Speed ​​characteristics: mean speed, variance speed, maximum / minimum speed; Acceleration characteristics: mean acceleration, variance of acceleration; Vibration characteristics: mean vibration, peak vibration, root mean square vibration; Braking frequency characteristics: mean, variance, total number of braking events.

[0088] (3) Model analysis: The time series data related to vehicle operation (i.e. vehicle operation monitoring data) is input into a neural network that combines a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM).

[0089] 1D-CNN: Uses one-dimensional convolutional kernels to slide along the time axis to extract local short-term dependent features (such as instantaneous patterns during emergency braking). LSTM: Through its gating mechanism (input gate, forget gate, output gate), it captures long-term dependencies of parameters such as speed and vibration (e.g., continuous overspeed or long-term abnormal vibration).

[0090] (4) Output: Vehicle operating state vector: composed of statistical features and deep neural network features, such as [mean speed, variance of acceleration, peak vibration, ...]; Vehicle operation risk index (i.e., the initial single risk index corresponding to vehicle operation monitoring data): a scalar from 0 to 1, output by the fully connected layer and sigmoid activation function at the end of the vehicle state analysis model. The higher the value, the greater the risk of abnormal vehicle state.

[0091] 2. Driver Behavior Analysis Model Function: Real-time monitoring and assessment of driver fatigue, distraction, and other behavioral risks.

[0092] Implementation steps: (1) Input: Video frame sequence from the cab camera (i.e., driver monitoring data).

[0093] (2) Key point detection: The OpenPose algorithm is used to detect key points of the driver's face (such as the outline of the eyes and mouth) and hands (wrist and knuckles), and output their pixel coordinates.

[0094] (3) Spatial Alignment: The detected face and hand region images are input into the spatial transformation network to unify scale and viewpoint, eliminating interference caused by changes in posture. The spatial transformation network uses bilinear interpolation to calculate the pixel values ​​of the sampling points, using the following formula: ; in, These are the target coordinates; These are the coordinates of the four surrounding pixels in the source image. It is its pixel value.

[0095] (4) Calculate behavioral characteristics: Behavioral characteristics include eye-aspect ratio (EAR), mouth-aspect ratio (MAR), and head posture. When the eye-aspect ratio is consistently lower than the corresponding preset threshold (e.g., 0.2), it is determined to be fatigue. Head posture estimation: The three-dimensional posture of the key points of the head is solved by the Perspective-n-Point (PnP) algorithm and converted into Euler angles (pitch angle, yaw angle, roll angle). By analyzing the changing trend of Euler angles, it is determined whether the driver has abnormal behaviors such as looking down or looking left and right.

[0096] (5) Output: Driver state vector: composed of the above features, such as [EAR, MAR, Euler angles, hand key point coordinates, etc.].

[0097] Driver Behavior Risk Index (i.e., the initial single risk index corresponding to driver monitoring data): a scalar from 0 to 1, output by the classifier at the end of the driver behavior analysis model. The higher the value, the greater the risk that the driver is in a state of fatigue or distraction.

[0098] 3. Environmental Risk Analysis Model Function: To assess the impact of the external environment on transportation safety.

[0099] Implementation steps: (1) Input: Park environmental sensor data (i.e., environmental monitoring data, including road congestion index, hazardous gas concentration, temperature, humidity, wind speed, wind direction, etc.).

[0100] (2) Key indicators: Road congestion index and average concentration of hazardous gases. The road congestion index is equal to the ratio of the free-flow velocity of a road segment to the actual travel speed of the road segment. The average concentration of hazardous gases refers to the average concentration data obtained from multiple sensors at the same time period.

[0101] (3) Data normalization: Map data of different dimensions to the [0,1] interval for easy fusion. Min-Max normalization is adopted: Xnorm=(X-Xmin) / (Xmax-min), where Xnorm is the normalized value, Xmin is the minimum value, and Xmax is the maximum value.

[0102] (4) Output: Park environmental factor vector: composed of normalized data, such as [congestion index, gas concentration, temperature, humidity, wind speed, wind direction].

[0103] Environmental risk index: a scalar from 0 to 1, which can be output by analyzing the correlation of multiple factors through models such as random forest or XGBoost. The higher the value, the more severe the external environment.

[0104] 4. Quantitative Model of Chemical Properties Function: Quantifies the inherent hazardous properties of chemicals into a computable vector.

[0105] Input: Data on toxicity (LD50), flammability (flash point), corrosivity, etc. (i.e., the characteristic data of the chemicals being transported by the vehicle) obtained from the Material Safety Data Sheet (MSDS). Interaction assessment: Based on chemical knowledge base or rules, determine whether there is an interaction between properties (e.g., if a toxic substance is corrosive, it will exacerbate the hazard) and assign an interaction weight value (e.g., 0.5).

[0106] Output: Dynamic characteristic vector of chemical transport: directly composed of attributes and interaction terms, for example, for an organic solvent [LD50=100, flash point=25, corrosivity=1, interaction term=0.5].

[0107] 5. The dynamic risk assessment model is the core of the system's decision-making. It does not directly process raw sensor data, but rather performs high-level fusion and inference on the outputs of the aforementioned specialized models. Its inputs all come from the outputs of upstream modules, such as: Vehicle operation risk index: derived from the vehicle condition analysis model, scalar, range [0,1].

[0108] Driver Behavior Risk Index: derived from the driver behavior analysis model, scalar, range [0,1].

[0109] Environmental risk index: derived from the environmental risk analysis model, scalar, range [0,1].

[0110] Vehicle operating state feature vector: from the feature extraction layer of the vehicle state analysis model, such as [mean speed, variance of acceleration, peak vibration, ...].

[0111] Driver state vector: derived from the feature extraction layer of the driver behavior analysis model, such as [EAR, MAR, head Euler angles, ..., hand key points].

[0112] Park environmental factor vector: derived from the feature extraction layer of the environmental risk analysis model, such as [road congestion index, average gas concentration, temperature, humidity, wind speed, wind direction].

[0113] Dynamic characteristic vector of chemical transportation: constructed based on MSDS data, such as [LD50 value, flash point temperature, corrosiveness index, interaction terms].

[0114] It is worth noting that the vehicle operating state feature vector, driver state vector, and park environmental factor vector are the initial feature vectors.

[0115] The dynamic risk assessment model comprises a five-layer architecture: a multi-scale temporal feature extraction layer, a causal risk inference layer, a context-aware fusion layer, a meta-learning adaptation layer, and a Bayesian uncertainty quantification layer. The causal risk inference layer uses a causal graph structure and graph neural network to correct the initial single risk index. The context-aware fusion layer introduces real-time contextual factors to fine-tune the basic weights, and uses the fine-tuned basic weights to perform a weighted summation of all corrected single risk indices to obtain a comprehensive risk index. The Bayesian uncertainty quantification layer introduces Bayesian uncertainty quantization, giving the output comprehensive risk index an uncertainty range. This section focuses on the node feature generation parts of the multi-scale temporal feature extraction layer and the causal risk inference layer, as well as the meta-learning adaptation layer. The working principles of other layers are detailed in the above embodiments and will not be elaborated here.

[0116] 1. Multi-scale temporal feature extraction layer Input: Receive feature vectors from the driver behavior analysis model, vehicle state analysis model, and environmental risk analysis model, and their time-varying sequences.

[0117] Suppose that at time step t, the input is a concatenated feature vector. It consists of the outputs of each upstream model. A time series of length T can be represented as: ; among them, each It is a multidimensional vector representing a comprehensive snapshot of the vehicle, driver, and environment states at time t.

[0118] Processing: Temporal Convolutional Network (TCN) is used to extract local short-term features (such as the instantaneous effects of sudden braking), while LSTM with attention mechanism enhancement is used to capture long-term dependent features (such as the cumulative effect of driver fatigue).

[0119] Specifically, it includes: (1) Temporal Convolutional Network (TCN) extracts short-term features Using one-dimensional dilated causal convolution, by introducing a dilation factor d, the convolution kernel skips some data points while covering the input, thereby exponentially expanding the receptive field without increasing computational cost. The dilation factor of the k1-th layer is typically... This ensures that the output at time t depends only on time step t and earlier, preventing future information leakage and meeting the requirements of real-time inference.

[0120] Formula expression: For an input sequence and convolution kernel The dilated convolution output at time t is: ; in, Indicates the expansion factor as Convolution operations; Is the convolution kernel at the 1st The weight of each position; Is the input sequence in The value at time; It is the size of the convolution kernel.

[0121] The final layer of the TCN network outputs a feature vector H_t^{TCN} (i.e., local short-term feature) for each time step t. This vector encodes local short-term context information centered at time t.

[0122] (2) Attention-enhanced LSTM captures long-term dependent features Its purpose is to capture dependencies that span a longer time period (such as the cumulative process of driver fatigue, continuous speeding, and long-term environmental degradation trends).

[0123] Its core mechanism is that LSTM itself learns long-term dependencies through its gating mechanism (input gate, forget gate, output gate, and cell state). By introducing an attention mechanism, the model can learn which hidden state of the LSTM should receive more attention at different times.

[0124] The specific implementation and formula are as follows: (a) LSTM cell computation (at each time step t): Input Gate: ; Forgotten Gate: ; Output gate: ; Candidate state cells: ; Cell status update: ; Hidden output: ; in, It's the sigmoid function, and ⊙ represents element-wise multiplication. It is the hyperbolic tangent function; , , , , , , , These are the weight matrices from input to the input gate, from hidden state to forget gate, from input to forget gate, from hidden state to forget gate, from input to output gate, from hidden state to output gate, from input to candidate cell state, and from hidden state to candidate cell state, respectively. , , , These are the bias vectors for the input gate, the forget gate, the output gate, and the candidate cell state, respectively. It's the input gate, which determines how much new information is updated; It is the Gate of Oblivion, which determines how much old information is discarded; It is the output gate, which determines how many cell states are output; It is a candidate cell state; It is a cellular state; It is in a hidden state.

[0125] (b) Attention mechanism computation (after all time steps are completed): Assuming we have the LSTM hidden states for all time steps .

[0126] The attention mechanism computes each hidden state using a small neural network. Importance score , It is the hidden state at time s; then it is normalized into weights using the softmax function. : ; ; in, and These are the weights and biases of the attention network; It is a weight vector, mainly used to map the hidden state after nonlinear transformation into a scalar score; It represents the importance of the hidden state at the s-th time step to the current comprehensive evaluation.

[0127] (c) Generate context vector: The final long-term feature vector c is a weighted sum of all LSTM hidden states: ; The vector c is no longer limited to the last state, but includes information about all the "important" moments in the entire sequence that the model considers "important", thus capturing long-term dependencies more effectively.

[0128] Output: Enhanced temporal features that fuse short-term and long-term contextual information. Specifically, the last state (or pooled features) of the short-term feature sequence extracted by TCN, H_T^{TCN}, is concatenated with the long-term context vector c generated by the LSTM attention mechanism to form the final enhanced temporal feature H_fused.

[0129] Output characteristics: The H_fused vector simultaneously encodes: short-term instantaneous patterns (captured by TCN): for example, whether there was sudden braking accompanied by severe vibration in the last few seconds. Long-term evolution trends (captured by attention LSTM): for example, the driver's fatigue index EAR showing a continuous downward trend over the past ten minutes, while the vehicle speed variance continues to increase.

[0130] 2. Causal Risk Reasoning Layer Input: Enhanced time series features H_fused and risk indices of each upstream model (initial vehicle operation risk index Rv', initial driver behavior risk index Rd', initial environmental risk index Re').

[0131] The risk index is concatenated with the enhanced features to form complete node features: F_nodes=[Rv',Rd',Re']⊕H_fused.

[0132] 3. Meta-learning adaptation layer Input: Historical risk assessment cases, including the input characteristics of each upstream model and the final verified real risk labels.

[0133] Solution: The Model-Agnostic Meta-Learning (MAML) algorithm is employed to enable the model to quickly adapt to new scenarios. When a new chemical is introduced into the park or a certain risk pattern recurs, the model can be quickly fine-tuned with only a small amount of new sample data, without the need for retraining from scratch.

[0134] Output: An updated risk assessment model that can quickly adapt to new risk patterns.

[0135] Compared with existing technologies, the chemical transportation risk perception method provided in this invention first acquires monitoring data related to the chemicals transported by the vehicle; wherein the monitoring data includes at least two of the following: vehicle operation monitoring data, driver monitoring data, and environmental monitoring data; then, based on the monitoring data, multi-dimensional node features and an initial single risk index are extracted and generated; next, on a causal graph structure pre-constructed based on domain knowledge, a pre-trained graph neural network is used to perform message passing and node state updates on the node features, and causal reasoning is used to correct the initial single risk index, outputting the corrected single risk index corresponding to each monitoring data; finally, all corrected single risk indices are combined to obtain a comprehensive risk index. Therefore, this invention, by acquiring multi-source monitoring data related to the chemicals transported by the vehicle and performing multi-dimensional feature fusion and causal reasoning correction based on a causal graph structure and graph neural network, effectively improves the accuracy of risk assessment in the chemical transportation process, contributing to risk prevention and control.

[0136] See Figure 2 , Figure 2 This is a schematic diagram of a chemical transportation risk sensing device provided in an embodiment of the present invention. The chemical transportation risk sensing device 20 includes: The data acquisition module 21 is used to acquire monitoring data related to the chemicals transported by the vehicle; wherein the monitoring data includes at least two of the following: vehicle operation monitoring data, driver monitoring data, and environmental monitoring data. The initial processing module 22 is used to extract and generate multi-dimensional node features and an initial single risk index based on the monitoring data. The risk correction module 23 is used to perform message passing and node state updates on the node features using a pre-trained graph neural network on a causal graph structure pre-constructed based on domain knowledge, to perform causal reasoning correction on the initial single risk index, and to output the corrected single risk index corresponding to each of the monitoring data. Risk fusion module 24 is used to synthesize all the individual risk indices after correction to obtain a comprehensive risk index.

[0137] In one implementation, a cause-effect graph construction module is also included, for: Acquire domain knowledge; Based on the domain knowledge, nodes and directed edges are predefined to obtain a causal graph structure; The nodes in the cause-effect graph structure include root cause nodes, intermediate nodes, and result nodes; the root cause nodes represent the root cause of the risk, the intermediate nodes represent the manifestation of the risk, and the result nodes represent the consequences of the risk.

[0138] In one implementation, a graph neural network training module is further included, for: Obtain historical samples corresponding to the monitoring data; The features in the historical samples are assigned to the corresponding nodes of the causal graph structure; The graph neural network is invoked to perform multiple rounds of message passing on the causal graph structure, learns the weights of the directed edges through a causal attention mechanism, and adjusts the parameters of the graph neural network.

[0139] In one embodiment, the data acquisition module is further configured to: acquire characteristic data of the chemicals transported by the vehicle; The risk fusion module 24 is specifically used for: Obtain a predefined correlation matrix; wherein the correlation matrix records the predefined weights of each chemical characteristic in each risk dimension; Based on the predefined weights, calculate the basic weights for each of the risk dimensions according to the characteristic data; Based on the aforementioned basic weights, the weighted sum of all the modified individual risk indices is used to obtain the comprehensive risk index.

[0140] In one implementation, the step of weighted summing of all the modified individual risk indices based on the fundamental weights to obtain a comprehensive risk index includes: Real-time contextual factors are introduced to fine-tune the basic weights; Based on all the adjusted basic weights, the weighted sum of all the corrected individual risk indices is used to obtain the comprehensive risk index.

[0141] In one implementation, an uncertainty quantification module is further included, for: The calculation process of the comprehensive risk index incorporates Bayesian uncertainty quantification, resulting in an output comprehensive risk index with an uncertainty range.

[0142] In one embodiment, the initial processing module 22 is specifically used for: Feature extraction is performed on each of the monitoring data to obtain an initial feature vector corresponding to each of the monitoring data; Based on the initial feature vectors, determine the initial single risk index corresponding to each monitoring data; All the initial feature vectors are concatenated to obtain a concatenated vector; and local short-term feature extraction and long-term dependency feature capture are performed on the concatenated vector to generate enhanced temporal features; The enhanced time-series features are concatenated with all the initial single risk indices to obtain multi-dimensional node features.

[0143] It is worth noting that the working principle of the device is the same as that of any of the above method embodiments, and will not be repeated here.

[0144] Compared with existing technologies, the chemical transportation risk perception device disclosed in this invention first acquires monitoring data related to the chemicals transported by the vehicle; wherein the monitoring data includes at least two of the following: vehicle operation monitoring data, driver monitoring data, and environmental monitoring data; then, based on the monitoring data, multi-dimensional node features and an initial single risk index are extracted and generated; next, on a causal graph structure pre-constructed based on domain knowledge, a pre-trained graph neural network is used to perform message passing and node state updates on the node features, and causal reasoning is used to correct the initial single risk index, outputting the corrected single risk index corresponding to each monitoring data; finally, all corrected single risk indices are combined to obtain a comprehensive risk index. Therefore, this invention, by acquiring multi-source monitoring data related to the chemicals transported by the vehicle and performing multi-dimensional feature fusion and causal reasoning correction based on a causal graph structure and graph neural network, effectively improves the accuracy of risk assessment in the chemical transportation process, contributing to risk prevention and control.

[0145] See Figure 3 , Figure 3 This is a schematic diagram of the structure of a chemical transportation risk sensing device 30 provided in an embodiment of the present invention. The chemical transportation risk sensing device 30 includes a processor 31, a memory 32, and a computer program stored in the memory 32 and configured to be executed by the processor 31. When the processor 31 executes the computer program, it implements the steps as described in the above embodiment of the chemical transportation risk sensing method, for example... Figure 1 The steps S1 to S4 described above; or, when the processor 31 executes the computer program, it implements the functions of each module in the above-described device embodiments.

[0146] For example, the computer program can be divided into one or more modules, which are stored in the memory 32 and executed by the processor 31 to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the chemical transportation risk sensing device. For example, the computer program can be divided into multiple modules, and the specific working process of each module can be referred to the working process of the chemical transportation risk sensing device described in the above embodiments, which will not be repeated here.

[0147] The chemical transportation risk sensing device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The chemical transportation risk sensing device may include, but is not limited to, a processor 31 and a memory 32. Those skilled in the art will understand that the chemical transportation risk sensing device may also include input / output devices, network access devices, buses, etc.

[0148] The processor 31 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor 31 is the control center of the chemical transportation risk sensing equipment, connecting all parts of the equipment via various interfaces and lines.

[0149] The memory 32 can be used to store the computer program and / or modules. The processor 31 realizes various functions of the chemical transportation risk sensing device by running or executing the computer program and / or modules stored in the memory 32 and calling the data stored in the memory 32. The memory 32 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory 32 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0150] If the integrated module of the chemical transportation risk perception device is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by the processor 31, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0151] This invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the chemical transportation risk perception method as described in any of the above embodiments.

[0152] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A chemical transportation risk perception method, characterized by, include: Acquire monitoring data related to the chemicals transported by the vehicle; wherein the monitoring data includes at least two of the following: vehicle operation monitoring data, driver monitoring data, and environmental monitoring data; Based on the monitoring data, multi-dimensional node features and an initial single risk index are extracted and generated; Based on a causal graph structure pre-constructed according to domain knowledge, a pre-trained graph neural network is used to perform message passing and node state updates on the node features, and causal reasoning is used to correct the initial single risk index, outputting the corrected single risk index corresponding to each of the monitoring data. The composite risk index is obtained by combining all the individual risk indices after correction.

2. The chemical transportation risk perception method of claim 1, wherein, The cause-effect graph structure is constructed in the following way: Acquire domain knowledge; Based on the domain knowledge, nodes and directed edges are predefined to obtain a causal graph structure; The nodes in the cause-effect graph structure include root cause nodes, intermediate nodes, and result nodes; the root cause nodes represent the root cause of the risk, the intermediate nodes represent the manifestation of the risk, and the result nodes represent the consequences of the risk.

3. The chemical transportation risk perception method of claim 2, wherein, The graph neural network is trained in the following way: Obtain historical samples corresponding to the monitoring data; The features in the historical samples are assigned to the corresponding nodes of the causal graph structure; The graph neural network is invoked to perform multiple rounds of message passing on the causal graph structure, learns the weights of the directed edges through a causal attention mechanism, and adjusts the parameters of the graph neural network.

4. The chemical transportation risk perception method as described in claim 1, characterized in that, Also includes: Obtain characteristic data of the chemicals transported by the vehicle; The comprehensive risk index is obtained by combining all the individual risk indices after comprehensive adjustment, including: Obtain a predefined correlation matrix; wherein the correlation matrix records the predefined weights of each chemical characteristic in each risk dimension; Based on the predefined weights, calculate the basic weights for each of the risk dimensions according to the characteristic data; Based on the aforementioned basic weights, the weighted sum of all the modified individual risk indices is used to obtain the comprehensive risk index.

5. The chemical transportation risk perception method as described in claim 4, characterized in that, Based on the aforementioned basic weights, a weighted sum of all the modified individual risk indices is performed to obtain a comprehensive risk index, including: Real-time contextual factors are introduced to fine-tune the basic weights; Based on all the adjusted basic weights, the weighted sum of all the corrected individual risk indices is used to obtain the comprehensive risk index.

6. The chemical transportation risk perception method as described in claim 1, characterized in that, Also includes: The calculation process of the comprehensive risk index introduces Bayesian uncertainty quantification, so that the output comprehensive risk index has an uncertainty range.

7. The chemical transportation risk perception method as described in claim 1, characterized in that, The step of extracting and generating multi-dimensional node features and an initial single risk index based on the monitoring data includes: Feature extraction is performed on each of the monitoring data to obtain an initial feature vector corresponding to each of the monitoring data; Based on the initial feature vectors, determine the initial single risk index corresponding to each monitoring data; All the initial feature vectors are concatenated to obtain a concatenated vector; and local short-term feature extraction and long-term dependency feature capture are performed on the concatenated vector to generate enhanced temporal features; The enhanced time-series features are concatenated with all the initial single risk indices to obtain multi-dimensional node features.

8. A chemical transportation risk sensing device, characterized in that, include: The data acquisition module is used to acquire monitoring data related to the chemicals transported by the vehicle; wherein the monitoring data includes at least two of the following: vehicle operation monitoring data, driver monitoring data, and environmental monitoring data. The initial processing module is used to extract and generate multi-dimensional node features and an initial single risk index based on the monitoring data. The risk correction module is used to perform message passing and node state updates on the node features using a pre-trained graph neural network on a causal graph structure pre-constructed based on domain knowledge, to perform causal reasoning correction on the initial single risk index, and to output the corrected single risk index corresponding to each of the monitoring data. The risk fusion module is used to synthesize all the individual risk indices after correction to obtain a comprehensive risk index.

9. A chemical transportation risk sensing device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the chemical transport risk perception method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the chemical transportation risk perception method as described in any one of claims 1 to 7.

11. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the chemical transportation risk perception method as described in any one of claims 1 to 7.

Images