Real-time early warning and evaluation method for risks in vehicle transportation process based on internet of things

Abstract

The present application relates to the technical field of risk analysis, and particularly relates to a risk real-time early warning evaluation method in a vehicle transportation process based on an Internet of Things. The method comprises the following steps: adopting a time-varying Kalman filter model to evaluate and quantify a driving state deviation degree representing a deviation between a driver's steering intention and a vehicle dynamic feedback; inputting vehicle IMU data and cargo pressure distribution data into a preset vehicle-cargo rigid-flexible coupling dynamics model to solve a vehicle-cargo system instability index representing the vehicle under current motion disturbance in real time; and utilizing a space-time graph convolution network to fuse real-time traffic flow and meteorological data as dynamic attributes and road geometry and historical accident data as static attributes to predict risk potential field values of each road section along the transportation line within a future time window. Compared with the prior art, the present application constructs a multi-dimensional collaborative risk evaluation system of a person-vehicle-cargo-road-environment, and significantly improves the accuracy and foresight of risk early warning.

Description

Technical Field

[0001] This invention relates to the field of risk analysis technology. More specifically, this invention relates to a real-time risk early warning and assessment method for vehicle transportation processes based on the Internet of Things (IoT). Background Technology

[0002] With the rapid development of IoT and sensor technologies, safety monitoring and risk management in vehicle transportation have become increasingly important. Existing technologies primarily collect real-time vehicle operating status information such as GPS location, speed, and acceleration via onboard terminals, and acquire some vehicle operating data through the Controller Area Network (CAN) bus. Some solutions also incorporate video surveillance equipment to identify driver fatigue (such as yawning or closing eyes). These technologies, by applying threshold judgments or simple analyses to single or a few dimensions of information, achieve basic monitoring and alarms for high-risk behaviors such as speeding, sudden acceleration / deceleration, and fatigued driving, thus improving road transportation safety to some extent.

[0003] However, existing technologies have significant shortcomings in the depth and breadth of risk assessment. First, most solutions analyze risk factors within the human-vehicle-freight-road-environment system in isolation, lacking a comprehensive assessment model that deeply integrates driver physiological and psychological states, the dynamic characteristics of vehicle-freight coupling, and time-varying road environmental risks, resulting in limited accuracy and predictability of early warnings. Second, monitoring of driver status often focuses on identifying external physiological manifestations such as fatigue, failing to quantify subtle deviations between driving intentions and vehicle responses caused by cognitive load, emotional fluctuations, etc. Third, in freight transportation scenarios, especially for liquids, bulk or easily shifted goods, existing technologies generally ignore the impact of the dynamic coupling effect between goods and vehicles on driving stability. Finally, assessments of external environmental risks often rely on static road attributes or real-time weather information, making it difficult to effectively predict dynamic risks such as future traffic flow and severe weather evolution along the route. Therefore, how to achieve deep integration of multi-source heterogeneous information and construct a method capable of accurately, dynamically, and proactively assessing comprehensive transportation risks is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0004] To address the technical challenge of accurately, dynamically, and proactively assessing integrated transportation risks, this invention provides the following solution.

[0005] A real-time risk warning and assessment method for vehicle transportation based on the Internet of Things (IoT) includes the following steps: S1: Real-time acquisition of vehicle IMU data, CAN bus data from the controller area network, heart rate variability (HRV) data, cargo pressure distribution data, and external environment and road network data; S2: Based on the HRV data and the driver's micro-operation sequence extracted from the CAN bus data, a time-varying Kalman filter model is used to evaluate and quantify the driving state deviation, which characterizes the deviation between the driver's operating intention and the vehicle's dynamic feedback; S3: Inputting the vehicle IMU data and cargo pressure distribution data into a preset vehicle-cargo rigid-flexible coupling dynamics model to calculate in real time the risk characteristics of the vehicle in the current transportation process. S4: Using a spatiotemporal graph convolutional network, real-time traffic flow and meteorological data (dynamic attributes) and road geometry and historical accident data (static attributes) are fused to predict the risk potential field value of each road segment along the transportation route within a future time window; wherein the real-time traffic flow, meteorological data, road geometry, and historical accident data are derived from the external environment and road network data; S5: Using a predefined hierarchical Bayesian network, the driving state deviation, vehicle-cargo system instability index, and risk potential field value corresponding to the current vehicle position are used as evidence input to perform probabilistic inference and output the comprehensive transportation risk probability; when the comprehensive transportation risk probability exceeds a preset classification threshold, a risk warning of the corresponding level is triggered.

[0006] This invention constructs a multi-dimensional collaborative risk assessment system involving people, vehicles, goods, roads, and the environment, which can accurately, dynamically, and proactively assess comprehensive transportation risks.

[0007] Furthermore, the step of evaluating and quantifying the deviation of driving state includes: constructing a vehicle dynamics model using the driver's steering wheel angle as input, and correcting the parameters of the vehicle dynamics model in real time by combining the heart rate variability (HRV) data; using the vehicle yaw rate and centroid sideslip angle obtained from the vehicle IMU data as actual observed values; and using a time-varying Kalman filter model to calculate the Mahalanobis distance between the vehicle state predicted by the time-varying Kalman filter model and the actual observed state as the deviation of driving state.

[0008] This invention, by integrating heart rate variability (HRV) data with vehicle micro-operation data, accurately quantifies the degree of deviation between driving intention and vehicle feedback, achieving a deep insight into the driver's cognitive and operational state.

[0009] Furthermore, the vehicle-cargo rigid-flexible coupling dynamics model includes: establishing a 14-DOF vehicle rigid body dynamics model; equating the cargo system to a multi-flexible body element based on the absolute nodal coordinate method, and coupling it with the vehicle rigid body dynamics model at the suspension system; using the vehicle's three-axis acceleration and angular velocity as model excitation inputs, converting cargo pressure distribution data into normal forces on the cargo-cargo contact surface, calculating the real-time vertical loads of the vehicle's left and right tires through the vehicle rigid body dynamics model, and calculating the vehicle-cargo system instability index based on the lateral load transfer rate formula; the vehicle's three-axis acceleration and angular velocity are derived from the vehicle's IMU data.

[0010] By introducing a rigid-flexible coupling dynamic model of vehicle and cargo, and taking into account the pressure distribution of cargo, the instability index of the vehicle-cargo system can be accurately calculated.

[0011] Furthermore, the step of predicting the risk potential field value of each road segment along the transportation route within a future time window includes: dividing the transportation route into continuous road segments according to a predetermined length, using them as nodes, and constructing a graph structure based on the physical topological relationship between the road segments; constructing a feature vector for each node that includes real-time traffic flow, meteorological data, road geometric parameters, and historical accident frequencies; and processing the data using a spatiotemporal graph convolutional network that includes graph convolutional layers and recurrent neural network layers to output the risk potential field value of each road segment within the predetermined future time window.

[0012] By integrating real-time traffic flow, meteorological data, road geometry parameters, historical accident frequencies, and other data, the forecasting of risk potential fields for future periods is enhanced.

[0013] Furthermore, the step of using a predefined hierarchical Bayesian network for probabilistic reasoning includes: constructing a Bayesian network with driving state deviation, vehicle-cargo system instability index, and risk potential field value as parent nodes, and comprehensive transportation risk as child nodes; setting a threshold based on the statistical distribution characteristics of historical data to discretize the continuous value input of each parent node into multiple state levels; generating a conditional probability table based on historical statistical data, and reasoning on the multiple state levels after input discretization to obtain the probability of comprehensive transportation risk.

[0014] Furthermore, the step of triggering the corresponding risk warning includes: setting multiple risk levels corresponding to different probability ranges; when the probability of the comprehensive transportation risk falls into the range of a certain level, triggering the warning corresponding to that level, the warning mode and intensity of which increase with the increase of the risk level, and the mode includes one or more combinations of sound prompts, body-sensing warnings and sending alarm information to the background monitoring center.

[0015] By setting up a risk warning registration system, data can be easily categorized and warnings can be issued.

[0016] Furthermore, the vehicle IMU data includes data collected by a six-axis IMU sensor installed inside the vehicle.

[0017] Furthermore, CAN bus data includes at least steering wheel angle, accelerator pedal opening, brake pedal pressure, and vehicle speed.

[0018] Furthermore, the vehicle dynamics model is a baseline two-degree-of-freedom vehicle dynamics model.

[0019] Furthermore, the spatiotemporal graph convolutional network is a spatiotemporal graph convolutional network (STGCN) model that includes graph convolutional layers and one-dimensional convolutional or GRU layers.

[0020] In summary, the beneficial effects of this invention are as follows: Compared with existing technologies, this invention constructs a multi-dimensional collaborative risk assessment system involving a person, vehicle, cargo, road, and environment, significantly improving the accuracy and foresight of risk warnings. Regarding driver state assessment, this invention overcomes the limitations of traditional fatigue detection by integrating heart rate variability (HRV) data with vehicle micro-operation data, accurately quantifying the deviation between driving intention and vehicle feedback, achieving a deep understanding of the driver's cognitive and operational state. Regarding vehicle stability, it introduces a vehicle-cargo rigid-flexible coupling dynamic model for the first time, incorporating cargo state (such as cargo pressure distribution) into consideration, accurately calculating the vehicle-cargo system instability index, effectively filling the gap in existing technologies that ignore the impact of cargo on driving safety. Regarding external environmental risks, by integrating real-time traffic flow, meteorological data, road geometric parameters, and historical accident frequencies, it proactively predicts the risk potential field along the transportation route for future periods. Finally, through hierarchical Bayesian networks, it performs probabilistic reasoning on the aforementioned multi-source heterogeneous risk factors, outputting a comprehensive risk probability, making the early warning decision-making basis more sufficient and the results more reliable, thereby effectively ensuring the safety of the entire transportation process. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the concept of a real-time risk early warning and assessment method for vehicle transportation based on the Internet of Things according to an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram illustrating the reasoning of driving state deviation according to an embodiment of the present invention;

[0023] Figure 3 This is a flowchart illustrating an Internet of Things-based real-time risk warning and assessment method for vehicle transportation processes according to an embodiment of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] like Figure 3 As shown, a real-time risk early warning and assessment method based on the Internet of Things (IoT) in vehicle transportation includes the following steps:

[0026] S1: Real-time acquisition of vehicle IMU data, CAN bus data from the controller area network, heart rate variability (HRV) data, cargo pressure distribution data, and external environment and road network data, indicating the vehicle's operating status. The HRV data represents the driver's physiological state, and the cargo pressure distribution data represents cargo status sensing data.

[0027] Specifically, a six-axis IMU sensor installed inside the vehicle collects the vehicle's three-axis acceleration and three-axis angular velocity. The IMU stands for Onboard Inertial Measurement Unit. A CAN data acquisition card connects to the vehicle's OBD interface, reading and analyzing CAN bus data such as steering wheel angle, accelerator pedal opening, brake pedal pressure, and vehicle speed at a frequency of, for example, 100 Hz—this is the vehicle's operational data. The driver wears a smart bracelet with an integrated photoplethysmography (PPG) sensor to monitor and calculate the heart rate interval in real time, and further calculate heart rate variability (HRV), which refers to the minute differences in the time interval between consecutive heartbeats. A flexible thin-film pressure sensor array is installed at the bottom of the cargo compartment to obtain a real-time pressure distribution map of the cargo during vehicle movement. By calling APIs such as those from Gaode Maps and China Weather Network, real-time road conditions, traffic flow density, future weather forecasts, road curvature, slope, and historical accident data—this is the external environment and road network data—can be obtained for the vehicle's current location.

[0028] like Figure 1 As shown, based on the technical problem addressed by this invention—a comprehensive assessment model that deeply integrates driver's physiological and psychological state, vehicle-cargo coupling dynamic characteristics, and time-varying road environment risks—the main challenge facing this invention is how to integrate information such as "driver's physiological and psychological state, vehicle-cargo coupling dynamic characteristics, and time-varying road environment" into the assessment model. The concept of this invention is:

[0029] The "human characteristics" representing the relationship between people and vehicles are transformed into driving state deviation, which is used to characterize the deviation between the driver's driving intention and the vehicle's feedback. The so-called "human characteristics" not only reflect the characteristics of the person, but also include the interaction with the vehicle, thus reflecting the "characteristics of the driver's driving state", so as to better mine and utilize information for subsequent reasoning.

[0030] The "vehicle features" representing the relationship between vehicles and goods are transformed into a vehicle system instability index, which is used to extract the coupling relationship between vehicles and goods to the greatest extent possible, so as to facilitate subsequent reasoning.

[0031] The "environmental characteristics" representing the road environment are transformed into risk potential field values. These risk potential field values ​​can integrate various types of multimodal information, such as the correlation between roads and weather, as well as the correlation with historical information, thereby enabling better information mining and utilization for subsequent reasoning.

[0032] Therefore, the focus of this invention is on the specific implementation of the above three features, which will be described below through steps S2 to S4. It should be noted that S2, S3 and S4 are parallel steps, so the order of S2 to S4 does not mean that they must be performed in the order of S2 to S4.

[0033] S2: Based on the heart rate variability (HRV) data and the driver micro-operation sequence extracted from the CAN bus data, a time-varying Kalman filter model is used to evaluate and quantify the driving state deviation, which characterizes the deviation between the driver's operating intention and the vehicle's dynamic feedback.

[0034] Specifically, micro-operation features such as the rate of change of steering wheel angle and the frequency of switching between accelerator and brake pedals within a 2-second time window are extracted from the CAN bus data to form an observation vector. The driver's ideal control input and the vehicle's expected response state, such as yaw rate, are used as the system state vector. A time-varying Kalman filter model is constructed, where the state transition matrix is ​​established based on the vehicle's two-degree-of-freedom dynamics model, and the observation matrix is ​​associated with the system state and the driver's micro-operations. The real-time HRV index is used as an adjustment factor for the model noise covariance. For example, when the HRV index decreases, indicating an increase in the driver's cognitive load, the process noise covariance is increased. The model continuously predicts the micro-operation observations at the next moment, and the Mahalanobis distance between the predicted value and the actual observation value is used as the deviation of the driving state. The larger the distance, the greater the deviation between the driver's actual operation and their intention and the vehicle's expected feedback.

[0035] In an optional embodiment, the step of assessing and quantifying the deviation from the driving state includes:

[0036] The vehicle dynamics model is constructed using the driver's steering wheel angle as input, and the model parameters are corrected in real time using the heart rate variability (HRV) data. The vehicle yaw rate and center of mass sideslip angle obtained from the vehicle IMU data are used as actual observation values. A time-varying Kalman filter model is used to calculate the Mahalanobis distance between the vehicle state predicted by the time-varying Kalman filter model and the actual observed state as the driving state deviation.

[0037] Specifically, a baseline two-degree-of-freedom vehicle dynamics model (also known as a bicycle model or sideslip model) can be established. This model receives the steering wheel angle applied by the driver in a conscious state as input and predicts the theoretical yaw rate and sideslip angle of the vehicle. This prediction represents the dynamic response the vehicle should have under ideal driving conditions. Simultaneously, real-time heart rate variability (HRV) data collected by a wearable device is used to dynamically adjust the parameters of this vehicle dynamics model. For example, if the HRV value drops from the normal 50 milliseconds to 25 milliseconds, indicating that the driver may be fatigued or stressed, the driver reaction time parameter in the model will increase by 0.1 seconds accordingly, making the model prediction closer to the driver's current physiological state.

[0038] like Figure 2 As shown, the real-time Kalman prediction model, at its core, is the aforementioned vehicle dynamics model. It can take input driver micro-operations (such as steering wheel angle) and output vehicle yaw rate and sideslip angle. The real-time Kalman prediction model can further fine-tune the parameters of the vehicle dynamics model based on real-time heart rate variability (HRV) data. These parameters include the equivalent tire sideslip stiffness of the front axle and the equivalent tire sideslip stiffness of the rear axle. The principle is that real-time HRV data reflects the driver's state. If the driver's state is poor, leading to slowed reaction time or inadequate operation, this can be equivalent to changes in the equivalent tire sideslip stiffness of the front axle and the rear axle. Finally, the real-time Kalman prediction model outputs the vehicle yaw rate and sideslip angle to reflect the driver's current physiological state.

[0039] Specifically, to quantify the degree of deviation, the predicted yaw rate and sideslip angle output by the vehicle dynamics model after HRV correction are used as the predicted state. The yaw rate measured in real time by the onboard inertial measurement unit (IMU) and the estimated sideslip angle are used as the actual observed state. Finally, the deviation of the driving state is obtained by calculating the Mahalanobis distance between these two state vectors. Mahalanobis distance comprehensively considers the variance and covariance of each state variable and is a dimensionless statistical distance (a metric used to measure the similarity between two data points; it considers the correlation and scale differences between features in the dataset and is more statistically reasonable than Euclidean distance). For example, a Mahalanobis distance value greater than a preset threshold, such as 5, can clearly indicate that the driver's handling behavior has significantly deviated from the normal range.

[0040] S3: Input the vehicle IMU data and cargo pressure distribution data into the preset vehicle-cargo rigid-flexible coupling dynamic model, and calculate the vehicle-cargo system instability index in real time, which characterizes the vehicle under the current motion disturbance.

[0041] Specifically, a multi-body dynamics model of the trailer or truck is pre-established in ADAMS software, including rigid components such as the frame, suspension, and tires. Using the finite element method, a flexible body dynamics model is established according to the type of cargo, such as liquid or bulk cargo. Real-time collected cargo pressure distribution data is used as load boundary conditions and applied to the flexible body dynamics model to simulate the real-time changes in cargo shape and center of gravity. Real-time vehicle acceleration and angular velocity collected by the IMU are used as motion excitations for the entire coupled model. The dynamic equations of the rigid-flexible coupled system are solved through co-simulation, and the lateral load transfer rate (LTR) of the vehicle is calculated in real time. This value is the instability index of the vehicle-cargo system. The closer the value is to 1, the higher the risk of vehicle rollover.

[0042] In an optional embodiment, the vehicle-cargo rigid-flexible coupling dynamics model includes:

[0043] Establish a 14-DOF rigid body dynamics model of the vehicle (14 DDOF typically includes: 6 DDOFs of the vehicle chassis: translation (longitudinal, lateral, vertical) and rotation (roll, pitch, yaw) about the X, Y, and Z axes; 1 vertical DDOF for each of the 4 wheels: describing the vertical runout of each wheel relative to the vehicle body; and 1 rotational DDOF for each of the 4 wheels: describing the rotation of each wheel).

[0044] The cargo system is represented as an equivalent of multiple flexible body elements (ANCF elements) based on the absolute nodal coordinate method, and coupled with the vehicle rigid body model at the suspension system. This is the key innovation, introducing consideration of cargo deformation and simulating the interaction between the cargo and road surface excitations through the vehicle's suspension system during actual vehicle operation. Representing the cargo system as multiple flexible body elements means that the cargo is no longer considered a simple rigid load, but a deformable entity. The absolute nodal coordinate method is an advanced flexible body dynamics modeling method, particularly suitable for handling large deformations and complex contact problems. The coordinates of the nodes in the ANCF elements are absolute during motion, effectively avoiding the limitations of traditional finite element methods when dealing with large rotations and simplifying the solution of dynamic equations; it also indicates that the cargo system can be decomposed into multiple ANCF elements to simulate its complex flexible behavior, rather than a single flexible body. Coupling with the vehicle rigid body model at the suspension system means that the motion and deformation of the cargo are transmitted to the vehicle through the suspension system, and vice versa.

[0045] The vehicle's triaxial acceleration and angular velocity measured by the IMU are used as excitation inputs to the vehicle-cargo rigid-flexible coupling dynamic model. The cargo pressure distribution data is converted into the normal force on the contact surface between the cargo and the vehicle body. The real-time vertical loads of the left and right tires of the vehicle are calculated through the model, and the instability index of the vehicle-cargo system is calculated based on the lateral load transfer rate formula.

[0046] Specifically, this vehicle-cargo rigid-flexible coupling dynamics model abstracts the entire vehicle as a 14-DOF rigid body system, encompassing longitudinal, lateral, and vertical motions of the vehicle body, as well as roll, pitch, and sway motions, and independent rotation and vertical hop of the four wheels. This accurately describes the macroscopic attitude changes of the vehicle during operation. For cargo within the cargo compartment, especially non-solid substances such as liquids or bulk particles, this method does not treat them as rigid bodies but discretizes them into multiple interconnected flexible elements using absolute nodal coordinates. This method accurately simulates the swaying, accumulation, and deformation of cargo during vehicle turning or acceleration / deceleration, and transmits the additional forces and torques generated by the cargo's swaying to the overall vehicle rigid body model through the suspension system, achieving dynamic mechanical coupling between the vehicle and cargo.

[0047] Specifically, the vehicle-cargo rigid-flexible coupling dynamics model requires real-time data input. Vehicle IMU data, including triaxial acceleration and angular velocity data (e.g., lateral acceleration reaching 4 m / s²), is used as the dynamic excitation for the entire model. Simultaneously, pressure distribution data of the cargo, collected by pressure sensor arrays mounted on the cargo floor and side walls (e.g., significantly higher pressure on the left side than the right), is converted into normal forces acting on the contact surfaces and used as boundary conditions for the cargo flexible body system. The model performs numerical solutions at each time step, outputting key vehicle state parameters, especially the real-time vertical loads on the left and right tires. For example, in an emergency avoidance maneuver, the model calculates a vertical load of 8000 N on the left tire and 18000 N on the right. Subsequently, the instability index calculated using the lateral load transfer rate formula reaches 0.8, indicating that the vehicle is approaching a critical rollover state.

[0048] S4: Utilizing a spatiotemporal graph convolutional network, real-time traffic flow and meteorological data (dynamic attributes) are fused with road geometry and historical accident data (static attributes) to predict the risk potential field value of each road segment along the transportation route within a future time window. The real-time traffic flow, meteorological data, road geometry, and historical accident data originate from the external environment and road network data. Real-time traffic flow refers to the actual operating status of vehicles on the road currently or in the recent period, including flow rate, speed, density, and occupancy. Risk potential field value is a concept in the fields of intelligent driving, environmental perception, and risk assessment, used to quantify and spatially represent potential hazards and uncertainties in the environment.

[0049] Specifically, the planned transportation route is first divided into continuous road segments of 500 meters in length, and these segments are constructed as a topological graph, with road segments as nodes and the connections between road segments as edges. The feature vector of each node includes static attributes such as road curvature, slope, number of lanes, and historical accident frequency, as well as dynamic attributes such as real-time traffic congestion index, average vehicle speed, rainfall, and visibility. The node feature sequence every five minutes over the past hour is used as input and fed into a spatiotemporal graph convolutional network (STGCN) model containing graph convolutional layers and one-dimensional convolutional or GRU layers. The STGCN model aggregates the spatial risk information of adjacent road segments through graph convolution and learns the law of risk evolution over time through temporal convolutional layers. Finally, the STGCN model outputs the risk prediction value of each road segment node in the next 15 minutes. The prediction value sequence of all road segments together constitutes the risk potential field along the transportation route.

[0050] In an optional embodiment, the step of predicting the risk potential field value of each segment along the transportation route within a future time window includes:

[0051] The transportation route is divided into continuous segments according to a predetermined length, which serve as nodes in a graph network. A graph structure is constructed based on the physical topological relationships between the segments. For each node, a feature vector containing real-time traffic flow, meteorological data, road geometric parameters, and historical accident frequencies is constructed. The constructed graph data (including the graph structure and node feature vectors) is input into a spatiotemporal graph convolutional network containing graph convolutional layers and recurrent neural network layers, and the risk potential field value of each segment within a predetermined future time window is output.

[0052] Specifically, a transportation route spanning hundreds of kilometers is first divided into segments, for example, every 500 meters, with each segment becoming a node in a graph network. Then, based on the geographical connections between these segments, links are established between nodes, forming a chain-like or more complex graph structure. For each segment node, the system aggregates multiple data sources to create a multi-dimensional feature vector. For example, for a segment of the Shanghai-Nanjing Expressway, its feature vector at 10:00 AM might include real-time traffic flow of 1500 vehicles per hour, light rain, a road curvature radius of 800 meters, and an average monthly accident rate of 0.5 incidents over the past three years.

[0053] Specifically, the constructed spatiotemporal graph data is input into a deep learning model capable of processing both spatial and temporal features. For example, in existing technologies, the core of a deep learning model is a spatiotemporal graph convolutional network, which comprises two key components: graph convolutional layers are responsible for capturing spatial interactions; for instance, traffic congestion on upstream road segments increases the risk on downstream road segments, and graph convolutional operations can learn and quantify this spatial dependency. Recurrent neural network layers, such as long short-term memory networks, are responsible for processing temporal patterns; for example, they can learn the periodic fluctuations in risk values ​​during morning and evening rush hours. After being trained on a large amount of historical data, the deep learning model can predict the risk potential value of each road segment node within a preset time window, such as the next 30 minutes, outputting a value between 0 and 10, where a higher value indicates a greater overall risk for that road segment in the future time period.

[0054] In the above embodiments, the calculation of the risk potential field incorporates multiple factors; in other embodiments, it can also be implemented in a simpler way, such as by removing some factors.

[0055] S5: Using a predefined hierarchical Bayesian network, the driving state deviation, the vehicle-cargo system instability index, and the risk potential field value corresponding to the vehicle's current position are used as evidence inputs to perform probabilistic reasoning and output the comprehensive transportation risk probability; when the comprehensive transportation risk probability exceeds the preset classification threshold, the corresponding level of risk warning is triggered.

[0056] Specifically, a three-layer Bayesian network is constructed. The top-level root node represents comprehensive transportation risk, with its state categorized into low, medium, and high levels. The second layer consists of parent nodes, including driver risk, vehicle-cargo system risk, and road environment risk. The third layer comprises leaf nodes, corresponding to three continuous input pieces of evidence: driving state deviation, vehicle-cargo system instability index, and risk potential field value. Based on historical data and expert knowledge, a conditional probability table is pre-defined between each node, such as P(driver risk|driving state deviation). During vehicle operation, the three indicator values ​​calculated by S2, S3, and S4 are input as evidence to the corresponding leaf nodes. A confidence propagation algorithm is used for probabilistic inference to calculate the posterior probability that the comprehensive transportation risk at the root node is at a high level. A first-level warning threshold of 0.6 and a second-level warning threshold of 0.8 are set. When the calculated comprehensive transportation risk probability is 0.7, the first-level threshold is exceeded, and the system issues a second-level risk warning to the driver via a pop-up window on the in-vehicle screen and a voice broadcast.

[0057] In an optional embodiment, the step of performing probabilistic reasoning using a predefined hierarchical Bayesian network includes:

[0058] A Bayesian network is constructed with driving state deviation, vehicle-cargo system instability index, and risk potential field value as parent nodes and comprehensive transportation risk as child nodes. Based on the statistical distribution characteristics of historical data, a threshold is set to discretize the continuous values ​​of each parent node into multiple state levels. A conditional probability table is generated based on historical statistical data, and the discretized evidence is input for reasoning to obtain the comprehensive transportation risk probability.

[0059] Specifically, this method establishes an intuitive probabilistic graphical model. The network structure clearly defines three independent risk sources as parent nodes: driving state deviation, vehicle-cargo system instability index, and risk potential field value. These three parent nodes all point to a single child node, namely, comprehensive transportation risk. Before inference, the continuous numerical inputs obtained from the preceding steps need to be discretized. For example, the Mahalanobis distance value of the driving state deviation can be divided into three levels: normal, slight deviation, and severe deviation; the vehicle-cargo system instability index can be divided into three levels: stable, critical, and unstable; and the risk potential field value can also be divided into three levels: low, medium, and high.

[0060] Specifically, the core of a hierarchical Bayesian network lies in its conditional probability table, which is derived through statistical analysis of massive amounts of historical transportation data. It stores the probability of child nodes taking different states under all combinations of parent node states. For example, the conditional probability table might contain a record stating that when the driving state is severely deviated, the vehicle-cargo system is critical, and the environmental risk is high, the probability of the overall transportation risk being a dangerous state is 95%. When the system receives real-time evidence, such as a severely deviated driving state, a stable vehicle-cargo system, and a moderate environmental risk, the Bayesian network uses probabilistic inference algorithms, combined with the entire conditional probability table, to calculate the posterior probability distribution of the overall transportation risk at each level, such as safe, warning, or dangerous. The final output might be a probability vector, such as a 70% probability of dangerous.

[0061] In an optional embodiment, the step of triggering the corresponding level of risk warning includes:

[0062] Multiple risk levels are set up to correspond to different probability ranges; when the probability of the comprehensive transportation risk falls into the range of a certain level, an early warning corresponding to that level is triggered. The mode and intensity of the early warning increase with the increase of the risk level. The mode includes one or more combinations of sound prompts, body-sensing warnings, and sending alarm information to the background monitoring center.

[0063] Specifically, the system predefines a tiered response mechanism. For example, the probability of hazardous states in integrated transportation is divided into three intervals to correspond to different risk levels. When the calculated probability of hazardous state is below 10%, it is defined as Level 1 risk or a safe state, and no proactive warning is triggered. When the probability of hazardous state is between 10% and 50%, it is defined as Level 2 risk or a warning state. When the probability of hazardous state exceeds 50%, it is defined as Level 3 risk or a hazardous state. These probability thresholds can be flexibly configured according to the importance of the transportation task, the value of the goods, or regulatory requirements.

[0064] Specifically, once the probability of overall transportation risk falls within a preset range, the system will automatically trigger the corresponding level of warning. For Level 2 risks, the system may issue a milder alert, such as a yellow warning icon illuminating the dashboard accompanied by a brief beep, to remind the driver. For the highest Level 3 risks, the warning measures will be significantly enhanced to ensure the driver can respond immediately. At this time, the system may issue a continuous and sharp alarm sound, while simultaneously activating the vibration function of the driver's seat or steering wheel to generate a strong tactile warning. While issuing the alarm to the driver, the system will also immediately and automatically send alarm information, including vehicle location, speed, risk level, and specific cause, to the company's back-end monitoring center via wireless network, enabling management personnel to remotely monitor and provide necessary intervention support.

[0065] Those skilled in the art will conceive of many modifications, alterations, and alternatives without departing from the spirit and essence of this invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for real-time risk early warning and assessment in vehicle transportation based on the Internet of Things, characterized in that, Includes the following steps: S1: Real-time acquisition of vehicle IMU data, CAN bus data of the controller area network, heart rate variability (HRV) data, cargo pressure distribution data, and external environment and road network data; S2: Based on heart rate variability (HRV) data and driver micro-operation sequences extracted from CAN bus data, a time-varying Kalman filter model is used to evaluate and quantify the driving state deviation, which characterizes the deviation between the driver's operating intention and the vehicle's dynamic feedback. The steps include: using the driver's steering wheel angle as input to construct a vehicle dynamics model, and combining heart rate variability (HRV) data to correct the vehicle dynamics model parameters in real time; using the vehicle yaw rate and center of mass sideslip angle obtained from the vehicle IMU data as actual observation values; and using a time-varying Kalman filter model to calculate the Mahalanobis distance between the vehicle state predicted by the time-varying Kalman filter model and the actual observed state as the driving state deviation. S3: Input the vehicle IMU data and cargo pressure distribution data into the preset vehicle-cargo rigid-flexible coupling dynamic model, and calculate the vehicle-cargo system instability index in real time, which characterizes the vehicle under the current motion disturbance. The vehicle-cargo rigid-flexible coupling dynamics model includes: establishing a 14-DOF vehicle rigid body dynamics model; equating the cargo system to a multi-flexible body element based on the absolute nodal coordinate method and coupling it with the vehicle rigid body dynamics model at the suspension system; using the vehicle's three-axis acceleration and angular velocity as model excitation inputs; converting cargo pressure distribution data into normal forces on the cargo-cargo contact surface; calculating the real-time vertical loads of the left and right tires of the vehicle through the vehicle rigid body dynamics model; and calculating the vehicle-cargo system instability index based on the lateral load transfer rate formula; the vehicle's three-axis acceleration and angular velocity are derived from vehicle IMU data. S4: Using a spatiotemporal graph convolutional network, real-time traffic flow and meteorological data, which are dynamic attributes, are fused with road geometry and historical accident data, which are static attributes, to predict the risk potential field value of each road segment along the transportation route in the future time window; wherein the real-time traffic flow, meteorological data, road geometry, and historical accident data are derived from the external environment and road network data. S5: Using a predefined hierarchical Bayesian network, the driving state deviation, the vehicle-cargo system instability index, and the risk potential field value corresponding to the vehicle's current position are used as evidence inputs to perform probabilistic reasoning and output the comprehensive transportation risk probability; when the comprehensive transportation risk probability exceeds the preset classification threshold, the corresponding level of risk warning is triggered.

2. The method according to claim 1, characterized in that, The steps for predicting the risk potential field value of each segment along the transportation route within a future time window include: The transportation route is divided into continuous segments according to a predetermined length, which are used as nodes. A graph structure is constructed based on the physical topological relationship between the segments. For each node, a feature vector containing real-time traffic flow, meteorological data, road geometric parameters, and historical accident frequency is constructed. The spatiotemporal graph convolutional network containing graph convolutional layers and recurrent neural network layers is used for processing to output the risk potential field value of each segment within a predetermined future time window.

3. The method according to claim 1, characterized in that, The steps for probabilistic reasoning using a predefined hierarchical Bayesian network include: A Bayesian network is constructed with driving state deviation, vehicle-cargo system instability index, and risk potential field value as parent nodes and comprehensive transportation risk as child nodes. Based on the statistical distribution characteristics of historical data, a threshold is set to discretize the continuous value input of each parent node into multiple state levels. A conditional probability table is generated based on historical statistical data, and reasoning is performed on the multiple state levels after input discretization to obtain the comprehensive transportation risk probability.

4. The method according to claim 1, characterized in that, The steps for triggering the corresponding level of risk warning include: Multiple risk levels are set up to correspond to different probability ranges; when the probability of the comprehensive transportation risk falls into the range of a certain level, an early warning corresponding to that level is triggered. The mode and intensity of the early warning increase with the increase of the risk level. The mode includes one or more combinations of sound prompts, body-sensing warnings, and sending alarm information to the background monitoring center.

5. The method according to claim 1, characterized in that, Vehicle IMU data includes data collected by a six-axis IMU sensor installed inside the vehicle.

6. The method according to claim 1, characterized in that, CAN bus data includes at least steering wheel angle, accelerator pedal opening, brake pedal pressure, and vehicle speed.

7. The method according to claim 1, characterized in that, The vehicle dynamics model is a baseline two-degree-of-freedom vehicle dynamics model.

8. The method according to claim 2, characterized in that, The spatiotemporal graph convolutional network is the STGCN model, which includes graph convolutional layers and one-dimensional convolutional or GRU layers.

Images