A hub damping online adjustment method based on multi-source sensing data fusion

Abstract

This invention relates to the field of vibration reduction control technology, specifically to an online adjustment method for wheel hub vibration reduction based on multi-source sensor data fusion. The method includes: simultaneously acquiring and aligning image sequences and point cloud data of the road ahead using visual sensors and lidar; constructing a rasterized elevation model using a terrain segmentation algorithm to extract road surface geometric features; generating a multi-source physical state matrix by combining vehicle dynamic parameters acquired by an inertial unit, and inputting this matrix into a classifier with an embedded spatiotemporal cross-attention mechanism to output driving scenario commands; triggering an electromagnetic adjustment link for flat road conditions, and optimizing motor control parameters online through a PID parameter mapping network; and triggering a mechanical compensation link for pulse impact road conditions, simultaneously outputting suspension damping coefficients and motor reverse compensation torque commands through frequency domain decoupling calculations, thus achieving dynamic adaptive vibration reduction in all scenarios.

Description

Technical Field

[0001] This invention relates to the field of vibration reduction control technology, specifically to an online adjustment method for wheel hub vibration reduction based on multi-source sensor data fusion. Background Technology

[0002] With the development of distributed drive technology, hub motor systems have been widely used due to their compact structure and high transmission efficiency. However, the introduction of the motor significantly increases the unsprung mass of the vehicle, causing the vertical acceleration of the wheel to surge when subjected to road impacts, which seriously deteriorates the ride comfort and NVH performance. At the same time, there is a complex coupling between low-frequency random road vibration and high-frequency electromagnetic vibration of the motor (such as unbalanced electromagnetic force and torque pulsation) within the system, making traditional vibration reduction methods ineffective. Currently, vibration damping control technology for in-wheel motors suffers from two significant shortcomings. First, the sensing dimension is singular and severely lagging. Existing systems generally rely on a single physical sensor (such as an accelerometer) for reactive, reactive responses, lacking high-precision pre-sensing of the geometric features of the road surface ahead, thus failing to secure a critical feedforward time window for chassis adjustment. Second, the electromechanical system is isolated, and the control logic is rigid. The feedback variables and optimization objectives of existing controllers are typically statically fixed, lacking electromechanical coordination and decoupling capabilities. When faced with complex switching of vehicle driving conditions, existing technologies cannot proactively intervene in the motor's underlying control to suppress high-frequency electromagnetic noise on smooth roads, nor can they instantaneously switch feedback objectives to smooth out sprung mass fluctuations when encountering pulse impacts, failing to meet the comprehensive suppression requirements of complex electromechanical excitation across a wide frequency range.

[0003] To address this, a method for online adjustment of wheel hub damping based on multi-source sensor data fusion is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide an online adjustment method for wheel hub damping based on multi-source sensor data fusion, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for online adjustment of wheel hub damping based on multi-source sensor data fusion includes: The system collects point cloud data and image sequence data of the road ahead using visual sensors and LiDAR; performs elevation mapping processing using terrain segmentation algorithms; and generates road surface geometric feature data. The vehicle body's three-axis acceleration parameters, motor's three-axis acceleration parameters, wheel bounce parameters, and current longitudinal speed are collected in real time by the inertial unit. These parameters are then spliced ​​with the road surface geometric feature data to generate a multi-source physical state matrix. This matrix is ​​then input into a road condition feature classifier for matching and outputs driving scenario instructions. When a driving scenario command indicating a flat road ahead is received, the electromagnetic regulation link is triggered, and the preset PID parameter mapping network is called to perform online forward inference on the speed loop PID parameters and torque loop PID parameters of the hub motor controller, generate the target current parameters and output them to the underlying inverter, and update the PWM modulation command stream. When a driving scenario command indicating a pulse impact load on the road ahead is obtained, the mechanical compensation link is triggered by the multi-source physical state matrix for calculation and processing, generating a suspension damping coefficient command and a motor reverse compensation torque command, which are then synchronously output to the suspension actuator and the motor driver, respectively.

[0006] Preferably, the road point cloud data and image sequence data are obtained by controlling the vision sensor and the lidar to jointly acquire the original image and the original lidar point cloud at the same timestamp; the original image is subjected to distortion correction processing to generate the image sequence data, and the original lidar point cloud is subjected to noise reduction and clutter filtering processing to generate a preprocessed point cloud; the pre-calibrated sensor extrinsic parameter matrix is ​​retrieved, and the three-dimensional coordinate system of the preprocessed point cloud is projected onto the camera coordinate system of the vision sensor to obtain the road point cloud data that is strictly aligned in time and space.

[0007] Preferably, the terrain segmentation algorithm is an elevation mapping algorithm based on a rasterized digital elevation model; the step of generating road surface geometric feature data by performing elevation mapping processing through the terrain segmentation algorithm includes: extracting the drivable road surface boundary ahead based on the image sequence data, extracting the corresponding target road surface point cloud from the road point cloud data ahead, and dividing the target road surface point cloud into a two-dimensional raster network; extracting the Z-axis elevation coordinates of the point cloud inside each raster, and calculating the range and variance of the elevation coordinates inside each raster as the road surface undulation degree characterizing the local roughness of the corresponding area; Calculate the mean elevation difference between adjacent grids, extract continuous grid groups with differences exceeding a preset threshold as abrupt change nodes, and calculate the concave-convex feature geometric dimensions based on the span and extreme values ​​of the continuous grid groups; then, perform matrix concatenation of the road surface undulation and the concave-convex feature geometric dimensions to generate the road surface geometric feature data.

[0008] Preferably, the specific process of generating the multi-source physical state matrix includes: performing time-stamp synchronization and filtering noise reduction on the real-time collected vehicle body three-axis acceleration parameters, motor three-axis acceleration parameters, wheel bounce parameters, and current vehicle longitudinal speed to construct a vehicle motion state sub-matrix within a set time sliding window; extracting the road surface geometric feature data, and serializing and expanding the road surface undulation and concave-convex feature geometric dimensions according to their longitudinal spatial position relative to the vehicle to construct a feedforward road surface excitation sub-matrix corresponding to the prediction range of the time sliding window; performing dimensionless processing on the vehicle motion state sub-matrix and the feedforward road surface excitation sub-matrix respectively; aligning and splicing the normalized vehicle motion state sub-matrix and the feedforward road surface excitation sub-matrix according to feature dimensions to generate the multi-source physical state matrix representing the coupling relationship between the vehicle's current dynamic attitude and the road conditions ahead, and inputting it into a road condition feature classifier for matching processing.

[0009] Preferably, the road condition feature classifier is a neural network model with an embedded spatiotemporal cross-attention mechanism. It decouples and extracts the multi-source physical state matrix into temporal features representing vehicle dynamics and spatial features representing the geometric distribution of the road surface ahead. The temporal features are input as a query vector to the spatiotemporal cross-attention layer, where adaptive weight allocation and feature aggregation are performed on the spatial features to generate a global context feature vector. This global context feature vector is then input to the classification output layer for calculation, generating classification confidence scores for each preset driving condition. The condition category corresponding to the maximum classification confidence score is extracted. When the category is determined to be a smooth condition, a driving scenario command representing a flat road surface ahead is output; when the category is determined to be an impact condition, a driving scenario command representing a pulse impact load on the road surface ahead is output.

[0010] Preferably, the preset PID parameter mapping network includes an input layer, a hidden layer, and an output layer; the vehicle body triaxial acceleration parameters, motor triaxial acceleration parameters, wheel bounce parameters, and current vehicle longitudinal speed at the end of the current time sliding window in the multi-source physical state matrix are extracted in real time and normalized to construct the input vector of the input layer; the input vector is input to the input layer, and a weighted summation operation based on a pre-stored weight matrix and bias term is performed through the hidden layer, and feature mapping is performed using a nonlinear activation function, which is then passed to the output layer; the output layer outputs six consecutive floating-point values, corresponding to the gain of the speed loop and the gain of the torque loop in the hub motor controller, respectively; the six gain parameters are written to the bottom control register of the hub motor controller in real time, and the speed deviation and current deviation are adjusted in a closed loop according to the updated gain parameters.

[0011] Preferably, the mechanical compensation link extracts the concave and convex feature geometric dimensions and the current vehicle longitudinal speed from the multi-source physical state matrix. Based on the longitudinal spatial position of the concave and convex feature geometric dimensions and the current vehicle longitudinal speed, it calculates the expected arrival time of the pulse impact load and generates a feedforward excitation amplitude according to the geometric elevation mapping. It extracts the three-axis acceleration parameters of the vehicle body and motor from the multi-source physical state matrix as real-time feedback quantities. Combining the expected arrival time and the feedforward excitation amplitude, it calculates the actual dynamic force deviation of the suspension system. According to the preset frequency band difference of the electromechanical actuator response, it decouples the actual dynamic force deviation in the frequency domain: for the low-frequency, high-amplitude vertical excitation component, it calculates and generates the suspension damping coefficient command for adjusting the active suspension dissipation channel; for the high-frequency wheel speed fluctuation and micro-longitudinal slip component caused by road impact, it calculates and generates the motor reverse compensation torque command for suppressing the torsional vibration of the unsprung mass of the hub motor.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention acquires raw data synchronously through hardware-level timestamps of visual sensors and lidar, and establishes the correlation between vehicle inertial parameters and the physical contour of the road surface in front in a unified spatiotemporal coordinate system by using an elevation mapping algorithm based on a rasterized digital elevation model. This mechanism changes the drawback of traditional wheel hub damping systems that rely solely on single-point physical sensors (such as accelerometers) for reactive response. It extends the sensing reach from the center of the vehicle body to a preset travel range, realizing a dimensional leap from instantaneous single point to a continuous spatiotemporal state matrix of control input source. This provides a crucial feedforward time window for chassis adjustment and effectively solves the technical defects of single sensing dimension and severe lag.

[0013] 2. This invention uses a neural network model with an embedded spatiotemporal cross-attention mechanism to decouple the features of the multi-source physical state matrix, and performs logical switching between the electromagnetic adjustment link and the mechanical compensation link based on the confidence classification results. This eliminates the limitation of the feedback target and the actuator being in a static and isolated state for a long time in the prior art. When the road condition is determined to be flat, the system calls the PID parameter mapping network to perform online forward inference on the speed loop and torque loop parameters of the hub motor controller. The adjustment target is changed from a single power output to multi-objective optimization including electromagnetic pulsation suppression. By actively intervening in the underlying control of the motor, high-frequency electromagnetic howling is suppressed, filling the technical gap that traditional mechanical vibration reduction methods cannot cover the electromagnetic excitation frequency band.

[0014] 3. This invention decouples the pulsed impact load in the frequency domain through a mechanical compensation link. Based on the difference in response frequency bands of the electromechanical actuators, it achieves a scientific distribution of the target load among different mechanisms. The active suspension handles the low-frequency energy excitation of the long stroke, while the rapid response of the drive motor generates a reverse compensation torque to suppress high-frequency wheel speed fluctuations and unsprung mass torsional vibrations caused by road impacts. This electromechanical collaborative architecture effectively resolves the physical contradiction between the surge in unsprung mass caused by the introduction of the motor and the response delay of a single actuator, significantly improving the ride comfort and NVH performance of hub motor vehicles under complex impact conditions. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an online adjustment method for wheel hub damping based on multi-source sensor data fusion. Figure 2 A schematic diagram illustrating the process of constructing a multi-source state matrix; Figure 3 This is a schematic diagram of the road condition feature classification and online adjustment control process of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: Please see Figure 1 This invention provides an online adjustment method for wheel hub vibration damping based on multi-source sensor data fusion, the technical solution of which is as follows: A method for online adjustment of wheel hub damping based on multi-source sensor data fusion includes: The system collects point cloud data and image sequence data of the road ahead using visual sensors and LiDAR; performs elevation mapping processing using terrain segmentation algorithms; and generates road surface geometric feature data. The vehicle body's three-axis acceleration parameters, motor's three-axis acceleration parameters, wheel bounce parameters, and current longitudinal speed are collected in real time by the inertial unit. These parameters are then spliced ​​with the road surface geometric feature data to generate a multi-source physical state matrix. This matrix is ​​then input into a road condition feature classifier for matching and outputs driving scenario instructions. When a driving scenario command indicating a flat road ahead is received, the electromagnetic regulation link is triggered, and the preset PID parameter mapping network is called to perform online forward inference on the speed loop PID parameters and torque loop PID parameters of the hub motor controller, generate the target current parameters and output them to the underlying inverter, and update the PWM modulation command stream. When a driving scenario command indicating a pulse impact load on the road ahead is obtained, the mechanical compensation link is triggered by the multi-source physical state matrix for calculation and processing, generating a suspension damping coefficient command and a motor reverse compensation torque command, which are then synchronously output to the suspension actuator and the motor driver, respectively.

[0018] The road point cloud data and image sequence data are obtained by jointly acquiring the visual sensor and the lidar at the same timestamp, respectively acquiring the original image and the original lidar point cloud; the original image is subjected to distortion correction processing to generate the image sequence data, and the original lidar point cloud is subjected to noise reduction and clutter filtering processing to generate the preprocessed point cloud; the pre-calibrated sensor extrinsic parameter matrix is ​​retrieved, and the three-dimensional coordinate system of the preprocessed point cloud is projected onto the camera coordinate system of the visual sensor to obtain the road point cloud data that is strictly aligned in time and space.

[0019] Specifically, a fixed-period synchronous pulse signal is generated by a hardware trigger built into the vehicle's central processing unit. This synchronous pulse signal is simultaneously transmitted via a wiring harness to the underlying image capture interface of the vision sensor and the scanning drive interface of the LiDAR. Upon receiving the synchronous pulse signal, the electronic shutter of the vision sensor and the laser emitter of the LiDAR simultaneously initiate the image exposure and point cloud scanning actions of the current cycle at the hardware level. The vehicle's central processing unit receives a single frame of raw image data generated by the vision sensor after one exposure and a single frame of raw laser point cloud data generated by the LiDAR after one scan, and writes the same system absolute time stamp on the raw image data and the raw laser point cloud data. The specific implementation process of generating image sequence data by performing distortion correction processing on the original image includes: first obtaining the internal parameter matrix of the visual sensor and the distortion parameter vector containing radial distortion coefficients and tangential distortion coefficients using a standard black and white alternating calibration checkerboard; after obtaining the original image with absolute time stamps, extracting the two-dimensional planar coordinate values ​​of each pixel in the original image; using the pre-obtained internal parameter matrix and distortion parameter vector, constructing a nonlinear mapping relationship model between distorted pixel coordinates and ideal distortion-free pixel coordinates; calculating the ideal pixel position corresponding to each distorted pixel point in reverse using this nonlinear mapping relationship model, and using a bilinear interpolation algorithm to fill the blank areas with missing pixels after position mapping with weighted neighboring gray values; and combining multiple frames of distortion-free images from consecutive acquisition periods in the order of absolute time stamps to generate image sequence data. The denoising and clutter filtering process involves constructing a rectangular spatial bounding box with the vehicle center as the origin in three-dimensional space. Scanning points whose three-dimensional coordinates fall within this rectangular spatial bounding box are directly removed, filtering out local clutter generated by reflections from the vehicle's hood or rearview mirrors. For the remaining point cloud data after filtering, a filtering algorithm is used to denoise outliers. Each spatial scanning point in the remaining point cloud data is traversed, and the average Euclidean distance between the spatial scanning point and a preset number of neighboring scanning points is calculated. Outliers are then removed based on the distance distribution characteristics. The remaining point cloud data after denoising and clutter filtering is cached as a preprocessed point cloud. The camera coordinate system projected onto the vision sensor is pre-calculated in a static test environment using a joint calibration target board with high reflectivity. The spatial rotation and translation matrices between the origin of the LiDAR's own 3D coordinate system and the origin of the vision sensor's camera coordinate system are calculated through feature point alignment. These matrices are then merged to construct a 4x4 sensor extrinsic parameter matrix, which is persistently stored in the vehicle's onboard memory. During driving, the 3D spatial coordinates of each scan point in the pre-processed point cloud are extracted. The 3D spatial coordinates of each scan point are then multiplied with the retrieved sensor extrinsic parameter matrix to calculate the new 3D coordinates of each scan point, originally under the LiDAR's relative viewpoint, within the vision sensor's camera coordinate system. The strict alignment in time and space dimensions involves, after completing the baseline transformation to the camera coordinate system, further utilizing the internal parameter matrix of the visual sensor to perform dimensionality reduction projection on the three-dimensional coordinate values ​​in the camera coordinate system, mapping them to the two-dimensional pixel plane coordinate system of the aforementioned image sequence data, and obtaining the precise pixel row and column index position of each laser scanning point in the image sequence data; based on the precise pixel row and column index position, extracting the color channel values ​​and local texture feature values ​​at that location in the image sequence data; and binding and merging the physical depth value and reflection intensity value carried by each laser scanning point with the corresponding extracted color channel values ​​and local texture feature values ​​in multiple dimensions. By achieving hardware-level synchronous triggering of visual sensors and LiDAR, and unified transformation of the physical space coordinate system, the time delay difference and installation posture deviation of heterogeneous sensors during data acquisition are effectively eliminated. Targeted image distortion correction and 3D coordinate projection processing solve the technical challenges of image distortion caused by the lens's physical structure and the misalignment of point clouds and image data, achieving a high degree of fusion of multi-source sensing data under the same spatiotemporal reference.

[0020] Elevation mapping is performed using terrain segmentation algorithms to generate road surface geometric feature data. The terrain segmentation algorithm is an elevation mapping algorithm based on a rasterized digital elevation model. The step of generating road surface geometric feature data by performing elevation mapping processing through the terrain segmentation algorithm includes: extracting the drivable road surface boundary ahead based on the image sequence data, extracting the corresponding target road surface point cloud from the road point cloud data ahead, and dividing the target road surface point cloud into a two-dimensional raster network; extracting the Z-axis elevation coordinates of the point cloud inside each raster, and calculating the range and variance of the elevation coordinates inside each raster as the road surface undulation degree characterizing the local roughness of the corresponding area. Calculate the mean elevation difference between adjacent grids, extract continuous grid groups with differences exceeding a preset threshold as abrupt change nodes, and calculate the concave-convex feature geometric dimensions based on the span and extreme values ​​of the continuous grid groups; then, perform matrix concatenation of the road surface undulation and the concave-convex feature geometric dimensions to generate the road surface geometric feature data.

[0021] Specifically, the process of extracting the drivable road surface boundary based on image sequence data and extracting the corresponding target road surface point cloud from the road point cloud data ahead involves using a pre-trained semantic segmentation deep learning model (in this embodiment, BiSeNet is used, but other deep learning models with real-time semantic segmentation capabilities such as DeepLabV3+ and HRNet can also be used) to perform frame-by-frame semantic annotation on the input image sequence data, identify the pixel regions representing the road plane in the image, and extract the closed contour lines of the edges of these regions as the drivable road surface boundary. Since the road point cloud data ahead has been projected onto the camera coordinate system of the vision sensor through the sensor extrinsic matrix, by establishing an index mapping relationship between pixel coordinates and spatial coordinates, laser scanning points whose spatial projection positions fall within the aforementioned drivable road surface boundary are selected from the 3D point cloud dataset, and extraneous points outside the road (such as green belts, buildings, or roadside trees) are removed as the target road surface point cloud. The specific implementation process of dividing the target road surface point cloud into a two-dimensional grid network includes: taking the projection point of the vehicle's current geometric center on the ground as the origin, establishing a two-dimensional local coordinate system in the horizontal ground direction that covers a preset detection range in front (e.g., 15 meters in front and 3 meters to the left and right); dividing the two-dimensional local coordinate system area into several uniform square grids according to a preset physical resolution, with the side length of each grid set between 0.05 meters and 0.1 meters, thereby constructing a two-dimensional grid network; traversing the horizontal coordinates of each scan point in the target road surface point cloud, and assigning it to the corresponding grid index according to the coordinate value, transforming the originally disordered spatial points into gridded data with topological positional relationships; For each assigned point cloud grid, the height values ​​of all laser scanning points within that grid in the direction of gravity are extracted, i.e., the Z-axis elevation data. The maximum and minimum Z-axis coordinates within that grid are identified, and the difference between them is calculated to obtain the elevation range, which characterizes the maximum severity of road surface undulation within that grid area. Simultaneously, the statistical variance of the elevation values ​​of all scanning points within that grid is calculated, i.e., the average of the sum of squares of the differences between the elevation values ​​of all points and their mean values, which characterizes the distribution of micro-texture within that grid. The range and variance values ​​of each grid are encapsulated to form the corresponding local road surface undulation parameters. The two-dimensional grid network is traversed row by row, and the absolute value of the difference between the mean heights of the point clouds inside any two adjacent grids is calculated. When the absolute value of the difference exceeds the preset abrupt change threshold (set to 3 cm in this embodiment), it is determined that there is a road surface height jump at that location. Multiple grids that are spatially connected and whose differences all exceed the threshold are marked as a continuous grid group, defined as abrupt change nodes, representing road surface structures such as speed bumps, protrusions, or manhole covers ahead. The number of grids covered by this continuous grid group in the vehicle's forward direction and width direction is counted, and multiplied by the side length of a single grid to obtain the length and width span of the concave-convex feature. At the same time, the maximum offset of the mean elevation of all grids in the continuous grid group relative to the mean of the surrounding flat road surface is extracted as the height of the concave-convex feature. All raster-calculated undulation parameters (including range and variance matrices) are organized into feature map tensors according to their original physical spatial arrangement. The geometric dimensions (length, width, height, or depth) of each identified abrupt change node are arranged into geometric feature vectors in order of distance from the vehicle. The undulation feature map tensor and the geometric feature vectors are then concatenated and normalized to generate road surface geometric feature data that describes the micro-roughness and macro-geometric configuration of the road surface. By introducing a terrain segmentation algorithm based on a rasterized digital elevation model, massive, unordered 3D point cloud data is transformed into a structured 2D raster feature matrix, significantly reducing the computational load on the onboard computing platform during real-time processing. Semantic segmentation technology is used to pre-define the boundaries of drivable road surfaces, effectively eliminating interference from non-road clutter such as curbs and green belts on vibration reduction decisions. Furthermore, through dual extraction of micro-undulations within the raster and macro-geometric dimensions between rasteres, a full-dimensional digital model of the road surface physical morphology, from texture roughness to obstacle contours, is achieved.

[0022] The vehicle body's three-axis acceleration parameters, motor's three-axis acceleration parameters, wheel bounce parameters, and current vehicle longitudinal speed are collected in real time by an inertial unit and then stitched together with the road surface geometric feature data to generate a multi-source physical state matrix. See Figure 2The specific process of generating the multi-source physical state matrix includes: performing time-stamp synchronization and filtering noise reduction on the real-time collected vehicle body three-axis acceleration parameters, motor three-axis acceleration parameters, wheel bounce parameters, and current vehicle longitudinal speed to construct a vehicle motion state sub-matrix within a set time sliding window; extracting the road surface geometric feature data, and serializing and expanding the road surface undulation and concave-convex feature geometric dimensions according to their longitudinal spatial position relative to the vehicle to construct a feedforward road surface excitation sub-matrix corresponding to the prediction range of the time sliding window; performing dimensionless processing on the vehicle motion state sub-matrix and the feedforward road surface excitation sub-matrix respectively; aligning and splicing the normalized vehicle motion state sub-matrix and the feedforward road surface excitation sub-matrix according to feature dimensions to generate the multi-source physical state matrix representing the coupling relationship between the vehicle's current dynamic attitude and the road conditions ahead, and inputting it into a road condition feature classifier for matching processing.

[0023] Specifically, the instantaneous acceleration of the vehicle's center of gravity in the three axes of front-rear, left-right, and vertical directions, output by the inertial measurement unit, is acquired via the vehicle bus (controller area network bus in this embodiment) at a fixed sampling period (10ms in this embodiment). The instantaneous acceleration on the motor side is also acquired from the triaxial accelerometer installed in the wheel hub, and the rate of change of wheel displacement relative to the vehicle body is acquired through the suspension displacement sensor. The original acceleration signal is processed by high-frequency noise removal and zero-point drift compensation using the extended Kalman filter algorithm, and the longitudinal speed of the vehicle in the current period is extracted. A time sliding window containing a preset number of sampling points (the most recent 50 sampling points are selected in this embodiment) is established. The acceleration set corresponding to each sampling moment in the window is serialized and arranged with the vehicle speed value to form a two-dimensional data block representing the recent dynamic evolution of the vehicle, which serves as the vehicle motion state sub-matrix. The rasterized road surface geometric features generated by the terrain segmentation algorithm are retrieved, and the physical distance that the vehicle will travel within the predicted time frame corresponding to the current time sliding window is calculated based on the vehicle's current longitudinal speed. In the two-dimensional raster network, the feature raster sequence corresponding to the predicted distance range is extracted by extending forward along the center line of the vehicle's current driving trajectory. The road surface undulation parameters (range and variance) recorded in each raster in the sequence and the geometric dimensions (height, width, etc.) of the marked abrupt change nodes are flattened according to the longitudinal spatial order from near to far from the vehicle's current position, and converted into a one-dimensional feature chain indexed by spatial distance. The spatial distance index is mapped to a time index consistent with the time sliding window step size through a linear interpolation algorithm to generate a feedforward road surface excitation sub-matrix. Historical limit values ​​of various physical parameters are pre-stored in the vehicle's memory, including the maximum fluctuation of vehicle acceleration, the maximum travel of suspension bounce, and the maximum elevation range of road surface undulations. For each element in the vehicle motion state sub-matrix, a maximum-minimum normalization method is used to convert its original physical value into a floating-point proportional value between 0 and 1, eliminating the difference in numerical magnitude caused by the different units of acceleration (meters per second squared), vehicle speed (kilometers per hour), and displacement (millimeters). The road surface geometry data in the feedforward road excitation sub-matrix is ​​normalized and scaled based on the maximum undulation threshold to ensure that the feedforward excitation features and the current motion feedback features are consistent in numerical distribution range. Obtain the one-dimensional expanded feature vectors of the normalized vehicle motion state sub-matrix and the one-dimensional expanded feature vectors of the feedforward road excitation sub-matrix; establish a composite matrix space, taking the motion feature vector representing the vehicle feedback response as the left component of the matrix and the feedforward feature vector representing the road input inducement as the right component of the matrix, and perform lateral tensor concatenation; insert the normalized adjustment factor of the current vehicle longitudinal speed at the concatenation point to characterize the dynamic coupling relationship between road excitation and vehicle response (i.e., the difference in excitation frequency of the same road surface at different vehicle speeds), forming a multi-source physical state matrix, which is then input into the road condition feature classifier for matching processing. The road condition feature classifier is a neural network model with an embedded spatiotemporal cross-attention mechanism; the step of inputting the road condition feature classifier for matching processing and outputting driving scenario instructions includes: The multi-source physical state matrix is ​​decoupled and extracted into temporal features representing vehicle dynamics and spatial features representing the geometric distribution of the road surface ahead. The temporal features are input as query vectors to the spatiotemporal cross-attention layer, where adaptive weight allocation and feature aggregation are performed on the spatial features to generate a global context feature vector. The global context feature vector is input to the classification output layer for calculation to generate classification confidence scores for each preset driving condition. The condition category corresponding to the maximum classification confidence score is extracted. When the category is determined to be a smooth condition, the driving scenario command representing a flat road surface ahead is output. When the category is determined to be an impact condition, the driving scenario command representing a pulse impact load on the road surface ahead is output.

[0024] Specifically, the multi-source physical state matrix is ​​reverse-segmented according to the previously spliced ​​feature dimensions using matrix slicing to reconstruct the vehicle motion state sub-matrix and the feedforward road excitation sub-matrix. For the segmented vehicle motion state sub-matrix, a one-dimensional convolutional layer is used for temporal encoding to extract the trend of vehicle motion posture changes and periodic oscillation patterns within a set time sliding window, mapping them into a high-dimensional time-dimensional feature vector. For the segmented feedforward road excitation sub-matrix, a multilayer perceptron is used for feature dimensionality reduction and abstraction to extract the spatial distribution features of road elevation range and geometric dimensions as a function of longitudinal distance, mapping them into a high-dimensional spatial feature sequence. The temporal feature vector is converted into a query vector through a linear mapping layer containing learnable parameters, while the spatial feature sequence is converted into a key matrix and a value matrix. The dot product similarity between the query vector and the key vector corresponding to each spatial location node in the key matrix is ​​calculated. All dot product similarity values ​​are transformed into an attention weight distribution column summing to one using a normalized exponential function, intuitively representing the vehicle's current specific motion posture's attention priority to road surface geometric features at different distances ahead. The attention weight distribution column is used to proportionally weight and sum the spatial feature vectors in the value matrix to obtain a targeted feature aggregation vector. The feature aggregation vector is then concatenated with the original temporal feature vector using residual concatenation and layer standardization to generate a global context feature vector. A classification output module is constructed, consisting of alternating cascaded fully connected network layers. Linear rectified activation layers and random deactivation layers are inserted between the fully connected network layers to enhance the nonlinear expressive power of the network and prevent overfitting during model training. The generated global context feature vector is input into this classification output module for layer-by-layer high-dimensional nonlinear mapping, reducing its dimensionality to a one-dimensional logical vector equal to the total number of preset driving condition categories. The preset driving conditions include at least smooth driving conditions, slightly rough driving conditions, and pulse impact driving conditions. The dimensionality-reduced one-dimensional logical vector is input into a normalized probability function layer for numerical transformation, calculating the percentage probability that the current input state feature belongs to each of the above preset driving conditions. These percentage probability values ​​are used as the classification confidence of each driving condition category. When the judgment category is an impact condition, the specific implementation process of outputting the driving scenario command indicating that there is a pulse impact load on the road surface ahead includes: in the underlying logic control unit, traversing and comparing the classification confidence of all preset driving conditions output in the current calculation cycle, selecting the one with the largest value, and locking the category index corresponding to the largest classification confidence as the judgment category that the vehicle is about to face; when the judgment category is locked as a smooth condition (such as a well-paved asphalt road or a flat cement road), the first type of trigger level flag is issued through the internal bus, that is, the driving scenario command indicating that the road condition ahead is flat is output, which is used to wake up and execute the subsequent electromagnetic parameter adaptive adjustment link for the hub motor driver; when the judgment category is locked as a pulse impact condition (such as the presence of speed bumps, protrusions or deep potholes ahead), the second type of trigger level flag is issued through the internal bus, that is, the driving scenario command indicating that there is a pulse impact load on the road surface ahead is output, which is used to urgently wake up and trigger the subsequent joint compensation link of suspension mechanical damping and motor reverse thrust torque; By introducing a neural network architecture with an embedded spatiotemporal cross-attention mechanism, the limitations of traditional road condition recognition that relies solely on a single environmental perception or a single vehicle body feedback are broken. The vehicle's current temporal dynamic posture is used as a query clue to actively focus on and assign different attention weights to the geometric features of the road surface in the spatial dimension ahead. This adaptive feature aggregation mechanism enables the classifier to capture the dynamic coupling effect between the vehicle and the road.

[0025] See Figure 3 When a driving scenario command indicating a flat road ahead is received, the electromagnetic regulation link is triggered, and the preset PID parameter mapping network is called to perform online forward inference on the speed loop PID parameters and torque loop PID parameters of the hub motor controller, generate the target current parameters and output them to the underlying inverter, and update the PWM modulation command stream. The preset PID parameter mapping network includes an input layer, a hidden layer, and an output layer. It extracts the vehicle body triaxial acceleration parameters, motor triaxial acceleration parameters, wheel bounce parameters, and current vehicle longitudinal speed from the multi-source physical state matrix at the end of the current time sliding window in real time, and performs normalization processing to construct the input vector for the input layer. This input vector is then input to the input layer and subjected to a weighted summation operation based on a pre-stored weight matrix and bias terms through the hidden layer. A nonlinear activation function is used for feature mapping, and the result is passed to the output layer. The output layer outputs six consecutive floating-point values, corresponding to the gain of the speed loop and the gain of the torque loop in the hub motor controller, respectively. These six gain parameters are written to the bottom-level control register of the hub motor controller in real time, and the speed deviation and current deviation are adjusted in a closed loop based on the updated gain parameters. Specifically, when the road condition feature classifier outputs a driving scenario command indicating that the road ahead is flat, the online parameter optimization process is immediately triggered; features are extracted from the multi-source physical state matrix maintained in memory in real time, and the data sequence at the end of the current time sliding window, that is, the one closest to the current sampling time, is accurately extracted. This includes the real-time parameters of the vehicle's vertical three-axis acceleration fed back by the inertial unit, the vibration acceleration parameters of the wheel hub motor housing, the displacement and bounce parameters of the suspension system, and the vehicle's current actual longitudinal driving speed. Perform normalization preprocessing. Using the maximum and minimum values ​​of each sensor's physical range pre-stored in a configuration table, the raw data with physical units is linearly mapped to a dimensionless numerical range between 0 and 1. The processed data is then concatenated in a preset order to construct a multidimensional column vector, which serves as the excitation signal for the input layer of a three-layer feedforward neural network. The system invokes the neural network model weights stored in non-volatile memory to initiate online forward inference. The input vector passes through the input layer into the hidden layer, where each neuron is fully connected to the input layer nodes. During inference computation, the controller performs a matrix-dot product operation between the input vector and the pre-stored hidden layer weight matrix, and accumulates the corresponding bias terms one by one. The result is processed by a non-linear activation function, which employs a hyperbolic tangent function to map the linearly combined signal into intermediate features with complex non-linear expressive capabilities. The feature signals processed by the hidden layer are further transmitted to the output layer; the output layer consists of 6 independent neurons, which output 6 consecutive floating-point values. These 6 values ​​correspond to the core decision variables in the control strategy, and are specifically divided into two groups: the first group of 3 values ​​corresponds to the proportional coefficient, integral coefficient, and derivative coefficient in the speed loop control algorithm of the hub motor controller, respectively; the second group of 3 values ​​corresponds to the proportional coefficient, integral coefficient, and derivative coefficient in the torque loop control algorithm, respectively. Through the interrupt mechanism of the underlying bus, these parameters are written into the underlying control register of the hub motor controller in real time, completing the overwrite update of the original fixed PID parameters; in the subsequent control cycle, the speed loop performs closed-loop adjustment on the deviation between the actual speed obtained by the resolver sensor and the target speed according to the updated gain parameters, calculates and outputs the target torque command. The torque loop receives the target torque command output by the speed loop and performs a secondary closed-loop calculation in conjunction with the real-time stator current signal collected by the current sensor. At this time, the torque loop uses the updated proportional, integral, and derivative gains to perform high-frequency correction on the torque deviation. By adjusting the dynamic characteristics of the torque loop, target current parameters are generated to counteract the high-frequency micro-vibrations of the road surface, which characterize the magnitude and phase of the compensation torque required by the electromagnetic field inside the motor to suppress vibration. The actual current signal inside the hub motor is collected in real time by a current sensor, and the difference between the actual current and the target current parameter is calculated to obtain the current deviation value. At this time, the controller uses the inferred torque loop proportional, integral and derivative gain to process the deviation at high speed and generate a voltage compensation vector. Based on the calculated voltage compensation vector, the conduction time of the six power switches in the underlying inverter within one modulation cycle is calculated using space vector control logic. In practice, these times are converted into specific digital quantities and written into the underlying comparison register of the microprocessor in real time. As the values ​​in the register are updated, the original PWM instruction stream changes accordingly, that is, the output voltage is controlled by adjusting the duty cycle of the width of each pulse signal. The underlying inverter receives the updated pulse width modulation command stream and drives its internal power devices to switch on and off at a frequency of tens of thousands of times per second.

[0026] By extracting multi-source dynamic features in real time through a three-layer feedforward neural network and performing nonlinear mapping, online adaptive tuning of the dual-loop control parameters of the hub motor is achieved, overcoming the shortcomings of traditional fixed PID parameters such as response lag and limited adjustment. The six gain parameters generated by inference are directly written into the underlying control register, enabling high-frequency updates of the PWM command stream and extremely fast closed-loop response of the compensation current. This scheme effectively smooths out high-frequency micro-amplitude vibrations caused by the road surface, improves vehicle ride comfort, and suppresses motor torque pulsation and additional electromagnetic energy consumption.

[0027] When a driving scenario command indicating that there is a pulse impact load on the road surface ahead is obtained, the mechanical compensation link is triggered by the multi-source physical state matrix to perform calculation processing, generating suspension damping coefficient command and motor reverse compensation torque command, which are synchronously output to suspension actuator and motor driver respectively. The mechanical compensation link extracts the concave and convex feature geometric dimensions and the current vehicle longitudinal speed from the multi-source physical state matrix. Based on the longitudinal spatial position of the concave and convex feature geometric dimensions and the current vehicle longitudinal speed, it calculates the expected arrival time of the pulse impact load and generates a feedforward excitation amplitude according to the geometric elevation mapping. It extracts the three-axis acceleration parameters of the vehicle body and motor from the multi-source physical state matrix as real-time feedback quantities. Combining the expected arrival time and the feedforward excitation amplitude, it calculates the actual dynamic force deviation of the suspension system. According to the preset frequency band difference of the electromechanical actuator response, the actual dynamic force deviation is decoupled in the frequency domain: for the low-frequency, high-amplitude vertical excitation component, it calculates and generates the suspension damping coefficient command for adjusting the active suspension dissipation channel; for the high-frequency wheel speed fluctuation and microscopic longitudinal slip component caused by road impact, it calculates and generates the motor reverse compensation torque command for suppressing the torsional vibration of the unsprung mass of the hub motor. Specifically, the controller retrieves the geometric dimensions of the concave and convex features identified in the feedforward road surface excitation sub-matrix, including the height or depth of the obstacle, its longitudinal span length, and the obstacle's current longitudinal spatial position relative to the vehicle. Simultaneously, it extracts the current vehicle longitudinal speed parameters in real time. By dividing the longitudinal distance between the obstacle and the vehicle by the current longitudinal speed of the vehicle, the countdown time before the wheel contacts the impact source is calculated. Based on a preset elevation-force mapping table, and according to the height and geometry of the obstacle, combined with the sprung mass and spring stiffness coefficient of the vehicle, the amplitude of the vertical excitation force generated by the road surface on the wheel is estimated. This amplitude serves as a reference value for feedforward control, used to pre-adjust the initial attitude of the actuator. The system monitors and extracts dynamic response parameters from a multi-source physical state matrix in real time. Specifically, this includes triaxial acceleration parameters of the vehicle body collected by acceleration sensors located at the four corners of the vehicle, as well as triaxial acceleration parameters of the motors integrated within the wheel hub motors. These real-time feedback quantities are normalized and transformed into dynamic load changes representing the vertical direction of the wheels and the vehicle body. The controller compares the predicted feedforward excitation force amplitude with the measured dynamic load feedback value, calculating the actual dynamic force deviation between the two. This deviation represents the residual impact force that the current suspension system has not yet been able to effectively absorb. Because the active suspension actuators (such as electronically controlled shock absorbers or air springs) and the hub motors have significantly different response frequencies, the controller uses a built-in frequency filter to decompose the actual dynamic force deviation into components of different frequency bands. For the low-frequency vertical excitation component below 10 Hz with high amplitude characteristics, it is determined that it mainly affects the vehicle's attitude stability. At this point, the controller calculates and generates suspension damping coefficient commands based on the magnitude of the force deviation. For high-frequency wheel speed fluctuations and microscopic longitudinal slip components above 15 Hz caused by road impacts, it is determined that they mainly affect the operating stability of the hub motor and the torsional vibration of the unsprung mass. Since the motor has microsecond-level torque response capability, the controller calculates and generates motor reverse compensation torque commands based on these high-frequency components. For the transition frequency band between 10 and 15 Hz, a weighted fusion method is adopted to simultaneously trigger mechanical damping fine-tuning and motor torque compensation. The suspension damping coefficient command is sent to the controller of the suspension actuator, which adjusts the damping force of the electromagnetic shock absorber to achieve mechanical energy buffering. The motor reverse compensation torque command is sent synchronously to the motor driver, which drives the hub motor to generate an instantaneous compensation torque opposite to the impact direction by changing the output phase and amplitude of the current loop. This achieves full-frequency online adjustment of the pulse impact load, ensuring the vehicle's ride comfort and handling stability under complex impact conditions.

[0028] By establishing a mechanical compensation link, feedforward predictive control and frequency-domain decoupling adjustment across actuators for pulsed impact loads were achieved. By utilizing obstacle geometry and vehicle speed to predict impact arrival time and excitation amplitude, the actuators were given advanced response margins, overcoming the lag of traditional feedback control. Through frequency-domain decoupling of dynamic force deviations, the dissipation advantages of active suspension in absorbing low-frequency, high-energy impacts, as well as the rapid response capabilities of hub motors in suppressing high-frequency speed fluctuations and unsprung mass torsional vibrations, were fully utilized.

[0029] By deeply fusing multi-source sensor data and extracting road surface geometric features online, the system achieves the identification and feedforward prediction of driving scenarios. This solution constructs a dual-link regulation mechanism that combines electromagnetic adjustment and mechanical compensation: under flat road conditions, the motor control parameters are dynamically optimized through a PID parameter mapping network to suppress high-frequency micro-amplitude vibrations; under impact road conditions, the coordinated action of suspension damping adjustment and motor reverse torque mitigates the interference of transient impacts on vehicle posture and suppresses transient wheel slippage. This scenario-specific online regulation method coordinates the dynamic response of the in-wheel motor drive system and the suspension system, enabling dynamic adaptation of the vehicle under different road conditions.

[0030] Example 2: This embodiment applies an online adjustment method for wheel hub damping based on multi-source sensor data fusion to a test vehicle equipped with a wheel hub motor traveling at a constant longitudinal speed of 36 km / h (10 m / s). The road ahead initially consists of a smooth asphalt surface, followed by a standard speed bump with a height of 0.04 m (4 cm) and a longitudinal span of 0.2 m. During the data acquisition phase, the vehicle-mounted central processing unit emits a 10Hz synchronization pulse to control the vision sensor to capture an image with a resolution of 1920×1080. At the same time, the lidar completes the acquisition of 50,000 scanning points in a single frame, retrieves the pre-stored 4×4-dimensional sensor extrinsic parameter matrix, and projects the scanning point with coordinates (5.0, 0.2, -1.2) in the original point cloud to the (960, 540) pixel position in the image coordinate system, ensuring that the visual texture and spatial depth information are strictly aligned in the spatiotemporal dimension within a 10ms acquisition cycle. In the terrain segmentation and feature generation stage, the BiSeNet semantic segmentation model identifies the drivable road surface boundary within a 15-meter range ahead and divides this area into a 0.1m × 0.1m square grid network. For the asphalt road surface grid, the elevation range is calculated to be 0.005m and the variance is 0.0002. When a speed bump is scanned, the mean elevation difference between adjacent grids reaches 0.035m, exceeding the abrupt change threshold of 0.03m. The abrupt change node is then identified, and the geometric dimension of the concave-convex feature is locked at a height of 0.04m, thus completing the quantitative modeling of the road surface features. A 500ms (50 sampling points) time sliding window is established to collect parameters such as vehicle vertical acceleration (e.g., 0.1g) and motor-side acceleration in real time. Based on a vehicle speed of 10m / s, a feature sequence of 5 meters in length is extracted from the grid network. Through maximum and minimum value normalization, the acceleration, vehicle speed and road surface undulation are converted into floating-point numbers in the range [0,1], generating a 50×N-dimensional multi-source physical state matrix. During the road condition classification decision-making stage, the neural network with an embedded spatiotemporal cross-attention mechanism decouples the state matrix. Before the vehicle approaches the speed bump, the spatial dimension feature sequence is given a higher attention weight. The classification output layer calculates that the confidence level of the impact condition is 92%, which is much higher than the 5% of the "smooth condition". It issues a second type of trigger level flag and urgently switches the control link from electromagnetic adjustment for flat roads to mechanical compensation link. During the electromagnetic regulation link execution phase (for the previously smoothed road surface), a three-layer feedforward neural network is invoked. The current vibration vector is input, and the network forward inferences to output 6 floating-point values ​​(proportional gain 2.5, integral gain 0.8), which are written to the controller's underlying register in real time through an interrupt mechanism. The hub motor driver updates the 20kHz PWM modulation command stream accordingly, generating a micro-compensation torque to eliminate high-frequency pulsations inside the motor and keep the vertical acceleration fluctuation of the vehicle body within ±0.02g. During the mechanical compensation link execution phase (for speed bump impact), calculations indicate that the impact will arrive in 0.5 seconds (distance 5m / vehicle speed 10m / s), and the estimated vertical excitation amplitude is 1500N. For low-frequency components below 10Hz, the output command adjusts the opening of the active suspension solenoid valve to 35% to increase instantaneous damping; for high-frequency speed fluctuations above 15Hz, the motor outputs a reverse compensation torque of 15N·m to ensure that the suspension travel is controlled within ±30mm when the vehicle crosses the speed bump.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online adjustment of wheel hub damping based on multi-source sensor data fusion, characterized in that, include: The system collects point cloud data and image sequence data of the road ahead using visual sensors and LiDAR; and performs elevation mapping processing using terrain segmentation algorithms. Generate road surface geometric feature data; The vehicle body's three-axis acceleration parameters, motor's three-axis acceleration parameters, wheel bounce parameters, and current longitudinal speed are collected in real time by the inertial unit. These parameters are then spliced ​​with the road surface geometric feature data to generate a multi-source physical state matrix. This matrix is ​​then input into a road condition feature classifier for matching and outputs driving scenario instructions. When a driving scenario command indicating a flat road ahead is received, the electromagnetic regulation link is triggered, and the preset PID parameter mapping network is called to perform online forward inference on the speed loop PID parameters and torque loop PID parameters of the hub motor controller, generate the target current parameters and output them to the underlying inverter, and update the PWM modulation command stream. When a driving scenario command indicating a pulse impact load on the road ahead is obtained, the mechanical compensation link is triggered by the multi-source physical state matrix for calculation and processing, generating a suspension damping coefficient command and a motor reverse compensation torque command, which are then synchronously output to the suspension actuator and the motor driver, respectively.

2. The method for online adjustment of wheel hub vibration damping based on multi-source sensor data fusion according to claim 1, characterized in that, The road point cloud data and image sequence data are obtained by jointly acquiring the visual sensor and the lidar at the same timestamp, respectively acquiring the original image and the original lidar point cloud; the original image is subjected to distortion correction processing to generate the image sequence data, and the original lidar point cloud is subjected to noise reduction and clutter filtering processing to generate the preprocessed point cloud; the pre-calibrated sensor extrinsic parameter matrix is ​​retrieved, and the three-dimensional coordinate system of the preprocessed point cloud is projected onto the camera coordinate system of the visual sensor to obtain the road point cloud data that is strictly aligned in time and space.

3. The method for online adjustment of wheel hub damping based on multi-source sensor data fusion according to claim 1, characterized in that, The terrain segmentation algorithm is an elevation mapping algorithm based on a rasterized digital elevation model; The step of generating road surface geometric feature data by performing elevation mapping processing through terrain segmentation algorithm includes: extracting the drivable road surface boundary ahead based on the image sequence data, extracting the corresponding target road surface point cloud from the road point cloud data ahead, and dividing the target road surface point cloud into a two-dimensional grid network; extracting the Z-axis elevation coordinates of the point cloud inside each grid, and calculating the range and variance of the elevation coordinates inside each grid as the road surface undulation degree characterizing the local roughness of the corresponding area. Calculate the mean elevation difference between adjacent grids, extract continuous grid groups with differences exceeding a preset threshold as abrupt change nodes, and calculate the geometric dimensions of the concave and convex features based on the span and extreme values ​​of the continuous grid groups. The road surface undulation and the geometric dimensions of the concave and convex features are matrix-concatenated to generate the road surface geometric feature data.

4. The method for online adjustment of wheel hub vibration damping based on multi-source sensor data fusion according to claim 1, characterized in that, The specific process of generating the multi-source physical state matrix includes: performing time-stamp synchronization and filtering noise reduction on the real-time collected vehicle body three-axis acceleration parameters, motor three-axis acceleration parameters, wheel bounce parameters, and current vehicle longitudinal speed to construct a vehicle motion state sub-matrix within a set time sliding window; extracting the road surface geometric feature data, and serializing and expanding the road surface undulation and concave-convex feature geometric dimensions according to their longitudinal spatial position relative to the vehicle to construct a feedforward road surface excitation sub-matrix corresponding to the prediction range of the time sliding window; performing dimensionless processing on the vehicle motion state sub-matrix and the feedforward road surface excitation sub-matrix respectively; aligning and concatenating the normalized vehicle motion state sub-matrix and the feedforward road surface excitation sub-matrix according to feature dimensions to generate the multi-source physical state matrix representing the coupling relationship between the vehicle's current dynamic attitude and the road conditions ahead, and inputting it into a road condition feature classifier for matching processing.

5. The method for online adjustment of wheel hub vibration damping based on multi-source sensor data fusion according to claim 4, characterized in that, The road condition feature classifier is a neural network model with an embedded spatiotemporal cross-attention mechanism, which decouples and extracts the multi-source physical state matrix into time-dimensional features representing vehicle dynamics and spatial-dimensional features representing the geometric distribution of the road surface ahead. The time-dimensional features are input as query vectors to the spatiotemporal cross-attention layer, where adaptive weight allocation and feature aggregation are performed on the spatial-dimensional features to generate a global context feature vector. The global context feature vector is then input to the classification output layer for calculation to generate classification confidence scores for each preset driving condition. Extract the operating condition category corresponding to the maximum classification confidence. When the category is determined to be a smooth operating condition, output the driving scenario command that represents a flat road ahead. When the category is determined to be an impact operating condition, output the driving scenario command that represents a pulse impact load on the road ahead.

6. The method for online adjustment of wheel hub damping based on multi-source sensor data fusion according to claim 1, characterized in that, The preset PID parameter mapping network includes an input layer, a hidden layer, and an output layer. It extracts the vehicle body triaxial acceleration parameters, motor triaxial acceleration parameters, wheel bounce parameters, and current vehicle longitudinal speed from the multi-source physical state matrix at the end of the current time sliding window in real time, and performs normalization processing to construct the input vector for the input layer. This input vector is then fed into the input layer, and a weighted summation operation based on a pre-stored weight matrix and bias terms is performed through the hidden layer. A non-linear activation function is then used for feature mapping, and the result is passed to the output layer. The output layer outputs six consecutive floating-point values, which correspond to the gain of the speed loop and the gain of the torque loop in the hub motor controller, respectively. The six gain parameters are written into the bottom control register of the hub motor controller in real time, and the speed deviation and current deviation are adjusted in a closed loop according to the updated gain parameters.

7. The method for online adjustment of wheel hub vibration damping based on multi-source sensor data fusion according to claim 1, characterized in that, The mechanical compensation link extracts the concave and convex feature geometric dimensions and the current vehicle longitudinal speed from the multi-source physical state matrix. Based on the longitudinal spatial position of the concave and convex feature geometric dimensions and the current vehicle longitudinal speed, it calculates the expected arrival time of the pulse impact load and generates a feedforward excitation amplitude according to the geometric elevation mapping. It extracts the body and motor triaxial acceleration parameters from the multi-source physical state matrix as real-time feedback quantities. Combining the expected arrival time and the feedforward excitation amplitude, it calculates the actual dynamic force deviation of the suspension system. According to the preset electromechanical actuator response frequency band difference, it decouples the actual dynamic force deviation in the frequency domain: for the low-frequency, high-amplitude vertical excitation component, it calculates and generates the suspension damping coefficient command for adjusting the active suspension dissipation channel. For the high-frequency wheel speed fluctuations and microscopic longitudinal slip components caused by road impacts, the reverse compensation torque command of the motor is calculated and generated to suppress the torsional vibration of the unsprung mass of the hub motor.

Images