Adaptive control method and system for autonomous driving trucks

Abstract

The present invention provides an adaptive control method and system for an autonomous driving truck, the method comprising: generating a standardized data packet containing vehicle state parameters and environmental information by filtering and structured processing multi-dimensional data collected by an on-board sensor network. According to the standardized data packet, a deep learning neural network model library is retrieved, a currently applicable basic model is selected, and its parameters are dynamically adjusted to generate an optimized adaptive control model. The standardized data packet is input into the adaptive control model operation to generate candidate control instructions, and an optimal control strategy is generated through rapid simulation and multi-objective optimization calculation. Finally, the optimal control strategy is executed by the vehicle controller, and the vehicle response is monitored in real time, and the controller parameters are adaptively adjusted based on the deviation. The present invention realizes the dynamic adaptability and strategy optimization of the control model, and improves the safety and efficiency of autonomous driving trucks in a changing environment.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an adaptive control method and system for an autonomous driving truck. Background Art

[0002] As an important application in the field of intelligent transportation and logistics, the performance of the control system of self-driving trucks directly affects the driving safety, fuel economy and operational efficiency of the vehicle. In actual applications, self-driving trucks need to face changing road conditions, complex traffic environments and dynamic changes in the vehicle's own state. Therefore, adaptive control of the vehicle has become one of the key technologies to achieve efficient and safe self-driving.

[0003] Existing autonomous driving control methods usually rely on pre-set rules or control models with fixed parameters. This approach can provide basic control capabilities in specific scenarios, but because the model is difficult to flexibly adapt to the changing actual environment, it is prone to insufficient control accuracy. For example, when the road is slippery or the traffic density is high, the fixed parameter control model may not be able to adjust the dynamic response of the vehicle in time, resulting in the vehicle being unable to drive safely. In addition, when generating control strategies, existing methods lack comprehensive consideration of multi-objective optimization, which easily leads to insufficient trade-offs between control stability, energy consumption optimization, and driving efficiency. Summary of the invention

[0004] The main purpose of the present invention is to solve the technical problems in the existing autonomous driving truck control method, that is, the control model parameters are fixed, it is difficult to adapt to complex dynamic environments, and the lack of multi-objective optimization strategy leads to insufficient control accuracy and low efficiency;

[0005] A first aspect of the present invention provides an adaptive control method for an autonomous driving truck, the adaptive control method for an autonomous driving truck comprising:

[0006] Filter and structure the multi-dimensional data collected by the onboard sensor network on the self-driving truck to obtain a standardized data package containing vehicle state parameters and environmental information;

[0007] Retrieving and selecting a pre-trained deep learning neural network model library according to the standardized data packet to obtain a basic model currently corresponding to the autonomous driving truck, and dynamically adjusting model parameters of the basic model according to the standardized data packet to obtain an optimized adaptive control model;

[0008] Inputting the standardized data packet into the adaptive control model for operation to obtain candidate control instructions, and performing fast simulation and multi-objective optimization calculation on the candidate control instructions to obtain an optimal control strategy;

[0009] The optimal control strategy is executed by a controller on the self-driving truck, and the actual response of the self-driving truck is monitored in real time during the execution process. The parameters of the controller are adaptively adjusted according to the deviation between the monitored actual response and the expected response.

[0010] Optionally, in a first implementation of the first aspect of the present invention, the retrieving and selecting a pre-trained deep learning neural network model library according to the standardized data packet to obtain a basic model currently corresponding to the autonomous driving truck, and dynamically adjusting model parameters of the basic model according to the standardized data packet to obtain an optimized adaptive control model includes:

[0011] Parsing the standardized data packet to obtain the current load value, road slope information and curvature information in the standardized data packet;

[0012] Searching a deep learning neural network model library according to the current load value, and selecting a basic model that matches the current load value;

[0013] Normalizing the road slope information and curvature information in the standardized data packet to obtain a normalized feature vector, and inputting the normalized feature vector into a preset parameter adjustment network to obtain a model parameter adjustment amount;

[0014] The model parameters of the basic model are updated according to the model parameter adjustment amount to obtain an optimized adaptive control model.

[0015] Optionally, in a second implementation of the first aspect of the present invention, updating the model parameters of the basic model according to the model parameter adjustment amount to obtain the optimized adaptive control model includes:

[0016] Decomposing the model parameter adjustment amount to obtain a parameter adjustment matrix corresponding to each layer of the basic model;

[0017] According to the parameter adjustment matrix and the preset parameter update rule, the weight matrix and the bias vector of the basic model are updated and calculated to obtain updated model parameters;

[0018] Normalizing the updated model parameters to obtain normalized model parameters, and applying the normalized model parameters to the basic model to obtain a preliminarily optimized adaptive control model;

[0019] The initially optimized adaptive control model is fine-tuned and trained to obtain an optimized adaptive control model.

[0020] Optionally, in a third implementation of the first aspect of the present invention, the step of inputting the standardized data packet into the adaptive control model for operation to obtain candidate control instructions, and performing fast simulation and multi-objective optimization calculation on the candidate control instructions to obtain the optimal control strategy includes:

[0021] Inputting the standardized data packet into an adaptive control model to obtain multiple groups of candidate control instructions, and quickly simulating each group of candidate control instructions to obtain corresponding predicted trajectories and state sequences;

[0022] The predicted trajectory and state sequence are evaluated according to the preset multi-objective optimization function to obtain the performance index of each group of candidate control instructions;

[0023] A non-inferior sorting genetic algorithm is used to iteratively optimize the candidate control instructions based on the performance indicators of each group of candidate control instructions to obtain a Pareto optimal solution set, and an optimal control strategy is generated according to the Pareto optimal solution set.

[0024] Optionally, in a fourth implementation of the first aspect of the present invention, the predicted trajectory and state sequence are evaluated according to a preset multi-objective optimization function to obtain a performance indicator for each group of candidate control instructions including:

[0025] Comparing the predicted trajectory with the reference trajectory generated by the high-level path planning module, calculating the weighted sum of lateral and longitudinal deviations, and obtaining a trajectory tracking accuracy index;

[0026] According to the speed change rate and lateral acceleration in the state sequence, the jitter of the autonomous driving truck is calculated to obtain the ride comfort index;

[0027] Based on the engine operating point data and vehicle mass in the state sequence, combined with the pre-calibrated fuel consumption model, the energy consumption per unit mileage is calculated to obtain the energy efficiency index;

[0028] Using the spatiotemporal relationship between the predicted trajectory and dynamic obstacles, the time margin for potential collisions is calculated to obtain driving safety indicators.

[0029] The trajectory tracking accuracy index, the ride comfort index, the energy efficiency index and the driving safety index are weighted and summed using weights determined by a hierarchy analysis method to obtain a performance index.

[0030] Optionally, in a fifth implementation of the first aspect of the present invention, executing the optimal control strategy by a controller on the autonomous driving truck, and monitoring the actual response of the autonomous driving truck in real time during the execution process, and adaptively adjusting the parameters of the controller according to the deviation between the monitored actual response and the expected response includes:

[0031] Converting the optimal control strategy into control instructions executable by a controller on the autonomous driving truck, and executing the control instructions through the controller;

[0032] Collect the actual status information of the autonomous driving truck in real time to obtain the actual response data, and calculate the deviation between the actual response data and the expected response to obtain the response error;

[0033] The adjustment amount of the controller parameters is calculated according to the response error to obtain the parameter update value, and the parameter update value is used to adaptively adjust the controller parameters.

[0034] Optionally, in a sixth implementation of the first aspect of the present invention, calculating the adjustment amount of the controller parameter according to the response error to obtain the parameter update value, and using the parameter update value to adaptively adjust the controller parameter includes:

[0035] Performing spectrum analysis on the response error to obtain the frequency distribution characteristics of the response error, and performing piecewise weighting on the response error according to the frequency distribution characteristics to obtain a weighted error vector;

[0036] The relationship between the weighted error vector and the controller parameters is modeled using the recursive least square method to obtain the parameter sensitivity matrix.

[0037] Based on the parameter sensitivity matrix and the preset learning rate, the gradient descent method is used to calculate the parameter adjustment amount to obtain the preliminary parameter update value;

[0038] The initial parameter update value is adaptively limited in dynamic range to obtain a bounded parameter adjustment amount, and the bounded parameter adjustment amount and the historical adjustment amount are exponentially averaged to obtain a parameter update value, and the parameter update value is superimposed on the parameters of the controller to complete the adaptive adjustment.

[0039] A second aspect of the present invention provides an adaptive control system for an autonomous driving truck, the adaptive control system for an autonomous driving truck comprising:

[0040] The data processing module is used to filter and structure the multi-dimensional data collected by the on-board sensor network on the self-driving truck to obtain a standardized data packet containing vehicle status parameters and environmental information;

[0041] A model optimization module, configured to retrieve and select a pre-trained deep learning neural network model library according to the standardized data packet to obtain a basic model currently corresponding to the autonomous driving truck, and dynamically adjust model parameters of the basic model according to the standardized data packet to obtain an optimized adaptive control model;

[0042] A strategy generation module, used for inputting the standardized data packet into the adaptive control model for operation to obtain candidate control instructions, and performing fast simulation and multi-objective optimization calculation on the candidate control instructions to obtain an optimal control strategy;

[0043] The execution adjustment module is used to execute the optimal control strategy through the controller on the self-driving truck, and monitor the actual response of the self-driving truck in real time during the execution process, and adaptively adjust the parameters of the controller according to the deviation between the monitored actual response and the expected response.

[0044] The above-mentioned adaptive control method and system for the autonomous driving truck generates a standardized data packet containing vehicle state parameters and environmental information by filtering and structured processing the multi-dimensional data collected by the on-board sensor network. According to the standardized data packet, the deep learning neural network model library is retrieved, the currently applicable basic model is selected, and its parameters are dynamically adjusted to generate an optimized adaptive control model. The standardized data packet is input into the adaptive control model operation to generate candidate control instructions, and the optimal control strategy is generated through fast simulation and multi-objective optimization calculation. Finally, the optimal control strategy is executed by the vehicle controller, and the vehicle response is monitored in real time, and the controller parameters are adaptively adjusted based on the deviation. The present invention realizes the dynamic adaptability and strategy optimization of the control model, and improves the safety and efficiency of autonomous driving trucks in a changing environment.

[0045] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic diagram of a first embodiment of an adaptive control method for an autonomous driving truck in an embodiment of the present invention;

[0048] Figure 2 Schematic diagram of an embodiment of an adaptive control system for an autonomous driving truck in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0051] To facilitate understanding of this embodiment, an adaptive control method for an autonomous driving truck disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps:

[0052] 101. Filter and structure the multi-dimensional data collected by the on-board sensor network on the self-driving truck to obtain a standardized data packet containing vehicle state parameters and environmental information;

[0053] In one embodiment of the present invention, the vehicle sensor network includes multiple sensors such as GPS, inertial measurement unit (IMU), laser radar, camera and on-board diagnostic system (OBD) interface. These sensors collect data at different frequencies, so the data needs to be time synchronized first. Time synchronization uses a timestamp alignment method to align all sensor data to a unified time base. Subsequently, the Kalman filter algorithm is applied to the synchronized data to eliminate sensor noise and measurement errors. The Kalman filter estimates the optimal state estimate based on the historical data and current measurement values ​​of the sensor. For GPS data, an extended Kalman filter is used to handle nonlinear problems and IMU data is fused to improve positioning accuracy. For laser radar and camera data, a particle filter algorithm is used for target tracking and environment modeling. These filtered data form a preliminary processed data set. Next, feature extraction and dimensionality reduction are performed on the preliminary processed data set. For vehicle state parameters, such as speed, acceleration, yaw rate, etc., they are directly extracted from the corresponding sensor data. For environmental information, such as road slope and curvature, it needs to be obtained through data fusion and calculation. For example, the slope of the road can be calculated by fusing GPS elevation data and IMU data, while the curvature can be inferred by analyzing the vehicle trajectory and steering wheel angle data. For information about surrounding obstacles, a clustering algorithm is used to process the lidar point cloud data to identify and track surrounding dynamic and static obstacles. At the same time, a deep learning model is used to perform semantic segmentation on camera images to identify key environmental elements such as roads, lane lines, and traffic signs. These extracted features form a high-dimensional feature vector. In order to reduce data redundancy and improve processing efficiency, the principal component analysis (PCA) algorithm is applied to this high-dimensional feature vector for dimensionality reduction, retaining the most representative principal components.

[0054] Specifically, the data after dimensionality reduction needs to be standardized to eliminate the dimensional differences between different features. For continuous data, such as speed and acceleration, the Z-score standardization method is used to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1. For discrete data, such as road type and traffic sign category, the One-Hot Encoding method is used to convert the data into a numerical representation. The standardization process also includes the processing of outliers, using the 3-sigma principle to identify and process abnormal data to ensure data consistency and reliability. The processed standardized data is organized according to a predefined data structure to form a unified format data packet. This data packet contains vehicle state parameters (such as position, speed, acceleration, yaw rate, body posture, current load, engine status, etc.) and environmental information (such as road slope, curvature, estimated road friction coefficient, position and speed information of surrounding vehicles, etc.). Finally, the standardized data packet is checked for integrity and consistency. The integrity check ensures that all necessary fields are filled. If there is missing data, it is supplemented using methods such as interpolation or nearest neighbor filling. The consistency check verifies the logical relationship between each field. For example, the vehicle's speed and position changes should be consistent, and the engine speed and vehicle speed should conform to a certain proportional relationship. If inconsistency is detected, data correction or marking as low-confidence data is required. After these processing steps, the final result is a structured, standardized, and highly reliable data packet. This data packet is encoded in binary format to optimize storage space and transmission efficiency, and checksum information is attached to ensure data integrity. This data packet serves as input for subsequent control strategy generation and decision-making processes.

[0055] 102. Retrieve and select a pre-trained deep learning neural network model library according to a standardized data package to obtain a basic model corresponding to the current autonomous driving truck, and dynamically adjust the model parameters of the basic model according to the standardized data package to obtain an optimized adaptive control model;

[0056] In one embodiment of the present invention, the pre-trained deep learning neural network model library is retrieved and selected according to the standardized data packet to obtain the basic model currently corresponding to the self-driving truck, and the model parameters of the basic model are dynamically adjusted according to the standardized data packet to obtain the optimized adaptive control model, including: parsing the standardized data packet to obtain the current load value, road slope information and curvature information in the standardized data packet; searching the deep learning neural network model library according to the current load value, and selecting the basic model that matches the current load value; normalizing the road slope information and curvature information in the standardized data packet to obtain a normalized feature vector, and inputting the normalized feature vector into a preset parameter adjustment network to obtain a model parameter adjustment amount; updating the model parameters of the basic model according to the model parameter adjustment amount to obtain the optimized adaptive control model.

[0057] Specifically, standardized data packets are usually stored in binary format, so they need to be decoded using a predefined data structure template. During the decoding process, the system extracts the current load value, road slope information, and curvature information in sequence according to the fixed byte offset and data type definition. The load value is usually expressed in floating point form in kilograms; the road slope information may contain multiple values, such as the current slope angle (expressed in degrees or radians) and the sequence of slope changes within a certain distance in the future; the curvature information may include parameters such as curvature or curvature change rate. The parsing process also includes data validity verification, such as checking whether the value is within a reasonable range, and handling possible data missing or abnormal situations. If a data anomaly is encountered, the system will use a preset default value or an interpolation result based on historical data to replace it to ensure the continuity of subsequent processing.

[0058] Specifically, based on the current load value obtained through analysis, the system searches the pre-trained deep learning neural network model library. This model library contains multiple basic models trained for different load ranges. The retrieval process uses fast search algorithms, such as binary search or hash mapping, to improve efficiency. Each basic model has a predefined applicable load range. The system compares the current load value with these ranges and selects the best matching model. If the current load value happens to fall between the applicable ranges of the two models, the system will select the model whose upper limit is closer to the current load value. The selection process also considers the performance indicators of the model, such as historical prediction accuracy and computational complexity, and weighs between multiple candidate models. The selected basic model contains pre-trained neural network structures and initial weights, which are trained based on a large amount of historical data and simulation data, and can better adapt to the control needs under general conditions.

[0059] Specifically, the road slope information and curvature information in the standardized data package are normalized to eliminate the dimensional differences between different features so that they have the same importance in the subsequent parameter adjustment network. The normalization process uses the min-max normalization method to map each feature value to the [0, 1] interval. Specifically, for the road slope, the maximum and minimum slopes in the historical data are taken as the normalization benchmarks; for the curvature, multiple parameters may need to be considered, such as the maximum and average values ​​of the curvature. The normalized features form a vector, the dimension of which depends on the complexity of the original data and may usually be between 10 and 50. Subsequently, this normalized feature vector is input into the preset parameter adjustment network. The parameter adjustment network is a small neural network, usually using a multi-layer perceptron (MLP) structure, containing 1-2 hidden layers. The number of nodes in the input layer of this network is the same as the dimension of the normalized feature vector, and the number of nodes in the output layer corresponds to the number of parameters that need to be adjusted in the basic model. The network uses ReLU or tanh activation functions to capture the nonlinear relationship between features and parameter adjustment. The output of the network is the adjustment of the model parameters, which reflects the impact of the current road conditions on the control strategy. Finally, the basic model is updated using the obtained model parameter adjustments. The update process uses a weighted summation method, that is, a linear combination of the original parameter values ​​and the adjustment amounts. Specifically, for each layer of the neural network in the basic model, its weight matrix W and bias vector b will be adjusted. After the update is completed, the system will normalize the new parameter set, such as L2 regularization, to prevent the parameter value from being too large and causing the model to overfit. The updated model is the optimized adaptive control model, which combines the pre-trained general knowledge and the adjustments in the current specific situation, and can better adapt to the current driving environment and vehicle status. The optimized model will be used to generate subsequent control strategies until the next model update.

[0060] Furthermore, the updating of the model parameters of the basic model according to the model parameter adjustment amount to obtain the optimized adaptive control model includes: decomposing the model parameter adjustment amount to obtain the parameter adjustment matrix corresponding to each layer of the basic model; updating and calculating the weight matrix and bias vector of the basic model according to the parameter adjustment matrix and the preset parameter updating rule to obtain the updated model parameters; normalizing the updated model parameters to obtain the normalized model parameters, and applying the normalized model parameters to the basic model to obtain the preliminary optimized adaptive control model; and fine-tuning the preliminary optimized adaptive control model to obtain the optimized adaptive control model.

[0061] Specifically, the base model is usually a deep neural network, containing multiple convolutional layers, fully connected layers, etc. Each layer has its own specific weight matrix and bias vector. The parameter adjustment is a one-dimensional vector, which needs to be decomposed into multiple matrices and vectors according to the predefined mapping relationship. For example, assuming that the base model has 3 convolutional layers and 2 fully connected layers, the parameter adjustment will be decomposed into 5 weight adjustment matrices and 5 bias adjustment vectors. The decomposition process uses operations such as reshape and split to ensure that each decomposed adjustment matrix or vector is consistent with the parameter size of the corresponding layer. In addition, the decomposition process also includes scale adjustment, which appropriately scales the original adjustment according to the number of parameters in each layer to balance the update amplitude of different layers. These decomposed adjustment matrices and vectors constitute the parameter adjustment matrix set, which provides direct input for subsequent parameter updates.

[0062] The core step of model optimization is to update the weight matrix and bias vector of the basic model according to the parameter adjustment matrix and the preset parameter update rule. The update rule is usually based on a variant of the gradient descent method, such as the Adam or RMSprop algorithm. In specific implementation, for each layer of the basic model, the following operations are performed: First, the original weight matrix of the layer is added to the corresponding adjustment matrix, and the original bias vector is added to the corresponding adjustment vector. This addition operation is not a simple element addition, but an update method with momentum, which can smooth parameter updates and reduce oscillations. Then, the learning rate decay strategy is applied to dynamically adjust the learning rate according to the current number of iterations to achieve more refined parameter adjustments in the later stages of training. The update process also includes gradient clipping to prevent instability caused by gradient explosion. After the update calculation is completed, a new set of weight matrices and bias vectors are obtained, which reflect the adaptive adjustment of the basic model to the current driving environment. The updated model parameters are normalized to prevent the parameter values ​​from being too large or too small, resulting in model performance degradation or numerical instability. The normalization process uses batch normalization technology. For each layer’s weight matrix, first calculate its mean and variance, then normalize it. For the bias vector, simply subtract the mean. After normalization, apply scaling and translation operations. This process not only standardizes the parameter distribution, but also retains the model’s expressive power. After the normalization process is completed, these parameters are applied to each layer of the base model to replace the original parameters. This process involves the reconstruction of the model structure, and the connection relationship of each layer needs to be carefully handled to ensure the correct mapping of the parameters. After applying the new parameters, the resulting model is a preliminarily optimized adaptive control model, which has adapted to the current driving environment and vehicle status to a certain extent.

[0063] Fine-tuning the initially optimized adaptive control model is the last step of the optimization process, which aims to further improve the adaptability of the model to the current specific scenario. Fine-tuning training uses the currently collected real-time data as the training set, which includes vehicle status, environmental information, and the corresponding optimal control strategy. The training process adopts the small batch stochastic gradient descent method, and each batch contains 100-200 data samples collected recently. The loss function is designed as a multi-objective optimization problem, including the weighted sum of multiple indicators such as trajectory tracking error, energy efficiency, and ride comfort. The weight coefficient is dynamically adjusted according to the current driving environment. For example, the weight of the safety indicator may be increased under complex road conditions. The early stopping strategy is used in the training process, that is, the training is stopped when the performance on the validation set does not improve for multiple consecutive rounds to prevent overfitting. At the same time, the learning rate decay strategy is adopted to gradually reduce the learning rate as the number of training rounds increases, so as to achieve more refined parameter adjustment. In addition, regularization techniques such as L2 regularization and Dropout are introduced to improve the generalization ability of the model. Fine-tuning training is usually performed for 10-20 epochs, and the model performance is evaluated on the validation set after each epoch. After the training is completed, the model with the best performance on the validation set is selected as the final optimized adaptive control model. This model not only inherits the general knowledge of the basic model, but also adapts to the current specific driving environment and can generate more accurate and efficient control strategies.

[0064] 103. Input the standardized data packet into the adaptive control model for operation to obtain candidate control instructions, and perform rapid simulation and multi-objective optimization calculation on the candidate control instructions to obtain the optimal control strategy;

[0065] In one embodiment of the present invention, the standardized data packet is input into the adaptive control model for operation to obtain candidate control instructions, and the candidate control instructions are quickly simulated and multi-objective optimized to obtain the optimal control strategy, which includes: inputting the standardized data packet into the adaptive control model to obtain multiple groups of candidate control instructions, and quickly simulating each group of candidate control instructions to obtain corresponding predicted trajectories and state sequences; evaluating the predicted trajectories and state sequences according to a preset multi-objective optimization function to obtain performance indicators for each group of candidate control instructions; using a non-inferior sorting genetic algorithm to iteratively optimize the candidate control instructions based on the performance indicators of each group of candidate control instructions to obtain a Pareto optimal solution set, and generating an optimal control strategy based on the Pareto optimal solution set.

[0066] Specifically, the standardized data packets contain vehicle state parameters and environmental information, which need to be converted into an input format acceptable to the model. Typically, this involves reshaping the data into a specific tensor shape. The adaptive control model is a deep neural network that contains multiple layers, such as convolutional layers, recurrent layers, and fully connected layers. The data passes through these layers in sequence, and each layer performs a specific operation. For example, the convolutional layer may be used to extract spatial features, the recurrent layer processes timing information, and the fully connected layer integrates all features to make decisions. The output of the model is multiple sets of candidate control instructions, each of which contains parameters such as steering angle, acceleration, and braking force. In order to increase diversity, random sampling or temperature scaling techniques are introduced in the model output layer to generate multiple sets of slightly different candidate instructions. These candidate instructions represent different possible control strategies, providing a rich selection space for the subsequent optimization process.

[0067] Specifically, the simulation process for fast simulation of each set of candidate control instructions uses a simplified vehicle dynamics model that includes parameters such as vehicle mass, moment of inertia, and tire characteristics. The simulation starts with the current vehicle state, applies control instructions for each time step (usually 0.1 seconds), and calculates the new state of the vehicle. This process is repeated many times, usually simulating driving conditions for the next 5-10 seconds. The simulation takes into account environmental factors such as road geometry, slope, friction coefficient, and also simulates the movement of other vehicles. In order to improve computational efficiency, parallel computing technology is used to simulate multiple sets of instructions simultaneously. The simulation results include predicted trajectories (a series of position coordinates) and state sequences (variations of parameters such as speed, acceleration, and yaw rate over time). These results reflect the expected effect of each set of control instructions in the current environment and provide basic data for subsequent performance evaluation.

[0068] Specifically, evaluating the predicted trajectory and state sequence according to the preset multi-objective optimization function is the process of quantifying the performance of each set of candidate control instructions. The multi-objective optimization function usually includes multiple aspects such as safety, efficiency, and comfort. Safety indicators may include the minimum distance to other vehicles, the distance to the road boundary, etc.; efficiency indicators may include driving distance, energy consumption, etc.; comfort indicators may involve acceleration rate of change (sudden acceleration or sudden braking), lateral acceleration, etc. Each indicator has its own specific calculation method. For example, the minimum distance to other vehicles can be obtained by calculating the Euclidean distance between the predicted trajectory and the predicted trajectory of other vehicles; energy consumption can be calculated based on the speed and acceleration sequence using a pre-calibrated energy consumption model. These indicators are normalized to the same scale (usually between 0 and 1), and then weighted summed according to the preset weights to obtain the comprehensive performance indicators of each set of candidate control instructions. The setting of weights reflects the relative importance of each goal in different situations. For example, the weight of safety may be increased in complex traffic environments.

[0069] Specifically, the optimization of candidate control instructions using the non-inferior sorting genetic algorithm (NSGA-II) is an iterative process. First, the candidate control instructions are encoded as a gene sequence, each gene representing a control parameter. The initial population consists of these candidate instructions. Then, each individual in the population is non-dominated and sorted to divide the population into different fronts. Within the same front, the crowding distance is calculated to maintain diversity. Next, the parent individuals are selected using the tournament selection method, and crossover and mutation operations are performed to generate offspring. The crossover operation may use simulated binary crossover (SBX), while mutation uses polynomial mutation. The new generation population consists of the best individuals in the parent and offspring, keeping the population size unchanged. This process is repeated many times, usually 50-100 generations. In each generation, the Pareto front, that is, the non-dominated solution set, is updated. Finally, the algorithm converges to a set of Pareto optimal solutions that achieve the best balance between different objectives. From this Pareto optimal solution set, a most suitable solution is selected as the final optimal control strategy based on the characteristics of the current driving scenario. The selection may be based on additional decision rules, such as preferring solutions that are safe, or solutions that are the most balanced across objectives.

[0070] Furthermore, the predicted trajectory and state sequence are evaluated according to a preset multi-objective optimization function to obtain performance indicators for each group of candidate control instructions, including: comparing the predicted trajectory with the reference trajectory generated by a high-level path planning module, calculating the weighted sum of lateral and longitudinal deviations, and obtaining a trajectory tracking accuracy indicator; calculating the jitter amount of the autonomous driving truck movement according to the speed change rate and lateral acceleration in the state sequence to obtain a ride comfort indicator; calculating the energy consumption per unit mileage based on the engine operating point data and vehicle mass in the state sequence, combined with a pre-calibrated fuel consumption model, to obtain an energy efficiency indicator; using the spatiotemporal relationship between the predicted trajectory and dynamic obstacles, calculating the time margin for potential collisions to obtain a driving safety indicator; and performing weighted summation of the trajectory tracking accuracy indicator, ride comfort indicator, energy efficiency indicator, and driving safety indicator using the weights determined by the hierarchical analysis method to obtain a performance indicator.

[0071] Specifically, first, the two trajectories need to be represented in the same coordinate system, usually the vehicle local coordinate system. The reference trajectory is the ideal driving path generated by the high-level path planning module based on the target location, road network and traffic rules, while the predicted trajectory is the actual possible driving path obtained by fast simulation. During the comparison process, for each time point, the shortest distance from the point on the predicted trajectory to the reference trajectory is calculated, and this distance can be decomposed into lateral deviation and longitudinal deviation. The lateral deviation reflects the degree to which the vehicle deviates from the expected driving route, and the longitudinal deviation indicates the advance or lag of the vehicle in the driving direction. The projection method is usually used to calculate these two deviations. The predicted trajectory point is projected onto the reference trajectory. The distance between the projection point and the origin is the longitudinal deviation, and the vertical distance between the projection point and the origin is the lateral deviation. In order to obtain a comprehensive trajectory tracking accuracy index, the lateral and longitudinal deviations are weighted and summed. The setting of the weight reflects the relative importance of lateral control and longitudinal control. Usually, the weight of the lateral deviation is slightly higher because it is directly related to the lane keeping ability. Finally, the weighted deviations over the entire prediction time period are averaged to obtain a single trajectory tracking accuracy index. The smaller the index, the higher the tracking accuracy.

[0072] Specifically, calculating the jitter of the autonomous truck's motion based on the velocity change rate and lateral acceleration in the state sequence is the key to evaluating ride comfort. The velocity change rate, that is, the rate of change of acceleration (commonly known as "sudden acceleration and sudden braking"), directly affects the comfort of passengers in the front and rear directions. The lateral acceleration affects the comfort of passengers in the left and right directions, especially when turning or changing lanes. The calculation process first numerically differentiates the velocity sequence to obtain the acceleration sequence, and then differentiates the acceleration sequence to obtain the jerk sequence. For the lateral acceleration, it is directly extracted from the state sequence. Then, the absolute values ​​of the jerk and lateral acceleration are compared with the human comfort threshold respectively. Generally, the comfort threshold of jerk is about 2.5m / s³, and the comfort threshold of lateral acceleration is about 2m / s². Values ​​exceeding these thresholds are accumulated, and the larger the accumulated value, the worse the ride comfort. In order to comprehensively consider these two factors, they can be normalized to the same scale (for example, between 0 and 1) and then weighted summed. The weight setting can be adjusted according to the specific application scenario. For example, on highways, longitudinal comfort may be more important, while on urban roads, lateral comfort may be more important. The final ride comfort index is a dimensionless value. The smaller the value, the better the ride comfort.

[0073] Specifically, the process of evaluating energy efficiency is to calculate the energy consumption per unit mileage based on the engine operating point data and vehicle mass in the state sequence combined with a pre-calibrated fuel consumption model. First, the engine speed and torque data are extracted from the state sequence. These two parameters together define the engine operating point. Then, these operating points are mapped to the pre-calibrated fuel consumption model. The fuel consumption model is usually a two-dimensional lookup table or a polynomial fitting function, which reflects the instantaneous fuel consumption rate at different speeds and torques. For each time step, the instantaneous fuel consumption rate is calculated by looking up the table or function, in grams per second. Next, the influence of vehicle mass is considered. Mass information can be obtained from on-board sensors or estimated values ​​can be used. The larger the mass, the more energy is required for the same acceleration. Therefore, the instantaneous fuel consumption rate is multiplied by a coefficient related to the mass to obtain the corrected fuel consumption rate. Then, the corrected fuel consumption rate over the entire prediction period is integrated to obtain the total fuel consumption. At the same time, the total driving distance of the predicted trajectory is calculated. Finally, the total fuel consumption is divided by the total driving distance to obtain the energy consumption per unit mileage, in liters per 100 kilometers or grams per kilometer. This indicator directly reflects energy efficiency, and the smaller the value, the higher the efficiency.

[0074] Specifically, using the spatiotemporal relationship between the predicted trajectory and dynamic obstacles to calculate the time margin for potential collisions is an important method for evaluating driving safety. First, it is necessary to predict the movement trajectory of surrounding dynamic obstacles (such as other vehicles and pedestrians) in the future time period. This is usually achieved through a Kalman filter or a more complex trajectory prediction algorithm. Then, the predicted trajectory of the autonomous truck is matched with the predicted trajectory of these obstacles in time and space. For each time step, the distance between the truck and each obstacle is calculated. If at a certain point in time, this distance is less than a predetermined safety threshold, it is considered that there is a potential collision risk. The time margin is defined as the time interval from the current moment to the first occurrence of a potential collision risk. The size of the vehicle and the obstacle must also be considered in the calculation, usually represented by a bounding box. In addition, speed differences must be considered, because objects approaching at high speed are more dangerous than objects approaching at low speed. Therefore, when calculating the time margin, a speed factor is introduced for adjustment. The larger the final time margin, the higher the driving safety. If no potential collision risk is detected during the entire prediction time period, the time margin is set to the maximum value (such as the length of the prediction time period).

[0075] Specifically, the weighted summation of the trajectory tracking accuracy index, ride comfort index, energy efficiency index and driving safety index using the weights determined by the hierarchical analysis method is the last step to obtain the comprehensive performance index. The analytic hierarchy process (AHP) is a structured technique used to determine the relative importance of each factor in complex decision-making problems. In this process, a hierarchical structure is first established, with the top layer being the overall goal (optimal control strategy) and the next layer being the four sub-goals (trajectory tracking accuracy, ride comfort, energy efficiency and driving safety). Then, through expert judgment or historical data analysis, the four sub-goals are compared in pairs to form a judgment matrix. The elements of the judgment matrix represent the importance of one goal relative to another, usually using a scale of 1-9. Next, the eigenvector of the judgment matrix is ​​calculated, which is the weight of each sub-goal. After obtaining the weights, a consistency check is required to ensure the rationality of the judgment. If the consistency check is passed, these weights can be used. Finally, the four indicators are multiplied by their corresponding weights and then summed to obtain the final comprehensive performance index.

[0076] 104. The optimal control strategy is executed by the controller on the self-driving truck, and the actual response of the self-driving truck is monitored in real time during the execution process. The parameters of the controller are adaptively adjusted according to the deviation between the monitored actual response and the expected response.

[0077] In one embodiment of the present invention, the optimal control strategy is executed by a controller on the autonomous driving truck, and the actual response of the autonomous driving truck is monitored in real time during the execution process. According to the deviation between the monitored actual response and the expected response, the parameters of the controller are adaptively adjusted, including: converting the optimal control strategy into control instructions executable by the controller on the autonomous driving truck, and executing the control instructions through the controller; collecting actual state information of the autonomous driving truck in real time to obtain actual response data, and calculating the deviation between the actual response data and the expected response to obtain a response error; calculating the adjustment amount of the controller parameters according to the response error to obtain a parameter update value, and using the parameter update value to adaptively adjust the parameters of the controller.

[0078] Specifically, the optimal control strategy is usually expressed in a high-level abstract form, such as the desired trajectory points, velocity profile, or acceleration profile. The conversion process first needs to decompose these high-level representations into specific control parameters, such as steering angle, throttle opening, and braking force. This decomposition process involves the inverse operation of the vehicle dynamics model, for example, calculating the required engine torque based on the desired acceleration and the current vehicle speed. The conversion also needs to consider the physical limitations of the vehicle, such as the maximum steering angle, maximum acceleration, etc., to ensure that the generated instructions are within the vehicle's capabilities. The converted control instructions are usually generated at a fixed frequency (such as 50Hz), and each control cycle contains a complete set of control parameters. These parameters are encoded as digital signals that can be recognized by the controller and are usually transmitted via the CAN bus. After receiving these instructions, the controller will perform safety checks to ensure that the instructions will not lead to dangerous operations, and then convert them into specific execution signals, such as sending a voltage signal to the steering motor to change the steering angle, or sending a torque request to the engine management system to adjust the speed.

[0079] Specifically, the real-time acquisition process of the actual state information of the autonomous truck utilizes various sensors on the vehicle, including but not limited to GPS, inertial measurement unit (IMU), wheel speed sensor, steering angle sensor, etc. These sensors operate at different frequencies, so data synchronization is first required, usually using a timestamp alignment method. The synchronized data is initially filtered to remove outliers and high-frequency noise. Then, a sensor fusion algorithm, such as an extended Kalman filter, is used to fuse the multi-source data to obtain a more accurate state estimate. This state estimate includes key parameters such as the vehicle's position, velocity, acceleration, and yaw rate. Next, these actual states are compared with the expected response. The expected response is predicted based on the previously executed control instructions and the vehicle dynamics model. The comparison process calculates the error of each state variable, such as position deviation, velocity deviation, etc. These deviations are organized into an error vector that reflects various aspects of the control effect. In order to obtain a comprehensive response error indicator, the error vectors are weighted and summed, and the weights reflect the importance of different state variables. For example, in some scenarios, position accuracy may be more important than velocity accuracy. The final response error is a scalar value that represents the overall deviation between the current control effect and the expected target.

[0080] Calculating the adjustment of controller parameters based on the response error and making adaptive adjustments requires building a performance indicator function based on the response error, usually in the form of square error or absolute error. Then, calculate the gradient of this performance indicator function with respect to the controller parameters. The gradient calculation can be done analytically (if the performance indicator function has a clear mathematical expression) or numerically (such as the finite difference method). After obtaining the gradient, the gradient descent method is used to calculate the adjustment of the parameters. The size of the adjustment is determined by the learning rate, and the selection of the learning rate needs to balance the adjustment speed and stability. In order to prevent sudden changes in parameters, the adjustment amount is usually limited to ensure that the adjustment amplitude is within a reasonable range each time. The calculated adjustment amount is added to the current controller parameters to obtain the updated parameter value. When updating the parameters, the coupling relationship between the parameters must also be considered to ensure that the updated parameter set still meets the stability conditions of the controller. After the update is completed, the new parameter set is immediately applied to the controller. This adaptive adjustment process is continuous, allowing the controller to continuously adapt to changing vehicle states and environmental conditions to maintain optimal control performance. In order to further improve the robustness of the adjustment, a sliding window mechanism can be introduced to comprehensively consider the response errors at multiple recent time points to avoid excessive adjustments due to instantaneous fluctuations.

[0081] Furthermore, the method of calculating the adjustment amount of the controller parameters according to the response error to obtain the parameter update value, and using the parameter update value to adaptively adjust the parameters of the controller includes: performing spectral analysis on the response error to obtain the frequency distribution characteristics of the response error, and weighting the response error in sections according to the frequency distribution characteristics to obtain a weighted error vector; using the recursive least squares method to model the relationship between the weighted error vector and the parameters of the controller to obtain a parameter sensitivity matrix; based on the parameter sensitivity matrix and a preset learning rate, using the gradient descent method to calculate the parameter adjustment amount to obtain a preliminary parameter update value; performing adaptive dynamic range limitation on the preliminary parameter update value to obtain a bounded parameter adjustment amount, and performing exponential sliding averaging on the bounded parameter adjustment amount and the historical adjustment amount to obtain a parameter update value, and superimposing the parameter update value on the parameters of the controller to complete the adaptive adjustment.

[0082] Specifically, the response error data in the time domain is converted to the frequency domain through the fast Fourier transform (FFT). The FFT algorithm usually requires the number of data points to be a power of 2, so the original data needs to be zero-filled or truncated. The spectrum obtained after the transformation contains amplitude and phase information, with the amplitude spectrum being the focus. The spectrum analysis results show the energy distribution of the error at different frequencies. Usually, the low-frequency part reflects the slow changes of the system, while the high-frequency part is related to noise and fast disturbances. Based on this frequency distribution feature, the response error is weighted in sections. Specifically, the spectrum is divided into multiple frequency bands (such as low frequency, medium frequency, and high frequency), and different weights are assigned to each frequency band. Low-frequency errors are usually more valued because they reflect the continuous deviation of the system, so they are assigned higher weights; while high-frequency errors may contain more noise and are given lower weights. This weighting scheme is applied to the original time domain error through the inverse Fourier transform to obtain a weighted error vector.

[0083] Specifically, the relationship between the weighted error vector and the controller parameters is modeled using recursive least squares (RLS). The advantage of the RLS algorithm is that it can learn and adapt to time-varying systems online, which is very suitable for dynamic environments such as autonomous driving. The implementation process first initializes the parameters of the RLS algorithm, including the initial estimated covariance matrix P and the forgetting factor λ. The forgetting factor is usually set between 0.95 and 0.99 to balance the algorithm's emphasis on new and old data. Then, for each time step of the data sample (including the weighted error vector and the current controller parameters), the following update steps are performed: calculate the gain vector K, update the parameter estimate θ, and update the estimated covariance matrix P. This process is iterative, and as more data is input, the model will continue to optimize. The final parameter estimate θ is the parameter sensitivity matrix, which reflects the degree of influence of the controller parameter change on the weighted error. Each row of this matrix corresponds to a controller parameter, and each column corresponds to a component of the weighted error vector. The larger the absolute value of the matrix element, the more significant the influence of the corresponding parameter on the error component. The parameter sensitivity matrix provides important guidance for subsequent parameter adjustments, so that the adjustment process can improve system performance more targeted.

[0084] Specifically, based on the parameter sensitivity matrix and the preset learning rate, the gradient descent method is used to calculate the parameter adjustment amount. First, an objective function needs to be defined, which is usually the norm of the weighted error vector (such as the Euclidean norm or the infinite norm). Then, the parameter sensitivity matrix is ​​used to calculate the gradient of the objective function with respect to each controller parameter. Specifically, the gradient vector g = S^T * e, where S is the parameter sensitivity matrix and e is the weighted error vector. After the gradient is calculated, according to the principle of the gradient descent method, the parameter adjustment amount Δθ = -η * g, where η is the preset learning rate. The choice of learning rate is crucial. Too large may lead to instability, and too small will lead to slow convergence. To solve this problem, an adaptive learning rate strategy can be used, such as the Adam optimizer, which can dynamically adjust the learning rate of each parameter based on the historical information of the gradient. The calculated parameter adjustment amount Δθ is the preliminary parameter update value. This update value reflects the steepest descent direction in the current state of the system, and theoretically can reduce the error the fastest. However, directly using this preliminary update value may cause the parameter to change too much, so further processing is required.

[0085] Specifically, adaptive dynamic range limiting of the preliminary parameter update values ​​is an important step to ensure the stability and safety of parameter adjustment. This process first defines a range of allowable changes for each controller parameter, which is usually based on the physical meaning of the parameter and the safety limits of the system. For example, for the gain parameter of the PID controller, it may be limited to no more than 10% of the current value at each adjustment. Then, the preliminary parameter update values ​​are checked one by one, and if they exceed the predetermined range, they are clipped to the range boundary. This range limit is dynamic and will be adjusted according to the current state and historical performance of the system. For example, if the system is stable for a long time, the limit range can be gradually relaxed; if violent fluctuations are detected, the limit is tightened. After the range limit, the bounded parameter adjustment amount is obtained. Next, this bounded adjustment amount is exponentially averaged with the historical adjustment amount. This averaging process can reduce the fluctuation of the parameter adjustment and make the adjustment process smoother. Finally, the average adjustment amount is used as the final parameter update value and superimposed on the current controller parameter to complete an adaptive adjustment.

[0086] In this embodiment, a standardized data packet containing vehicle state parameters and environmental information is generated by filtering and structuring the multi-dimensional data collected by the vehicle sensor network. According to the standardized data packet, the deep learning neural network model library is retrieved, the currently applicable basic model is selected, and its parameters are dynamically adjusted to generate an optimized adaptive control model. The standardized data packet is input into the adaptive control model operation to generate candidate control instructions, and the optimal control strategy is generated through fast simulation and multi-objective optimization calculation. Finally, the optimal control strategy is executed by the vehicle controller, and the vehicle response is monitored in real time, and the controller parameters are adaptively adjusted based on the deviation. The present invention realizes the dynamic adaptability and strategy optimization of the control model, and improves the safety and efficiency of autonomous driving trucks in a changing environment.

[0087] The above describes the adaptive control method of the autonomous driving truck in the embodiment of the present invention. The following describes the adaptive control system of the autonomous driving truck in the embodiment of the present invention. Figure 2 , an embodiment of the adaptive control system of the autonomous driving truck in the embodiment of the present invention includes:

[0088] The data processing module 201 is used to filter and structure the multi-dimensional data collected by the on-board sensor network of the self-driving truck to obtain a standardized data packet containing vehicle state parameters and environmental information;

[0089] A model optimization module 202, configured to retrieve and select a pre-trained deep learning neural network model library according to the standardized data packet, obtain a basic model currently corresponding to the autonomous driving truck, and dynamically adjust model parameters of the basic model according to the standardized data packet to obtain an optimized adaptive control model;

[0090] A strategy generation module 203 is used to input the standardized data packet into the adaptive control model for operation to obtain candidate control instructions, and to perform fast simulation and multi-objective optimization calculation on the candidate control instructions to obtain an optimal control strategy;

[0091] The execution adjustment module 204 is used to execute the optimal control strategy through the controller on the self-driving truck, and monitor the actual response of the self-driving truck in real time during the execution process, and adaptively adjust the parameters of the controller according to the deviation between the monitored actual response and the expected response.

[0092] In an embodiment of the present invention, the adaptive control system of the autonomous driving truck runs the adaptive control method of the autonomous driving truck, and the adaptive control system of the autonomous driving truck generates a standardized data packet containing vehicle state parameters and environmental information by filtering and structured processing the multi-dimensional data collected by the on-board sensor network. According to the standardized data packet, the deep learning neural network model library is retrieved, the currently applicable basic model is selected, and its parameters are dynamically adjusted to generate an optimized adaptive control model. The standardized data packet is input into the adaptive control model operation to generate candidate control instructions, and the optimal control strategy is generated through fast simulation and multi-objective optimization calculation. Finally, the optimal control strategy is executed by the vehicle controller, and the vehicle response is monitored in real time, and the controller parameters are adaptively adjusted based on the deviation. The present invention realizes the dynamic adaptability and strategy optimization of the control model, and improves the safety and efficiency of the autonomous driving truck in a changing environment.

[0093] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device or unit described above can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0094] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0095] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive control method for an autonomous driving truck, characterized in that: The adaptive control method of the autonomous driving truck includes: Filter and structure the multi-dimensional data collected by the onboard sensor network on the self-driving truck to obtain a standardized data package containing vehicle state parameters and environmental information; Retrieving and selecting a pre-trained deep learning neural network model library according to the standardized data packet to obtain a basic model currently corresponding to the autonomous driving truck, and dynamically adjusting model parameters of the basic model according to the standardized data packet to obtain an optimized adaptive control model; Inputting the standardized data packet into the adaptive control model for operation to obtain candidate control instructions, and performing fast simulation and multi-objective optimization calculation on the candidate control instructions to obtain an optimal control strategy; The optimal control strategy is executed by a controller on the self-driving truck, and the actual response of the self-driving truck is monitored in real time during the execution process. The parameters of the controller are adaptively adjusted according to the deviation between the monitored actual response and the expected response.

2. The adaptive control method of an autonomous driving truck according to claim 1, characterized in that: The method of retrieving and selecting a pre-trained deep learning neural network model library according to the standardized data packet to obtain a basic model currently corresponding to the autonomous driving truck, and dynamically adjusting the model parameters of the basic model according to the standardized data packet to obtain an optimized adaptive control model includes: Parsing the standardized data packet to obtain the current load value, road slope information and curvature information in the standardized data packet; Searching a deep learning neural network model library according to the current load value, and selecting a basic model that matches the current load value; Normalizing the road slope information and curvature information in the standardized data packet to obtain a normalized feature vector, and inputting the normalized feature vector into a preset parameter adjustment network to obtain a model parameter adjustment amount; The model parameters of the basic model are updated according to the model parameter adjustment amount to obtain an optimized adaptive control model.

3. The adaptive control method of the autonomous driving truck according to claim 2, characterized in that: The updating of the model parameters of the basic model according to the model parameter adjustment amount to obtain the optimized adaptive control model includes: Decomposing the model parameter adjustment amount to obtain a parameter adjustment matrix corresponding to each layer of the basic model; According to the parameter adjustment matrix and the preset parameter update rule, the weight matrix and the bias vector of the basic model are updated and calculated to obtain updated model parameters; Normalizing the updated model parameters to obtain normalized model parameters, and applying the normalized model parameters to the basic model to obtain a preliminarily optimized adaptive control model; The initially optimized adaptive control model is fine-tuned and trained to obtain an optimized adaptive control model.

4. The adaptive control method of an autonomous driving truck according to claim 1, characterized in that: The step of inputting the standardized data packet into the adaptive control model for operation to obtain candidate control instructions, and performing fast simulation and multi-objective optimization calculation on the candidate control instructions to obtain the optimal control strategy includes: Inputting the standardized data packet into an adaptive control model to obtain multiple groups of candidate control instructions, and quickly simulating each group of candidate control instructions to obtain corresponding predicted trajectories and state sequences; The predicted trajectory and state sequence are evaluated according to the preset multi-objective optimization function to obtain the performance index of each group of candidate control instructions; A non-inferior sorting genetic algorithm is used to iteratively optimize the candidate control instructions based on the performance indicators of each group of candidate control instructions to obtain a Pareto optimal solution set, and an optimal control strategy is generated according to the Pareto optimal solution set.

5. The adaptive control method of the autonomous driving truck according to claim 4, characterized in that: The predicted trajectory and state sequence are evaluated according to the preset multi-objective optimization function to obtain the performance indicators of each group of candidate control instructions, including: Comparing the predicted trajectory with the reference trajectory generated by the high-level path planning module, calculating the weighted sum of lateral and longitudinal deviations, and obtaining a trajectory tracking accuracy index; According to the speed change rate and lateral acceleration in the state sequence, the jitter of the autonomous driving truck is calculated to obtain the ride comfort index; Based on the engine operating point data and vehicle mass in the state sequence, combined with the pre-calibrated fuel consumption model, the energy consumption per unit mileage is calculated to obtain the energy efficiency index; Using the spatiotemporal relationship between the predicted trajectory and dynamic obstacles, the time margin for potential collisions is calculated to obtain driving safety indicators. The trajectory tracking accuracy index, the ride comfort index, the energy efficiency index and the driving safety index are weighted and summed using weights determined by a hierarchy analysis method to obtain a performance index.

6. The adaptive control method of an autonomous driving truck according to claim 1, characterized in that: The optimal control strategy is executed by a controller on the autonomous driving truck, and the actual response of the autonomous driving truck is monitored in real time during the execution process. According to the deviation between the monitored actual response and the expected response, the parameters of the controller are adaptively adjusted, including: Converting the optimal control strategy into control instructions executable by a controller on the autonomous driving truck, and executing the control instructions through the controller; Collect the actual status information of the autonomous driving truck in real time to obtain the actual response data, and calculate the deviation between the actual response data and the expected response to obtain the response error; The adjustment amount of the controller parameters is calculated according to the response error to obtain the parameter update value, and the parameter update value is used to adaptively adjust the controller parameters.

7. The adaptive control method of the autonomous driving truck according to claim 6, characterized in that: The step of calculating the adjustment amount of the controller parameter according to the response error to obtain the parameter update value, and using the parameter update value to adaptively adjust the controller parameter includes: Performing spectrum analysis on the response error to obtain the frequency distribution characteristics of the response error, and performing piecewise weighting on the response error according to the frequency distribution characteristics to obtain a weighted error vector; The relationship between the weighted error vector and the controller parameters is modeled using the recursive least square method to obtain the parameter sensitivity matrix. Based on the parameter sensitivity matrix and the preset learning rate, the gradient descent method is used to calculate the parameter adjustment amount to obtain the preliminary parameter update value; The initial parameter update value is adaptively limited in dynamic range to obtain a bounded parameter adjustment amount, and the bounded parameter adjustment amount and the historical adjustment amount are exponentially averaged to obtain a parameter update value, and the parameter update value is superimposed on the parameters of the controller to complete the adaptive adjustment.

8. An adaptive control system for an autonomous driving truck, characterized in that: The adaptive control system of the autonomous truck includes: The data processing module is used to filter and structure the multi-dimensional data collected by the on-board sensor network on the self-driving truck to obtain a standardized data packet containing vehicle status parameters and environmental information; A model optimization module, configured to retrieve and select a pre-trained deep learning neural network model library according to the standardized data packet to obtain a basic model currently corresponding to the autonomous driving truck, and dynamically adjust model parameters of the basic model according to the standardized data packet to obtain an optimized adaptive control model; A strategy generation module, used for inputting the standardized data packet into the adaptive control model for operation to obtain candidate control instructions, and performing fast simulation and multi-objective optimization calculation on the candidate control instructions to obtain an optimal control strategy; The execution adjustment module is used to execute the optimal control strategy through the controller on the self-driving truck, and monitor the actual response of the self-driving truck in real time during the execution process, and adaptively adjust the parameters of the controller according to the deviation between the monitored actual response and the expected response.

Images