A delivery vehicle linkage control method and system based on AI intelligent perception

Abstract

The present application relates to the technical field of intelligent distribution vehicle control, in particular to a distribution vehicle linkage control method and system based on AI intelligent sensing. The present application builds a complete link from multi-source sensor data acquisition, impact risk prediction to vehicle motion control instruction generation, takes the real-time attitude information of the goods as the constraint condition and optimization target of the dynamic programming module, realizes the full-linkage control, and solves the problem that the goods state monitoring system and the vehicle dynamic programming system are independent of each other, lack information fusion and cooperative linkage mechanism, and the vehicle cannot actively adjust the motion state according to the real-time attitude of the goods in the cabin.

Description

Technical Field

[0001] This invention belongs to the field of intelligent delivery vehicle control technology, specifically relating to a delivery vehicle linkage control method and system based on AI intelligent perception. Background Technology

[0002] With the development of e-commerce, instant retail, and last-mile logistics, unmanned delivery vehicles have been deployed on a large scale in urban roads, industrial parks, campuses, and communities. When performing delivery tasks, these unmanned vehicles typically carry high-value or highly sensitive goods such as fragile items, fresh produce, medicines, and electronic products. During operation, these vehicles frequently traverse uneven road surfaces such as speed bumps, manhole covers, potholes, and seams. The resulting instantaneous acceleration impacts and continuous vibrations are among the main causes of cargo damage, misalignment, and collapse.

[0003] Autonomous vehicles have relatively mature environmental perception, localization, and trajectory planning capabilities. However, current planning algorithms typically construct their objective functions around safety, comfort, and timeliness, without incorporating the cargo's attitude as a constraint or objective into the planning process. Furthermore, existing cargo status monitoring systems are mostly used for local recording or anomaly alarms, and their outputs do not interact with the vehicle's dynamic planning module. This results in the cargo perception system and the vehicle planning system operating independently, lacking information fusion and collaborative mechanisms. Consequently, when facing uneven road surfaces, the vehicle cannot proactively adjust its motion based on the real-time attitude of the cargo inside the cargo.

[0004] To address the aforementioned issues, this invention proposes a delivery vehicle linkage control method and system based on AI intelligent perception. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, the present invention aims to provide a delivery vehicle linkage control method and system based on AI intelligent perception.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A delivery vehicle linkage control method based on AI intelligent perception includes: Cargo status data is collected in real time by a multi-source sensor array deployed in the cargo hold. The cargo status data includes at least the cargo attitude angle, angular velocity, linear acceleration, deformation coefficient, and offset of the bottom pressure center. The vehicle's external sensing module scans the road surface ahead to obtain information on the road surface unevenness distribution. Based on cargo status data and road surface unevenness distribution information, the pre-trained impact risk prediction model outputs cargo impact risk indicators within a future preset time window. Based on the cargo impact risk index, and under the premise of meeting the time constraints of delivery tasks and vehicle dynamics constraints, the longitudinal speed, longitudinal acceleration, steering angle and reference trajectory of the vehicle are jointly optimized to generate linkage control commands. The linkage control command is sent to the vehicle chassis actuator to realize real-time adjustment of the vehicle's motion state.

[0007] Preferably, the multi-source sensor group includes at least an inertial measurement unit, a vision sensor array, and a distributed pressure sensor array; the inertial measurement unit is used to collect the attitude angle, angular velocity, and linear acceleration of the cargo; the vision sensor array is used to identify the cargo contour and output the deformation coefficient and pose offset; and the distributed pressure sensor array is used to obtain the pressure distribution on the bottom surface of the cargo and output the offset of the bottom pressure center; the three types of sensors achieve timestamp alignment through an external trigger synchronization mechanism.

[0008] Preferably, the cargo impact risk prediction model adopts a hybrid structure of temporal convolutional network and long short-term memory network. It takes the cargo state characteristics of a preset number of time steps in the past and the road surface unevenness sequence of a preset number of time steps in the future as input, and outputs the impact risk index sequence within the preset time window in the future.

[0009] Preferably, the joint optimization adopts a multi-objective nonlinear model predictive control method, whose objective function includes at least a cargo safety term, a speed tracking term, a steering stability term, and a time constraint term. Each objective term is normalized with its corresponding preset benchmark value and then weighted and summed. The cargo safety term is determined by the square of the ratio of the impact risk index to the vehicle speed divided by the maximum vehicle speed, and the time constraint term is determined by the ratio of the deviation between the expected arrival time and the target arrival time divided by the target arrival time. The weight coefficients are dynamically adjusted according to the impact risk index. When the impact risk increases, the weight of the cargo safety term is increased and the weight of the time constraint term is decreased.

[0010] Preferably, the vehicle external perception module includes at least two of the following: a forward-facing camera, millimeter-wave radar, and three-dimensional lidar. It is used to perform semantic segmentation of the road ahead to identify different road surface categories, and to output the unevenness level and boundary position of each road segment in combination with the height difference of lidar point cloud. When a road segment with an unevenness level not lower than a preset threshold is detected ahead, the reference trajectory is adjusted in advance to detour through it, or a gradual deceleration is performed when the detour conditions are not met.

[0011] Preferably, the shock risk prediction model is trained using a supervised learning method. The training dataset comes from the operating data of multiple delivery vehicles with different load configurations in actual operation scenarios, covering various road conditions and loading states. The shock risk label is marked by events in which the actual posture of the goods exceeds the safety threshold. To address the imbalance between positive and negative samples, data augmentation strategies such as time axis scaling, random masking of sensor channels, and adding Gaussian noise are used to expand the positive samples. The impact risk label is defined as follows: when the actual change in the cargo's posture exceeds the safety threshold, that moment is marked as a positive sample.

[0012] Preferably, it also includes closed-loop feedback and online learning, including: after the linkage control command is sent to the vehicle chassis actuator, the cargo compartment perception subsystem continuously monitors the actual cargo status; calculates the root mean square error between the risk prediction value and the actual observation value; when the root mean square error continues to exceed a preset threshold for a preset number of times, the new data collected within the most recent preset time period is used as training samples to fine-tune the parameters of the impact risk prediction model.

[0013] Preferably, the method also includes desensitizing the road surface unevenness data collected by the vehicle and uploading it to the cloud, where it is aggregated and distributed to the entire fleet to achieve crowdsourced updates of road information; for high-frequency delivery routes, historical unevenness detection results are written into the map cache, and the historical risk distribution is directly called when the vehicle passes through again.

[0014] This invention also provides a delivery vehicle linkage control system based on AI intelligent perception, comprising: The cargo hold perception subsystem is used to collect cargo status data in real time through a multi-source sensor group deployed in the cargo hold. The cargo status data includes at least the cargo attitude angle, angular velocity, linear acceleration, deformation coefficient, and bottom pressure center offset. The external perception subsystem is used to scan the road surface ahead and obtain information on the road surface unevenness distribution; The impact risk prediction module is used to output cargo impact risk indicators within a preset time window based on cargo status data and road unevenness distribution information through a pre-trained impact risk prediction model. The dynamic programming linkage control module is used to jointly optimize the longitudinal speed, longitudinal acceleration, steering angle and reference trajectory of the vehicle based on the cargo impact risk index, under the premise of meeting the time constraints of the delivery task and the vehicle dynamics constraints, and generate linkage control commands. The execution and feedback subsystem is used to send linkage control commands to the vehicle chassis actuators to achieve real-time adjustment of the vehicle's motion state.

[0015] Preferably, the dynamic programming linkage control module adopts a multi-objective nonlinear model predictive control method, and its objective function includes at least cargo safety, speed tracking, steering smoothness, and time constraint terms; the weight coefficients are dynamically adjusted according to the impact risk index. The system also includes a system expansion and deployment module, which is used to interface the above modules with the original autonomous driving planning stack of the unmanned delivery vehicle in a modular manner, and to enable and degrade control through configurable switches; And it uploads road surface unevenness data collected by individual vehicles to the cloud to achieve crowdsourced updates for the entire fleet.

[0016] Beneficial effects 1. This invention constructs a complete link from multi-source sensor data acquisition and impact risk prediction to vehicle motion control command generation. It uses real-time cargo attitude information as the constraint and optimization target of the dynamic planning module, thereby realizing full-link linkage control. This solves the problems in the prior art where the cargo status monitoring system and the vehicle dynamic planning system are independent of each other, lack information fusion and collaborative linkage mechanism, and the vehicle cannot actively adjust its motion state according to the real-time attitude of the cargo in the cabin.

[0017] 2. This invention uses an external sensing module to scan the road surface ahead in real time, identify unevenness information such as speed bumps and potholes in advance, and uses a hybrid model of temporal convolutional network and long short-term memory network to predict the impact risk of goods in a temporal manner. Before the vehicle actually reaches the uneven road surface, it generates optimized control commands and adjusts the vehicle speed and steering in advance, transforming the traditional post-event passive compensation into pre-event active intervention. This solves the defects of being unable to accurately predict and respond in advance to upcoming speed bumps or potholes, and the protection effect being lagging and insufficient.

[0018] 3. This invention constructs a multi-objective weighted architecture that includes cargo impact risk and delivery timeliness in the objective function of nonlinear model predictive control, and designs an adaptive strategy that dynamically adjusts the weight coefficients according to the risk index. When the impact risk increases, the weight of cargo safety is automatically increased and the weight of delivery timeliness is reduced. At the same time, a hard constraint of forced vehicle speed triggered by impact risk is set to ensure that the vehicle is forced to slow down in high-risk scenarios, thus achieving a dynamic balance between cargo safety and delivery timeliness. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely for explaining the invention and are not intended to limit the scope of protection of the invention.

[0021] Example 1 Please see Figure 1 As shown in the figure, this embodiment provides a delivery vehicle linkage control method and system based on AI intelligent perception. The method includes the following steps: S1. Real-time data collection of cargo status is achieved through a multi-source sensor array deployed within the cargo hold; The multi-source sensor array includes at least an inertial measurement unit, a vision sensor array, and a distributed pressure sensor array.

[0022] Six-axis or nine-axis inertial measurement units are arranged on the top, side walls and bottom of the cargo hold, with a sampling rate of not less than 200Hz, preferably 200-1000Hz. The high-frequency band is used to capture transient impacts. Multiple inertial measurement units are fused through Kalman filtering to output the overall and local attitude angles (roll, pitch, yaw), angular velocity and linear acceleration of the cargo.

[0023] The visual sensor array is equipped with wide-angle or fisheye cameras at the four corners and top of the cargo hold, with infrared illumination and a frame rate of no less than 30fps, preferably 30-60fps, to balance computing power and recognition accuracy. The lightweight object detection network identifies the cargo outline, deformation and stacking status in real time, and outputs the deformation coefficient and pose offset. The lightweight object detection network preferably uses YOLOv8-nano.

[0024] The deformation coefficient is defined as the ratio of the current outline pixel area of ​​the cargo to the outline pixel area at the initial loading stage, i.e.: In the formula, The pixel area of ​​the cargo outline at the initial moment is automatically collected and stored by a vision sensor at a static moment after loading is completed and before the vehicle starts. This represents the pixel area of ​​the cargo outline at the current moment. When The greater the deviation from 1, the more severe the compression or deformation of the goods.

[0025] The pose offset includes the lateral and longitudinal offsets of the cargo center point in the cargo hold plane coordinate system. When the visual detection confidence level is lower than a preset threshold, Kalman filtering is used to track and predict the cargo center point to maintain continuous estimation of the pose offset until the visual detection confidence level is restored. The preset threshold is preferably 0.6.

[0026] A distributed pressure sensor array is constructed by laying a thin-film pressure sensor array on the cargo hold floor. The resolution is no less than 16×16, and the refresh rate is no less than 50Hz, preferably 50-100Hz, to meet slippage detection requirements. The pressure distribution map can identify whether slippage, tilting, or partial suspension has occurred on the cargo floor. Specifically, the offset of the floor pressure center is measured. Defined as the offset of the center of the weighted average of the position coordinates and pressure values ​​of each pressure detection point relative to the center of the cargo hold floor. Let the first... The coordinates of each detection point are: The corresponding pressure value is The total pressure on the bottom surface of the cargo is The coordinates of the pressure center are:

[0027] but ,in The coordinates are the center coordinates of the cargo hold floor.

[0028] The multi-source sensor array employs an externally triggered synchronization mechanism for timestamp alignment in both time and space, ensuring that the data timestamp deviation is less than 1ms. Raw data is preprocessed by the edge computing unit and then uploaded to the central computing platform as a structured data stream. The edge computing unit is preferably NVIDIA Jetson Orin.

[0029] S2. Scan the road ahead using the external sensing module to obtain information on the road surface unevenness distribution; The vehicle external perception module includes at least two of the following: a forward-facing camera, millimeter-wave radar, and 3D lidar. The effective detection distance is not less than 80m, preferably 60-120m, which is not less than the safe parking distance on urban roads, and the data refresh rate is not less than 20Hz.

[0030] The external perception module is used for road surface unevenness detection. Specifically, it uses a deep learning model to segment the road area ahead into categories such as "normal road surface," "speed bumps," "manhole covers," "potholes," and "joints." Combining the height difference of the LiDAR point cloud with acceleration signals, it outputs the unevenness level (0-5) and boundary location for each road segment. The preferred deep learning model is an improved LaneNet combined with a road surface semantic segmentation network.

[0031] The unevenness levels are classified as follows: Level 0 is a smooth road surface (height difference ≤ 2cm, vertical acceleration change ≤ 0.5m / s²); Level 1 is slightly uneven (height difference 2-4cm, vertical acceleration change 0.5-1.0m / s²); Level 2 is moderately uneven (height difference 4-6cm, vertical acceleration change 1.0-1.8m / s²); Level 3 is significantly uneven (height difference 6-8cm, vertical acceleration change 1.8-2.8m / s²); Level 4 is severely uneven (height difference 8-10cm, vertical acceleration change 2.8-4.0m / s²); and Level 5 is extremely uneven (height difference > 10cm, vertical acceleration change > 4.0m / s²).

[0032] Furthermore, the height difference is the maximum relative height of local bumps or depressions on the road surface detected by the lidar; the vertical acceleration change is the estimated vertical acceleration response value of the vehicle body when passing through at a reference speed (e.g., 20 km / h). The above grading standards are obtained based on a benchmark vehicle model (e.g., a typical delivery vehicle with a curb weight of approximately 1000 kg and a suspension stiffness of approximately 30 N / mm) through actual vehicle calibration on a standard test road surface.

[0033] Furthermore, in practical applications, if key characteristics such as the vehicle's curb weight, suspension stiffness, and tire parameters differ significantly from the benchmark vehicle, adaptive calibration correction is required according to the aforementioned classification logic. Specifically, this can be achieved by driving the actual vehicle on a standard road surface with a known roughness level at a benchmark speed, collecting the vehicle's vertical acceleration response data, establishing a mapping relationship between "height difference and vertical acceleration" for the vehicle model, and redetermining the acceleration thresholds corresponding to each level accordingly to ensure the consistency and accuracy of roughness level classification across different vehicle models. The benchmark speed is preferably 20 km / h.

[0034] The external perception module is also used for target obstacle recognition. Specifically, it simultaneously performs routine obstacle detection for pedestrians, vehicles, non-motorized vehicles, etc., to provide obstacle avoidance information for path planning.

[0035] The road surface unevenness distribution information is divided into multiple levels according to road segments; when a road segment with an unevenness level not lower than the preset threshold (such as ≥3) is detected ahead, the system can adjust the reference trajectory in advance to detour through it, or perform gradual deceleration to reduce the impact when the detour conditions are not met.

[0036] S3. Based on cargo status data and road unevenness distribution information, construct a cargo impact risk prediction model and output cargo impact risk indicators within a future preset time window; Extracting the instantaneous attitude angle of cargo from cargo hold sensing data. angular velocity Linear acceleration Visual deformation coefficient Offset from the center of pressure on the bottom surface Extracting the front view from external perception data Road surface roughness sequence within meters .

[0037] The cargo impact risk prediction model employs a hybrid structure of Temporal Convolutional Network (TCN) and Long Short-Term Memory Network (LSTM); the input is past data. Cargo status characteristics at each time step and future The road surface unevenness sequence at each time step is output as the future... A sequence of shock risk indicators for each time step, namely:

[0038]

[0039] Among them, risk indicators The higher the value, the higher the probability that the goods will slip, tip over or break in the future.

[0040] Furthermore, the structure of the TCN-LSTM hybrid network is as follows: the input cargo state features and road surface unevenness sequence are concatenated into a multi-dimensional temporal feature vector, which is then processed by a one-dimensional temporal convolutional layer for local feature extraction (the kernel size is 3, the dilation rate increases layer by layer to 1, 2, and 4, and the number of channels is 64, 128, and 256, respectively); the output of TCN is used as the input sequence of the LSTM layer, which contains two layers, each with a hidden state dimension of 128; the output is mapped to a scalar value of the impact risk index through a fully connected layer (dimensions 128, 64, and 1).

[0041] The model training employs a supervised learning approach. The training dataset consists of operational data from at least 10 delivery vehicles with different load configurations in real-world scenarios. Each vehicle accumulates no less than 200 hours of cargo compartment perception data and external road surface perception data, covering various road conditions such as urban roads, park roads, speed bump areas, and pothole areas, as well as various loading states such as empty, half-loaded, and fully loaded, to ensure the diversity and representativeness of the training samples. Impact risk labels are generated based on events where the actual posture change of the cargo exceeds a safety threshold (e.g., roll angle > 15° or pressure center offset exceeding 20% ​​of the cargo's bottom surface dimension).

[0042] To address the severe imbalance between positive and negative samples caused by the extremely small number of high-risk events such as cargo tipping or significant slippage in actual operations (typically accounting for less than 0.1% of the total sample), the following strategy is adopted to construct a balanced training set: The data augmentation strategy involves scaling the collected raw time-series data along the time axis (scaling factor 0.8-1.2), randomly masking sensor channels (masking ratio not exceeding 10%), and adding Gaussian noise (signal-to-noise ratio not less than 20dB), expanding the number of effective positive samples to more than 50 times the original number. The simulation data supplementation strategy, based on vehicle dynamics and cargo physics models, simulates the cargo attitude response of delivery vehicles at different speeds (5-40km / h), different road surface unevenness levels (0-5), and different cargo stacking methods (single-layer, double-layer, staggered stacking), generating simulated time-series data with accurate risk labels. The ratio of the total simulated data to the actual collected data does not exceed 1:1 to avoid the model over-reliance on the simulation distribution. Through these strategies, the proportion of high-risk positive samples in the final training set is increased to 5%-10%, with the remainder being negative and medium-risk samples, ensuring the model can effectively learn the temporal characteristics of high-risk events.

[0043] The loss function uses weighted mean squared error loss, assigning higher weights to high-risk samples (weight coefficients range from 3 to 5) to enhance the model's prediction accuracy for high-risk events; the optimizer is Adam; the initial learning rate is 10. -3 Training continues until the loss on the validation set converges. Furthermore, to verify the model's generalization ability, the training, validation, and test sets are divided in a 6:2:2 ratio, with data from different vehicles and different operating periods to avoid continuous data from the same vehicle appearing simultaneously in both the training and test sets.

[0044] Regarding thresholds and alarms, when When a high-risk alarm is triggered; when If the risk is high, a medium-risk warning will be triggered; otherwise, normal operation will continue. The risk prediction time domain will be consistent with the subsequent NMPC prediction time domain, taking 2-3 seconds.

[0045] S4. Based on the cargo impact risk index, and under the premise of meeting the time constraints of the delivery task and the vehicle dynamics constraints, the longitudinal speed, longitudinal acceleration, steering angle and reference trajectory of the vehicle are jointly optimized to generate linkage control commands. The joint optimization adopts the multi-objective nonlinear model predictive control (NMPC) method; the NMPC controller is used as an extended constraint in the "comfort and cargo safety" dimension of the autonomous driving planning stack for joint optimization.

[0046] State quantities include the vehicle's longitudinal position. Longitudinal velocity Longitudinal acceleration Heading angle Front wheel steering angle Horizontal position The control parameters include longitudinal acceleration commands. With steering angle command .

[0047] The objective function of NMPC contains at least four weighting terms, namely:

[0048] in, To preset the maximum longitudinal speed, To preset the maximum steering angle, For the target arrival time, , , and These are weighting coefficients, representing the relative importance of four dimensions: cargo safety, speed tracking, steering smoothness, and timeliness constraints. When the impact risk increases, the controller automatically adjusts its weighting. weight and reduce The weighting of these factors allows for a trade-off between protecting goods and ensuring timeliness.

[0049] Furthermore, the dynamic adjustment strategy for the weighting coefficients is as follows: Let the average risk index in the current prediction time domain be:

[0050] but ; Other weights , Remain unchanged. and The preset adjustment coefficient, The value range is 0.1-0.5. The value range is 0.1-0.3; and These are the initial calibration values, and it is necessary to ensure that all weights are non-negative after the weight adjustment.

[0051] when When it approaches 1, Prioritize ensuring cargo safety; when When it approaches 0, Revert to the normal delivery time-priority mode.

[0052] NMPC constraints include: longitudinal velocity constraints (Preferred) =0km / h, =60km / h, further limited to urban delivery scenarios. =40km / h); longitudinal acceleration constraint (Preferred) = 3m / s 2 , =2m / s 2 (including upper and lower limits for comfort); steering angle constraints. (Preferred) =35°); path tracking constraints (Preferred) =0.3m); obstacle safety distance constraints (maintaining minimum longitudinal and lateral distances from pedestrians, vehicles, etc.); and cargo safety constraints (when Forced constraint when >0.7 , A speed of 5-15 km / h can be used, adjusted according to the fragility level of the goods. For fragile items, a speed of 5 km / h is used, and for general goods, a speed of 10-15 km / h is used.

[0053] The NMPC solver uses the IPOPT or ACADO toolkit; discretization step size (sampling time). Taking 0.1s, predict the time domain. Taking 2-3 seconds corresponds to 20-30 grid points in the prediction time domain, controlling the time domain. The time interval is set to 0.5-1s, corresponding to 5-10 control steps in the control time domain. The NMPC solution cycle is 30-50ms, matching the discretization step size to ensure a complete optimization solution is completed within each sampling cycle. The single solution time is controlled within 30ms. This real-time performance is achieved on the following hardware platform: The central computing platform uses NVIDIA Jetson Orin (or an embedded AI computing platform with equivalent computing power, such as Horizon Journey 5), equipped with 64GB LPDDR5 memory, and the NMPC solver runs on the ARM Cortex-A78AE core, with IPOPT's linear solver MUMPS enabled for acceleration. On this platform, the 30ms single-solve time constraint corresponds to the aforementioned discretization step size and grid point configuration, which can meet the real-time control requirements of the vehicle at a maximum speed of 40km / h.

[0054] It should be noted that if the computing power of the deployment platform is lower than the above configuration (such as Jetson Xavier NX), the discretization step size can be increased to 0.15s or the prediction time domain can be shortened to 1.5-2s to reduce the computational burden, but the control parameters need to be adjusted accordingly to ensure control quality.

[0055] S5. Send the linkage control command to the vehicle chassis actuator and the upper planning module to realize the real-time adjustment of the vehicle's motion state; The linkage control command is sent to the vehicle actuators, including the drive motor, steering motor, and braking system, via the drive chassis interface; the actuators provide feedback on the actual execution results, and the cargo compartment perception subsystem continuously monitors the actual effect, forming a closed loop of perception, planning, execution, and feedback.

[0056] When the model prediction error is detected to exceed the threshold, an online learning process is triggered to fine-tune the risk prediction model parameters using new data.

[0057] Furthermore, the prediction error is calculated as follows: At the present moment Compared to the past Risk prediction value at each time step Compared with actual observed risk values Calculate the root mean square error:

[0058] The actual observed risk value is calculated based on the physical quantities of the cargo status collected in real time by the cargo hold sensing subsystem at the current moment. Specifically, the cargo roll angle, angular velocity, vertical acceleration, and bottom pressure center offset at the current moment are obtained, and corresponding safety thresholds are preset for each of the above physical quantities. The ratio of the current measured value of each physical quantity to its corresponding safety threshold is calculated, and the values ​​greater than 1 in each ratio are truncated to 1, while the other values ​​remain unchanged, so that each ratio falls within a closed interval of 0 to 1. The maximum value among the truncated ratios is taken as the actual observed risk value at the current moment.

[0059] When RMSE continuously exceeds the threshold When the value reaches a preset number of times (e.g., 0.15) (e.g., 10 consecutive sampling periods), the online learning process is triggered. Using newly collected data from the last 24 hours as training samples, the parameters of the TCN-LSTM network were fine-tuned using an online gradient descent algorithm; the fine-tuning learning rate was set to 1 / 10 of the initial training learning rate (i.e., 10). -4 Each fine-tuning iteration should not exceed 100 rounds to avoid overfitting.

[0060] The linkage control command is simultaneously sent to the fixed mechanism control interface inside the cargo hold to control the electromagnetic lock, pneumatic clamping device or active buffer platform to adjust the clamping force or support stiffness, so as to enhance the passive protection of the cargo when the vehicle decelerates or turns.

[0061] S6. Enables crowdsourced updates and sharing of road information across the entire fleet through vehicle-cloud collaboration; By modularizing the above steps and integrating them with the existing autonomous driving planning stack of the driverless delivery vehicle, they can be embedded into the original planning process as extended constraints of the "comfort and cargo safety" dimensions. This approach minimizes intrusion into the original system and facilitates rapid deployment on existing driverless delivery vehicle fleets.

[0062] Furthermore, the modular integration method is as follows: the above steps are run as an independent ROS node, receiving the planning reference trajectory and vehicle status information published by the original planning stack through the standard topic interface, and simultaneously publishing speed and steering correction commands optimized by NMPC to the chassis control topic of the original planning stack; the original planning stack selects whether to enable this module through a configurable switch. When this module is not enabled or malfunctions, the system automatically downgrades to the original planning mode.

[0063] After anonymizing the road surface unevenness data collected by individual vehicles, the data is uploaded to the cloud. The cloud then aggregates newly discovered road surface unevenness information such as speed bumps and potholes and distributes it to the entire fleet, enabling crowdsourced updates of road information. Simultaneously, for high-frequency delivery routes, pre-loading optimization is performed using high-precision maps: historical unevenness detection results are written to the map cache, and when a vehicle passes through the route again, the historical risk distribution is directly retrieved to shorten the first frame response time.

[0064] In addition, the linkage control command is simultaneously sent to the fixed mechanism control interface inside the cargo hold to control the electromagnetic lock, pneumatic clamping device or active buffer platform to adjust the clamping force or support stiffness, so as to enhance the passive protection of the cargo when the vehicle decelerates or turns.

[0065] Example 2 Please see Figure 2 As shown, this embodiment provides an AI-based intelligent perception-based delivery vehicle linkage control system, including a cargo compartment perception subsystem, an external perception subsystem, an impact risk prediction module, a dynamic planning linkage control module, an execution and feedback subsystem, and a system expansion and deployment module. The modules are interconnected via in-vehicle Ethernet or CAN-FD bus, and data between modules is transmitted using a timestamp-aligned publish-subscribe mode, with end-to-end latency controlled within 50ms.

[0066] The cargo hold perception subsystem comprises an inertial measurement unit (IMU), a vision sensor array, and a distributed pressure sensor array deployed within the cargo hold. The IMU is positioned on the top, side walls, and floor of the cargo hold to collect the cargo's attitude angles, angular velocities, and linear accelerations. The vision sensor array uses wide-angle or fisheye cameras at the four corners and top of the cargo hold to identify the cargo's outline, deformation, and stacking status, outputting deformation coefficients and pose offsets. The distributed pressure sensor array uses thin-film pressure sensors on the cargo hold floor to acquire the pressure distribution on the cargo's bottom surface, identifying whether the cargo has slipped, tilted, or is partially suspended. The three types of sensors achieve timestamp alignment through an external trigger synchronization mechanism. Raw data is preprocessed by the edge computing unit before being uploaded to the central computing platform.

[0067] The vehicle exterior perception subsystem includes at least two of the following: a forward-facing camera, millimeter-wave radar, and 3D LiDAR, used to acquire information on the unevenness distribution of the road surface ahead. This module uses a deep learning model to segment the road area ahead into different road surface categories, and combines the height difference of the LiDAR point cloud with acceleration signals to output the unevenness level and boundary position of each road segment; simultaneously, it performs obstacle detection for pedestrians, vehicles, non-motorized vehicles, etc., providing obstacle avoidance information for path planning.

[0068] The impact risk prediction module communicates with the cargo compartment perception subsystem and the vehicle exterior perception subsystem. Based on received cargo status data and road unevenness information, it uses a hybrid model combining temporal convolutional networks and long short-term memory networks to predict a sequence of cargo impact risk indicators within a preset time window. This module also outputs corresponding risk warning prompts based on the comparison results between the risk indicators and preset thresholds.

[0069] The dynamic programming linkage control module communicates with the impact risk prediction module. Based on cargo impact risk indicators, it employs a multi-objective nonlinear model predictive control method, with cargo safety, speed tracking, smooth steering, and timeliness constraints as optimization objectives. It jointly optimizes the vehicle's longitudinal speed, longitudinal acceleration, steering angle, and reference trajectory, generating linkage control commands. The weighting coefficients in this module dynamically adjust according to the impact risk indicators; when the risk increases, the weighting for cargo safety is increased, while the weighting for timeliness is decreased.

[0070] The execution and feedback subsystem includes the vehicle chassis drive-by-wire interface, the steering actuator interface, and the upper-level planning module interface. It is used to receive and execute linkage control commands, and send the commands to the drive motor, steering motor, and braking system. The actuator provides feedback on the actual execution results, and the cargo compartment perception subsystem continuously monitors the actual effect, forming a closed loop. When the model prediction error continues to exceed the threshold, an online learning process is triggered to fine-tune the risk prediction model parameters using new data.

[0071] The system expansion and deployment module modularly integrates the aforementioned modules with the existing autonomous driving planning stack of the unmanned delivery vehicle, embedding them as extended constraints into the original planning process. Enable and degrade control are achieved through configurable switches. This module is also responsible for anonymizing road unevenness data collected by individual vehicles and uploading it to the cloud. The cloud then aggregates the data and distributes it to the entire fleet for crowdsourced updates. Furthermore, it can pre-load and optimize historical risk distribution using high-precision maps, and synchronously send control commands to the control interface of the cargo compartment's internal fixing mechanisms to achieve coordinated adjustment of clamping force or support stiffness.

[0072] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for coordinated control of delivery vehicles based on AI intelligent perception, characterized in that, include: Cargo status data is collected in real time by a multi-source sensor array deployed in the cargo hold. The cargo status data includes at least the cargo attitude angle, angular velocity, linear acceleration, deformation coefficient, and offset of the bottom pressure center. The vehicle's external sensing module scans the road surface ahead to obtain information on the road surface unevenness distribution. Based on cargo status data and road surface unevenness distribution information, the pre-trained impact risk prediction model outputs cargo impact risk indicators within a future preset time window. Based on the cargo impact risk index, and under the premise of meeting the time constraints of delivery tasks and vehicle dynamics constraints, the longitudinal speed, longitudinal acceleration, steering angle and reference trajectory of the vehicle are jointly optimized to generate linkage control commands. The linkage control command is sent to the vehicle chassis actuator to realize real-time adjustment of the vehicle's motion state.

2. The delivery vehicle linkage control method based on AI intelligent perception according to claim 1, characterized in that, The multi-source sensor array includes at least an inertial measurement unit, a vision sensor array, and a distributed pressure sensor array. The inertial measurement unit is used to acquire the attitude angle, angular velocity, and linear acceleration of the cargo. The vision sensor array is used to identify the cargo contour and output the deformation coefficient and pose offset. The distributed pressure sensor array is used to acquire the pressure distribution on the bottom surface of the cargo and output the offset of the bottom pressure center. The three types of sensors achieve timestamp alignment through an external trigger synchronization mechanism.

3. The delivery vehicle linkage control method based on AI intelligent perception according to claim 1, characterized in that, The cargo impact risk prediction model adopts a hybrid structure of temporal convolutional network and long short-term memory network. It takes the cargo state characteristics of a preset number of time steps in the past and the road surface unevenness sequence of a preset number of time steps in the future as input, and outputs the impact risk index sequence within the preset time window in the future.

4. The delivery vehicle linkage control method based on AI intelligent perception according to claim 1, characterized in that, The joint optimization adopts a multi-objective nonlinear model predictive control method. Its objective function includes at least cargo safety, speed tracking, steering smoothness and time constraint terms. Each objective term is normalized with its corresponding preset benchmark value and then weighted and summed. The cargo safety item is determined by the square of the ratio of the impact risk index to the vehicle speed divided by the maximum vehicle speed, and the time constraint item is determined by the ratio of the deviation between the expected arrival time and the target arrival time divided by the target arrival time. Each weighting coefficient is dynamically adjusted according to the impact risk index. When the impact risk increases, the weight of the cargo safety item is increased and the weight of the timeliness constraint item is decreased.

5. The delivery vehicle linkage control method based on AI intelligent perception according to claim 1, characterized in that, The vehicle's external perception module includes at least two of the following: a forward-facing camera, millimeter-wave radar, and 3D LiDAR. It is used to perform semantic segmentation of the road ahead to identify different road surface categories, and to output the unevenness level and boundary position of each road segment by combining the height difference of the LiDAR point cloud. When a road segment with an unevenness level not lower than a preset threshold is detected ahead, the reference trajectory is adjusted in advance to detour through it, or a gradual deceleration is performed when the detour conditions are not met.

6. The delivery vehicle linkage control method based on AI intelligent perception according to claim 1, characterized in that, The impact risk prediction model is trained using a supervised learning approach. The training dataset is derived from the operational data of multiple delivery vehicles with different load configurations in actual operating scenarios, covering various road conditions and loading states. Impact risk labels are used to mark events in which the actual posture of the cargo exceeds a safety threshold; to address the imbalance between positive and negative samples, data augmentation strategies such as time axis scaling, random masking of sensor channels, and addition of Gaussian noise are used to expand the positive samples. The impact risk label is defined as follows: when the actual change in the cargo's posture exceeds the safety threshold, that moment is marked as a positive sample.

7. The delivery vehicle linkage control method based on AI intelligent perception according to claim 1, characterized in that, It also includes closed-loop feedback and online learning, including: after the linkage control command is sent to the vehicle chassis actuator, the cargo compartment perception subsystem continuously monitors the actual cargo status; calculates the root mean square error between the risk prediction value and the actual observation value, and when the root mean square error continues to exceed the preset threshold for a preset number of times, the new data collected within the most recent preset time period is used as training samples to fine-tune the parameters of the impact risk prediction model.

8. The delivery vehicle linkage control method based on AI intelligent perception according to claim 1, characterized in that, This also includes uploading anonymized road surface unevenness data collected by individual vehicles to the cloud, where it is aggregated and distributed to the entire fleet, enabling crowdsourced updates of road information; for high-frequency delivery routes, historical unevenness detection results are written into the map cache, and the historical risk distribution is directly retrieved when the vehicle passes through again.

9. A delivery vehicle linkage control system based on AI intelligent perception, characterized in that, include: The cargo hold perception subsystem is used to collect cargo status data in real time through a multi-source sensor group deployed in the cargo hold. The cargo status data includes at least the cargo attitude angle, angular velocity, linear acceleration, deformation coefficient, and bottom pressure center offset. The external perception subsystem is used to scan the road surface ahead and obtain information on the road surface unevenness distribution; The impact risk prediction module is used to output cargo impact risk indicators within a preset time window based on cargo status data and road unevenness distribution information through a pre-trained impact risk prediction model. The dynamic programming linkage control module is used to jointly optimize the longitudinal speed, longitudinal acceleration, steering angle and reference trajectory of the vehicle based on the cargo impact risk index, under the premise of meeting the time constraints of the delivery task and the vehicle dynamics constraints, and generate linkage control commands. The execution and feedback subsystem is used to send linkage control commands to the vehicle chassis actuators to achieve real-time adjustment of the vehicle's motion state.

10. A delivery vehicle linkage control system based on AI intelligent perception according to claim 9, characterized in that, The dynamic programming linkage control module adopts a multi-objective nonlinear model predictive control method. Its objective function includes at least cargo safety, speed tracking, steering smoothness, and time constraint terms. The weight coefficients are dynamically adjusted according to the impact risk index. The system also includes a system expansion and deployment module, which is used to interface the above modules with the original autonomous driving planning stack of the unmanned delivery vehicle in a modular manner, and to enable and degrade control through configurable switches; And it uploads road surface unevenness data collected by individual vehicles to the cloud to achieve crowdsourced updates for the entire fleet.

Images