A method and related device for detecting roll hazard in multi-axle off-road vehicles
By integrating multi-source data to predict road surface adhesion coefficient and constructing a multi-agent architecture, the problem of lag in the detection of roll hazard of multi-axle vehicles is solved, and real-time and accurate assessment of roll hazard and dynamic adaptation of safety boundaries are realized in off-road environments.
Patent Information
- Application Number
- CN202611416959.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-14
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-09-14
AI Technical Summary
Existing technologies struggle to effectively integrate environmental perception information with vehicle interior status in complex off-road environments, resulting in delayed and inaccurate detection of roll hazard in multi-axle vehicles and a lack of prediction of future moments.
By acquiring ontological perception data, visual geometric data, and vehicle speed information, and combining filtering algorithms and multilayer perception neural networks, the spatial distribution of road surface adhesion coefficient is predicted. Furthermore, the dynamic vertical load of the wheels is estimated in real time through a multi-agent architecture, and a roll risk evaluation index and dynamic safety boundary are constructed to achieve a comprehensive roll risk score.
It enables real-time and accurate detection of roll hazard in multi-axle off-road vehicles, adapts to changes in complex off-road terrain, provides reliable safety boundaries, reduces central computing load, and improves system robustness and adaptability.
Smart Images

Figure CN122915148B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of vehicle safety control technology, and in particular to a method and related device for detecting the roll hazard of multi-axle off-road vehicles. Background Technology
[0002] With the rapid development of autonomous control and collaborative technologies, distributed multi-axle unmanned special vehicles have demonstrated significant advantages in complex off-road environments such as disaster relief and field transportation, thanks to their autonomous perception and decision-making capabilities and multi-actuator collaborative control. However, in typical off-road terrains such as soft ground, slopes, and gravel, vehicles are prone to multi-directional instability, including sideslip, longitudinal slip, and lateral tilt. Once instability occurs, it can lead to loss of vehicle maneuverability, interruption of the mission chain, and in extreme cases, even equipment overturning and mission failure.
[0003] In distributed multi-axle vehicles, each wheel can be controlled independently. While this increases the controllable dimensions, the redundancy of controlling multiple wheels exacerbates the coupling between vehicle dynamic states, increasing the likelihood of instability in complex off-road terrain. In off-road terrain, vehicles are subjected to greater lateral forces, more complex adhesion variations, and unstable support surfaces, making roll instability one of the most prevalent and dangerous failure modes. Therefore, implementing roll hazard detection for multi-axle vehicles in off-road scenarios, providing a reliable safety boundary for subsequent vehicle control, is crucial for ensuring vehicle stability and safety.
[0004] Traditional vehicle roll hazard detection primarily relies on vehicle dynamics models to estimate roll angles and roll moments to determine hazard levels. While these methods offer good real-time computation and are relatively mature in implementation, their predictive capabilities for undulating terrain and non-uniform tire grip are limited, making them difficult to directly apply to off-road vehicle roll stability prediction. Furthermore, these methods are mostly limited to four-wheeled vehicles, lacking adaptability to multi-axle distributed drive vehicles and failing to effectively utilize the advantages of distributed wheels in terms of multi-dimensional information acquisition and multi-degree-of-freedom control.
[0005] With advancements in intelligent environmental perception algorithms, environmental perception has provided new data sources and technical pathways for assessing vehicle stability. However, most current environmental perception algorithms remain limited to target recognition and active collision avoidance, failing to effectively integrate external perception information with the internal vehicle motion state. Consequently, these algorithms struggle to adapt to changing vehicle operating conditions. Furthermore, traditional roll hazard assessment algorithms rely solely on the vehicle's current state, only evaluating the hazard level of the current condition and lacking predictions for future moments, resulting in lag in practical applications.
[0006] Therefore, how to effectively integrate environmental perception information from off-road scenarios with vehicle internal status information to achieve accurate assessment of vehicle roll risk within a future time window is a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0007] The purpose of this application is to provide a method and related device for detecting the roll hazard of multi-axle off-road vehicles, which can achieve accurate and real-time detection of the roll hazard of multi-axle off-road vehicles.
[0008] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for detecting the roll hazard of a multi-axle off-road vehicle, comprising the following steps: Acquire the target vehicle's body perception data, visual geometry data, and vehicle speed information during its journey across off-road terrain.
[0009] Based on the ontological perception data, the visual geometry data, and the vehicle speed information, the terrain traversal risk cost of the off-road terrain where the target vehicle is located is determined.
[0010] Based on the aforementioned ontological perception data and combined with the vehicle dynamics model, a filtering algorithm is used to calculate the road surface adhesion coefficient of the target vehicle's current driving area in real time.
[0011] Based on the terrain travel risk cost and the road surface adhesion coefficient of the current driving area, the spatial distribution prediction result of the road surface adhesion coefficient in the undriven area in front of the target vehicle is predicted.
[0012] The vehicle control system of the target vehicle is divided into one vehicle intelligent agent and multiple wheel intelligent agents, and a communication topology between the vehicle intelligent agent and the multiple wheel intelligent agents is constructed.
[0013] A tire model for off-road scenarios is constructed, and the dynamic vertical load of each wheel in the multiple wheel agents is estimated in real time by combining the communication topology and the spatial distribution prediction results of the road surface adhesion coefficient.
[0014] The roll hazard assessment index is calculated based on the dynamic vertical load of each wheel and the body sensing data, and the dynamic safety boundary of the roll hazard assessment index is determined according to the road surface adhesion coefficient of the current driving area and the vehicle speed information.
[0015] Based on the relative distance between the roll hazard assessment index and the dynamic safety boundary, the comprehensive roll hazard score of the target vehicle is determined.
[0016] Secondly, this application provides a roll hazard detection device for multi-axle off-road vehicles, comprising: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the roll hazard detection method for multi-axle off-road vehicles.
[0017] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for detecting the roll hazard of multi-axle off-road vehicles.
[0018] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method and related apparatus for detecting roll hazard in multi-axle off-road vehicles. First, by acquiring ontological perception data, visual geometric data, and vehicle speed information, and determining the terrain traversal risk cost based on the aforementioned multi-source data, a quantitative characterization of the hazard level of off-road terrain on vehicle passage can be achieved, providing terrain-level input for subsequent road condition prediction and roll hazard detection. Second, based on ontological perception data combined with a vehicle dynamics model and filtering algorithm, the road surface adhesion coefficient of the current driving area is calculated in real time. Furthermore, based on the terrain traversal risk cost and the current road surface adhesion coefficient, the spatial distribution of the road surface adhesion coefficient in the untraveled area ahead is predicted, thereby achieving advance perception of the physical characteristics of the road surface in the future driving area, overcoming the shortcomings of traditional methods that can only assess the current state and have a lag. Third, by dividing the vehicle control system into one vehicle agent and multiple wheel agents and constructing a communication topology, combined with a tire model for off-road scenarios and the prediction results of the spatial distribution of the road surface adhesion coefficient, the dynamic vertical load of each wheel can be estimated in real time, achieving real-time estimation of the dynamic vertical load of each wheel. Simultaneously, the distributed architecture helps reduce the central computing load, improving the system's robustness and adaptability to multi-axle vehicles. Furthermore, by combining the dual-source calculation evaluation index with the dynamic safety boundary, the roll hazard evaluation index is calculated based on the dynamic vertical load of each wheel and the vehicle's perception data. The dynamic safety boundary is dynamically determined according to the road surface adhesion coefficient and vehicle speed information in the current driving area. This can adapt to changes in road conditions and vehicle operating conditions in off-road scenarios and avoid misjudgments based on fixed thresholds. Finally, the comprehensive roll hazard score is determined by the relative distance between the evaluation index and the dynamic safety boundary, realizing real-time detection and accurate quantification of the roll hazard of multi-axle off-road vehicles, providing a reliable safety boundary for subsequent vehicle control. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 An application environment diagram for a method for detecting the roll hazard of a multi-axle off-road vehicle provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for detecting the roll hazard of a multi-axle off-road vehicle according to an embodiment of this application; Figure 3 This is a schematic diagram of the complete process for detecting the roll hazard of a multi-axle off-road vehicle according to an embodiment of this application; Figure 4 This is a schematic diagram of the nominal vertical load estimation result provided in an embodiment of this application, wherein, Figure 4 (a) in the diagram is a schematic diagram of the nominal vertical load estimation result corresponding to wheel agent 1. Figure 4 (b) in the diagram is a schematic diagram of the nominal vertical load estimation result corresponding to wheel agent 2. Figure 4 (c) in the diagram is a schematic diagram of the nominal vertical load estimation result corresponding to wheel agent 3. Figure 4 (d) in the diagram is a schematic diagram of the nominal vertical load estimation result corresponding to wheel agent 4. Figure 4 (e) in the diagram is a schematic representation of the nominal vertical load estimation result corresponding to wheel agent 5. Figure 4 (f) in the diagram is a schematic representation of the nominal vertical load estimation result corresponding to wheel agent 6. Figure 4 (g) in the diagram represents the nominal vertical load estimation result corresponding to wheel agent 7. Figure 4 (h) in the figure is a schematic diagram of the nominal vertical load estimation result corresponding to wheel agent 8; Figure 5 This is a schematic diagram of the total vertical load estimation result provided in an embodiment of this application, wherein, Figure 5 (a) in the diagram is a schematic diagram of the total vertical load estimation result corresponding to wheel agent 1. Figure 5 (b) in the diagram is a schematic diagram of the total vertical load estimation result corresponding to wheel agent 2. Figure 5 (c) in the diagram is a schematic diagram of the total vertical load estimation result corresponding to wheel agent 3. Figure 5 (d) in the diagram is a schematic representation of the total vertical load estimation result corresponding to wheel agent 4. Figure 5 (e) in the diagram is a schematic representation of the total vertical load estimation result corresponding to wheel agent 5. Figure 5(f) in the diagram is a schematic representation of the total vertical load estimation result corresponding to wheel agent 6. Figure 5 (g) in the diagram represents the total vertical load estimation result corresponding to wheel agent 7. Figure 5 (h) in the figure is a schematic diagram of the total vertical load estimation result corresponding to wheel agent 8; Figure 6 This is a schematic diagram of the structure of a multi-axle off-road vehicle roll hazard detection device provided in an embodiment of this application. Detailed Implementation
[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0022] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0023] The roll hazard detection method for multi-axle off-road vehicles provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the target vehicle's body perception data, visual geometry data, and vehicle speed information during its off-road driving process to server 104. After receiving these data, server 104 determines the terrain traversal risk cost of the off-road terrain where the target vehicle is located, based on the body perception data, visual geometry data, and vehicle speed information. The terrain traversal risk cost represents the degree of danger posed by the off-road terrain to vehicle passage. Based on the body perception data and combined with a vehicle dynamics model, a filtering algorithm is used to calculate the road surface adhesion coefficient of the target vehicle's current driving area in real time. Based on the terrain traversal risk cost and the road surface adhesion coefficient of the current driving area, the spatial distribution prediction result of the road surface adhesion coefficient in the untraversed area ahead of the target vehicle is predicted. The vehicle control system of the target vehicle is divided into one vehicle agent (VA) and multiple wheel agents (Wheel Agents). The system constructs a communication topology between the vehicle agent and the multiple wheel agents. It then builds a tire model for off-road scenarios, combining the communication topology and the spatial distribution prediction results of the road surface adhesion coefficient to estimate the dynamic vertical load of each wheel in the multiple wheel agents in real time. Based on the dynamic vertical load of each wheel and the vehicle perception data, it calculates a roll hazard evaluation index and determines the dynamic safety boundary of the roll hazard evaluation index according to the road surface adhesion coefficient of the current driving area and the vehicle speed information. Based on the relative distance between the roll hazard evaluation index and the dynamic safety boundary, it determines the comprehensive roll hazard score of the target vehicle. The server 104 can feed back the obtained comprehensive roll hazard score to the terminal 102. Furthermore, in some embodiments, the multi-axle off-road vehicle roll hazard detection method can also be implemented separately by the server 104 or the terminal 102. For example, the terminal 102 can directly perform vehicle roll hazard detection processing based on the body perception data, visual geometry data, and vehicle speed information of the target vehicle during off-road terrain driving. Alternatively, the server 104 can obtain the body perception data, visual geometry data, and vehicle speed information of the target vehicle during off-road terrain driving from the data storage system, and perform vehicle roll hazard detection processing based on the body perception data, visual geometry data, and vehicle speed information of the target vehicle during off-road terrain driving.
[0024] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, and IoT devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.
[0025] In one exemplary embodiment, such as Figure 2 As shown, a method for detecting roll hazard in multi-axle off-road vehicles is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the following steps are included: S1: Acquire the target vehicle's body perception data, visual geometry data, and vehicle speed information during its off-road driving process.
[0026] S2: Based on the ontological perception data, the visual geometry data, and the vehicle speed information, determine the terrain traversal risk cost of the off-road terrain where the target vehicle is located. The terrain traversal risk cost represents the degree of danger posed by the off-road terrain to vehicle passage.
[0027] S3: Based on the aforementioned body perception data and combined with the vehicle dynamics model, a filtering algorithm is used to calculate the road surface adhesion coefficient of the target vehicle's current driving area in real time.
[0028] S4: Based on the terrain travel risk cost and the road surface adhesion coefficient of the current driving area, predict the spatial distribution of the road surface adhesion coefficient in the undriven area in front of the target vehicle.
[0029] S5: Divide the vehicle control system of the target vehicle into one vehicle intelligent agent and multiple wheel intelligent agents, and construct the communication topology between the vehicle intelligent agent and the multiple wheel intelligent agents.
[0030] S6: Construct a tire model for off-road scenarios, and combine the communication topology and the spatial distribution prediction results of the road surface adhesion coefficient to estimate the dynamic vertical load of each wheel in the multiple wheel agents in real time.
[0031] S7: Calculate the roll hazard assessment index based on the dynamic vertical load of each wheel and the body perception data, and determine the dynamic safety boundary of the roll hazard assessment index according to the road surface adhesion coefficient of the current driving area and the vehicle speed information.
[0032] S8: Based on the relative distance between the roll hazard evaluation index and the dynamic safety boundary, determine the comprehensive roll hazard score of the target vehicle.
[0033] By implementing steps S1 to S8 above, multi-source perception data is fused to achieve a quantitative assessment of the risk of off-road terrain travel. By combining terrain assessment and dynamic calculation, the road surface condition of untraveled areas ahead can be predicted in advance. By adopting a multi-agent distributed architecture, the dynamic vertical load of each wheel can be estimated in real time. By combining dual-source computation evaluation indicators with dynamic safety boundaries, real-time monitoring and accurate quantification of the roll hazard of multi-axle off-road vehicles can be achieved.
[0034] As an optional implementation, in step S2, based on the ontology perception data, the visual geometry data, and the vehicle speed information, the terrain traversal risk cost of the off-road terrain where the target vehicle is located is determined, specifically including the following steps: S21: Using the inertial measurement unit (IMU) data in the body perception data, extract the vertical acceleration of the target vehicle.
[0035] S22: Calculate and integrate the power spectral density of the vertical acceleration within a preset frequency band to obtain the frequency band power; the frequency band power is used to determine the terrain access risk cost sequence, and the terrain access risk cost sequence is used as a supervision label for training the convolutional neural network model.
[0036] S23: Register the image information and dense point cloud information in the visual geometric data and project them into the bird's-eye view (BEV) space to generate a multi-channel local map containing color features, elevation features and effective area masks.
[0037] S24: Perform Fourier feature mapping on the vehicle speed information to obtain speed features.
[0038] S25: Based on the convolutional neural network model, extract the visual geometric features of the multi-channel local map, concatenate the velocity features with the visual geometric features and input them into a fully connected layer, outputting a normalized scalar terrain access risk cost as the terrain access risk cost.
[0039] As an optional implementation, in step S3, based on the ontological perception data and combined with the vehicle dynamics model, a filtering algorithm is used to calculate the road surface adhesion coefficient of the target vehicle's current driving area in real time, specifically including the following steps: S31: Define a vehicle state vector containing the road surface adhesion coefficient, and construct a nonlinear state-space model of the target vehicle; the tire force in the nonlinear state-space model is related to the road surface adhesion coefficient.
[0040] S32: Select the physical quantities that can be directly measured by the vehicle-mounted sensors in the body perception data as observation vectors, use the extended Kalman filter algorithm to estimate the state of the nonlinear state-space model, correct the estimated value of the road adhesion coefficient by the residual between the observed vehicle's actual acceleration and the model's predicted acceleration, and extract the road adhesion coefficient of the current driving area from the updated state vector.
[0041] As an optional implementation, in step S4, based on the terrain travel risk cost and the road surface adhesion coefficient of the current driving area, the spatial distribution prediction result of the road surface adhesion coefficient in the untraveled area ahead of the target vehicle is made, specifically including the following steps: S41: Time-align the sequence of terrain passage risk costs with the sequence of road surface adhesion coefficients in the current driving area to obtain synchronous training samples.
[0042] S42: Construct a multilayer perceptron neural network model, and train the multilayer perceptron neural network model based on the synchronous training samples, establish a spatial mapping relationship from the terrain travel risk cost to the road surface adhesion coefficient, and predict the road surface adhesion coefficient of the untraveled area ahead based on the multilayer perceptron neural network model and the spatial mapping relationship.
[0043] S43: Project the road surface adhesion coefficient of the untraveled area ahead onto a bird's-eye view elevation map to generate an enhanced terrain map containing elevation information, the terrain traffic risk cost, and the road surface adhesion coefficient, which serves as the spatial distribution prediction result of the road surface adhesion coefficient. Specifically, the enhanced terrain map is the carrier of the spatial distribution prediction result of the road surface adhesion coefficient, and the prediction result includes elevation, terrain traffic risk cost, and road surface adhesion coefficient information.
[0044] As an optional implementation, in step S5, the vehicle control system of the target vehicle is divided into one vehicle intelligent agent and multiple wheel intelligent agents, and a communication topology between the vehicle intelligent agent and the multiple wheel intelligent agents is constructed, specifically including the following steps: S51: Divide the six-degree-of-freedom dynamic state of the target vehicle into a vehicle intelligent agent, and divide the rotational dynamic state of each wheel into a wheel intelligent agent.
[0045] S52: Based on the vehicle agent and each of the wheel agents, construct the communication topology; wherein, the wheel agents located on the left side of the target vehicle can communicate with each other, the wheel agents located on the right side of the target vehicle can communicate with each other, the wheel agents located on the left side of the target vehicle and the wheel agents located on the right side of the target vehicle cannot communicate directly, and all wheel agents can communicate with the vehicle agent. However, the wheel agents on the left and right sides of the target vehicle cannot communicate directly; their interaction information needs to be relayed through the vehicle agent.
[0046] As an optional implementation, in step S6, a tire model for off-road scenarios is constructed. Combining the communication topology and the spatial distribution prediction results of the road surface adhesion coefficient, the dynamic vertical load of each wheel in the multiple wheel agents is estimated in real time. This specifically includes the following steps: S61: Construct a tire model for off-road scenarios, decomposing the vertical load on each wheel into a nominal vertical load and an additional vertical load; the nominal vertical load is the load transfer caused by vehicle acceleration on a flat road surface, and the additional vertical load is the change in external additional load caused by off-road terrain.
[0047] S62: Based on the principle of virtual displacement, the nominal vertical load of each wheel is calculated according to the longitudinal acceleration, lateral acceleration and vertical acceleration of the target vehicle.
[0048] S63: The additional vertical load of each wheel is estimated using a distributed state observer based on the communication topology, and the additional vertical load is corrected by combining the spatial distribution prediction results of the road surface adhesion coefficient to obtain the corrected additional vertical load.
[0049] S64: Add the nominal vertical load to the modified additional vertical load to obtain the dynamic vertical load of each wheel.
[0050] As an optional implementation, step S63, which estimates the additional vertical load on each wheel using a distributed state observer based on the communication topology, specifically includes the following steps: S631: Construct a fixed-time state observer for each wheel agent to estimate the change in the additional vertical load on the wheel using the wheel's local rotational speed and slip ratio information.
[0051] S632: Based on the estimated trend of the additional vertical load change of each wheel agent, divide all wheel agents into a set of decreasing vertical load and a set of increasing vertical load.
[0052] S633: Based on the communication topology, obtain the total load reduction of all wheels in the vertical load reduction set, and use the principle of minimum potential energy to redistribute the total load reduction to the wheels in the vertical load increase set to obtain the additional vertical load of each wheel.
[0053] As an optional implementation, in step S7, the roll risk evaluation index includes: lateral load transfer rate (LTR) and rollover time (TTR); the lateral load transfer rate is used to characterize the current static load distribution of the target vehicle, and the rollover time is used to characterize the dynamic stability prediction capability of the target vehicle.
[0054] As an optional implementation, in step S7, a roll hazard assessment index is calculated based on the dynamic vertical load of each wheel and the body sensing data, and the dynamic safety boundary of the roll hazard assessment index is determined according to the road surface adhesion coefficient of the current driving area and the vehicle speed information. This specifically includes the following steps: S71: Based on the dynamic vertical load of each wheel, determine the total vertical load of the left wheel and the total vertical load of the right wheel of the target vehicle.
[0055] S72: The lateral load transfer rate is calculated based on the ratio of the difference between the total vertical load of the left wheel and the total vertical load of the right wheel to the sum of the total vertical loads of the left wheel and the right wheel.
[0056] S73: Based on the inertial measurement unit data in the body perception data, obtain the vehicle roll angle, roll velocity and roll acceleration, and calculate the rollover time according to the vehicle roll angle, the roll velocity and the roll acceleration.
[0057] S74: Determine the dynamic safety boundary of the lateral load transfer rate based on the road surface adhesion coefficient of the current driving area; the dynamic safety boundary of the lateral load transfer rate decreases as the road surface adhesion coefficient of the current driving area decreases.
[0058] S75: Determine the dynamic safety boundary of the rollover time based on the vehicle speed information; the dynamic safety boundary of the rollover time increases as the vehicle speed increases.
[0059] As an optional implementation, in step S8, based on the relative distance between the roll hazard evaluation index and the dynamic safety boundary, the comprehensive roll hazard score of the target vehicle is determined, specifically including the following steps: S81: Calculate the relative distances of the lateral load transfer rate and the rollover time to their respective dynamic safety boundaries, and convert the relative distances into single-index risk contribution values between 0 and 1, thereby obtaining the single-index risk contribution values corresponding to the lateral load transfer rate and the single-index risk contribution values corresponding to the rollover time.
[0060] S82: The single-index risk contribution corresponding to the lateral load transfer rate and the single-index risk contribution corresponding to the rollover time are weighted and fused, and a contradiction correction term is introduced to correct the risk assessment result when the two rollover hazard evaluation indicators show significant deviation, so as to obtain the comprehensive rollover hazard score result.
[0061] S83: Divide the hazard level according to the numerical range of the comprehensive roll hazard score result, and map the hazard level to the corresponding vehicle control strategy.
[0062] This application adopts a vehicle roll stability assessment framework that integrates vehicle internal state perception and external terrain environment perception information. It incorporates external intelligent perception results into the roll hazard detection framework. At the same time, based on the advantages of distributed vehicle wheels being independent, autonomous, controllable, and observable, it introduces a distributed state observer based on multi-agents, so that each wheel can perform state observation as an independent agent and detect the roll hazard state of the whole vehicle through mutual communication, which can improve the accuracy of roll hazard detection for multi-axle off-road vehicles.
[0063] In an exemplary embodiment, to illustrate the technical solutions provided in the embodiments of this application in detail, the following are provided: Figure 3 The method shown is a roll hazard detection method for multi-axle off-road vehicles. This method is essentially a roll hazard detection algorithm for distributed drive vehicles in off-road terrain that integrates internal and external perception. The complete process includes the following steps: Step 1: By fusing ontological perception data, visual geometry data, and vehicle speed information of the target vehicle during its off-road driving, a convolutional neural network model is used to predict the continuously normalized passage risk cost of the terrain. Specific sub-steps are as follows: Step 1.1: Using inertial measurement unit data during the target vehicle's driving process, construct a traffic risk cost that describes terrain roughness, bumpiness, and deformation capability, and use it as a supervision label for training the convolutional neural network model.
[0064] In this embodiment, the traffic risk cost of the target vehicle utilizes inertial measurement data during the target vehicle's movement, by extracting the target vehicle's vertical acceleration. The Welch method is used to calculate its power spectral density, and the band power is obtained by integration within a specific frequency band, thereby constructing a description of the terrain access risk cost. Ultimately, a terrain access risk cost sequence over a period of time is obtained. .
[0065] The formula for calculating the terrain access risk cost of the target vehicle is as follows: ; in, The cost of terrain-related risks for the target vehicle. In the first Within a time window, the vertical acceleration (i.e., the vehicle's Z-axis linear acceleration) at a frequency... Power spectral density at that point and These are the minimum and maximum frequencies for calculating the bandwidth power, representing the frequency range for calculating the bandwidth power. In this embodiment, the power variation range of the power spectral density is set to 1-30Hz, which is highly correlated with terrain roughness. The calculated terrain-based travel risk cost for the target vehicle is... It will be used as a supervisory label for training convolutional neural network models.
[0066] Step 1.2: Aggregate externally perceived information into a local map from a bird's-eye view. Use a stereo matching network to obtain a disparity map, and combine visual odometry to register RGB data with dense point cloud information and project it into the bird's-eye view space. The local map contains stacked multi-channel features: (1) RGB map channel: contains the average RGB color value of each grid cell; (2) Height map channel: contains the minimum, maximum, mean and standard deviation of the point cloud height in each grid cell, and removes suspended points above 2 meters above the ground to eliminate interference; (3) Mask channel: contains Boolean masks to identify areas without information due to occlusion or sensor field of view limitations.
[0067] In this embodiment, while acquiring inertial measurement unit data, terrain image information is collected through a visual sensor, a disparity map is obtained using a stereo matching network, and RGB data is registered with dense point clouds and projected onto a bird's-eye view space using visual odometry. The local map contains stacked multi-channel features, specifically including: RGB map channel features, height map channel features, and mask channel features.
[0068] Step 1.3: Perform Fourier feature mapping on the target vehicle's speed scalar and encode it as a high-dimensional vector. Extract visual geometric features and speed features separately, concatenate them, and input them into a fully connected layer. Combine the supervised labels to calculate the backpropagation cost and output a normalized scalar terrain passage risk cost.
[0069] In this embodiment of the application, the target vehicle speed information is extracted, and the target vehicle speed is scalarized. Fourier feature mapping is performed to encode the features into high-dimensional vectors. Then, visual geometric features of the local map are extracted using a ResNet backbone network, and velocity features are extracted using a three-layer fully connected neural network. The concatenated features are then input into a fully connected layer with a Sigmoid activation function, combined with the terrain access risk cost sequence from step 1.1. (As a supervisory label sequence) Calculate the backpropagation cost and finally output the normalized scalar terrain access risk cost.
[0070] In this embodiment of the application, the calculation formula for the target vehicle speed scalar through Fourier feature mapping is as follows: ; in, Let the target vehicle speed be a scalar. Represents the target vehicle speed scalar Fourier feature mapping, To adjust the scale The frequency coefficients of the samples in the Gaussian distribution, The number of frequencies used for mapping will ultimately result in a target vehicle speed scalar. Mapped to 3D eigenvectors.
[0071] Step 2: Using target vehicle motion state sensor data and combining it with the vehicle dynamics model, a filtering algorithm is used to calculate the road surface adhesion coefficient in real time, providing crucial support for subsequent road surface adhesion coefficient prediction. Specific sub-steps are as follows: Step 2.1: Define the vehicle state vector, construct an extended distributed multi-axle vehicle state space model that includes the road adhesion coefficient, adopt a simplified tire force model, establish the mapping relationship between tire force and road adhesion coefficient, so that the filter can reverse correct the road adhesion coefficient estimate by observing the residual between the actual acceleration of the target vehicle and the acceleration predicted by the model.
[0072] In this embodiment, vehicle-related parameters, including wheelbase, track width, and tire stiffness, are obtained through direct measurement, parameter identification, and consultation of relevant design manuals; [Definition] Vehicle state vector at time (current time) ,in and The target vehicles The longitudinal and lateral velocities at time t. For the target vehicle yaw rate at time t, for The estimated value of the road surface adhesion coefficient at time t. The discrete state transition equation for the target vehicle is established as follows: ,in, for The vehicle state vector at time t. for The control input vector at any given time includes the front wheel steering angle. Drive or braking torque, etc.; This is the state transition function; for The process noise at any given time follows a mean of 0 and a covariance of 0. The Gaussian distribution.
[0073] Step 2.2: Design an extended Kalman filter algorithm to perform state estimation for the nonlinear system.
[0074] In this embodiment, a physical quantity that can be directly measured by the vehicle-mounted sensor is selected as the observation vector to construct the observation function. The state transition function and the measurement function are calculated with respect to the state vector. The Jacobian matrix is then calculated. The prior state, prior covariance, and Kalman gain are then computed. Finally, the state estimate and posterior error covariance are updated using the measurement residuals. The updated state vector is then used to calculate the prior state, prior covariance, and Kalman gain. Extract the fourth component as Estimated value of road surface adhesion coefficient at time 1 , which serves as a dynamic constraint parameter for terrain hazard assessment and route planning in step 3.
[0075] In this embodiment of the application, an observation vector is constructed, which consists of physical quantities directly measured by vehicle-mounted sensors. Its measurement equation is ,in, For measurement function, for Measurement noise at time (covariance is) Calculate the Jacobian matrix: ,in, and These are the state transition Jacobian matrix and the observation Jacobian matrix, respectively. and For the Jacobian operator, and They are respectively Post-hoc state estimation and Prior state estimation at time step. During the prediction phase, calculate the prior state estimate and the prior error covariance: , ,in, Given the posterior covariance matrix, calculate the Kalman gain. And update the state estimate and posterior error covariance: , , extract The fourth component is used as the current road surface adhesion coefficient estimate. ,in, It is an identity matrix.
[0076] Step 3: By fusing visual geometric features with dynamic estimates, predict the road adhesion coefficient of the untraveled area ahead and quantify the roll risk on the future path, addressing the limitation of single-sensor perception in off-road environments. Specific sub-steps are as follows: Step 3.1: Time-align the sequence of terrain passage risk costs obtained in Step 1 and the sequence of road surface adhesion coefficients of the current driving area obtained in Step 2.
[0077] In this embodiment of the application, the terrain access risk cost sequence output in step 1 is set as follows: The road surface adhesion coefficient estimation sequence output in step 2 is as follows: An interpolation method is used to synchronize the two to a global timestamp. This yields the synchronization sequence data.
[0078] Step 3.2: Construct a multilayer perceptual neural network model to achieve spatial mapping from off-road terrain drivability assessment to road surface adhesion coefficient. Then, project the estimation results onto an off-road elevation map to generate a terrain representation of distributed vehicle roll estimation for future time windows, serving as the spatial distribution prediction result of the road surface adhesion coefficient. Here, the terrain representation refers to the result after projecting the road surface adhesion coefficient onto the off-road elevation map, including both road surface adhesion coefficient information and elevation information.
[0079] In this embodiment, the multilayer perceptron model employs a multilayer perceptron (MLP). The MLP model is trained based on synchronous data. The input to this MLP model is a normalized scalar terrain access risk cost, and the output is the corresponding road surface adhesion coefficient. First, a real-world dataset is constructed, containing access risk cost input features and road surface adhesion coefficient label pairs. Then, by designing a loss function for the neural network and utilizing a backpropagation parameter update strategy, the neural network parameters are updated to meet the mapping accuracy requirements, achieving a spatial mapping from off-road terrain accessibility assessment to road surface adhesion coefficients. The predicted road surface adhesion coefficients are then used to... Projected onto a bird's-eye view elevation map, an enhanced terrain map is generated. Since the target vehicle is a multi-axle off-road vehicle, for the first... An axis, whose attribute vector is defined as ,in, For multi-axle off-road vehicles Axis tilt eigenvector, For multi-axle off-road vehicles The roll angle state quantity corresponding to the axis. For multi-axle off-road vehicles Shaft side tilt rate, For multi-axle off-road vehicles The road surface adhesion coefficient corresponding to the axis.
[0080] In step 3, due to differences in sampling frequency and transmission delay between the visual and inertial measurement unit data in step 1 and the vehicle dynamics observation data in step 2.1, a unified time reference needs to be established. Therefore, the terrain access risk cost sequence output in step 1 is set as follows: The terrain access risk cost and its timestamp are respectively and The road surface adhesion coefficient estimation sequence output in step 2 is as follows: The estimated values of the road surface adhesion coefficient and their timestamps are respectively... and Linear interpolation or spline interpolation methods are used to synchronize the two to the global time series. , is represented as: ; ; in, and Synchronization time Interpolated roll angular velocity and road adhesion coefficient This is the interpolation function.
[0081] In the construction of the multilayer perceptron neural network model, the training set needs to be obtained through testing in a known terrain scenario. First, the target vehicle drives in the current terrain, acquiring data from its internal and external sensors, and obtaining the terrain hazard assessment data according to step 1; then, using step 2, the road surface adhesion coefficient estimate of the off-road terrain is obtained, and the synchronized data of the currently driven area is obtained through timestamp alignment. and As training samples, they are input into a multilayer perceptual neural network for training, and output the road adhesion coefficient of the untraveled area ahead. .
[0082] The road adhesion coefficient predicted in step 3.2 Projecting this onto the bird's-eye view elevation map constructed in step 1 generates an enhanced terrain map containing both geometric and physical attributes. For enhanced terrain maps Each grid cell in Its attribute vector is defined as ,in, This is the average elevation of the grid cell. The cost of the original terrain access risk for this grid cell. This is the mapped road surface adhesion coefficient for this grid cell.
[0083] Step 4: Decompose the vehicle control system into a hierarchical multi-agent system. By defining communication rules between agents, distributed collaborative computing is achieved, reducing the central computing load and improving system robustness, serving as the foundation for subsequent distributed observer design. Specific sub-steps are as follows: Step 4.1: Divide the state space model of the distributed drive off-road vehicle into two categories with a total of nine agents. The first category of agents is the vehicle agent, which is used to represent the six-degree-of-freedom dynamic state of the entire vehicle and is responsible for global state coordination; the second category of agents is the wheel agent, which is used to represent the rotational dynamic state of each wheel.
[0084] In this embodiment, a typical off-road vehicle with four axles and eight wheels (four axles, one wheel on each side of each axle, for a total of eight wheels) is used as an example. Its vehicle control system is divided into one vehicle agent (responsible for the global six-degree-of-freedom state) and eight wheel agents (responsible for the rotational dynamics of a single wheel). The dynamic equations of the wheel agents are modeled as follows: ; in, Indicates the first The moment of inertia of each wheel Indicates the first The angular acceleration of each wheel, Indicates the first The driving or braking torque (control input) of each wheel. Indicates the first The longitudinal tire force of each wheel For the wheel radius, Indicates the first The rotational damping coefficient of each wheel, Indicates the first The angular velocity of each wheel Indicates the first Additional control inputs or disturbance compensation items for each wheel.
[0085] Step 4.2: Construct the communication topology between agents. The four wheel agents on the left can communicate with each other, and the four wheel agents on the right can communicate with each other. The wheel agents on the left and right sides cannot communicate directly with each other. All wheel agents can communicate with the vehicle agent. This communication network architecture reduces the complexity of the communication network between agents.
[0086] In this embodiment of the application, an undirected graph is constructed. This represents the topological connections between agents. Left wheel set. Internally communicable, right-side wheel assembly Internal communication is possible, but direct communication is not possible between the left and right wheels; however, both wheels can communicate with the vehicle's intelligent agent (node 0). This communication mechanism is bidirectional. The vehicle's intelligent agent model is represented as follows: ; in, and These represent the longitudinal velocity, lateral velocity, and yaw rate of the target vehicle, respectively. and They are respectively and Time derivative, Indicates the heading angle of the target vehicle. for Time derivative, and These represent the X-axis velocity and Y-axis velocity of the vehicle's center of mass in the geodetic inertial coordinate system, respectively, and L represents the wheelbase. Indicates the first The distance from each axle to the center of gravity. For the turning angle of the wheel, The curb weight of the target vehicle. For mechanical transmission efficiency, For the wheel radius, Indicates the air drag coefficient. The rolling resistance coefficient, and These are the sum of the output torques of all wheels on the left and all wheels on the right of the target vehicle, respectively. Indicates the first Lateral stiffness of axle, wheel, and tire. It is the acceleration due to gravity. Let yaw moment be the moment of inertia of the vehicle about the vertical z-axis of its center of mass.
[0087] No. The dynamic equations of the individual wheel-shaped intelligent agent are modeled as follows: ; in, For the first The rotational speed of each wheel To apply to the first Torque of each wheel and They are respectively and The derivative, Indicates the first The moment of inertia of each wheel It is a time constant. To control the input, The coefficient of friction between the road surface and the tire is the slip ratio. With road surface feature parameter vector The function, For the first Vertical load on each wheel.
[0088] To reduce communication overhead between agents, the topological communication connection between the vehicle agent and the wheel agent is constructed using an undirected graph. It indicates. Among them. This represents a set of intelligent agents, where agent 0 represents the vehicle intelligent agent. This represents eight wheel agents: Wheel Agent 1, Wheel Agent 2, Wheel Agent 3, Wheel Agent 4, Wheel Agent 5, Wheel Agent 6, Wheel Agent 7, and Wheel Agent 8. The wheels on the left and right sides of the target vehicle are divided into two sets. and .in, The set of intelligent agents for the left wheel. This is the set of intelligent agents for the right wheel. Edges represent communication connections between agents, and the communication architecture between agents is achieved through adjacency matrices. express: ; in, Let be the adjacency matrix corresponding to the communication topology between agents, and , For the number of agents, if but ,like but , Represents the adjacency matrix of the nth element. line, number Columns of matrix elements.
[0089] Step 5: Construct a tire model for off-road scenarios. Based on the Burckhardt tire model, the vertical load on the wheel is divided into two parts: nominal vertical load and additional vertical load. The nominal vertical load is the load transfer caused by the lateral and longitudinal acceleration of the target vehicle on a flat road surface. The additional vertical load is the change in external additional vertical load caused by off-road terrain. The sum of the two is the total vertical load on the wheel. Based on the principle of virtual displacement, a nominal vertical load distribution model for the vehicle agent is constructed to obtain the calculation results of the nominal vertical load.
[0090] In this embodiment of the application, a Burckhardt tire model oriented towards off-road terrain is established, and the longitudinal tire force of this model is expressed as follows: ,in, For the first The road-tire adhesion coefficient of each wheel is the result of calculation using the Burckhardt tire model. This is a vector of road surface feature parameters, belonging to the parameters of the Burckhardt tire model. For the first The slip ratio of each wheel Output the deviation between the tire force and the actual tire force for the Burckhardt tire model. Indicates the cause of the terrain. Vertical load variation of each wheel For the first The nominal vertical load of each wheel. The lateral acceleration of the target vehicle is collected. and longitudinal acceleration Based on the curb weight of the target vehicle Calculate the virtual work caused by external forces on the target vehicle. Then, based on the deformation of the suspension, the virtual work caused by the suspension deformation is calculated. The nominal vertical load on the target vehicle is calculated using the principle of virtual displacement. ,in, For dynamic observation vectors, For virtual displacement, This represents the transpose of the observation mapping matrix. This is the stiffness coefficient of the suspension. This represents the product of the observation mapping matrix and the suspension stiffness coefficient.
[0091] In step 5, the Burckhardt tire model is modified to adapt to off-road terrain, and the longitudinal tire force of each wheel is adjusted. Represented as Assuming Sufficiently small and negligible, the Burckhardt tire model is rewritten as: ; in, Indicates the first The change in the number of wheels.
[0092] The calculation of the nominal vertical load of a multi-axle independently driven vehicle is a statically indeterminate problem. Based on the principle of virtual displacement, the gravity acting on the target vehicle on a flat road surface is defined as a dynamic observation vector containing roll moment and pitch moment. ,in, These are the lateral acceleration and longitudinal acceleration of the target vehicle, respectively. Let the height of the target vehicle's center of gravity be defined. Define the virtual displacement. ,in, Let represent the yaw, pitch, and roll angles of the target vehicle, respectively, which can be obtained through an inertial measurement unit. Then, the virtual work caused by external forces is... The virtual work caused by suspension deformation is ,in, This represents the transpose of the observation mapping matrix, which is mainly determined by the size of the target vehicle, and is expressed as: ; in, Indicates the first The distance from each axle to the center of gravity is based on the principle of virtual displacement. The nominal vertical load of each wheel was calculated. .
[0093] Step 6: Using the communication topology from Step 4 and the Burckhardt tire model from Step 5, design a distributed observer to calculate the dynamic vertical loads under off-road terrain in real time. Specific sub-steps are as follows: Step 6.1: For ,in, This represents the set of wheels deemed valid for risk assessment, where each wheel is primarily influenced by its own motion state. A fixed-time convergent state observer is constructed, utilizing only the rotational speed. and slip ratio Information estimates the load drop. In this embodiment, the nominal vertical load is... By incorporating the Burckhardt tire model, a fixed-time convergent state observer is established to... Observations were conducted to obtain the additional vertical load caused by the off-road terrain.
[0094] Step 6.2: Based on the real-time load change trend estimated by the wheel agent. All wheel agents are dynamically divided into two sets: the set of wheel agents with reduced vertical load. and the collection of intelligent agents with increased vertical load Among them, the wheel intelligent agent set with reduced vertical load. : Typically, this refers to a wheel suspended in mid-air over an obstacle; or an ensemble of intelligent agents experiencing increased vertical loads. : This is usually the wheel that runs over an obstacle.
[0095] This application embodiment divides the wheel smart agents into sets based on the load change trend, indicating a decrease in vertical load. (Such as suspended wheels) and an ensemble of intelligent agents with increased vertical loads. (e.g., a wheel running over an obstacle). Compare the calculated vertical load of all wheels with the nominal vertical load. If the load decreases, classify them as wheels with reduced vertical loads. Otherwise, it is classified as a set of agents with increased vertical load. middle.
[0096] The nominal vertical load estimation results are as follows Figure 4 As shown, where, Figure 4 (a) in the figure represents the nominal vertical load estimation result corresponding to wheel agent 1. Figure 4 (b) in the figure represents the nominal vertical load estimation result corresponding to wheel agent 2. Figure 4 (c) in the figure represents the nominal vertical load estimation result corresponding to wheel agent 3. Figure 4 In the figure, (d) represents the nominal vertical load estimation result corresponding to wheel agent 4. Figure 4 In the figure, (e) represents the nominal vertical load estimation result corresponding to wheel agent 5. Figure 4 In the figure, (f) represents the nominal vertical load estimation result corresponding to wheel agent 6. Figure 4 In the figure, (g) represents the nominal vertical load estimation result corresponding to wheel agent 7. Figure 4 In the equation (h), the nominal vertical load estimation result corresponds to wheel agent 8. Figure 4 The blue line represents the actual value, and the red line represents the estimated value.
[0097] In step 6.2, the first step is to define the... The observation error between the actual wheel speed and the estimated wheel speed of each wheel. ,in This represents the estimated wheel speed. This represents the actual wheel speed. To ensure observer convergence, the observation error... The design is as follows: ; in, As an auxiliary error state variable, , They are respectively , Time derivative, These are all design parameters for the observer, and All are greater than 0. Greater than 1; This is the time derivative of the lumped disturbance in the first channel.
[0098] Definition of the first The observation error between the actual torque and the observed torque of each wheel , Indicates the first The observed torque of each wheel Indicates the first The actual torque of each wheel. Observation error. and The derivative is: ; in, For second-order auxiliary error state variables, , They are respectively , The time derivative; These are all design parameters for the observer, and , ; This is the time derivative of the lumped disturbance in the second channel.
[0099] definition , The observer is designed as follows: ; in, , These are the auxiliary states inside the first channel of the observer and the auxiliary states inside the second channel of the observer, respectively. , , , They are respectively , , , The time derivative; This is the time derivative of the equivalent disturbance torque.
[0100] Based on the designed observer, the first The longitudinal tire force error of each wheel can be calculated. The result is obtained and converges within a fixed time. The range of vertical load error variation caused by terrain is calculated as follows: ; in, and They are respectively The upper and lower bounds can be determined experimentally. This represents the predicted spatial distribution of the road surface adhesion coefficient. The wheel agent, through agent communication, sends the calculated vertical load variation to the vehicle agent's nominal vertical load result for compensation, combining this with the predicted spatial distribution of the road surface adhesion coefficient. The compensated vertical load can be obtained. : ; in, This is the tire slip ratio threshold.
[0101] Step 6.3: For The increase in load originates from the load transfer of other wheels. Using the communication topology defined in step 4, the total vertical load decrease calculated in step 6.2 is calculated. Then, using the principle of minimum potential energy, the total decrease in vertical load is redistributed to other wheel agents. Combined with the spatial distribution prediction results of the road surface adhesion coefficient, the real-time vertical load of each wheel running in off-road terrain is finally obtained.
[0102] In this embodiment of the application, a set of wheel smart agents with reduced vertical load is included. The sum of the load decreases on all wheels is used to construct an optimization problem based on the principle of minimum potential energy, and then rationally distributed among the intelligent agents whose vertical loads are increasing. This ensures that the overall vertical load remains constant.
[0103] In step 6.3, the redistribution of vertical loads first redefines the vertical loads based on the agent partitioning in step 6.1. and balance matrix ,in, This indicates that the load on a wheel in a low slip ratio state needs to be redistributed, i.e. Vertical load on the wheel; The vertical load calculated in step 6.2 does not need to be redistributed. Vertical loads in the system. This represents the first dimension of the roll risk characteristic component, corresponding to the vehicle roll dynamics characteristics. Let be the second-dimensional roll risk characteristic component, corresponding to the wheel load transfer characteristic. Based on the principle of minimum potential energy, the following optimization problem is constructed to minimize the elastic potential energy of the suspension: ; Among them, the stiffness matrix of the suspension Suspension stiffness and Also in accordance with and The partitioned diagonal matrix. Solving the optimization problem using the Lagrange multiplier method, the required redistribution of the wheel vertical loads is obtained as follows: ; in, This is a regularization term used to avoid numerical instability during matrix inversion. It is combined with the spatial distribution prediction results of the road surface adhesion coefficient. The final compensated vertical load The calculation formula is: ; in, Indicates the first The load is redistributed to a wheel that is in a low slip ratio state.
[0104] Figure 5 The total vertical load estimation results are shown, where, Figure 5 (a) in the figure represents the total vertical load estimation result corresponding to wheel agent 1. Figure 5 (b) in the figure represents the total vertical load estimation result corresponding to wheel agent 2. Figure 5 (c) in the figure represents the total vertical load estimation result corresponding to wheel agent 3. Figure 5 In the figure, (d) represents the total vertical load estimation result corresponding to wheel agent 4. Figure 5 In the figure, (e) represents the total vertical load estimation result corresponding to wheel agent 5. Figure 5 In the figure, (f) represents the total vertical load estimation result corresponding to wheel agent 6. Figure 5 In the figure, (g) represents the total vertical load estimation result corresponding to wheel agent 7. Figure 5 In the figure, (h) represents the total vertical load estimation result corresponding to wheel agent 8. Figure 5 The blue line represents the true value, and the red line represents the estimated value. This represents the actual vertical load on the vehicle. This indicates the vehicle vertical load calculated by the method in the embodiments of this application.
[0105] Step 7: Integrate off-road surface perception results, construct a roll hazard assessment index, establish the index's boundary range, and build a roll hazard scoring model. Specific sub-steps are as follows: Step 7.1 Definition and calculation of roll safety index.
[0106] In this embodiment, rollover time and lateral load transfer rate are selected as core evaluation indicators to characterize the dynamic stability prediction capability and current static load distribution of the target vehicle, respectively.
[0107] In this embodiment of the application, inertial measurement unit data of the target vehicle is collected, including roll rate. and roll acceleration Then, the roll angle can be obtained by integration. Based on the recorded current roll angle The rollover time is calculated in real time, and the calculation formula is as follows: ; in, Indicates the time of rollover. The current roll angle of the target vehicle. , These are the target vehicle's current roll rate and roll acceleration, respectively. This represents the maximum permissible roll threshold. Roll angle, roll rate, and roll acceleration can all be directly measured or indirectly calculated using the onboard inertial measurement unit. The smaller the rollover time of the target vehicle, the closer the target vehicle is to a rollover state. This is based on the real-time vertical loads of each wheel output by the distributed cooperative observer in step 6. The lateral load transfer rate is obtained by dividing the difference in the total vertical reaction force between the left and right tires by the total value, and is expressed as: ; in, For lateral load transfer rate, These represent the wheel numbers. , , The first Vertical load of the first wheel, the first wheel Vertical load of the first wheel, the first wheel The vertical load of each wheel. The lateral load transfer rate ranges from [0, 1]. When the lateral load transfer rate approaches 1, it indicates that one wheel is about to leave the ground.
[0108] Step 7.2: Considering the friction coefficient of off-road surfaces The direct impact of the target vehicle's current speed on the vehicle's limit state is considered, and instead of using a fixed threshold, the road adhesion coefficient estimated in step 2 is used. Boundary determination is introduced based on the target vehicle speed to achieve dynamic adaptive adjustment of the boundary.
[0109] In this embodiment of the application, in the dynamic adaptive boundary design of the lateral load transfer rate, the reference boundary under a flat and dry road surface is first established. Then, based on real vehicle experiments, a low-addition coefficient penalty factor was established. The road surface observation results are imported into the dynamic safety boundary design to obtain the dynamic safety boundary for lateral load transfer rate. In the design of the rollover time dynamic safety boundary, the minimum safe response time window is first established. Then, the penalty factor was established based on real-vehicle experiments. longitudinal vehicle speed Importing this into the dynamic safety boundary design, the final result is the rollover time dynamic safety boundary.
[0110] In dynamic adaptive boundary design, the boundaries of lateral load transfer rate and rollover time dynamically change with the target vehicle's speed and road adhesion coefficient, among which the dynamic safety boundary of lateral load transfer rate... Represented as: ; in, As a reference boundary under a flat, dry road surface, the embodiments of this application The value is 0.9; This is a low-adjustment penalty factor, calibrated through real vehicle tests. The slipperier the road surface, the smaller the allowable lateral load transfer rate margin.
[0111] In the design of the rollover time dynamic safety boundary, the faster the vehicle speed, the faster the target vehicle will experience irreversible rollover, and the less time the control system has to react. Therefore, a larger prediction time window needs to be reserved as the rollover time dynamic safety boundary. , is represented as: ; in, For the minimum safe response time window, As a penalty factor, The longitudinal speed is the maximum speed required to allow for a longer rollover time.
[0112] Step 7.3: Calculate the relative distance between the current indicator value and the aforementioned dynamic safety boundary, and convert it into a standardized single-indicator risk contribution. A higher value indicates a higher risk. To address the differences in sensor signal-to-noise ratio under various working conditions in off-road scenarios, a dynamic weighting allocation ratio is constructed. This ratio is then used to fuse the risk contribution of a single indicator corresponding to the lateral load transfer rate and the risk contribution of a single indicator corresponding to the rollover time, resulting in a final comprehensive rollover risk score. Based on the comprehensive roll risk score... The numerical range is used to classify the danger level and map it to the corresponding control strategy.
[0113] In this embodiment, based on the dynamic safety boundary designed in step 7.2 and the current state of the target vehicle, the lateral load transfer rate and the risk contribution of the single index of rollover time are calculated, and then the signal quality weight of the inertial measurement unit is calculated. and confidence level of vertical load estimation ,ensure The final comprehensive roll risk score was obtained. .
[0114] Among them, the risk contribution of a single indicator based on the lateral load transfer rate Represented as: ; In the design of a single-indicator risk contribution based on rollover time, since a smaller rollover time is more dangerous, it is reversed to a single-indicator risk contribution based on rollover time. , is represented as: ; in, and All values are uniformly within the interval [0, 1], where 0 represents absolute safety and 1 represents reaching the limit boundary.
[0115] In this embodiment, the signal quality weight of the inertial measurement unit is set as follows: The confidence level for the vertical load estimation is ,and Overall roll risk score results The calculation formula is: ; in, and This represents a linear weighted fusion of the risk contribution of a single indicator corresponding to the lateral load transfer rate and the risk contribution of a single indicator corresponding to the rollover time. This is a contradictory correction term. This is a correction factor. When there is a significant deviation between the calculated rollover time and lateral load transfer rate (i.e., they cannot corroborate each other), this contradictory correction term increases, indicating that the system may have sensor malfunctions or be in a highly unstable transient state, requiring increased control conservatism. When the difference between the single-index risk contribution obtained by converting the lateral load transfer rate and the rollover time exceeds a preset deviation threshold, it is determined that there is a significant deviation between the two evaluation indicators, and the contradictory correction term is activated at this time. The contradictory correction term increases as the deviation between the two increases, thereby improving the conservatism of the final comprehensive rollover risk score.
[0116] In this embodiment of the application, the comprehensive roll risk score is used as the basis. The numerical range is used to classify the hazard level and map it to the corresponding control strategy, specifically including: (1) When At that time, the area is divided into safety zone levels, and the corresponding control strategy is to maintain normal driving control.
[0117] (2) When When the vehicle is in a high-risk area, it is classified as a warning zone, and the corresponding control strategy is to limit the maximum speed and activate the active suspension pre-adjustment.
[0118] (3) When When the area is classified as a danger zone, the corresponding control strategy is to trigger the emergency avoidance logic and directly intervene in differential braking and motor torque vector control through the multi-agent framework in step 4 to suppress roll.
[0119] This application's embodiments establish a road surface condition prediction model based on deep neural networks, based on the deep fusion of measurement results from external and internal sensors. By establishing a distributed state observer, it achieves accurate perception and prediction of the state of each wheel of a distributed multi-axle vehicle, constructs a dynamic rollover warning boundary, and lays the foundation for anti-rollover design of multi-axle vehicles. It has the following technical advantages: (1) By using the vertical acceleration power spectral density of the inertial measurement unit to construct a physically reliable risk monitoring label, and combining vehicle speed and bird's-eye view visual geometric features, the cost of continuous passage through off-road terrain without manual labeling is accurately quantified.
[0120] (2) The extended Kalman filter algorithm was used to realize the online real-time calculation of the road surface adhesion coefficient. By fusing visual geometry and dynamic estimation values, a neural network mapping model from terrain accessibility to road surface adhesion coefficient was constructed.
[0121] (3) Based on a distributed fixed-time state observation network of multiple agents, the results of terrain perception are integrated to realize the real-time acquisition of the vertical load of each wheel.
[0122] (4) The rollover time and lateral load transfer rate are selected as multi-dimensional core evaluation indicators. The road surface adhesion coefficient and vehicle speed are used to realize the dynamic adaptive adjustment of the safety boundary. By considering the dynamic weight fusion mechanism of the working condition signal-to-noise ratio, the standardized rollover hazard is accurately quantified and classified according to the control strategy.
[0123] In one exemplary embodiment, a roll hazard detection device for multi-axle off-road vehicles is provided. This device can be a computer device, which may be a server or a terminal, and its internal structure diagram can be as follows: Figure 6 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores the vehicle's propagation perception data, visual geometry data, and vehicle speed information during off-road driving. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for detecting the roll hazard of a multi-axle off-road vehicle.
[0124] Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0125] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0126] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0127] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.
[0128] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0129] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0130] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0131] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for detecting the roll hazard of a multi-axle off-road vehicle, characterized in that, include: Acquire the target vehicle's body perception data, visual geometry data, and vehicle speed information during its journey through off-road terrain; Based on the ontology perception data, the visual geometry data, and the vehicle speed information, the terrain access risk cost of the off-road terrain where the target vehicle is located is determined. Specifically, this includes: extracting the vertical acceleration of the target vehicle using inertial measurement unit data from the ontology perception data; calculating and integrating the power spectral density of the vertical acceleration within a preset frequency band to obtain the frequency band power; using the frequency band power to determine the terrain access risk cost sequence, which is used as a supervision label for training the convolutional neural network model; registering and projecting the image information and dense point cloud information from the visual geometry data onto a bird's-eye view space to generate a multi-channel local map containing color features, elevation features, and an effective area mask; performing Fourier feature mapping on the vehicle speed information to obtain speed features; extracting the visual geometry features of the multi-channel local map based on the convolutional neural network model; concatenating the speed features with the visual geometry features and inputting the concatenated features into a fully connected layer to output a normalized scalar terrain access risk cost as the terrain access risk cost. Based on the aforementioned body perception data, combined with the vehicle dynamics model, a filtering algorithm is used to calculate the road surface adhesion coefficient of the target vehicle's current driving area in real time. Based on the terrain travel risk cost and the road surface adhesion coefficient of the current driving area, predict the spatial distribution of the road surface adhesion coefficient in the untraveled area in front of the target vehicle. The vehicle control system of the target vehicle is divided into one vehicle intelligent agent and multiple wheel intelligent agents, and a communication topology between the vehicle intelligent agent and the multiple wheel intelligent agents is constructed. A tire model for off-road scenarios is constructed, and the dynamic vertical load of each wheel in the multiple wheel agents is estimated in real time by combining the communication topology and the spatial distribution prediction results of the road surface adhesion coefficient. The roll risk assessment index is calculated based on the dynamic vertical load of each wheel and the body sensing data, and the dynamic safety boundary of the roll risk assessment index is determined according to the road adhesion coefficient of the current driving area and the vehicle speed information. Based on the relative distance between the roll hazard assessment index and the dynamic safety boundary, the comprehensive roll hazard score of the target vehicle is determined.
2. The method for detecting the roll hazard of multi-axle off-road vehicles according to claim 1, characterized in that, Based on the aforementioned ontological perception data and combined with the vehicle dynamics model, a filtering algorithm is used to calculate the road surface adhesion coefficient of the target vehicle's current driving area in real time, specifically including: Define a vehicle state vector containing the road adhesion coefficient, and construct a nonlinear state-space model of the target vehicle; the tire force in the nonlinear state-space model is related to the road adhesion coefficient. The physical quantities that can be directly measured by the vehicle-mounted sensors in the ontological perception data are selected as observation vectors. The extended Kalman filter algorithm is used to estimate the state of the nonlinear state-space model. The estimated value of the road adhesion coefficient is corrected by the residual between the actual acceleration of the observed vehicle and the acceleration predicted by the model. The road adhesion coefficient of the current driving area is extracted from the updated state vector.
3. The method for detecting the roll hazard of multi-axle off-road vehicles according to claim 1, characterized in that, Based on the terrain-related traffic risk cost and the road surface adhesion coefficient of the current driving area, the spatial distribution prediction result of the road surface adhesion coefficient in the untraveled area ahead of the target vehicle is specifically included: The sequence of terrain passage risk costs is time-aligned with the sequence of road surface adhesion coefficients in the current driving area to obtain synchronous training samples; A multilayer perceptron neural network model is constructed, and the multilayer perceptron neural network model is trained based on the synchronous training samples. A spatial mapping relationship is established from the terrain passage risk cost to the road surface adhesion coefficient. Based on the multilayer perceptron neural network model and the spatial mapping relationship, the road surface adhesion coefficient of the untraveled area ahead is predicted. The road surface adhesion coefficient of the untraveled area ahead is projected onto a bird's-eye view elevation map to generate an enhanced terrain map containing elevation information, the terrain traffic risk cost, and the road surface adhesion coefficient, which serves as the spatial distribution prediction result of the road surface adhesion coefficient.
4. The method for detecting the roll hazard of multi-axle off-road vehicles according to claim 1, characterized in that, The vehicle control system of the target vehicle is divided into one vehicle agent and multiple wheel agents, and a communication topology is constructed between the vehicle agent and the multiple wheel agents, specifically including: The six-degree-of-freedom dynamic state of the target vehicle is divided into a vehicle intelligent agent, and the rotational dynamic state of each wheel is divided into a wheel intelligent agent. Based on the vehicle agent and each of the wheel agents, the communication topology is constructed; wherein, the wheel agents located on the left side of the target vehicle can communicate with each other, the wheel agents located on the right side of the target vehicle can communicate with each other, the wheel agents located on the left side of the target vehicle and the wheel agents located on the right side of the target vehicle cannot communicate directly, and all wheel agents can communicate with the vehicle agent.
5. The method for detecting the roll hazard of a multi-axle off-road vehicle according to claim 1, characterized in that, A tire model for off-road scenarios is constructed. Combining the communication topology and the spatial distribution prediction results of the road surface adhesion coefficient, the dynamic vertical load of each wheel in the multiple wheel agents is estimated in real time. Specifically, this includes: A tire model for off-road scenarios is constructed, and the vertical load on each wheel is decomposed into nominal vertical load and additional vertical load. The nominal vertical load is the load transfer caused by vehicle acceleration on a flat road surface, and the additional vertical load is the change in external additional load caused by off-road terrain. Based on the principle of virtual displacement, the nominal vertical load of each wheel is calculated according to the longitudinal acceleration, lateral acceleration and vertical acceleration of the target vehicle. The additional vertical load of each wheel is estimated using a distributed state observer based on the communication topology, and the additional vertical load is corrected by combining the spatial distribution prediction results of the road adhesion coefficient to obtain the corrected additional vertical load. The nominal vertical load is added to the modified additional vertical load to obtain the dynamic vertical load of each wheel.
6. The method for detecting the roll hazard of a multi-axle off-road vehicle according to claim 1, characterized in that, The rollover hazard assessment indicators include: lateral load transfer rate and rollover time; the lateral load transfer rate is used to characterize the current static load distribution of the target vehicle, and the rollover time is used to characterize the dynamic stability prediction capability of the target vehicle. The roll hazard assessment index is calculated based on the dynamic vertical load of each wheel and the vehicle body sensing data. The dynamic safety boundary of the roll hazard assessment index is determined based on the road surface adhesion coefficient of the current driving area and the vehicle speed information. Specifically, this includes: Based on the dynamic vertical load of each wheel, the total vertical load of the left wheel and the total vertical load of the right wheel of the target vehicle are determined. The lateral load transfer rate is calculated based on the ratio of the difference between the total vertical load of the left wheel and the total vertical load of the right wheel to the sum of the total vertical loads of the left wheel and the right wheel. The vehicle roll angle, roll velocity, and roll acceleration are obtained based on the inertial measurement unit data in the body perception data, and the rollover time is calculated based on the vehicle roll angle, the roll velocity, and the roll acceleration. The dynamic safety boundary of the lateral load transfer rate is determined based on the road surface adhesion coefficient of the current driving area; the dynamic safety boundary of the lateral load transfer rate decreases as the road surface adhesion coefficient of the current driving area decreases. Based on the vehicle speed information, the dynamic safety boundary of the rollover time is determined; the dynamic safety boundary of the rollover time increases as the vehicle speed increases.
7. The method for detecting the roll hazard of a multi-axle off-road vehicle according to claim 6, characterized in that, Based on the relative distance between the roll hazard assessment index and the dynamic safety boundary, the comprehensive roll hazard score of the target vehicle is determined, specifically including: The relative distances between the lateral load transfer rate and the rollover time and their respective dynamic safety boundaries are calculated, and the relative distances are converted into single-index risk contribution values between 0 and 1 to obtain the single-index risk contribution values corresponding to the lateral load transfer rate and the single-index risk contribution values corresponding to the rollover time. The risk contribution of the single index corresponding to the lateral load transfer rate and the risk contribution of the single index corresponding to the rollover time are weighted and fused, and a contradiction correction term is introduced to correct the risk assessment result when there is a significant deviation between the two rollover hazard evaluation indicators, so as to obtain the comprehensive rollover hazard score result. The risk level is determined based on the numerical range of the comprehensive roll risk score, and the risk level is mapped to the corresponding vehicle control strategy.
8. A roll hazard detection device for multi-axle off-road vehicles, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the method for detecting the roll hazard of a multi-axle off-road vehicle according to any one of claims 1-7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for detecting the roll hazard of a multi-axle off-road vehicle as described in any one of claims 1-7.
Citation Information
Patent Citations
Forest environment unmanned vehicle navigation method and system based on multi-source data fusion perception
CN120651263A
Active anti-rollover control method and system for six-axle large vehicle
CN122519244A
