Intelligent supervision system for road transport key operating vehicles

Abstract

The application relates to the technical field of intelligent transportation, and discloses a road transportation key operating vehicle intelligent supervision system, which comprises the following modules: a data acquisition and fusion module, which is used for collecting multi-source heterogeneous data such as vehicle terminals and external environments in real time and performing standardized processing; a cognitive digital twin construction module, which is used for constructing a dynamic heterogeneous graph to represent a man-vehicle-road-environment system and outputting a cognitive state vector through a graph neural network; a risk prediction and evaluation module, which is used for predicting a risk evolution trend based on a state vector sequence and quantifying the uncertainty of the prediction by using a Monte Carlo discard method; and an adaptive intervention decision module, which is used for combining the risk prediction and the uncertainty, deciding and executing optimal active intervention instructions through a reinforcement learning model. By constructing a cognitive digital twin, introducing uncertainty quantification and closed-loop self-checking, the application realizes forward-looking prediction and adaptive intervention of driving risks, and improves the accuracy and robustness of supervision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation, in particular to a road transport key operating vehicle intelligent supervision system. BACKGROUND

[0002] The road transport industry is the lifeblood of the national economy, among which the key operating vehicles represented by highway passenger transport vehicles, tourist passenger transport vehicles and dangerous goods transport vehicles are directly related to people's life and property safety and social public safety. Due to the long running time, high driving intensity and high potential risk of such vehicles, how to use advanced technical means to effectively supervise them, identify and intervene driving risks in advance, has become a key problem to be solved in the field of traffic safety.

[0003] To address this challenge, the prior art usually adopts the way of deploying vehicle-mounted sensing devices on the vehicle terminal for real-time monitoring. These devices mainly include: a driver state monitoring (DSM) system for monitoring the state of the driver such as fatigue and distraction, an advanced driving assistance system (ADAS) for sensing the surrounding environment of the vehicle and providing collision warning, lane departure warning and other functions, and a global positioning system (GPS) for recording the trajectory and speed of the vehicle. These systems identify and immediately alarm the dangerous behavior of the driver by collecting various types of data and based on pre-set rules or thresholds, which constitutes the basic technical framework of current road transport safety supervision.

[0004] Although the prior art has achieved immediate response to a single dangerous event to some extent by deploying various sensors and alarm systems, there are still some deficiencies:

[0005] The risk assessment model of the prior art is static and fragmented in nature. This is because its technical core relies on independent, rule-based logical judgments, for example, a fatigue alarm is triggered only when the driver's eyes are closed for more than a certain pre-set number of seconds. This way treats the driver, vehicle, road, environment and other multiple core elements that influence each other as independent analysis objects, and cannot model the complex and dynamic coupling relationship between them. Therefore, it cannot distinguish between "mild fatigue on an open straight road at high speed" and "mild fatigue on a rainy winding mountain road", which are two completely different risk levels, resulting in a lack of depth and situational awareness in its risk assessment.

[0006] In addition, the prior art generally lacks the ability to predict the evolution trend of risks and the ability to evaluate the confidence of its own judgment. The goal of its system design is to intervene when the risk behavior "occurs", rather than to predict before the risk state "forms", which makes it essentially a passive and lagging response mechanism. More importantly, since it uses deterministic rules, the output alarm signal is binary (alarm or no alarm), and the system cannot quantify the reliability of its judgment. When faced with ambiguous or marginal scenarios, the system either misses the report due to the inability to meet the strict triggering rules, or produces a large number of false alarms due to excessive sensitivity, reducing the driver's trust in the system.

[0007] Furthermore, the prior art solution is usually a set of fixed, open-loop system. Its model parameters and intervention logic are difficult to change after deployment, and cannot adapt to individual differences of drivers, changes in vehicle state or unexpected driving scenarios. When the system model faces new situations beyond its knowledge range, causing its prediction results to deviate from the real-world state, it lacks an effective mechanism to perceive this deviation and use new information for self-correction and online learning. This "one-time deployment, no growth" feature limits its long-term effectiveness and robustness in complex and changing real-world environments.

[0008] Finally, the analysis scope of the prior art is limited to a single vehicle itself, ignoring regional and systematic risks caused by external factors. Each vehicle's monitoring system is an independent information silo, and it cannot perceive the group behavior patterns of other vehicles in the surrounding area. Therefore, for a "risk field" formed in a specific road section due to factors such as bad weather, icy road surface, sudden accidents, etc., which may lead to multiple consecutive dangerous situations, the prior art cannot effectively identify and warn, and cannot provide proactive and coordinated intervention for vehicles entering the area. SUMMARY

[0009] In view of the shortcomings of the prior art, the present application provides an intelligent monitoring system for key operating vehicles in road transport, which solves the problem of static fragmentation, lack of predictability and adaptability of the risk evaluation model of the vehicle safety monitoring system in the prior art, leading to incomplete, non-prospective and low reliability of risk identification.

[0010] To achieve the above purpose, the present application is realized by the following technical solutions:

[0011] The present application provides an intelligent monitoring system for key operating vehicles in road transport, which includes a data acquisition and fusion module, a cognitive digital twin construction module, a risk prediction and evaluation module, and an adaptive intervention decision module.

[0012] The data acquisition and fusion module is used to acquire multimodal heterogeneous data from terminal devices deployed in the vehicle and external data sources. Specifically, the module collects driver status monitoring (DSM) data, advanced driver assistance (ADAS) data, vehicle bus (CAN) data and positioning (GPS) data through the vehicle terminal; at the same time, it accesses geographic information system (GIS) data and environmental weather (ENV) data through the cloud platform. The module aligns the timestamps of all collected data, cleans and standardizes them, and generates a real-time fusion of the collected data at each moment. Output a standardized multidimensional data set .

[0013] The cognitive digital twin construction module is connected to the output end of the data acquisition and fusion module to convert discrete data sets into Converted into a cognitive state vector that can represent the internal operating state of the system. The specific implementation of this module includes:

[0014] First, at every moment , based on the data set Building a dynamic heterogeneous graph , where the node set Represents entities such as drivers, vehicles, roads and their states, edge sets Represents the pre-set physical or logical relationship between these entities.

[0015] Secondly, a graph neural network (GNN) is used to learn the dynamic heterogeneous graph. In the graph neural network, any node In the The hidden state vector of the layer Updates are made using the following formula to aggregate neighbor node information and update their own status:

[0016] ;

[0017] Where, is the node in the graph neural network In the The hidden state vector of the layer; For nodes In the The hidden state vector of the layer; For nodes The set of neighbor nodes of is an aggregate function; It is an update function used to combine the node's own information and neighbor information.

[0018] Finally, a readout function is used to integrate the final state representations of all nodes in the graph to generate the moment The cognitive state vector , this vector is the core state of the cognitive digital twin.

[0019] The risk prediction and assessment module is connected to the output of the cognitive digital twin construction module and is used to deduce future risk trends based on historical state vector sequences. The specific implementation of this module includes:

[0020] First, the historical time window The sequence of cognitive state vectors within The input is fed into a long short-term memory (LSTM) model, which learns the pattern of state evolution over time and predicts the future time-step state vector sequence , this prediction sequence constitutes the risk evolution trend.

[0021] Secondly, in order to evaluate the reliability of the prediction results, the Monte Carlo discard method is used to random forward propagation, by computing The variance of the prediction results is used to obtain a quantitative prediction uncertainty score .

[0022] The adaptive intervention decision module is connected to the output end of the risk prediction and assessment module, and is used to generate and issue intervention instructions. In a preferred embodiment, the module is implemented using a reinforcement learning model. The composite state vector input of the model includes at least the risk evolution trend and the prediction uncertainty score. The action space of the model includes multiple graded intervention actions with different intervention intensities. The model is optimized by a reward function, which is designed to impose penalties on actions that cause the driver to manually cancel the intervention, while rewarding actions that effectively reduce risks, so that the system can generate an intervention strategy that takes into account both safety and driver acceptance.

[0023] In another preferred embodiment, the system also includes a regional risk field perception module. This module is set up in parallel with the risk prediction and assessment module and is connected to the adaptive intervention decision module. The specific implementation of this module includes:

[0024] First, based on the real-time geographic location of the vehicles, the physically adjacent vehicles are clustered into the same area.

[0025] Secondly, extract a key risk scalar for each vehicle in the area , calculate the mean of the risk scalar for all vehicles in the area and variance .

[0026] Finally, the existence of a regional risk field is determined by the following conditions:

[0027] ;

[0028] wherein, is a mean threshold value; is a variance threshold value; is a mean value of the critical risk scalar of all vehicles located in the region at time ; is a logical AND operator; is a variance of the critical risk scalar of all vehicles located in the region at time .

[0029] The determination result of the regional risk field is delivered to the adaptive intervention decision module. When the vehicle is in the risk field, the module raises the intervention level; when the vehicle is about to enter the risk field, the module performs a forward-looking warning.

[0030] In another preferred embodiment, the system further comprises a closed-loop model self-checking module. The module is connected with the risk prediction and evaluation module and the adaptive intervention decision module. The specific implementation of the module includes:

[0031] When the prediction uncertainty score exceeds the preset uncertainty threshold value, the module triggers and initiates a micro-interaction task to the driver, and collects the feedback data of the driver.

[0032] Then, the state at the time of triggering the task and the feedback data constitute a real-time labeling sample, and the upstream model is calibrated online through a total loss function including a detection loss term:

[0033] ;

[0034] wherein, is a total loss; is an original loss; is a detection loss calculated based on the model output under the current state and the feedback data ; is a hyperparameter for balancing the original loss and the detection loss; is the output of the model under the current state when the micro-interaction task is triggered; is the feedback data of the driver collected for the micro-interaction task.

[0035] The present application provides a road transport key operating vehicle intelligent supervision system. It has the following beneficial effects:

[0036] 1、The cognitive digital twin construction module can integrate and construct real-time data of multiple sources and heterogeneous into a cognitive state vector that can represent the comprehensive state of "person-vehicle-road-environment", and then the risk prediction and evaluation module uses a time series model to predict the future evolution trend of the vector sequence. This way changes the traditional technology which relies on threshold judgment of isolated and lagging dangerous driving behavior, and changes to the future development of the comprehensive risk state, so that intervention can be performed before the dangerous state is fully formed, improving the effectiveness and initiative of safety management.

[0037] 2、The regional risk field perception module can analyze the vehicle cluster based on geographic location, and identify group and systematic risks caused by adverse weather, sudden congestion and other external environmental factors by calculating the mean and variance of the risk states of multiple cognitive digital twins in the region. Therefore, the regulatory capacity of the system is expanded from isolated individual drivers to regional vehicle clusters, and the system can perform proactive warning for vehicles entering the risk area, effectively dealing with systematic risks that traditional technology cannot handle.

[0038] 3、The closed-loop model self-checking module can actively initiate micro-interaction tasks to obtain real state feedback from the driver when the prediction uncertainty score output by the risk prediction and evaluation module is high. The feedback is used as high-value real-time labeled samples to calibrate the upstream cognitive digital twin construction module and risk prediction and evaluation module online. This mechanism solves the problem of performance degradation of static models in complex and variable scenarios, enabling the system to continuously optimize itself in actual operation, thereby continuously improving the accuracy of its predictions and the effectiveness of its interventions. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a schematic diagram of the system structure of the present application;

[0040] Figure 2 is a schematic diagram of the workflow of the present application. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0042] Please refer to the drawings in the specification of the present application Figure 1 , Figure 1Fig. 1 is a schematic diagram of a system structure of an intelligent supervision system for road transport key operating vehicles according to an embodiment of the present application. The present application provides an intelligent supervision system for road transport key operating vehicles, which can include a data acquisition and fusion module 100, a cognitive digital twin construction module 200, a risk prediction and evaluation module 300, and an adaptive intervention decision module 400.

[0043] The data acquisition and fusion module 100 is used to obtain multi-dimensional raw data from vehicle terminals and external data sources, and to standardize the data. The output of the module is a data set containing driver state, vehicle dynamics, road environment, and other multi-dimensional information, which is time-synchronized and normalized. The data set is the input data for the subsequent modules.

[0044] The cognitive digital twin construction module 200 is connected to the output of the data acquisition and fusion module 100. The module receives the standardized multi-dimensional data set and constructs a dynamic heterogeneous graph at each time based on the data set. Then, the module uses a graph neural network to learn the dynamic heterogeneous graph and updates the feature representation of the nodes in the graph through iterative message passing.

[0045] Finally, the module outputs a cognitive state vector that can comprehensively represent the current system state.

[0046] The risk prediction and evaluation module 300 is connected to the output of the cognitive digital twin construction module 200. The module receives the real-time updated cognitive state vector sequence and uses a time series prediction model to predict the future evolution trend of the sequence. At the same time, the module quantitatively evaluates the reliability of the prediction results and generates a prediction uncertainty score. The output of the module is the risk evolution trend and the corresponding prediction uncertainty score.

[0047] The adaptive intervention decision module 400 is connected to the output of the risk prediction and evaluation module 300. The module receives the risk evolution trend and the prediction uncertainty score, and based on this information, generates the optimal active intervention instruction through a preset decision model (such as a reinforcement learning model). The output of the module is a specific intervention instruction, which is sent to the vehicle terminal for execution.

[0048] In a preferred embodiment, the system can also include a regional risk field perception module 500. The module is connected to the cognitive digital twin construction module 200 to obtain the cognitive states of multiple vehicles, and is connected to the adaptive intervention decision module 400 to deliver its analysis results. The module analyzes the coordination consistency of the risk states of multiple vehicles in a specific geographic area to determine whether there is a regional risk field dominated by external environmental factors.

[0049] In another preferred embodiment, the system can further comprise a closed-loop model self-checking module 600. The input of this module is connected to the risk prediction and assessment module 300 to obtain the prediction uncertainty score; its output is connected to the cognitive digital twin construction module 200 and the risk prediction and assessment module 300 to perform model calibration. When the prediction uncertainty score exceeds the preset threshold, the module initiates a micro-interaction task to the driver and collects feedback data, and then uses the feedback data to perform online calibration of the upstream model, and the calibration process is defined by the total loss function.

[0050] The overall data flow of the system is as follows: the data collection and fusion module 100 serves as the data source, and the data set output by it flows through the cognitive digital twin construction module 200 and the risk prediction and assessment module 300 in turn for state representation and risk prediction. The adaptive intervention decision module 400 makes the final decision based on the output of the risk prediction and assessment module 300 and the output of the regional risk field perception module 500 (if configured). The closed-loop model self-checking module 600 (if configured) forms a feedback calibration loop from the risk prediction and assessment module 300 to the cognitive digital twin construction module 200 and the risk prediction and assessment module 300 itself under certain conditions.

[0051] The data collection and fusion module 100 is the basic data input unit of the system of the present application. The function of this module is to obtain multi-modal heterogeneous data from multiple sensors deployed on the vehicle terminal and external cloud services in real time and in parallel, and to uniformly preprocess these data of different sources and different formats, finally generating a structured and standardized multi-dimensional data set for subsequent analysis and processing by the modules.

[0052] In a specific embodiment, the data collection and fusion module 100 can be further functionally divided into a vehicle-mounted multi-source data collection unit 110, an external information access unit 120, and a data standardization and synchronization unit 130.

[0053] The vehicle-mounted multi-source data collection unit 110 is deployed in the terminal hardware of the key operating vehicles. This unit integrates the data output interfaces of multiple vehicle-mounted sensors to collect the following four types of data:

[0054] The first type is driver state monitoring (DSM) data, such as eye closure, gaze direction vector, head posture deflection angle (pitch angle, yaw angle, roll angle), yawning frequency, and recognition signs of specific behaviors such as making a phone call or smoking, obtained through camera sensors.

[0055] The second type is advanced driver assistance system (ADAS) data, such as forward collision time (TTC), headway (HMW), lane departure warning (LDW) signals, etc. obtained through millimeter wave radar or cameras.

[0056] The third type is vehicle controller area network (CAN) bus data, such as steering wheel angle and angular velocity, throttle pedal opening, brake pedal depth, current vehicle speed, engine speed, gear information, etc.

[0057] The fourth type is global positioning system (GPS) data, such as high-precision longitude, latitude, altitude, real-time speed, heading angle, etc. spatio-temporal information.

[0058] The external information access unit 120 is deployed in the cloud server. Through the network interface protocol, it obtains external information related to the current location or driving path of the vehicle from third-party service providers. These information specifically includes:

[0059] One type is geographic information system (GIS) data, such as road curvature, road slope, speed limit value, road grade, and whether there is a tunnel or bridge, etc. static road attribute information of the road segment where the vehicle is currently located.

[0060] Another type is environmental meteorological (ENV) data, such as real-time or short-term predicted weather conditions (sunny, rainy, snowy, foggy), temperature, wind speed, and visibility, etc. dynamic environmental information obtained according to the vehicle location.

[0061] The data standardization and synchronization unit 130 receives data streams from the vehicle-mounted multi-source data acquisition unit 110 and the external information access unit 120. The unit performs the following preprocessing procedures:

[0062] First, time stamp alignment. The unit takes a unified, high-precision time source as the reference to assign or correct the time stamp of all received data points, ensuring the comparability of data from different sources in the time dimension.

[0063] Second, data cleaning. The unit checks the data stream, and for partially missing data, uses methods such as linear interpolation, adjacent value filling, etc. to complete; for obviously abnormal data points, they are rejected or corrected.

[0064] Third, data normalization. Since the physical dimensions and numerical ranges of the collected data are different, the unit uses methods such as Min-Max Scaling to linearly map all numerical features to the interval [0, 1] or [-1, 1], to eliminate the influence of different dimensions on subsequent model training.

[0065] After the above processing, the data acquisition and fusion module 100 is , and finally output a standardized multidimensional data set This data set is the direct basis for calculation and analysis of all subsequent modules.

[0066] The input end of the cognitive digital twin construction module 200 is connected to the output end of the data acquisition and fusion module 100. The function of this module is to receive standardized multidimensional data sets. , and convert it into a cognitive state vector that can compactly and comprehensively characterize the current integrated state of the driver-vehicle-road-environment system In a specific embodiment, the module can be further functionally divided into a dynamic graph construction unit 210 , a graph neural network representation unit 220 , and a state vector generation unit 230 .

[0067] Dynamic graph construction unit 210, which is used at each discrete moment , based on the data set , build a dynamic heterogeneous graph .

[0068] In this heterogeneous graph, the node set Used to represent the core entity of the system. For example, you can define a "driver" node whose initial feature vector is Initialization of the driver state monitoring (DSM) data in; a "vehicle" node, whose initial feature vector is given by Initialization of vehicle bus (CAN) data and advanced driver assistance system (ADAS) data in;

[0069] A "road" node, whose initial eigenvector is given by and an "Environment" node whose initial feature vector is given by Initialize the environmental meteorological (ENV) data in the edge set. Used to indicate the preset, objectively existing physical or logical relationship between the above entities.

[0070] For example, an edge is established between the "driver" node and the "vehicle" node to represent the driver's control relationship over the vehicle; an edge is established between the "vehicle" node and the "road" node to represent the position relationship of the vehicle on the road.

[0071] The graph neural network representation unit 220 receives the heterogeneous graph generated by the dynamic graph construction unit 210 As input, the unit adopts a graph neural network model and learns the high-dimensional feature representation of nodes in the graph through an iterative message passing and update mechanism, thereby capturing the deep nonlinear dependencies between different entity data.

[0072] In each layer (or each iteration) of a graph neural network, each node in the graph performs two steps: first, it aggregates information about its neighboring nodes; then, it combines this aggregated information with its current state to update its own state. This process is mathematically defined by the following formula, where any node In the The hidden state vector of the layer Updated through an iterative process that performs a preset This allows information to propagate over a longer distance on the graph structure, thereby obtaining a final node state representation that can fully characterize the node and its surrounding environment. .

[0073] The state vector generation unit 230 receives the final state representation of all nodes output from the graph neural network representation unit 220 The function of this unit is to integrate the representation information of multiple nodes in the graph into a single, fixed-dimensional graph-level representation vector, namely the cognitive state vector .

[0074] In one embodiment, this function is implemented through a readout function. The readout function can be an operation that performs a global summation, a global average, or a global maximum on the final state representation vectors of all nodes. In another embodiment, the attention mechanism can be used to assign different weights to different nodes and then perform a weighted summation. The cognitive state vector finally output by this unit is is transmitted to the risk prediction and assessment module 300 .

[0075] The input of the risk prediction and assessment module 300 is connected to the output of the cognitive digital twin construction module 200. This module's function is to receive the continuous-time cognitive state vector sequence generated by the cognitive digital twin construction module 200, deduce future risk states based on this sequence, and quantitatively assess the reliability of the deduction results. In a specific embodiment, this module can be further functionally divided into a time series risk prediction unit 310 and a prediction uncertainty quantification unit 320.

[0076] Time series risk prediction unit 310, which receives a time-step cognitive state vector sequence As input. This unit uses a long short-term memory network (LSTM) model, which is a recurrent neural network that can learn and memorize long-term dependencies in time series data. This unit predicts the future by feeding the input sequence into the LSTM model for processing. The state vector sequence of time steps is the predicted sequence, which is the risk evolution trend. The process is , For the model prediction, from time arrive The cognitive state vector sequence.

[0077] Predicted state vector It includes the predicted values ​​of specific risk indicators such as future fatigue, distraction probability, and irregular driving behavior probability.

[0078] The prediction uncertainty quantification unit 320 is used to evaluate the reliability of the prediction made by the time series risk prediction unit 310.

[0079] In one embodiment, the unit is implemented using a Monte Carlo Dropout method. Specifically, when performing prediction, the unit keeps the dropout layers in the LSTM model of the time series risk prediction unit 310 active.

[0080] This unit takes the same input sequence Repeatedly fed into the LSTM model independent forward propagation calculations. Due to the randomness of the dropout layer, each forward propagation will produce a slightly different sequence of prediction results .

[0081] The unit then calculates the The degree of dispersion between the prediction results, such as calculating their variance, is used as a quantitative prediction uncertainty score. The calculation of this score is defined by the following formula:

[0082] ;

[0083] Where, For the moment The calculated forecast uncertainty score; is the total number of random forward propagation executions; is the number of time steps for the prediction; For the The random forward propagation of the next The prediction vector of the state at time steps; For the future the mean of the predicted vectors of the previous time step.

[0084] Finally, the risk evolution trend and the prediction uncertainty score output by this module are delivered to the adaptive intervention decision module 400.

[0085] In a preferred embodiment, the system further comprises a regional risk field perception module 500. This module is connected to the data acquisition and fusion module 100 to obtain the vehicle location information, and is arranged in parallel with the risk prediction and assessment module 300, and its output end is connected to the adaptive intervention decision module 400.

[0086] The function of this module is to identify the systematic driving risk dominated by the external environmental factors of a specific region from a macroscopic and group perspective, independent of the risk state of individual vehicles. In a specific embodiment, this module can be further functionally divided into a twin spatio-temporal clustering unit 510, a risk resonance analysis unit 520, and a risk field judgment unit 530.

[0087] The twin spatio-temporal clustering unit 510 obtains high-precision GPS data of all online vehicles in the system from the data acquisition and fusion module 100 in real time. Based on these spatio-temporal location information, this unit uses a spatial clustering algorithm, such as density-based spatial clustering (DBSCAN) or a simple geographic grid division method, to dynamically divide vehicles that are geographically adjacent into the same regional set , where is the unique identifier of the region.

[0088] The risk resonance analysis unit 520 obtains the risk prediction results of each vehicle in the region set divided by the twin spatio-temporal clustering unit 510 from the risk prediction and assessment module 300. This unit extracts one or more pre-set key risk scalars from the risk prediction of each vehicle, such as the predicted probability of fatigue driving or the frequency of sudden acceleration / sudden deceleration events.

[0089] Subsequently, this unit calculates the statistical characteristics of the key risk scalars of all vehicles in the region set at time , specifically the mean and the variance , whose calculation process is defined by the following formula:

[0090] ;

[0091] ​ ;

[0092] wherein, is the mean of the critical risk scalar of all vehicles located within the region at time ; is the variance of the critical risk scalar of all vehicles located within the region at time ; is the total number of vehicles in the region set ; is the critical risk scalar value of the th vehicle in the region set at time ;

[0093] a risk field decision unit 530, which receives the mean and variance computed by the risk resonance analysis unit 520. This unit compares these two statistical values with preset thresholds to determine whether a risk field exists in the region.

[0094] The input of the adaptive intervention decision module 400 is connected to the output of the risk prediction and assessment module 300, and in the preferred embodiment, it is also connected to the output of the regional risk field perception module 500. The function of this module is to generate and execute an optimal active intervention instruction that matches the current risk level and risk type according to the input risk information.

[0095] In a specific embodiment, this module implements adaptive decision-making through a reinforcement learning (RL) model. The key components of this reinforcement learning model are defined as follows:

[0096] The state space of the model is defined as a composite state vector, which is constructed at each decision-making time , and includes at least: the risk evolution trend output by the risk prediction and assessment module 300, i.e., the predicted state vector sequence for the next time steps; and the predicted uncertainty score output by the module . In embodiments configured with the regional risk field perception module 500, the composite state vector also includes a binary flag indicating whether a regional risk field currently exists.

[0097] The action space A of the model is defined as a discrete, hierarchical set, where each action corresponds to a pre-defined, specific intervention instruction with different intervention intensity. For example, the action space can include action a0 (no intervention performed), action a1 (performing a level-1 intervention, e.g., displaying a text prompt on the in-vehicle screen), action a2 (performing a level-2 intervention, e.g., issuing a gentle voice prompt), and action a3 (performing a level-3 intervention, e.g., issuing a loud and flashing light alarm accompanied by seat vibration). The module selects an action according to the current state and generates the corresponding instruction, which is then issued to the human-machine interface of the vehicle for execution.

[0098] The model learns and optimizes through a reward function that evaluates the effect of the selected action on the state at the next time step. The reward function is designed as a weighted sum of multiple components to balance safety and driver acceptance. Its specific form is defined by the following equation:

[0099]

[0100] where is the total reward value at time step t; is the risk reduction amount, which is calculated as the risk assessment value before intervention minus the risk assessment value at the next time step after intervention; is the inherent cost of the performed action , different intervention intensity actions are assigned different cost values, e.g., c(a0)=0, c(a1)=0.1, c(a2)=0.3; is a binary flag indicating whether the driver manually canceled the intervention within a short time window after intervention execution; if the driver canceled the intervention, =1, otherwise =0. , , , are the weight coefficients of risk reduction amount, action cost, and manual cancellation behavior, respectively, which are pre-set hyperparameters.

[0101] The design of the reward function enables actions that effectively reduce risk to receive positive rewards, while unnecessary or driver-averse intervention actions receive negative rewards.

[0102] ​​The module internally contains a policy network, such as a deep Q network (DQN) or similar structure. The network receives the composite state vector as input and outputs an estimate of the value of each available action in the current state. The module selects the action to perform according to the value estimate. The policy network is trained offline on a large amount of historical or simulated data to learn an intervention policy that maximizes long-term cumulative reward.

[0103] In a preferred embodiment of the invention, the system further comprises a closed-loop model self-checking module 600. The input of the module is connected to the risk prediction and assessment module 300 to obtain the prediction uncertainty score; its output is connected to the cognitive digital twin construction module 200 and the risk prediction and assessment module 300 to perform online calibration on the model parameters inside these two modules. The function of this module is to obtain real-world feedback by actively interacting with the driver when the system doubts the reliability of its own risk prediction, and to use this feedback to make immediate corrections to the upstream models, in order to improve the adaptability and robustness of the entire system.

[0104] In a specific embodiment, the module can be further functionally divided into an active probing trigger unit 610, a micro-interaction task unit 620, and an online calibration unit 630.

[0105] Active probing trigger unit 610, this unit continuously monitors the prediction uncertainty score output by the risk prediction and assessment module 300. The unit internally presets an uncertainty threshold . At each time, the unit compares the current prediction uncertainty score with the threshold . When and only when , the unit is activated and sends a trigger signal to the micro-interaction task unit 620.

[0106] Micro-interaction task unit 620, this unit performs operations after receiving the trigger signal from the active probing trigger unit 610. The function of this unit is to design and perform a micro-interaction task related to the driving task, but which will not interfere with driving safety, in order to obtain feedback that reflects the current real state of the driver.

[0107] For example, the unit can control the human-machine interface of the vehicle, ask the driver a simple environmental perception question that requires his confirmation, or perform a short reaction ability test. The unit will record the driver's feedback behavior, such as the driver's key input or voice reply, simultaneously, and mark this feedback data as .

[0108] an online calibration unit 630, which obtains feedback data from the micro-interaction task unit 620 performs an operation. This unit pairs the feedback data with the system state at the time when the task was triggered (e.g. the cognitive state vector at that time ) to form a new training sample with real labels .

[0109] Subsequently, this unit performs a parameter update of the upstream models (i.e. the models in the cognitive digital twin construction module 200 and / or the risk prediction and assessment module 300) based on this sample, using a total loss function that contains a probing loss term.

[0110] This unit uses the computed values to compute gradients using backpropagation algorithms and performs an iterative update of the parameters of the upstream models. Through this mechanism, the system is able to continuously correct itself using real feedback from high-uncertainty scenarios during operation.

[0111] Please refer to the attached Figure 2 , Figure 2 is a workflow diagram of the intelligent supervision system for road transport key operating vehicles according to an embodiment of the present application. The overall workflow of the present application will be described in detail below in combination with a specific application scenario.

[0112] After the system is started, it first enters the continuous monitoring and risk prediction process. At each time step , the data acquisition and fusion module 100 acquires multi-dimensional data such as driver state, vehicle dynamics, road properties, and environmental meteorological data from the vehicle terminal and the cloud platform, and processes them into a standardized data set .

[0113] The cognitive digital twin construction module 200 receives the data set, generates a cognitive state vector that can represent the current comprehensive state of the system by constructing a dynamic heterogeneous graph and using a graph neural network to learn.

[0114] Subsequently, the risk prediction and assessment module 300 receives the cognitive state vector and sends it together with the historical state vector sequence to its internal time series prediction model to generate a prediction of the future risk evolution trend, while calculating the prediction uncertainty score of this prediction .

[0115] After obtaining the risk prediction results and uncertainty scores, the system enters different branch processes according to the uncertainty scores. In one embodiment, when the predicted risk level output by the risk prediction and assessment module 300 exceeds the preset intervention threshold and its predicted uncertainty score exceeds the threshold, the system will automatically select the intervention level. Below a preset uncertainty threshold , it indicates that the system has a higher confidence level in the prediction results of high risks.

[0116] At this point, the risk evolution trend and low uncertainty score are transmitted to the adaptive intervention decision module 400. Based on these inputs, the module uses its internal reinforcement learning model to determine the optimal intervention action that matches the current risk level and generates corresponding instructions, such as sound and light alarms or seat vibration prompts, which are sent to the vehicle terminal for execution.

[0117] In another embodiment, when the prediction uncertainty score output by the risk prediction and assessment module 300 is Above a preset uncertainty threshold When , it indicates that the system currently has doubts about the reliability of its prediction results. This situation will trigger the closed-loop model self-checking module 600.

[0118] The module first initiates an active detection task to the driver through the micro-interaction task unit 620 and collects the feedback data Then, the online calibration unit 630 uses the feedback data The system status when the task is triggered The pairing forms a high-value real-time labeled sample, and the detection loss is calculated based on it, and the model parameters inside the cognitive digital twin construction module 200 and the risk prediction and assessment module 300 are updated online, thereby completing a closed-loop self-calibration.

[0119] While the above processes are executing in parallel, the regional risk field perception module 500 also operates continuously. This module continuously performs spatiotemporal clustering of all vehicles in the system and analyzes the consistency of vehicle risk status within each region. When this module determines that a risk field has formed in a certain region (for example, if the average risk level of vehicles in the region has significantly increased and the risk status is highly consistent), it sends a signal indicating the presence of a regional risk field to the adaptive intervention decision module 400.

[0120] After receiving this signal, the adaptive intervention decision module 400 will adjust its intervention strategy accordingly: for vehicles already in the risk field, its intervention level will be increased; for vehicles whose driving path is about to pass through the risk field, even if their own risk prediction level is not high, forward-looking warning instructions will be generated and executed for them.

[0121] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. Intelligent supervision system for key road transport vehicles, characterized by: The system comprises: Data collection and fusion module, used to collect and fuse vehicle terminal data and external environment data in real time to generate standardized multi-dimensional data sets; a cognitive digital twin construction module, connected to the data acquisition and fusion module, for constructing a cognitive digital twin representing the comprehensive state of the vehicle and its driver based on the standardized multidimensional data set, and outputting a real-time updated cognitive state vector; a risk prediction and assessment module, connected to the cognitive digital twin construction module, for predicting the risk evolution trend over a period of time in the future based on the time series of the cognitive state vector, and evaluating the reliability of the prediction to generate a prediction uncertainty score; an adaptive intervention decision module, connected to the risk prediction and assessment module, for generating and executing adaptive proactive intervention instructions based on the risk evolution trend and the prediction uncertainty score; The risk prediction and assessment module includes: a time series risk prediction unit, configured to model the time series of the cognitive state vector using a long short-term memory network to generate the risk evolution trend; a prediction uncertainty quantification unit, configured to generate the prediction uncertainty score by performing multiple random forward propagations on the time series risk prediction unit and calculating the dispersion of the prediction results; The system further includes a closed-loop model self-verification module, which is connected to the risk prediction and assessment module and the adaptive intervention decision module and is used to: When the prediction uncertainty score exceeds a preset uncertainty threshold, triggering and executing a micro-interaction task related to the driving task; Collecting feedback data from the driver on the micro-interaction task; Based on the feedback data, the models in the cognitive digital twin construction module and the risk prediction and assessment module are calibrated online.

2. The intelligent supervision system for key road transport vehicles according to claim 1 is characterized in that: The cognitive digital twin building blocks include: A dynamic graph construction unit is configured to construct a dynamic heterogeneous graph comprising a node set and an edge set at each moment based on the standardized multidimensional data set, wherein the node set represents the vehicle, driver, and road environment entities, and the edge set represents the intrinsic relationship between the entities; A graph neural network representation unit is used to learn the dynamic heterogeneous graph using a graph neural network and aggregate neighbor node information through an iterative message passing mechanism to update the node status; A state vector generation unit is used to integrate the states of all nodes after learning by the graph neural network to generate the cognitive state vector.

3. The intelligent supervision system for key road transport vehicles according to claim 2 is characterized in that: In the graph neural network representation unit, any node In the The hidden state vector of the layer Updated by the following formula: ； Where, is the node in the graph neural network In the The hidden state vector of the layer; For nodes In the The hidden state vector of the layer; For nodes The set of neighbor nodes of is an aggregate function; It is an update function used to combine the node's own information and neighbor information.

4. The intelligent supervision system for key road transport vehicles according to claim 1 is characterized in that: The online model calibration in the closed-loop model self-verification module is achieved by a total loss function including a detection loss term: ； Where, is the total loss; For original loss; Based on the current state The model output and the feedback data Calculated detection loss; is a hyperparameter used to balance the original loss and the detection loss; The current state of the model when the micro-interaction task is triggered The output below; is the collected feedback data of the driver on the micro-interaction task.

5. The intelligent supervision system for key road transport vehicles according to claim 1 is characterized in that: The system further includes a regional risk field perception module, which is provided in parallel with the risk prediction and assessment module and connected to the adaptive intervention decision module, and is used to: clustering the cognitive digital twins based on geographic location; By analyzing the collaborative consistency of risk prediction status of multiple cognitive digital twins in the clustered region, we can identify whether there is a regional risk field. The recognition result is transmitted to the adaptive intervention decision module to influence the generation of the active intervention instruction.

6. The intelligent supervision system for key road transport vehicles according to claim 5 is characterized in that: The regional risk field perception module calculates the key risk scalars of all vehicles in the region. The mean and variance , and determine whether the regional risk field exists based on the following conditions: ； Where, is the mean threshold; is the variance threshold; For the moment Located in the area The mean of the key risk scalars for all vehicles in ; is the logical AND operator; For the moment Located in the area The variance of the key risk scalar for all vehicles within .

7. The intelligent supervision system for key road transport vehicles according to claim 5 is characterized in that: The adaptive intervention decision module is further configured to: When it is determined that the vehicle is in the regional risk field, based on the risk evolution trend of the vehicle itself and in combination with the existence of the regional risk field, the intervention level of the active intervention instruction is increased or the intervention mode is changed; When it is determined that the vehicle's driving path will pass through the regional risk field, even if the risk evolution trend of the vehicle itself does not reach the preset intervention threshold, a forward-looking warning instruction for the regional risk field is generated and executed for it.

8. The intelligent supervision system for key road transport vehicles according to claim 1 is characterized in that: The adaptive intervention decision module is configured to employ a reinforcement learning model to generate the proactive intervention instruction, wherein: The reinforcement learning model receives a composite state vector as input, the composite state vector including at least the risk evolution trend and the prediction uncertainty score; The action space of the reinforcement learning model includes a plurality of preset graded intervention actions with different intervention intensities; The reinforcement learning model is optimized using a reward function designed to enable the system to learn intervention strategies that avoid alarm fatigue by penalizing actions that result in the driver manually canceling intervention, while rewarding actions that effectively reduce risk.

Images