Mine area vehicle speed planning method fusing AI algorithm and multi-modal data

Abstract

This invention relates to the field of mining area traffic planning and discloses a method for planning vehicle speed in mining areas that integrates AI algorithms and multimodal data. The method includes the following steps: S1, determining that the current vehicle is in a following state and collecting the current vehicle's driving parameters, the position of the leading vehicle, and the distance between the two vehicles; S2, based on the positions of the current vehicle and the leading vehicle, determining whether the leading vehicle is traveling on a slope or about to leave the slope; S3, if the leading vehicle is on a slope or about to leave the slope, determining a safe following distance based on the distance between the two vehicles and the leading vehicle's slope travel distance; S4, planning the current vehicle's driving speed. This invention identifies vehicle following patterns and collects multimodal data in real time and comprehensively. It utilizes AI algorithms to perform high-precision perception and forward-looking prediction of the leading vehicle's slope travel status, enabling the solution to fully understand the complex terrain environment currently and in the future, thus allowing for pre-emptive risk assessment.

Description

Technical Field

[0001] This invention relates to the field of mining area traffic planning technology, and in particular to a mining area vehicle speed planning method that integrates AI algorithms and multimodal data. Background Technology

[0002] With the deepening of intelligent and unmanned construction in mines, autonomous driving technology for mining vehicles, especially the automatic following technology for mining transport vehicles, has become a core direction for improving transportation efficiency and safety. However, the complexity, variability, and harshness of the mining environment pose severe challenges to existing vehicle following technologies that are not applicable to traditional road environments.

[0003] Currently, in the field of autonomous vehicles in mining, common following systems often involve the following vehicle maintaining a fixed distance or time interval to follow the lead vehicle. The main drawback of this approach is that it ignores the unique and complex terrain of mining areas, particularly the impact of slopes on vehicle driving characteristics and braking performance. On or off slopes, a vehicle's power response and braking distance change non-linearly, and a fixed following strategy cannot effectively adapt to these dynamic changes. This leads to insufficient safe distance between the following and lead vehicles, increasing the risk of collisions. Especially in transitional areas such as slope entrances and exits, where vehicle states change most drastically, traditional following methods struggle to provide adequate safety margins in such high-risk scenarios.

[0004] Furthermore, existing vehicle-following technologies also have limitations in environmental perception. Some systems rely on a single sensor to acquire limited environmental information, making it difficult to comprehensively and accurately perceive the complex real-time road conditions in mining areas, such as the precise geometric features of slopes, the rate of change of slope, and road surface adhesion conditions. In harsh environments such as dusty, muddy, and drastically changing lighting conditions in mining areas, the perception capability of a single sensor is further limited, easily leading to delayed or inaccurate acquisition of environmental information. This, in turn, affects the timeliness and reliability of vehicle-following decisions, making it impossible to effectively predict risks before they occur, thus increasing potential vehicle-following risks.

[0005] Furthermore, in scenarios involving multiple vehicles operating collaboratively or traveling in convoys in mines, achieving efficient coordination and global optimization among vehicles within the convoy is crucial for improving overall transportation efficiency and safety. Traditional single-vehicle intelligent following strategies lack information sharing and collaborative optimization capabilities at the convoy level, leading to braking waves caused by local adjustments, thereby reducing the overall smoothness and throughput of the convoy. More importantly, if a following vehicle in the convoy experiences an anomaly, the system, lacking global optimization and coordination, struggles to respond effectively, drastically increasing the risk of collisions between the abnormal vehicle and other vehicles, thus severely impacting the stability and safety of the entire system. Summary of the Invention

[0006] The purpose of this invention is to provide a method for mine vehicle speed planning that integrates AI algorithms and multimodal data, which solves the problems of insufficient following safety, poor overall fleet coordination, and limited efficiency of existing mine vehicle following systems in complex terrain.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for planning vehicle speed in mining areas that integrates AI algorithms and multimodal data includes the following steps:

[0009] S1. Determine that the current vehicle is in a following state, and obtain the initial driving parameters of the current vehicle, the position information of the leading vehicle, and the distance between the current vehicle and the leading vehicle.

[0010] S2. Based on the current location of the vehicle and the location information of the leading vehicle, determine whether the leading vehicle is driving on the slope or whether the distance from the slope is less than a preset distance.

[0011] S3. When the distance between the leading vehicle traveling on the slope and leaving the slope is less than the preset distance, the following distance between the current vehicle and the leading vehicle is determined based on the distance between the current vehicle and the leading vehicle and the distance the leading vehicle travels on the slope.

[0012] S4. Based on the following distance, plan the speed of the current vehicle.

[0013] Preferably, in step S1, the initial driving parameters of the current vehicle, the position information of the leading vehicle, and the distance between the current vehicle and the leading vehicle are obtained in the following way:

[0014] The precise location of the lead vehicle and surrounding terrain information are obtained through lidar and visual sensors.

[0015] The terrain information includes the straight-line distance between the current vehicle and the leading vehicle.

[0016] Preferably, in step S2, determining whether the leading vehicle is traveling on the slope or has left the slope at a distance less than a preset distance specifically includes the following steps:

[0017] S21. Using the terrain information around the leading vehicle obtained by lidar and visual sensors, a local terrain model around the leading vehicle is constructed in real time.

[0018] S22. Using AI algorithms and combining them with the local terrain model around the leading vehicle, extract the current slope, slope angle, and slope length on the driving trajectory of the leading vehicle.

[0019] S23. Combining the speed and heading of the leading vehicle with the terrain model ahead, and based on the dynamic thresholds adjusted in real time by the preset model according to the speed, gradient change rate and current following distance of the leading vehicle, predict the time and position when or when the leading vehicle is about to enter or leave the slope.

[0020] S24. Determine whether the leading vehicle is within the predicted slope-related area.

[0021] Preferably, in step S3, determining the following distance between the current vehicle and the leading vehicle specifically includes:

[0022] By training a machine learning model, and taking the current vehicle speed, acceleration, the speed and acceleration of the leading vehicle, the slope angle, the slope length, the position and progress of the leading vehicle on the slope, the predicted distance to the slope change point, and the road conditions as input data, the model outputs a dynamic risk adjustment factor and an optimal following distance.

[0023] Preferably, the training of the machine learning model includes the following steps:

[0024] S31. Train the machine learning model using data including mining area scene data, historical accident data, and simulated scene data;

[0025] S32, to capture the nonlinear effect of slope steepness on braking distance, the effect of ramp length on the duration of changes in the behavior of the preceding vehicle, and the risks of the ramp entrance / exit transition zone.

[0026] Preferably, in step S3, the following distance between the current vehicle and the leading vehicle is determined in the following way:

[0027] If the leading vehicle is going uphill, the following distance will be determined as the uphill following distance.

[0028] If the leading vehicle is going downhill, the following distance will be defined as the downhill following distance.

[0029] The uphill following distance is calculated using the following formula:

[0030] ;

[0031] In the formula, This is the pre-set safety distance for a flat road surface; The uphill risk factor output by the machine learning model; The function is a distance increment function that takes into account the slope, the speed and acceleration of the leading vehicle, and the remaining slope length. The function is a safety margin function based on the reaction time and speed of the following vehicle; The slope angle; Speed ​​of the leading vehicle; For acceleration; This represents the remaining distance between the leading vehicle and the top of the ramp. It is a safety margin function based on the current vehicle reaction time and speed; This indicates the total response time of the current vehicle system; Indicates the current vehicle speed;

[0032] The downhill following distance is calculated using the following formula:

[0033] ;

[0034] In the formula, The downhill risk factor output by the machine learning model; The function is a distance increment function that takes into account the slope, the speed of the leading vehicle, and the braking performance. The function is the incremental function of the braking distance of the rear vehicle, which takes into account the speed, braking performance and slope of the rear vehicle. Indicates vehicle braking performance parameters; This represents the incremental function for calculating the braking distance of the following vehicle, taking into account the current vehicle speed, braking performance, and gradient. Indicates the current vehicle speed.

[0035] Preferably, in step S4, planning the current vehicle speed specifically includes the following steps:

[0036] S41. The determined following distance is taken as the target following distance of the current vehicle;

[0037] S42. Based on the target following distance, the current following distance, the speed of the leading vehicle, and the speed of the current vehicle, a cooperative adaptive cruise control algorithm is used to generate the target speed and target acceleration command for the current vehicle.

[0038] Preferably, in step S42, the cooperative adaptive cruise control algorithm includes the following steps:

[0039] S421. Calculate the distance error between the current actual following distance and the target following distance, and the speed error between the current vehicle and the leading vehicle;

[0040] S422. Based on the distance error and the velocity error, apply the control law to generate the target acceleration command;

[0041] S423. Calculate the target speed based on the target acceleration command and the current vehicle speed.

[0042] In summary, the present invention has at least one of the following beneficial technical effects:

[0043] 1. This invention accurately identifies vehicle following patterns and collects multimodal data in real time and comprehensively. Then, it uses AI algorithms to perform high-precision perception and forward-looking prediction of the leading vehicle's slope driving status, including constructing a local terrain model, extracting slope features, and predicting slope behavior. This allows the solution to fully understand the complex terrain environment it will face now and in the future, enabling it to make predictions before risks occur and providing sufficient and timely data support for subsequent decisions. This effectively avoids the risk of following other vehicles due to lagging or insufficient environmental information.

[0044] 2. This invention introduces a pre-trained machine learning model and incorporates its output dynamic risk adjustment factor into the calculation of following distance. It designs refined following distance calculation methods for uphill and downhill scenarios, achieving dynamic, context-aware, and adaptive adjustment of the following distance. Simultaneously, by learning from a large amount of mining area scenario data, the machine learning model can capture the nonlinear influence of high-risk factors such as slope steepness, slope length, and transition zones at slope entrances and exits on the safe distance. This makes the calculated optimal following distance more consistent with the actual safety requirements of driving on slopes in mining areas. This solves the problem of fixed or untimely following distance adjustments in complex slope environments using traditional methods, thus improving vehicle safety when following other vehicles on slopes.

[0045] 3. This invention transforms the dynamically determined optimal following distance into specific vehicle speed commands and employs a cooperative adaptive cruise control algorithm to generate target speed and target acceleration, achieving smooth and efficient control of following behavior. In multi-vehicle platooning scenarios, this solution can achieve global optimization and coordination of following distance through platoon management strategies, enabling vehicles in the platoon to share information and make collaborative decisions. This effectively avoids global instability caused by local adjustments, thereby improving the overall mobility and operational efficiency of the entire platoon and ensuring the overall safety and economy of platoon operation in multi-vehicle mining environments. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0047] Figure 2 This is a schematic diagram illustrating the multimodal data and dynamic adjustment of following distance under slope conditions according to the present invention.

[0048] Figure 3 This is a schematic diagram of the ramp status sensing of the present invention;

[0049] Figure 4 This is a schematic diagram of the cooperative adaptive cruise control of the present invention. Detailed Implementation

[0050] The following is in conjunction with the appendix Figure 1 -Appendix Figure 4 The present invention will be further described in detail below.

[0051] This invention provides a method for mine vehicle speed planning that integrates AI algorithms and multimodal data, comprising the following steps:

[0052] S1. Determine that the current vehicle is in a following state, and obtain the initial driving parameters of the current vehicle, the position information of the leading vehicle, and the distance between the current vehicle and the leading vehicle.

[0053] S2. Based on the current location of the vehicle and the location information of the leading vehicle, determine whether the leading vehicle is driving on the slope or has left the slope by less than a preset distance.

[0054] S3. When the leading vehicle is traveling on the slope or the distance from the slope is less than the preset distance, the following distance between the current vehicle and the leading vehicle is determined based on the distance between the current vehicle and the leading vehicle and the distance the leading vehicle travels on the slope.

[0055] S4. Based on the following distance, plan the current vehicle speed.

[0056] The specific implementation details of step S1 in the plan are as follows:

[0057] Specifically, step S1 aims to determine whether the vehicles in front and behind are in a following position, which is a prerequisite for triggering subsequent refined following distance planning.

[0058] It monitors the vehicle's motion parameters in real time, such as its speed, acceleration, relative distance and speed to the vehicle ahead, and combines this with a preset following mode determination logic. This determination logic is typically based on empirical thresholds or machine learning models. When the current vehicle's speed maintains a certain correlation with the potential leading vehicle ahead, and the distance between them is within a preset safe following range, it is determined that the current vehicle has entered the following mode. Once the following mode is confirmed, the subsequent data collection and planning process begins.

[0059] Subsequently, in order to support subsequent slope perception, risk assessment and following distance calculation, it is necessary to acquire multimodal data comprehensively and accurately.

[0060] Specifically, the initial driving parameters of the current vehicle, such as the vehicle's real-time speed and acceleration, are obtained by reading data from the vehicle's onboard controller area network. The CAN bus can provide accurate, real-time digital information on the vehicle's own kinematic state.

[0061] Simultaneously, to perceive specific information about the environment ahead and the leading vehicle, the method utilizes multimodal sensing sensors for data acquisition. Specifically, a lidar is used to acquire precise 3D position information of the leading vehicle and high-precision point cloud data of its surrounding environment. The point cloud data provides rich geometric information about the terrain ahead. Working in conjunction with the lidar is a vision sensor, which acquires image data of the leading vehicle and its surrounding environment. The image data provides visual features such as texture and color, effectively supplementing the point cloud data, and playing a particularly important role in scenarios such as identifying road conditions.

[0062] By fusing data acquired from LiDAR and visual sensors, this embodiment can accurately calculate the straight-line distance between the current vehicle and the leading vehicle. Straight-line distance is one of the fundamental inputs for determining following distance. Furthermore, the point cloud data acquired by LiDAR and the image data acquired by the visual sensors together constitute the surrounding terrain information of the leading vehicle, laying the foundation for subsequent construction of a local terrain model and identification of slope features.

[0063] In summary, step S1, through intelligent pattern recognition and multimodal sensor data fusion, provides comprehensive, real-time, and high-precision environmental perception and vehicle status information for subsequent accurate slope judgment and dynamic following distance planning.

[0064] The specific implementation details of step S2 in the plan are as follows:

[0065] Specifically, step S2 aims to determine, based on the current vehicle's location and the location information of the leading vehicle, whether the leading vehicle is traveling on the slope or has left the slope by a preset distance. This determination is the core element of this technical solution, enabling it to dynamically adjust the following distance and thus improve safety in complex terrain.

[0066] The judgment process preferably includes the following detailed steps:

[0067] S21. Using lidar and visual sensors to obtain terrain information around the leading vehicle, construct a local terrain model around the leading vehicle in real time.

[0068] This step utilizes the high-precision point cloud data from the LiDAR and image data from the visual sensor acquired in step S1 to perform real-time terrain modeling of a local area ahead of the leading vehicle's trajectory. The construction of the local terrain model typically involves point cloud segmentation algorithms to separate ground points from non-ground points; subsequently, ground extraction algorithms identify ground points in drivable areas; finally, techniques such as plane fitting or meshing are used to construct a high-precision local 3D terrain model that represents the terrain's undulations. The model accurately reflects the current position of the leading vehicle and the geometric characteristics of the road ahead, such as slope and aspect.

[0069] S22. Using AI algorithms and combining local terrain models around the leading vehicle, extract the current slope, its slope angle, and slope length on the leading vehicle's driving trajectory.

[0070] After acquiring the local terrain model, this embodiment further utilizes artificial intelligence (AI) algorithms for deep analysis of the model to accurately extract slope features. AI algorithms, such as, but not limited to, deep learning-based semantic segmentation networks, can identify slope areas along the trajectory of the leading vehicle. Once a slope is identified, the AI ​​algorithm further calculates and extracts key parameters of the slope: the slope angle, which represents the angle between the slope and the horizontal plane, obtained by calculating the angle between the normal vector of the slope surface and the direction of gravity; and the slope length, which represents the total length of the slope currently occupied by the leading vehicle. The slope angle and slope length are crucial inputs for risk assessment in subsequent following distance calculations.

[0071] S23. Combining the speed and heading of the leading vehicle with the terrain model ahead, and based on the dynamic thresholds adjusted in real time by the preset model according to the speed of the leading vehicle, the rate of change of the slope and the current following distance, predict the time and position when the leading vehicle is about to enter or leave the slope.

[0072] To achieve proactive following distance adjustment, this embodiment introduces a leading vehicle ramp behavior prediction mechanism. The prediction preferably combines multiple pieces of information: the real-time speed and heading of the leading vehicle, which reveal its current driving trend through kinematics; and the terrain model constructed in step S21, which provides potential ramp information. Prediction algorithms, such as, but not limited to, Kalman filtering, particle filtering, or deep learning-based temporal prediction models, can integrate this information to predict the specific time and location at which the leading vehicle is about to enter or leave the ramp.

[0073] It is worth noting that this embodiment introduces a dynamic threshold adjustment mechanism during the prediction process. The dynamic threshold, such as the prediction distance threshold or time threshold, is not fixed but is adjusted in real time by a preset model. This preset model is preferably a lightweight machine learning model, specifically a Gradient Boosting Decision Tree (GBDT) model trained using a comprehensive dataset containing over 100,000 sample points. This dataset can come from various sources, including:

[0074] 1) Real accident cases extracted from the historical accident database of the mining area;

[0075] 2) Simulate various critical risk scenarios using high-fidelity simulation platforms (such as PreScan and CARLA);

[0076] 3) Real road test data recorded by combining GPS and IMU data.

[0077] During training, the label for each sample point is defined as the optimal warning distance, verified by expert drivers or through offline simulation; that is, the latest safe distance at which a reaction can be made without a collision. The goal of model training is to minimize the root mean square error (RMSE) between its prediction threshold and this optimal warning distance label.

[0078] The model's input features include the speed of the leading vehicle. (Range 0-30m / s), Slope Change Rate (For example, the rate at which the slope changes abruptly from flat to steep, ranging from 0 to 0.3 rad / s) and the current following distance. (Range 10-200m), etc. These raw inputs are normalized before being fed into the model (e.g., Min-Max scaling to the [0, 1] interval) to ensure the stability and convergence speed of model training.

[0079] The GBDT model consists of 100 decision tree estimators with a learning rate of 0.1 and a maximum depth of 5 for each tree. To facilitate deployment and analysis in resource-constrained in-vehicle systems, we perform knowledge distillation on the trained model, approximating its complex nonlinear decision boundary as a continuous and interpretable mathematical expression. This achieves a precise description of its input-output relationship and adjustment rules.

[0080] ;

[0081] In the formula, The dynamic warning distance threshold is the core result calculated and output by the preset model. The real-time speed of the leading vehicle is a key input feature of the model; The slope change rate is one of the most sensitive input features for the model; This represents the actual following distance between the current vehicle and the leading vehicle. It is the hyperbolic tangent function.

[0082] This formula elaborates on the model's decision-making logic:

[0083] Basic and dynamic range: The formula is based on a 55-meter central reference, through... The dynamic adjustment of the meter makes the output threshold... It covers a working range of 30 to 80 meters. The median value of 55 meters references international safety standards and takes into account the specific characteristics of mining operations.

[0084] S-curve adjustment: The hyperbolic tangent function tanh is chosen to simulate the non-linear characteristics of risk perception. Risk does not increase linearly with speed indefinitely, but tends to saturate after reaching a certain level. The S-curve of tanh can perfectly fit this smooth transition and saturation at both ends.

[0085] Feature weights: The weight coefficients (0.2, 0.5, -0.1) in the formula are approximate values ​​obtained by performing feature importance analysis (e.g., SHAP value analysis) on the trained GBDT model. Among them, the slope change rate... With the highest positive weight (0.5), it indicates that drastic changes in slope are the most critical factor in raising the warning level. Leading vehicle speed This is also a positive impact. And the current following distance... The weight is negative, which is logical: when the vehicle has maintained a large safe distance, the additional need for warning distance will be appropriately reduced.

[0086] Finally, to prevent unstable fluctuations in the model output due to sensor noise or environmental transients, which could lead to frequent fine-tuning of the vehicle control system, this embodiment also employs a Kalman filter to smooth the output threshold. This filter uses the model output as an observation and, through state prediction and updating, obtains a more stable estimate that more closely approximates the actual risk state.

[0087] By quantizing the model with 8-bit integers, its memory footprint and computational load were significantly reduced, enabling it to run efficiently on computationally limited automotive embedded controllers. This resulted in a single inference time of less than 5ms and ensured that the frequency of the entire prediction update chain was no less than 10Hz, meeting the stringent real-time requirements for environmental perception and decision-making when vehicles are traveling at high speeds. This dynamic threshold adjustment aims to ensure the timeliness and accuracy of predictions, avoiding premature or late warnings.

[0088] S24. Determine whether the leading vehicle is within the predicted slope-related area.

[0089] Based on the prediction result of step S23, this embodiment ultimately determines whether the leading vehicle is within the predicted slope-related area. The slope-related area refers to the slope area that the leading vehicle is currently traveling on, or the distance between it and the slope area it is about to enter or leave is less than the preset distance threshold dynamically adjusted in step S23. This determination result is a key signal that triggers the dynamic following distance determination in the subsequent step S3.

[0090] In summary, step S2, by integrating multimodal perception data, applying AI algorithms for terrain feature extraction and leading vehicle behavior prediction, and combining a dynamic threshold adjustment mechanism, achieves accurate perception and forward-looking judgment of the leading vehicle's slope driving status, providing reliable input for the next step of dynamically adjusting the following distance.

[0091] The specific implementation details of step S3 in the plan are as follows:

[0092] Specifically, the S3 procedure aims to dynamically adjust the following distance between the current vehicle and the leading vehicle based on the unique driving risks of the ramp, thereby ensuring driving safety.

[0093] To achieve this goal, this solution introduces a pre-trained machine learning model for context awareness and risk assessment of the current vehicle-following scenario. The training process of the machine learning model is crucial, ensuring that the model can accurately capture the nonlinear relationships in the complex environment of the mining area.

[0094] Preferably, the training process of the machine learning model includes the following steps:

[0095] S31. Train the machine learning model using data including mining area scene data, historical accident data, and simulated scene data;

[0096] First, the data collection period for the mining area scenarios was 18 months, with data sourced from 10 main mining trucks (such as the CAT797F) in a large open-pit coal mine (e.g., a large coal mine in Inner Mongolia). Data was collected using lidar (continuously recording 3D point clouds at a 10Hz sampling rate) and forward-looking vision sensors deployed on the trucks (acquiring high-definition images at 30fps), accumulating over 5,000 hours of continuous operational data covering more than 100,000 kilometers of driving. This data comprehensively covers the mining trucks under both empty and fully loaded conditions, experiencing different road surfaces such as dry, wet, and muddy conditions, as well as various weather combinations including sunny, rainy, and snowy days. In the data preprocessing stage, the data streams of LiDAR and visual sensors are first aligned with sub-millisecond timestamps using NTP (Network Time Protocol). Then, the DBSCAN (Density-based Noise Application Spatial Clustering) algorithm is used to denoise the LiDAR point cloud data, effectively removing flying point noise caused by dust, rain, and snow. Finally, a semi-automatic annotation tool is used to annotate the road surface conditions and weather conditions frame by frame, forming structured scene labels.

[0097] Secondly, to enable the model to learn from high-risk events in the real world, this embodiment integrates historical accident data from 37 collision and near-collision accidents related to driving on slopes recorded in the mining area over the past five years. Each accident includes complete vehicle dynamic time-series data, such as vehicle speed, acceleration, steering wheel angle, and brake master cylinder pressure, recorded at a high sampling rate of 100Hz within 10 seconds prior to the collision, stored by the vehicle's black box or data recorder. In data preprocessing, instantaneous spikes or outliers caused by severe bumps in the sensor data are smoothed and interpolated using a moving average method based on five consecutive sampling points to ensure the continuity and validity of the data sequence. Furthermore, key derived features are extracted through feature engineering, such as the average deceleration within 3 seconds prior to the collision, the peak braking pressure, and the gradient of the slope angle change. These features are crucial for characterizing the suddenness of the accident.

[0098] Finally, considering the difficulty and high cost of collecting data on extreme and dangerous operating conditions in the real world, this embodiment further utilizes the high-fidelity simulation platform CARLA to generate massive amounts of high-fidelity simulation data. The vehicle dynamics parameters of the simulation model (such as engine torque curves, air resistance, and brake performance parameters) are all configured based on bench test calibration data of real mining trucks, and verified by comparing key indicators (such as 0-100 km / h acceleration time and fully loaded braking distance) with real road test data using root mean square error (RMSE), ensuring a simulation fidelity greater than 95%. The simulation focuses on constructing three types of extreme operating conditions:

[0099] 1) On a steep slope with an incline greater than 25°, the lead vehicle shall apply emergency braking;

[0100] 2) On long downhill sections with a slope length exceeding 500 meters, simulate the thermal fade phenomenon caused by continuous braking;

[0101] 3) Set up a wet and slippery surface in the transition area 50 meters before and after the ramp entrance and exit and simulate the rapid acceleration / deceleration behavior of vehicles.

[0102] By performing Latin hypercube sampling on the core parameters of these scenarios (such as initial velocity, slope, and road friction coefficient), more than 20,000 simulation sequences, each lasting 30 seconds, were generated. From these, 5,000 of the most representative extreme condition samples were selected. After cleaning, aligning, and labeling the data from these three sources, a comprehensive training dataset containing over 500,000 valid data frames—multi-source, heterogeneous, and fully labeled—was ultimately formed. This laid a solid foundation for training highly robust and generalizable machine learning models.

[0103] S32, to capture the nonlinear effect of slope steepness on braking distance, the effect of ramp length on the duration of changes in the behavior of the preceding vehicle, and the risks of the ramp entrance / exit transition zone.

[0104] In setting the training objectives, this embodiment goes beyond simple end-to-end fitting. Instead, it designs a specific model structure and training strategy to ensure the model accurately captures and understands the unique physical laws and risk patterns encountered while driving on mining slopes. This embodiment employs a hybrid neural network architecture that integrates a temporal convolutional network (TCN) for processing temporal data and a deep feedforward network (DFN) for processing static contextual features. The DFN is responsible for processing static or slowly varying features, such as vehicle load status (empty / fully loaded), current road surface type, and total slope length. It consists of three fully connected layers with 256, 128, and 64 neurons respectively. The activation function uniformly uses linear units with leakage correction to prevent neuron inactivation during training.

[0105] The TCN (Tracking Networking) component specifically processes high-frequency time-series data, such as the speed, acceleration, and braking pressure sequences of the preceding and following vehicles over the past 5 seconds. This TCN module contains four residual blocks, each with two convolutional layers featuring causal convolution and weight normalization. It employs exponentially increasing expansion factors (specifically 1, 2, 4, and 8) to expand the receptive field, thereby efficiently capturing long-term temporal dependencies. The output feature vectors of the DFN (Deep Flow Network) and TCN (both 64-dimensional) are concatenated and then deeply fused through two fully connected layers (128 and 32 neurons respectively). Finally, a linear output layer regresses and predicts the dynamic risk adjustment factor and the optimal following distance.

[0106] To capture the nonlinear effect of slope steepness on braking distance, this embodiment introduces a regularization term based on a physical model into the loss function.

[0107] Specifically, regarding the slope angle The nonlinear relationship with braking distance, especially the sharp increase in braking distance when the slope angle is greater than 15°, is explicitly modeled using the hyperbolic tangent function. The formula for calculating this regularization term is as follows:

[0108] ;

[0109] In the formula, The slope angle (unit: degrees) is calculated in real time using lidar point clouds. The standard braking distance for a flat road surface is determined by referring to the speedometer based on the current vehicle speed. This represents the braking distance increment caused entirely by changes in slope. Total loss function Loss due to mean square error and the loss of the physical regularization term Composition, namely:

[0110] ;

[0111] In the formula, the weighting coefficient The cross-validation value was set at 0.15, which forces the model to follow basic physical laws while optimizing prediction accuracy.

[0112] To capture the impact of ramp length on the duration of changes in the behavior of the vehicle in front, the aforementioned TCN module, through its exponentially growing receptive field, can effectively establish dependencies spanning hundreds of time steps (corresponding to tens of seconds of travel time). This allows for accurate capture of the characteristic pattern of a vehicle in front exhibiting a continuous and slow decrease in speed due to power decay or braking heat capacity limitations on long ramps (e.g., ramp length > 500m). Finally, to focus on capturing the high-risk characteristics of the transition zone at ramp entrances and exits, this embodiment employs a weighted loss function training strategy. When calculating the total loss, the system identifies whether the training samples are located within a 50-meter transition zone before and after the ramp entrances and exits, assigning samples in this zone a weight coefficient as high as 3.5 times. Through this optimization, the contribution of this type of sample to the total loss function significantly increases from the original 12% to 38%, forcing the model to invest more learning resources to optimize its prediction performance in these high-risk areas, ultimately improving the model's prediction accuracy for such scenarios from 78% to 91%.

[0113] In actual operation, the machine learning model takes several key parameters as input data, all of which are collected and perceived in real time from the preceding steps: including the current vehicle speed and acceleration; the speed and acceleration of the leading vehicle; the slope angle and slope length; the position and progress of the leading vehicle on the slope; the predicted distance to the slope change point; and road conditions, such as dry, wet, or muddy. The machine learning model, through its internal logic obtained during training, comprehensively analyzes these inputs and ultimately outputs a dynamic risk adjustment factor or directly outputs the optimal following distance. The risk adjustment factor is a value greater than 1, used to amplify the basic safe distance in subsequent formulas to cope with higher-risk driving scenarios.

[0114] Depending on the specific slope condition (uphill or downhill) of the leading vehicle, this plan uses different methods to determine the following distance between the current vehicle and the leading vehicle.

[0115] If the leading vehicle is going uphill, the following distance is defined as the uphill following distance. The uphill following distance is designed to provide sufficient safety space to account for potential risks such as deceleration, stalling, or insufficient power of the leading vehicle while going uphill. It is preferably calculated using the following formula:

[0116] ;

[0117] In the formula, This represents the preset basic safety distance on a flat road surface, which is a constant and is a fundamental component of the following distance. The uphill risk factor output by the machine learning model has a value greater than 1 and is used to dynamically adjust the distance increment based on the risk level of the current situation. The function is a distance increment function that takes into account the gradient, the speed and acceleration of the leading vehicle, and the remaining slope length. It reflects the risk of deceleration or stalling of the leading vehicle when going uphill due to performance degradation and is a distance compensation based on the state of the leading vehicle. The function is a safety margin function based on the reaction time and speed of the following vehicle (the current vehicle), which aims to ensure that the current vehicle has enough distance to avoid a collision with the leading vehicle in the event of emergency braking; The slope angle of the ramp; and These represent the speed and acceleration of the leading vehicle, respectively. Indicates the remaining distance between the leading vehicle and the top of the ramp; This represents the total reaction time of the current vehicle system, including delays in perception, decision-making, and execution. Indicates the current vehicle speed.

[0118] If the leading vehicle is going downhill, the following distance is defined as the downhill following distance. The downhill following distance is designed to account for the risk that the leading vehicle may accelerate due to gravity or that its braking performance may decrease due to the slope. It is preferably calculated using the following formula:

[0119] ;

[0120] In the formula, The downhill risk factor output by the machine learning model has a value greater than 1 and is used to dynamically adjust the distance increment according to the risk level of the current situation. The function is a distance increment function that takes into account the gradient, the speed of the leading vehicle, and the braking performance. It reflects the risk that the leading vehicle may accelerate or have reduced braking efficiency due to gravity when going downhill. It is a distance compensation based on the state of the leading vehicle. The function is a function that calculates the braking distance increment of the following vehicle (current vehicle) considering the speed, braking performance and gradient of the following vehicle (current vehicle). It reflects the braking demand of the current vehicle when going downhill and ensures that it can safely decelerate or stop. This indicates vehicle braking performance parameters, such as the vehicle's maximum braking deceleration.

[0121] In summary, step S3 uses AI algorithms to transform complex slope driving scenarios into specific risk factors, and combines them with physical model formulas to achieve dynamic, situational awareness, and accurate calculation of following distance under different slope conditions, providing a safe and reliable basis for subsequent speed planning.

[0122] The specific implementation details of step S4 in the plan are as follows:

[0123] Specifically, step S4 aims to translate the results of preceding perception, prediction, and decision-making into executable vehicle motion commands, thereby achieving actual control over following behavior.

[0124] Vehicle speed planning preferably includes the following detailed steps:

[0125] S41. Use the determined following distance as the target following distance for the current vehicle;

[0126] First, the optimal following distance (whether uphill or downhill) obtained through complex scenario evaluation and calculation in step S3 is set as the expected distance that the current vehicle needs to maintain in real time. This expected distance, called the target following distance, is a dynamically changing parameter that can adaptively adjust according to real-time changes in the mining environment. Using it as the target provides a clear basis for subsequent speed planning.

[0127] S42. Based on the target following distance, current following distance, leading vehicle speed, and current vehicle speed, a cooperative adaptive cruise control algorithm is used to generate the target speed and target acceleration command for the current vehicle.

[0128] To ensure the current vehicle can smoothly and accurately approach and maintain the target following distance, this embodiment preferably employs a cooperative adaptive cruise control algorithm to generate the target speed and target acceleration commands for the current vehicle. The cooperative adaptive cruise control algorithm not only considers the individual vehicle's state and its relative relationship with the vehicle ahead, but also incorporates fleet-level information sharing and collaborative optimization, aiming to achieve higher following efficiency and smoothness.

[0129] The generation process of the cooperative adaptive cruise control algorithm preferably includes the following more specific sub-steps:

[0130] S421. Calculate the distance error between the current actual following distance and the target following distance, as well as the speed error between the current vehicle and the leading vehicle;

[0131] In each control cycle, the adaptive cruise control algorithm first obtains the current actual following distance, which is acquired in step S1 using LiDAR and vision sensors. Then, it calculates the distance error between the current actual following distance and the target following distance set in step S41. This distance error directly reflects the deviation between the current following state and the desired state.

[0132] ;

[0133] In the formula, This indicates the distance error, which is the difference between the current actual following distance and the target following distance. This indicates the current actual following distance, obtained in real time by sensors; Indicates the following distance of the target vehicle.

[0134] Meanwhile, the cooperative adaptive cruise control algorithm also obtains the real-time speed of the current vehicle and the real-time speed of the leading vehicle, and calculates the speed error between them:

[0135] ;

[0136] In the formula, This indicates the speed error, which is the difference between the speed of the leading vehicle and the speed of the current vehicle. The speed of the leading vehicle is indicated by real-time measurements from sensors. This indicates the current vehicle speed, which is obtained through real-time measurement by sensors.

[0137] Distance error and speed error together constitute the key feedback signals for the cooperative adaptive cruise control algorithm to make control decisions.

[0138] S422. Based on the distance error and velocity error, apply the control law to generate the target acceleration command;

[0139] Based on the distance and speed errors calculated in step S421, the cooperative adaptive cruise control algorithm applies a preset control law to generate the target acceleration command for the current vehicle. The control law preferably employs a proportional-integral-derivative controller, model predictive control, or other advanced nonlinear control algorithms. A simplified PID control law can be expressed as:

[0140] ;

[0141] In the formula, This indicates the target acceleration command for the current vehicle; This represents the proportional gain coefficient, used to adjust the response strength to the current distance error; This represents the integral gain coefficient, used to eliminate long-term accumulated distance errors; The integral term represents the distance error, i.e., the accumulation of the distance error over time; This represents the differential gain coefficient, used to suppress the rate of change of distance error and improve system stability; The differential term represents the distance error, i.e., the rate of change of the distance error over time; This represents the speed error gain coefficient, used to adjust the response intensity to speed errors in order to quickly match the speed of the vehicle ahead.

[0142] S423. Calculate the target speed based on the target acceleration command and the current vehicle speed.

[0143] After generating the target acceleration command, the cooperative adaptive cruise control algorithm further combines the vehicle's real-time speed to calculate the target speed the vehicle should reach in the next moment or the next control cycle. The calculation of the target speed is usually based on simple kinematic equations:

[0144] ;

[0145] In the formula, Indicates the current target speed of the vehicle; Indicates the current real-time speed of the vehicle; Indicates the target acceleration command; This indicates the time step of the control cycle, which is the time interval for calculating update instructions.

[0146] The target speed command is then sent to the vehicle's underlying execution systems, such as the powertrain and braking systems, to actually drive the vehicle to adjust its speed and distance.

[0147] In summary, step S4 transforms complex environmental perception and high-level decision-making into precise and executable vehicle motion commands. By dynamically adjusting the target following distance and applying an advanced cooperative adaptive cruise control algorithm, it achieves real-time, efficient, and safe planning of vehicle following speed in the complex environment of the mining area, thereby optimizing the vehicle's response speed and overall driving efficiency.

[0148] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for planning vehicle speed in mining areas that integrates AI algorithms and multimodal data, characterized in that: Includes the following steps: S1. Determine that the current vehicle is in a following state, and obtain the initial driving parameters of the current vehicle, the position information of the leading vehicle, and the distance between the current vehicle and the leading vehicle. S2. Based on the current location of the vehicle and the location information of the leading vehicle, determine whether the leading vehicle is driving on the slope or whether the distance from the slope is less than a preset distance. S3. When the distance between the leading vehicle traveling on the slope and leaving the slope is less than the preset distance, the following distance between the current vehicle and the leading vehicle is determined based on the distance between the current vehicle and the leading vehicle and the distance the leading vehicle travels on the slope. S4. Based on the following distance, plan the speed of the current vehicle.

2. The mining area vehicle speed planning method integrating AI algorithm and multimodal data according to claim 1, characterized in that, In step S1, the initial driving parameters of the current vehicle, the position information of the leading vehicle, and the distance between the current vehicle and the leading vehicle are obtained in the following ways: The precise location of the lead vehicle and surrounding terrain information are obtained through lidar and visual sensors. The terrain information includes the straight-line distance between the current vehicle and the leading vehicle.

3. The mining area vehicle speed planning method integrating AI algorithm and multimodal data according to claim 1, characterized in that, In step S2, determining whether the leading vehicle is traveling on the slope or has left the slope at a distance less than a preset distance specifically includes the following steps: S21. Using the terrain information around the leading vehicle obtained by lidar and visual sensors, a local terrain model around the leading vehicle is constructed in real time. S22. Using AI algorithms and combining them with the local terrain model around the leading vehicle, extract the current slope, slope angle, and slope length on the driving trajectory of the leading vehicle. S23. Combining the speed and heading of the leading vehicle with the terrain model ahead, and based on the dynamic thresholds adjusted in real time by the preset model according to the speed, gradient change rate and current following distance of the leading vehicle, predict the time and position when or when the leading vehicle is about to enter or leave the slope. S24. Determine whether the leading vehicle is within the predicted slope-related area.

4. The mining area vehicle speed planning method integrating AI algorithm and multimodal data according to claim 1, characterized in that, In step S3, determining the following distance between the current vehicle and the leading vehicle specifically includes: By training a machine learning model, and taking the current vehicle speed, acceleration, the speed and acceleration of the leading vehicle, the slope angle, the slope length, the position and progress of the leading vehicle on the slope, the predicted distance to the slope change point, and the road conditions as input data, the model outputs a dynamic risk adjustment factor and an optimal following distance.

5. The mining area vehicle speed planning method integrating AI algorithm and multimodal data according to claim 4, characterized in that, The training of the machine learning model includes the following steps: S31. Train the machine learning model using data including mining area scene data, historical accident data, and simulated scene data; S32, to capture the nonlinear effect of slope steepness on braking distance, the effect of ramp length on the duration of changes in the behavior of the preceding vehicle, and the risks of the ramp entrance / exit transition zone.

6. The mining area vehicle speed planning method integrating AI algorithm and multimodal data according to claim 1, characterized in that, In step S3, the following distance between the current vehicle and the leading vehicle is determined in the following way: If the leading vehicle is going uphill, the following distance will be determined as the uphill following distance. If the leading vehicle is going downhill, the following distance will be defined as the downhill following distance. The uphill following distance is calculated using the following formula: ； In the formula, This is the pre-set safety distance for a flat road surface; The uphill risk factor output by the machine learning model; The function is a distance increment function that takes into account the slope, the speed and acceleration of the leading vehicle, and the remaining slope length. The function is a safety margin function based on the reaction time and speed of the following vehicle; The slope angle; Speed ​​of the leading vehicle; For acceleration; This represents the remaining distance between the leading vehicle and the top of the ramp. It is a safety margin function based on the current vehicle reaction time and speed; This indicates the total response time of the current vehicle system; Indicates the current vehicle speed; The downhill following distance is calculated using the following formula: ； In the formula, The downhill risk factor output by the machine learning model; The function is a distance increment function that takes into account the slope, the speed of the leading vehicle, and the braking performance. The function is the incremental function of the braking distance of the rear vehicle, which takes into account the speed, braking performance and slope of the rear vehicle. Indicates vehicle braking performance parameters; This represents the incremental function for calculating the braking distance of the following vehicle, taking into account the current vehicle speed, braking performance, and gradient. Indicates the current vehicle speed.

7. The mining area vehicle speed planning method integrating AI algorithm and multimodal data according to claim 1, characterized in that, In step S4, planning the current vehicle speed specifically includes the following steps: S41. The determined following distance is taken as the target following distance of the current vehicle; S42. Based on the target following distance, the current following distance, the speed of the leading vehicle, and the speed of the current vehicle, a cooperative adaptive cruise control algorithm is used to generate the target speed and target acceleration command for the current vehicle.

8. The mining area vehicle speed planning method integrating AI algorithm and multimodal data according to claim 7, characterized in that, In step S42, the cooperative adaptive cruise control algorithm includes the following steps: S421. Calculate the distance error between the current actual following distance and the target following distance, and the speed error between the current vehicle and the leading vehicle; S422. Based on the distance error and the velocity error, apply the control law to generate the target acceleration command; S423. Calculate the target speed based on the target acceleration command and the current vehicle speed.

Images