Intelligent wheelchair control method and system

Abstract

The present invention relates to the field of intelligent wheelchair control technology, and more particularly to an intelligent wheelchair control method and system. The method comprises the following steps: constructing a three-dimensional ground feature model by extracting historical environmental data from the memory of an electronic monitoring device on the wheelchair; then, performing a geometric analysis of potholes to obtain pothole data, and based on this data, performing a wheel jam scenario simulation to simulate wheel breakaway behavior; finally, constructing an optimization model using breakaway behavior learning data, and optimizing it by training feature importance parameters to form an optimized jam and breakaway control model; this model is then transmitted to an intelligent wheelchair control terminal. This present invention further improves intelligent wheelchair control technology by optimizing it.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent wheelchair control, and in particular to an intelligent wheelchair control method and system. Background Art

[0002] As an auxiliary travel tool, smart wheelchairs can provide these people with a more convenient and comfortable mobility experience. Smart wheelchairs not only have basic driving functions, but also have a variety of intelligent features, such as path planning, obstacle avoidance, automatic navigation, etc. Although modern smart wheelchairs have made significant progress, they still face many challenges. The most prominent of these is that on complex terrain or uneven ground, the wheels will fall into potholes, causing the wheelchair to be unable to travel smoothly or even get stuck. This situation not only increases the difficulty of using a wheelchair for users with limited mobility. However, a traditional smart wheelchair control method has a weak ability to respond to wheel jam situations, resulting in poor adaptive ability of smart wheelchairs to get out of wheel jams. Summary of the Invention

[0003] Based on this, it is necessary to provide a smart wheelchair control method and system to solve at least one of the above technical problems.

[0004] To achieve the above object, a method for controlling an intelligent wheelchair is provided, the method comprising the following steps:

[0005] Step S1: extracting historical environmental monitoring data on the wheelchair's path from the memory of the electronic monitoring device carried by the wheelchair; and constructing a three-dimensional ground feature model of the path based on the historical environmental monitoring data;

[0006] Step S2: performing a ground pothole geometry analysis on the three-dimensional ground feature model to obtain ground pothole geometry data; performing a wheel jam scenario simulation based on the ground pothole geometry data to obtain wheel jam scenario simulation data; performing a wheel jam and release behavior simulation learning on the wheel jam scenario simulation data based on the ground pothole geometry data to obtain wheel jam and release behavior learning data;

[0007] Step S3: construct a model for the wheel jamming and escaping behavior learning data, and then optimize the training feature importance parameters to obtain a jamming and escaping control optimization model; send the jamming and escaping control optimization model to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

[0008] Preferably, step S1 includes the following steps:

[0009] Step S11: extracting historical environmental monitoring data on the wheelchair's running path from the memory of the electronic monitoring device carried on the wheelchair;

[0010] Step S12: Analyzing the path ground characteristics of the historical environmental monitoring data to obtain the historical path ground characteristics;

[0011] Step S13: Marking the spatial coordinates of the historical path ground features to obtain the historical path coordinate ground features;

[0012] Step S14: constructing a three-dimensional ground feature model of the path based on the ground features of the historical path coordinates.

[0013] Preferably, step S2 includes the following steps:

[0014] Step S21: performing ground pothole geometry analysis on the three-dimensional ground feature model to obtain ground pothole geometry data;

[0015] Step S22: Obtain the size and structure of the wheelchair;

[0016] Step S23: performing wheel jam scenario simulation based on the ground pothole geometry data, thereby obtaining wheel jam scenario simulation data;

[0017] Step S24: performing wheel jamming and breaking-freeing behavior simulation learning on the wheel jamming scenario simulation data according to the ground pothole geometry data and the size and structure of the wheelchair to obtain wheel jamming and breaking-freeing behavior learning data.

[0018] Preferably, step S24 includes the following steps:

[0019] Step S241: Evaluate the pothole friction characteristics of the ground pothole geometry data to generate pothole friction characteristic data;

[0020] Step S242: Deducing the load-force deflection of the stuck wheel based on the wheel jam scenario simulation data based on the size and structure of the wheelchair, to obtain the load-force deflection data of the stuck wheel;

[0021] Step S243: extracting the depth, width, and inclination of the potholes in the ground; performing a wheel steering angle force fulcrum analysis based on the depth, width, inclination, and friction characteristic data of the potholes to obtain steering angle force fulcrum data;

[0022] Step S244: performing wheelchair center of gravity fluctuation deflection distribution analysis on the steering angle force fulcrum data based on the stuck wheel load force deflection data to obtain wheelchair center of gravity fluctuation deflection distribution data;

[0023] Step S245: Based on the wheelchair center of gravity fluctuation and deflection distribution data and the steering angle force fulcrum data, the multi-wheel drive output torque is controlled based on the pothole friction characteristic data and the depth, width, and pothole inclination of the ground pothole geometry data to obtain the multi-wheel drive output torque; and the breakaway acceleration kinetic energy of the stuck wheel is extracted from the multi-wheel drive output torque.

[0024] Step S246: performing sideslip-related kinetic energy differential control on the pothole friction characteristic data and the depth, width, and pothole inclination in the ground pothole geometry data based on the breakaway acceleration kinetic energy, thereby obtaining kinetic energy differential control data;

[0025] Step S247: performing wheel jamming and breaking-free behavior simulation learning based on the multi-wheel drive output torque and kinetic energy differential control data to obtain wheel jamming and breaking-free behavior learning data.

[0026] Preferably, step S244 includes the following steps:

[0027] Perform spatial vector decomposition on the load force deflection data of the stuck wheel to obtain the three-axis load offset vector of the stuck wheel;

[0028] According to the three-axis load offset vector of the stuck wheel, the transient reaction torque between the stuck wheel and the support contact surface is deduced from the steering angle force support data to obtain the transient reaction torque between the contact surfaces;

[0029] The instantaneous reaction torque between the contact surfaces is solved by calculating the angular acceleration and the mass center line acceleration between the stuck wheel steering angles to obtain the instantaneous center of gravity offset vector;

[0030] Performing center of gravity offset trajectory iterative processing on the instantaneous center of gravity offset vector to obtain center of gravity offset trajectory iterative data;

[0031] The wheelchair center of gravity fluctuation deviation distribution is analyzed based on the center of gravity offset trajectory iterative data to obtain the wheelchair center of gravity fluctuation deviation distribution data.

[0032] Preferably, step S246 includes the following steps:

[0033] The non-constant acceleration variance is calculated for the breakaway acceleration kinetic energy to obtain the non-constant acceleration variance;

[0034] The depth, width and inclination of the potholes in the ground are analyzed based on the non-constant acceleration variance to identify the lateral component of the steering angle, and obtain the lateral component of the steering angle increment data.

[0035] Based on the non-constant acceleration variance and the lateral component force increment data of the steering angle, the slip lateral adhesion loss data is quantified for the friction characteristic data of the pothole morphology to obtain the slip lateral adhesion loss data;

[0036] Perform slip distance change rate analysis based on slip lateral adhesion loss data to generate slip distance change rate;

[0037] Performing torque vector distribution on both wheels according to the slip lateral adhesion loss data and the slip distance change rate to obtain torque vector distribution data on both wheels;

[0038] The sideslip-related kinetic energy differential control is performed based on the torque vector distribution data of the wheels on both sides, thereby obtaining the kinetic energy differential control data.

[0039] Preferably, step S3 includes the following steps:

[0040] Step S31: performing feature selection processing on the wheel jamming and breaking free behavior learning data to obtain jamming and breaking free behavior feature selection data;

[0041] Step S32: constructing a model based on the selected data of the stuck-and-release behavior characteristics based on the deep Q network algorithm to obtain a stuck-and-release control model;

[0042] Step S33: Optimizing the training feature importance parameters based on the stuck-and-release control model to obtain a stuck-and-release control optimization model;

[0043] Step S34: Send the stuck and escape control optimization model to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

[0044] Preferably, step S32 includes the following steps:

[0045] Step S321: performing state feature encoding processing on the selected data of the stuck and escape behavior features, constructing a continuous state space representation, and obtaining an escape behavior state feature set;

[0046] Step S322: performing deep Q network value parameter training on the discrete feature set of the breakaway behavior state based on the deep Q network algorithm to obtain the breakaway behavior state value parameter;

[0047] Step S323: constructing a model based on the breaking free behavior state value parameter to obtain a stuck breaking free control model.

[0048] Preferably, step S33 includes the following steps:

[0049] Step S331: collecting operation response parameters based on the sticking and breaking free control model, thereby obtaining the sticking and breaking free operation response parameters;

[0050] Step S332: performing parameter sensitivity disturbance analysis on the stuck-and-break-free operation response parameters to obtain parameter sensitivity disturbance data;

[0051] Step S332: assigning feature importance weights to the stuck-and-break-free operation response parameters according to the parameter sensitivity disturbance data to obtain the response parameter feature importance weights;

[0052] Step S333: Optimizing the feature importance parameters of the stuck-and-release control model according to the feature importance weights of the response parameters to obtain an optimized stuck-and-release control model.

[0053] Preferably, the present invention further provides an intelligent wheelchair control system for executing the intelligent wheelchair control method described above, the intelligent wheelchair control system comprising:

[0054] A historical monitoring model extraction module is used to extract historical environmental monitoring data on the wheelchair's path from the memory of the electronic monitoring device carried by the wheelchair; based on the historical environmental monitoring data, a three-dimensional ground feature model of the path is constructed;

[0055] The breakaway behavior learning module is configured to analyze the geometric morphology of potholes on a three-dimensional ground feature model to obtain geometric morphology data of the potholes; simulate wheel jam scenarios based on the geometric morphology data of the potholes to obtain wheel jam scenario simulation data; and simulate learning of wheel jam breakaway behavior based on the wheel jam scenario simulation data using the geometric morphology data of the potholes to obtain wheel jam breakaway behavior learning data.

[0056] The escape control model construction module is used to construct a model for the wheel jamming and escape behavior learning data, and then optimize the training feature importance parameters to obtain the jamming and escape control optimization model; the jamming and escape control optimization model is sent to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

[0057] The present invention provides a beneficial effect by extracting historical environmental monitoring data from the memory of the wheelchair's onboard electronic monitoring device and constructing a three-dimensional ground feature model of the path based on this data, enabling comprehensive and accurate modeling of the ground environment traversed by the wheelchair. This lays the foundation for subsequent ground feature analysis, wheel jam scenario simulation, and control optimization. By utilizing historical environmental data, the system can obtain realistic and representative path information, avoiding the errors caused by relying solely on real-time monitoring data and ensuring the accuracy and reliability of the model. This process enables the wheelchair to understand the path terrain characteristics in advance, providing strong data support for predicting and preventing wheel jams. Ground pothole geometry analysis of the three-dimensional ground feature model accurately identifies the specific location and morphology of potholes, cracks, or other obstacles on the ground. This step, through precise geometric analysis, obtains detailed data on ground potholes, providing a deep understanding of terrain variations and providing detailed input data for subsequent wheel jam scenario simulations. This pothole data not only simulates jam scenarios but also enhances the intelligent wheelchair's ability to cope with complex ground environments by learning wheel jam and breakaway behaviors. By constructing a model for learning data on wheel jamming and breaking free and optimizing the importance parameters of training features, the smart wheelchair can obtain a highly adaptive and optimized jamming and breaking free control model. By learning the actual ground environment and jamming scenarios, the model can adjust the control strategy in real time according to the specific motion state of the wheel, quickly respond when jamming occurs, automatically adjust the wheel drive mode or take other measures to break free. The optimized control model ensures the reliability of the smart wheelchair in complex environments and avoids the situation where a traditional wheelchair cannot be freed due to jamming. Therefore, the present invention is an optimization process for a traditional smart wheelchair control method, which solves the problem that a traditional smart wheelchair control method has a weak ability to respond to wheel jamming scenarios, thereby causing the smart wheelchair to have poor adaptive ability to get out of wheel jams, improves the ability to respond to wheel jamming scenarios, and strengthens the adaptive ability of the smart wheelchair to get out of wheel jams. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A flowchart of a method for controlling an intelligent wheelchair;

[0059] Figure 2 for Figure 1 Detailed implementation steps of step S2 in FIG.

[0060] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG. DETAILED DESCRIPTION

[0061] See also Figures 1 to 3, a smart wheelchair control method, the method comprising the following steps:

[0062] Step S1: extracting historical environmental monitoring data on the wheelchair's path from the memory of the electronic monitoring device carried by the wheelchair; and constructing a three-dimensional ground feature model of the path based on the historical environmental monitoring data;

[0063] Step S2: performing a ground pothole geometry analysis on the three-dimensional ground feature model to obtain ground pothole geometry data; performing a wheel jam scenario simulation based on the ground pothole geometry data to obtain wheel jam scenario simulation data; performing a wheel jam and release behavior simulation learning on the wheel jam scenario simulation data based on the ground pothole geometry data to obtain wheel jam and release behavior learning data;

[0064] Step S3: construct a model for the wheel jamming and escaping behavior learning data, and then optimize the training feature importance parameters to obtain a jamming and escaping control optimization model; send the jamming and escaping control optimization model to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

[0065] In the embodiment of the present invention, reference Figure 1 The above is a flowchart of the steps of a smart wheelchair control method of the present invention. In this example, the smart wheelchair control method includes the following steps:

[0066] Step S1: extracting historical environmental monitoring data on the wheelchair's path from the memory of the electronic monitoring device carried by the wheelchair; and constructing a three-dimensional ground feature model of the path based on the historical environmental monitoring data;

[0067] In an embodiment of the present invention, an instruction set is called through the local memory interface of the electronic monitoring device carried by the smart wheelchair to retrieve a preset historical environmental monitoring data cache block in the storage path. The storage path is a partitioned data page structure inside the EEPROM structure. A method of matching the path number and the timestamp segment number based on the key-value index table is adopted to sequentially extract the continuous environmental data segments on the wheelchair's running path according to the path segment number sorting rules. The data content includes the ground height difference point cloud data along the path recorded by the high-precision laser ranging module, the ground vibration response data collected by the three-axis acceleration sensor, the posture change data collected by the angular velocity sensor, and the ground surface state data collected by the temperature and humidity module. The various types of raw data extracted are synchronized with time. After the markers are aligned, they are sent to the data cache channel, and the local linear interpolation method is used to fill the missing points in the time series to ensure data continuity. Subsequently, the three-dimensional space reconstruction processing flow is used to model the above historical environmental monitoring data. The processing flow includes four steps: data denoising, resampling, registration and point cloud gridding. The denoising process uses the bilateral filtering method to set the neighborhood kernel size to 9×9, the resampling uses the Poisson reconstruction algorithm to control the reconstruction depth to 8, the point cloud registration uses the ICP algorithm for position and posture adjustment, and the gridding process uses a resolution of 0.02m to construct a three-dimensional grid. The final generated path three-dimensional ground feature model is in the form of multi-segment splicing, each segment contains a terrain height map, a surface slope map, an obstacle edge map and a roughness distribution map.

[0068] Step S2: performing a ground pothole geometry analysis on the three-dimensional ground feature model to obtain ground pothole geometry data; performing a wheel jam scenario simulation based on the ground pothole geometry data to obtain wheel jam scenario simulation data; performing a wheel jam and release behavior simulation learning on the wheel jam scenario simulation data based on the ground pothole geometry data to obtain wheel jam and release behavior learning data;

[0069] In the embodiment of the present invention, based on the three-dimensional ground feature model of the path constructed in step S1, a geometric partitioning processing method based on the watershed algorithm is used to segment the three-dimensional ground data into pothole areas. The pothole area identification standard is that the height gradient is greater than 10° and the depression depth exceeds 2 cm. The original ground height distribution is smoothed by a three-dimensional Gaussian convolution kernel and the local lowest point area is calculated. The continuous depression area is identified by a clustering algorithm (DBSCAN density clustering, eps is set to 0.03, and the minimum sample number is 5), thereby extracting the ground pothole geometric data. Each extracted pothole area contains parameters such as depth, width, length, and edge inclination angle. Subsequently, the wheelchair structural parameters are obtained, including a wheel diameter of 0.2m, a total wheelchair mass of 120kg, a wheelbase of 0.45m, and a center of gravity height of 0.55m. A finite element simplified simulation scene is constructed based on the above-mentioned ground geometric data and the wheelchair structural parameters. The static friction coefficient in the interactive contact model is 0.6, and the rolling friction coefficient is 0.02. The Euler-Lagrangian simulation method is used to simulate the force state of the wheelchair moving into the pothole area. The simulation step size is set to 5ms. The data such as the degree of wheel group sinking, the stuck time period, the wheel rolling radius compression ratio, and the wheelchair posture change in each step are output as the wheel stuck scenario simulation data. The stuck and escape behavior simulation learning is further carried out based on the ground geometry data, wheelchair structure data and the stuck scenario simulation data. The wheel propulsion force, acceleration change trajectory, steering angle and center of gravity offset during the historical stuck process are statistically reviewed. The sequence regression analysis method is combined with the stuck trajectory data to establish the escape action feature sequence. Each sequence contains parameters such as the lateral and longitudinal force distribution under the time step, the output torque fluctuation amplitude, the wheel slip angle and the road adhesion boundary change as the wheel stuck and escape behavior learning data.

[0070] Step S3: construct a model for the wheel jamming and escaping behavior learning data, and then optimize the training feature importance parameters to obtain a jamming and escaping control optimization model; send the jamming and escaping control optimization model to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

[0071] In the embodiment of the present invention, the wheel sticking and breaking free behavior learning data obtained in step S2 is subjected to feature selection processing, and the selected dimensions include: maximum lateral output force, torque change rate, critical fluctuation value of wheel slip angle, longitudinal force balance time, front and rear wheel adhesion ratio and other features. The feature dimension is reduced to the first 6 principal components by the principal component analysis (PCA) method as the stuck and breaking free behavior feature selection data, and then the deep Q network algorithm is used to construct the stuck and breaking free control model. The network structure adopted is an input layer dimension of 6, two hidden layers with 128 and 64 ReLU neurons respectively, and the output layer is 5 action codes in the action space (in-place twisting, left and right unilateral steering, bilateral low-speed rolling, and rapid retreat). The greedy strategy and experience replay mechanism are used for training. The sample size of each training round is 3000, and the target value update frequency is set to every 10 The training process is repeated once every 0 rounds, and the convergence criterion is that the average reward exceeds the threshold of 0.85. After the training is completed, a preliminary stuck-and-breakaway control model is obtained, and the training feature importance parameter optimization processing is continued. The response data of the model in the simulation environment is collected, including four parameters: actual action success rate, time consumption, action frequency, and system stability. The disturbance sensitivity of each parameter to the control effect is calculated, and the gradient normalization method is used to obtain the disturbance sensitivity weight. Based on the weight, the influence of each feature on the model is regressed, and the input feature dimension of the deep Q network is weightedly optimized to form the final stuck-and-breakaway control optimization model. The model is encapsulated in ONNX format and sent to the embedded control unit of the smart wheelchair via the CAN communication interface protocol. It is deployed in the real-time processing module of the main control chip for real-time action decision-making, completing the execution deployment of the smart wheelchair control method.

[0072] Step S1 includes the following steps:

[0073] Step S11: extracting historical environmental monitoring data on the wheelchair's running path from the memory of the electronic monitoring device carried on the wheelchair;

[0074] Step S12: Analyzing the path ground characteristics of the historical environmental monitoring data to obtain the historical path ground characteristics;

[0075] Step S13: Marking the spatial coordinates of the historical path ground features to obtain the historical path coordinate ground features;

[0076] Step S14: constructing a three-dimensional ground feature model of the path based on the ground features of the historical path coordinates.

[0077] In an embodiment of the present invention, by invoking the built-in memory interface instruction set of the electronic monitoring device onboard the wheelchair, historical environmental monitoring data stored in a predetermined data page is accurately read from the internal EEPROM partition data page based on the path number and timestamp index. This data includes point cloud data acquired by high-precision laser scanning, vibration response data recorded by a triaxial accelerometer, attitude change data recorded by an angular velocity sensor, and environmental status data recorded by a temperature and humidity sensor. High-speed data exchange is achieved using the SPI bus data transmission protocol. All extracted data is linearly sorted by timestamp, and CRC32 redundancy check is performed on data blocks to ensure data integrity. The reading process sets the memory cache size to 512KB and stores data in the memory buffer at a data refresh rate of 500 times per second. During the extraction operation, the data format is binary, and the accuracy of laser scanning data is controlled to 0.005m, the accuracy of inertial data is controlled to 0.001 radians, and the accuracy of temperature and humidity data is controlled to 0.1 units. During the interface call process, strict parsing is performed according to the memory data structure, thereby achieving complete extraction of historical environmental monitoring data for continuous sections along the wheelchair's travel path during operation, and the data format meets the requirements of subsequent path ground feature analysis. The historical environmental monitoring data extracted in step S11 is first preprocessed, including applying Gaussian filtering to the raw laser scanning point cloud data to eliminate noise and performing median filtering to remove outliers. Subsequently, multi-sensor data is synchronized based on the data timestamp. The laser point cloud is registered using the iterative closest point algorithm (ICP algorithm). The data is segmented based on the registration results. Edge detection and fitting are performed on the segmented ground point cloud using the RANSAC plane fitting algorithm. The fitting error is set to an upper limit of 0.02m. The ground height mean, local slope, and roughness information are calculated for each data block. The edge identification of continuous sampling areas is performed using the region growing method. The texture information is further quantitatively described using the local binary pattern (LBP) statistical method. The roughness parameters within the ground sampling area are calculated using the point cloud density analysis method, thereby generating historical path ground feature data containing the elevation, slope, flatness, and obstacle edge information of each sampling area. During the processing, the data resolution of each sampling area is controlled within 0.03m and all operations are performed according to preset parameters to ensure that the data output format is unified into a two-dimensional matrix structure.A spatial coordinate marking operation is performed on the historical path ground feature data obtained in step S12. During the operation, the geographical coordinates of each sampling point are converted using the simultaneously collected GPS data and inertial navigation data. The Euler angle rotation matrix is ​​combined with the quaternion transformation algorithm to achieve a strict transformation between the local coordinate system and the global coordinate system. The accuracy of the rotation matrix is ​​set to four decimal places. Each sampling data is matched according to the acquisition time. If there is no GPS signal, the position is calculated based on the inertial navigation data. During the conversion operation, each sampling point obtains accurate longitude, latitude and altitude parameters. The marking results are stored in a standardized CSV format. Each record data contains longitude, latitude, altitude and corresponding feature index. The error is controlled within 0.1m during the entire operation to ensure that the historical path ground feature data corresponds one-to-one with the spatial coordinate information and the data conversion operation is accurately implemented according to the preset conversion matrix. Based on the historical path coordinate ground feature data obtained in step S13, a path three-dimensional ground feature model construction operation is implemented. During the operation, the spatial mark data is first gridded, and the size of each grid unit is set to 0.02m². The data in each grid is smoothed by the bilinear interpolation method for the height value. The local slope change is calculated by Lagrange interpolation at each grid boundary. After gridding, the octree data structure is used to realize global data fast indexing. The height, slope, and texture roughness parameters of each grid are stored in a sparse matrix storage method. During the construction process, the three-dimensional reconstruction algorithm is called to parallelize the massive data. The output result is a high-resolution digital elevation model and three-dimensional point cloud data. The model is accompanied by global coordinate transformation parameter information. The construction operation sets the reconstruction depth to 8 and the data processing resolution reaches the centimeter level. All operations are strictly performed according to the preset data grid algorithm and data smoothing algorithm, thereby realizing the construction of a path three-dimensional ground feature model with high precision and high consistency.

[0078] Step S2 includes the following steps:

[0079] Step S21: performing ground pothole geometry analysis on the three-dimensional ground feature model to obtain ground pothole geometry data;

[0080] Step S22: Obtain the size and structure of the wheelchair;

[0081] Step S23: performing wheel jam scenario simulation based on the ground pothole geometry data, thereby obtaining wheel jam scenario simulation data;

[0082] Step S24: performing wheel jamming and breaking-freeing behavior simulation learning on the wheel jamming scenario simulation data according to the ground pothole geometry data and the size and structure of the wheelchair to obtain wheel jamming and breaking-freeing behavior learning data.

[0083] As an example of the present invention, refer to Figure 2As shown, in this example, step S2 includes:

[0084] Step S21: performing ground pothole geometry analysis on the three-dimensional ground feature model to obtain ground pothole geometry data;

[0085] In the embodiment of the present invention, in the obtained three-dimensional ground feature model of the path, a three-dimensional grid segmentation processing method is first used to divide the three-dimensional model into multiple ground analysis areas with each 1m length as the division unit. The first-order and second-order derivatives of the ground elevation data in each analysis area are calculated to extract the elevation gradient, boundary slope and internal depth change value of the local depression area. A three-dimensional morphology recognition algorithm based on regional extreme value search is used to identify the minimum point in each grid unit and perform depression boundary closure processing. The boundary closure method is based on the connectivity threshold setting between the local minimum value and the surrounding maximum slope point. The threshold is set to a slope greater than 10° and a relative height difference greater than 20mm. Then, the method is used to identify the minimum point in each grid unit and perform depression boundary closure processing. The DBSCAN density clustering algorithm was used to cluster and integrate adjacent depression structures. The clustering parameter EPS was set to 0.05 m, and the minimum sample number was 10. Finally, each cluster area was defined as an independent ground pothole structure, and its maximum depth, effective width, projected area, boundary slope, and edge continuity were calculated. The depth calculation used the difference between the lowest point and the average elevation of the boundary as the evaluation basis. The boundary slope was calculated by the angle between the boundary normal and the horizontal plane. The edge continuity was calculated by the standard deviation of the side length of the edge triangulation structure. All extracted geometric parameters were stored in a structured array according to the block number, and finally the ground pothole geometric morphology data was formed.

[0086] Step S22: Obtain the size and structure of the wheelchair;

[0087] In the embodiment of the present invention, the wheelchair structure design drawings are standardized and data is collected according to ISO The 7176 standard documents the key dimensions of wheelchairs and uses vernier calipers and laser rangefinders to precisely measure the main components of the wheelchair. The data obtained includes a front wheel radius of 0.1m, a rear wheel radius of 0.2m, a wheelbase of 0.45m, a front-to-rear wheel spacing of 0.6m, a vehicle height of 0.95m, a vehicle length of 1.05m, a wheelchair center of mass height of 0.5m, and a vehicle mass of 110kg. The wheel set has a front-wheel steering and rear-wheel drive structure, and the wheel frame is a welded aluminum alloy structure. The vehicle body contour is sampled using a point cloud using 3D laser scanning with a resolution of 4 sampling points per millimeter. A spatial registration algorithm is used to calculate the 3D coordinates of the wheel position, bearing support, center of gravity position, and structural contact points. This data is entered into the CAD modeling software in a hierarchical structure. The material properties, connection methods, and rotation restrictions of all structural components are also marked. This allows the accurate use of mechanical property parameters consistent with the actual structure during subsequent motion simulation, forming a wheelchair size and structure database with structural stiffness, mass distribution, and motion restrictions.

[0088] Step S23: performing wheel jam scenario simulation based on the ground pothole geometry data, thereby obtaining wheel jam scenario simulation data;

[0089] In the embodiment of the present invention, the ground pothole geometry data obtained in step S21 and the wheelchair structure data constructed in step S22 are imported as input parameters into the multi-rigid body dynamics simulation module. The simulation module uses a forward integration solver based on the Euler method, the time step is set to 0.002 seconds, the total simulation time is 2 seconds, the initial speed of the wheelchair is set to 0.8 m / s, the initial attitude angle is 0°, and the drive wheel output torque is 12 N·m. The Penalty method is used to process the contact and collision process between the tire and the ground. The friction model adopts the Coulomb friction model and sets the static friction coefficient to 0.6 and the dynamic friction coefficient to 0.4. In each simulation step, the front and rear wheels are recorded in the pothole area. The sinking depth, tilt angle, slip distance, slip time period, overall pitch angle change of the wheelchair, and the output counter-torque of the drive wheel are calculated. The rate of change of the normal contact force between the wheel group and the ground is detected to determine whether a stuck state has occurred. The sticking judgment standard is that the wheel group contact force drops by more than 20% and the vehicle body speed is less than 0.1m / s for more than 0.3s. The simulation output includes multi-dimensional stuck scenario simulation data consisting of the starting time, duration, stuck wheel number, posture change trajectory, and wheel group force state of the wheelchair in different pothole shapes. The simulation process is performed under the boundary conditions of constant gravity acceleration and stationary ground. All variables are normalized and recorded and output in time series.

[0090] Step S24: performing wheel jamming and breaking-freeing behavior simulation learning on the wheel jamming scenario simulation data according to the ground pothole geometry data and the size and structure of the wheelchair to obtain wheel jamming and breaking-freeing behavior learning data.

[0091] In the embodiment of the present invention, based on the wheel jam scenario simulation data obtained in step S23, the wheelchair structural parameters and ground geometric parameters are introduced to construct a jam escape behavior sample sequence. For each set of jam scenario data, the driving wheel torque change value, the wheelchair center of mass lateral displacement, longitudinal displacement, steering angular velocity, ground pothole boundary morphological parameters and the wheel group slip speed during the jam period are extracted with a time resolution of 0.01s. The mapping relationship between the torque output fluctuation and the wheelchair posture response is modeled using the time series difference analysis method. The continuous action features of each escape process are constructed as a high-dimensional time series vector. The features include the driving strategy (such as the left and right wheel differential input), the front wheel slip, and the front wheel slip. The k-means clustering method is used to perform cluster analysis on all the break-away behaviors. The number of clusters is set to 10, and the Euclidean distance is used as the similarity index. The average feature extraction is performed on the behavior sequence of each cluster center. The extracted features include instantaneous lateral power peak, average rolling friction work, maximum slip angle, stuck duration and posture recovery time. Finally, a set of learning data on the break-away behavior covering various stuck situations is formed. In each set of data, the ground shape, wheel structure, input action and output response are strictly matched one by one, thereby providing a high-stability learning data foundation for the subsequent stuck control model training.

[0092] Step S24 includes the following steps:

[0093] Step S241: Evaluate the pothole friction characteristics of the ground pothole geometry data to generate pothole friction characteristic data;

[0094] Step S242: Deducing the load-force deflection of the stuck wheel based on the wheel jam scenario simulation data based on the size and structure of the wheelchair, to obtain the load-force deflection data of the stuck wheel;

[0095] Step S243: extracting the depth, width, and inclination of the potholes in the ground; performing a wheel steering angle force fulcrum analysis based on the depth, width, inclination, and friction characteristic data of the potholes to obtain steering angle force fulcrum data;

[0096] Step S244: performing wheelchair center of gravity fluctuation deflection distribution analysis on the steering angle force fulcrum data based on the stuck wheel load force deflection data to obtain wheelchair center of gravity fluctuation deflection distribution data;

[0097] Step S245: Based on the wheelchair center of gravity fluctuation and deflection distribution data and the steering angle force fulcrum data, the multi-wheel drive output torque is controlled based on the pothole friction characteristic data and the depth, width, and pothole inclination of the ground pothole geometry data to obtain the multi-wheel drive output torque; and the breakaway acceleration kinetic energy of the stuck wheel is extracted from the multi-wheel drive output torque.

[0098] Step S246: performing sideslip-related kinetic energy differential control on the pothole friction characteristic data and the depth, width, and pothole inclination in the ground pothole geometry data based on the breakaway acceleration kinetic energy, thereby obtaining kinetic energy differential control data;

[0099] Step S247: performing wheel jamming and breaking-free behavior simulation learning based on the multi-wheel drive output torque and kinetic energy differential control data to obtain wheel jamming and breaking-free behavior learning data.

[0100] In the embodiment of the present invention, in the extracted ground pothole geometric data, a surface contact grid unit is first established for each independent pothole area, and the grid side length is set to 0.01m. The corresponding boundary inclination angle, elevation gradient and surface curvature parameters of each grid unit are extracted. Then, combined with the measured tire slip distance, wheel idling rate and the resistance reaction value of the wheel group after ground contact collected from the on-site wheelchair passing through the pothole area, the kinetic friction coefficient and static friction coefficient of the pothole area are inversely calculated using the Coulomb friction theory. The measured tire contact surface material is a polyurethane composite rubber material with a baseline friction coefficient of 0.55. By performing a single-wheel low-speed climbing test under actual road conditions, the wheelchair propulsion speed is set. The speed was 0.3 m / s, the output torque of the driving wheel was controlled to 9 N·m, and the duration was 4 s. Mechanical feedback was collected at the peak point of the resistance after the wheel group entered the pothole unit with different slopes. The friction response coefficient of the tire at different boundary slopes was calculated by back-calculation. The slip speed-friction attenuation relationship curve was fitted and analyzed using polynomial regression. Finally, the friction coefficients corresponding to different grid cells in each pothole area were spatially interpolated. The regional friction characteristic field was constructed using the inverse distance weighted interpolation (IDW) method, and the spatial interpolation weight exponent was set to 2. The friction characteristic data of the pothole morphology, including the static friction coefficient, kinetic friction coefficient, friction change rate, and friction anisotropic gradient distribution, were obtained. Based on the wheel axle position, vehicle mass distribution, wheel geometry parameters and vehicle center of mass position coordinates in the wheelchair structure, the stuck wheel load force deflection deduction is performed with the support of the stuck scenario simulation data. First, during the simulation process, the mechanical data of all stuck moments are analyzed frame by frame, and the wheel normal force, tangential force, ground normal support force and wheelchair center of mass vector trajectory changes within the range of 1s before the start of the stuck and 0.5s after the end of the stuck are extracted. The wheel normal force sampling frequency is 500Hz, and the three-axis decomposition of each frame of force data is performed. The coordinate axis conversion matrix is ​​used to map all wheel forces to the vehicle body inertial coordinate system. The mapping relationship between the wheel support reaction force and the inertia trajectory of the vehicle body's center of mass is established in the linear coordinate system. The normal force value of the stuck wheel is further compared with the theoretical uniform load value. The difference is the load force deviator. At the same time, the boundary inclination and friction anisotropy of the pothole area where the wheel is located are considered. The projection force of the gravity component in the slope direction is superimposed and calculated, and the change of the friction force component in the actual side slip trend direction of the wheel is considered. Therefore, the three-axis force distribution vector sequence of the stuck wheel is established in each time frame, and the frequency domain characteristics of the force fluctuation are extracted using Fourier transform. Its frequency principal component is analyzed and output as the load force deviator data of the stuck wheel.Based on the friction characteristic data of the pothole shape obtained in step S241, the three-dimensional geometric parameters of the pothole area are extracted, including depth, width and boundary inclination angle, where the depth is defined as the difference between the lowest point of the area and the average elevation of the boundary, the width is defined as the maximum span of the area projected on the axis of the vehicle's running direction, and the inclination is defined as the average angle between the normal line of the boundary point and the horizontal plane. The three parameters are obtained by a gradient calculation method based on elevation data. The extracted parameters are then standardized and introduced as input parameters into the wheel steering angle force support point analysis process. This process uses the force support point deduction method to establish a six-degree-of-freedom moment balance equation group for the wheel group and the ground contact surface in three-dimensional space. The Newton-Euler rigid body equation is used for solution. The initial position of the wheel set is based on the posture coordinates in the simulation frame before jamming. The normal reaction force and lateral friction force vector generated by each fulcrum on the wheel set under different steering angles are calculated respectively. The fulcrum is defined as the point with high force on the tire surface in the ground contact area. During the analysis, the angle step value is set to 2°, and the angle range is plus or minus 30°. The force center position, torsional moment, slip trend angle and torque synthesis direction angle are calculated at each step. The calculation results are clustered according to the steering angle, and the key angles with maximum normal support force and minimum slip risk in the jam-breaking process and the corresponding force fulcrum coordinates are extracted to obtain the steering angle force fulcrum data. The load force deviation data of the stuck wheel calculated in step S242 is used as the basic input, combined with the steering angle force fulcrum data obtained in step S243, and the center of gravity trajectory reverse projection method is used to spatially map the offset trajectory of the wheelchair's overall center of gravity during the stuck process. In the specific operation, the center of mass force triangle rule is used to establish a connection relationship between the wheelchair's center of mass, the force fulcrum, and the wheel contact surface in three-dimensional space, and the center of mass offset direction and velocity vector caused by the torque imbalance during the stuck process are analyzed. The time series integration method is used to calculate the center of mass trajectory change within each 0.01s time step, and the center of mass offset vector field is constructed based on the trajectory change sequence. The center of gravity vectors within the time step are normalized in amplitude and then superimposed to form a center of mass fluctuation distribution map. Subsequently, the Gaussian kernel density estimation method is used to extract the offset probability density function. The kernel function bandwidth is set to 0.02m, and the center of gravity distribution in the two-dimensional plane is probabilistically processed. The distribution center is combined with the offset trend direction to quantify the center of gravity deviation trend of the wheelchair in the stuck state. Finally, the wheelchair center of gravity fluctuation deviation distribution data is formed, which includes four groups of core parameters such as the main offset direction angle, the extreme value of the offset distance, the center of mass fluctuation frequency and the fluctuation deviation change rate. The data structure is organized in the form of time series frame + vector group to ensure that the subsequent center of gravity control strategy has a stable analytical basis.

[0101] On the basis of the wheelchair center of gravity fluctuation deviation distribution data obtained in step S244, combined with the steering angle force fulcrum data extracted in step S243, a multi-input multi-output MIMO control structure is established to perform torque distribution analysis on the four wheels of the wheelchair. The control structure adopts a discrete time stepping control method with a control period of 0.01s. The input includes the center of gravity offset vector, the main offset direction angle, the fluctuation frequency and the offset extreme value. The output is the instantaneous driving torque value of the front left wheel, the front right wheel, the rear left wheel and the rear right wheel. First, according to the position of the steering angle fulcrum, the wheelchair dynamic response distribution is divided into five typical support domains. The stability margin and reaction force peak change of each support domain under different drive distribution strategies are calculated respectively. Then, the depth, width and boundary inclination in the ground pothole geometry data are used as dynamic correction factors to introduce The torque distribution function is incorporated into the function, which uses an empirical weighted regression method to establish the optimal matching relationship between each torque response and ground disturbance. The minimum driving torque of the wheel group is set to 4 N·m, and the maximum driving torque is set to 18 N·m. These constraints are used to iteratively solve the closed-loop control of the multi-wheel torque. The control goal is to make the angle between the wheelchair's center of gravity fluctuation vector and the center of gravity stability direction less than 15°, and to align the direction of the total vehicle torque force with the normal vector of the maximum force support. During the control process, the output torque increment of the stuck wheel is monitored in real time, and the instantaneous acceleration of the stuck wheel is estimated by backward velocity differentiation. The stuck wheel acceleration data is multiplied by the equivalent wheel mass to obtain the breakaway acceleration kinetic energy. This kinetic energy data is stored by time step, and the corresponding input torque, wheel position number, ground friction parameter, and boundary slope parameter are recorded for subsequent analysis and learning processing.The kinetic energy of the breakaway acceleration obtained in step S245 is used as the main input variable. The kinetic energy is vector-decomposed along the lateral direction of the wheelchair through the three-dimensional vector direction. Combined with the isotropic friction change gradient stored in the friction characteristic data of the pothole morphology, a side slip risk function is constructed. This function performs an inner product operation on the lateral component of the kinetic energy and the direction of the maximum friction gradient to calculate the energy projection value, which is then superimposed with the pothole boundary inclination in the ground geometric parameters as an offset factor of the risk index, where the boundary inclination is the angle between the normal direction of the pothole and the direction of the gravity vector. It is converted into an effective sliding component force distribution vector using trigonometric functions. On this basis, a kinetic energy differential control function is established. The control function adopts a linear proportional differential strategy, linearly matches the torque difference between the left and right wheels with the kinetic energy lateral projection value, and sets the proportional The factor K is 0.12 times the magnitude of the kinetic energy side vector. After each matching, a slip angle change prediction is performed. If the predicted slip angle is greater than 8°, wheelset modulation adjustment is performed, that is, a 1 N·m load reduction is performed on the high-torque side wheel and a 1 N·m load increase is performed on the low-torque side wheel. The maximum number of adjustments per wheel during the entire process is set to 3. After each adjustment, it is necessary to recalculate whether the angle between the wheel set output acceleration and the vehicle center of mass offset direction remains within 15°. If not, the adjustment value is rolled back and the left and right wheel differential speed is forced to not exceed 3 N·m. At the end of each wheel control cycle, the final torque combination under differential control, the corresponding pothole boundary slope, the various friction factors, and the output kinetic energy residual are recorded to form a kinetic energy differential control data sequence. The data is organized into a standard structure sequence according to wheel set number, timestamp, and input and output parameter format.The multi-wheel drive output torque data in step S245 and the kinetic energy differential control data in step S246 are jointly processed to construct a sample set for learning the simulation of the stuck and escape behavior. First, the stuck behavior learning time window is set to 1 s before and after, and the time step is 0.01 s. In each time window, the torque input value of each wheelchair, the slip rate, the rolling resistance change rate, the acceleration vector, the center of mass offset path, the fulcrum direction change trend and the actual displacement are extracted. The causal relationship between the kinetic energy input value and the posture change response is processed by first-order differentiation to obtain three types of indicators: acceleration response rate, angular velocity response rate and center of mass return speed as learning label data. After the sample construction is completed, all feature sequences are normalized. The normalization method adopts the range normalization method to linearly compress the feature distribution to between 0 and 1, and then the dynamic time warping algorithm (Dynamic Time Warping) is used. Warping is used to time-align all breakaway behavior sequences to correct for feature drift errors caused by differences in behavior duration. A multidimensional index tree is constructed based on each set of behavior features in the combined space of steering angle, kinetic energy change, friction characteristics, and torque for subsequent behavior clustering and control strategy back-checking. All behavior samples are clustered using K-means with 8 cluster centers. Based on the jam-release action type, they are classified into four action mechanisms: push-pull breakaway, in-place reverse differential breakaway, kinetic slip breakaway, and diagonal wheelset breakaway. This ultimately forms a jam-release behavior learning dataset containing structured fields such as feature trajectory sequence, key input variable combination, response rate parameter, posture recovery time, and final state stability description, which is used to support subsequent control model training and strategy generation.

[0102] Step S244 includes the following steps:

[0103] Perform spatial vector decomposition on the load force deflection data of the stuck wheel to obtain the three-axis load offset vector of the stuck wheel;

[0104] According to the three-axis load offset vector of the stuck wheel, the transient reaction torque between the stuck wheel and the support contact surface is deduced from the steering angle force support data to obtain the transient reaction torque between the contact surfaces;

[0105] The instantaneous reaction torque between the contact surfaces is solved by calculating the angular acceleration and the mass center line acceleration between the stuck wheel steering angles to obtain the instantaneous center of gravity offset vector;

[0106] Performing center of gravity offset trajectory iterative processing on the instantaneous center of gravity offset vector to obtain center of gravity offset trajectory iterative data;

[0107] The wheelchair center of gravity fluctuation deviation distribution is analyzed based on the center of gravity offset trajectory iterative data to obtain the wheelchair center of gravity fluctuation deviation distribution data.

[0108] In an embodiment of the present invention, the stuck wheel load force deflection data obtained in step S242 includes the three-axis components of the normal load, tangential load, and contact surface force distribution of the wheelset in each stuck time frame. First, the overall structural coordinate system of the wheelchair is converted into the vehicle body inertial coordinate system, and the spatial vector decomposition of each frame of the stuck wheel force data is performed based on the wheel set installation position and the support structure coordinates. The three-dimensional coordinate system projection matrix is ​​used to decompose all loads into load vectors along the X-axis (forward and backward direction of the wheelchair), the Y-axis (left and right direction of the wheelchair), and the Z-axis (vertical direction). The projection step size is set to 0.01s. At each moment, a three-axis force state table of the wheelset is established to record the offset direction, offset amplitude, and offset duration. On this basis, a three-axis load offset vector sequence of the stuck wheel is established. The vector structure is a time tag index + a three-axis load data group. Subsequently, based on the steering angle force fulcrum data obtained in step S243, the three-axis load offset vector of the stuck wheel is cross-producted with the normal unit vector of the fulcrum position to calculate the lever arm length vector between the tire and the ground contact surface. This is then vector-multiplied with the load vector to obtain the transient reaction torque between the contact surfaces. This calculation is performed using the vector cross product method. The torque unit in each frame is N·m. The output includes the transient reaction torque values ​​in the three directions around the X-axis, Y-axis, and Z-axis, and the angle between the instantaneous reaction torque and the normal direction of the corresponding fulcrum is recorded for analyzing its overturning trend and torque stability changes. After obtaining the transient reaction torque, the angular acceleration and center-of-mass linear acceleration are calculated for the tire response under different steering angles. First, a dynamic rigid-body model is established based on the wheelchair's structural parameters. The front and rear wheels are connected to the vehicle body using a fixed-axle connection. The wheel inertia moments are calibrated experimentally: 0.032 kg·m² for the front wheel and 0.048 kg·m² for the rear wheel. The transient reaction torque values ​​are substituted into the Euler dynamic equations. Combined with the actual eccentricity parameters of each wheel assembly, the rate of change of torque applied to the stuck wheel in different directions is analyzed. The angular acceleration curve of the stuck wheel is then calculated using the vector differential method. Based on this, the linear acceleration generated by the wheelchair's center of mass along the normal direction of the contact surface is inferred. The center of gravity height from the ground is set to 0.42 m, and the horizontal projection distance from the wheel axle to the center of gravity is set to 0.35 m. Combined with these parameters, a transient mechanical response channel is established, outputting the instantaneous center-of-gravity acceleration vector for each frame, which is the instantaneous center-of-gravity offset vector. Based on the continuous instantaneous center of gravity offset vector data, an iterative path of the wheelchair's overall center of gravity trajectory is constructed. The vector integration method is used for processing, and the offset vectors at each moment are cumulatively superimposed with an overlay step of 0.01s. The total cumulative time window is set to 3s. After each iteration, the accumulated offset position is recorded and a time-series trajectory coordinate set is generated. The trajectory structure is a three-dimensional coordinate sequence + timestamp index. This trajectory is used to describe the dynamic evolution process of the center of gravity offset. The trajectory data is then smoothed using a fifth-order Bezier curve fitting method to eliminate high-frequency jitter components and ensure that the trajectory changes reflect the actual force response trend.After completing the iterative construction of the center of gravity offset trajectory data, the spatial distribution density analysis of the data set was performed. The Gaussian kernel function was used to estimate the offset probability of each point. The kernel function bandwidth was set to 0.025m. The kernel density superposition method was used to calculate the offset distribution of the entire offset trajectory space. Six statistical indicators, including the main direction angle of the center of gravity offset, the extreme value of the offset range, and the fluctuation frequency change rate, were extracted to establish the wheelchair center of gravity fluctuation deviation distribution data. This data structure includes the main deviation direction vector, the fluctuation frequency sequence, the extreme value of the center of mass offset distance in each time period, the direction angle change rate curve, and the trajectory stability index. It serves as the data input basis for the subsequent multi-wheel torque scheduling and steering angle response coupling control.

[0109] Step S246 includes the following steps:

[0110] The non-constant acceleration variance is calculated for the breakaway acceleration kinetic energy to obtain the non-constant acceleration variance;

[0111] The depth, width and inclination of the potholes in the ground are analyzed based on the non-constant acceleration variance to identify the lateral component of the steering angle, and obtain the lateral component of the steering angle increment data.

[0112] Based on the non-constant acceleration variance and the lateral component force increment data of the steering angle, the slip lateral adhesion loss data is quantified for the friction characteristic data of the pothole morphology to obtain the slip lateral adhesion loss data;

[0113] Perform slip distance change rate analysis based on slip lateral adhesion loss data to generate slip distance change rate;

[0114] Performing torque vector distribution on both wheels according to the slip lateral adhesion loss data and the slip distance change rate to obtain torque vector distribution data on both wheels;

[0115] The sideslip-related kinetic energy differential control is performed based on the torque vector distribution data of the wheels on both sides, thereby obtaining the kinetic energy differential control data.

[0116] In this embodiment of the present invention, in the breakaway acceleration kinetic energy data extracted in step S245, each frame of acceleration kinetic energy is composed of the linear acceleration of the stuck wheel in the direction perpendicular to the main slope of the pothole shape and the equivalent mass borne by the wheel. First, this kinetic energy sequence is discretized according to the wheelchair control time period, with a processing period set to 0.02s and a total processing period length of 2s. A sliding window segmentation method is used with a window width of every 100 frames. The acceleration values ​​within each window segment are subjected to non-constant stability statistical analysis. The sample variance calculation formula is used to calculate the variance value of the acceleration sequence to obtain the degree of non-constancy of the acceleration kinetic energy within each period. The critical recognition value is set to 0.6m² / s. 4, the sections exceeding this value are marked as high-fluctuation areas, and each area is assigned a corresponding variance level label, ultimately forming a data pair comparing the time series and the non-constant acceleration variance value. The stuck wheel number, torque input, vehicle body tilt angle, and torque change rate corresponding to each set of variance are recorded, and stored in the acceleration variance index table after sorting according to the time label. The non-constant acceleration variance obtained in step S2461 is used as the main driving factor. Corresponding to each high-fluctuation area segment, the depth, width, and boundary tilt angle of the ground pothole geometry data within this time period are extracted. The spatial geometric distribution sampling radius is set to 0.25 meters, and the three-dimensional ground point cloud data of the area where the stuck wheel is located is fitted. The weighted least squares fitting method is used to reconstruct the surface of the point cloud height data. After reconstruction, the slope distribution function is obtained. The main direction of the surface is derived by the gradient operator to extract the main tilt direction of the pothole. The angle with the wheelchair's forward direction is used as the steering angle offset input. On this basis Calculate the lateral force increment. The lateral force increment is determined by the angle θ between the pothole boundary and the wheel force vector, as well as the friction characteristics. θ is obtained through projection calculation. The increment is calculated based on the lever arm length from the stuck wheel contact point to the center of gravity multiplied by the angular deviation ratio. Assuming a wheelbase of 0.56 meters and an offset angle of 12°, the lateral force increment is calculated as the lateral projection acceleration component of the wheel group multiplied by the equivalent sideslip friction coefficient μ (μ=0.38). This type of data is feature assembled, and finally a steering angle lateral force increment dataset is constructed with the stuck wheel number, pothole feature triplet, offset angle, and lateral force increment as fields. Use the non-constant acceleration variance data from step S2461 and the steering angle lateral force increment data from step S2462 to quantify the lateral adhesion loss caused by slip on the pothole morphology friction characteristic data. First, the friction characteristic data is divided according to the properties of the tire and ground materials, and the basic friction coefficient of each contact area is extracted. and the directional friction variation factor λ, where The value range is between 0.3 and 0.45. λ is the slope factor that describes the change of the friction coefficient in different slip directions. It is defined as the rate of change of the friction coefficient corresponding to a unit angle change. The angle step is set to 5°. λ is obtained from the actual ground tire contact test. For example, the measured λ on the hard asphalt surface is -0.012 / °. On this basis, the expected friction reduction in the current slip direction is calculated according to the lateral angle and force value in the lateral component force increment data of the steering angle. The non-constant acceleration variance is used as the amplification factor. The actual adhesion loss value is the friction direction reduction rate multiplied by λ and the slip angle, and then subtracted The residual adhesion is obtained. If the residual adhesion is lower than 0.1, it is marked as a high-risk loss state. Each set of calculation results records the stuck wheel number, non-constant acceleration variance value, slip angle, friction drop value and residual adhesion value. All records are organized into slip lateral adhesion loss data according to the wheelchair operation timeline.

[0117] In the slip lateral adhesion loss data obtained in step S2463, each set of data corresponds to the remaining lateral adhesion capacity and slip angle change information of the stuck wheel within a specific time period. First, a slip distance integral function is established based on the adhesion loss sequence. The time segment interval is set to a fixed sliding window of 0.02 seconds. The lateral slip displacement value of the stuck wheel is recorded in each sliding window. This displacement is obtained by multiplying the lateral velocity component calculated from the angle between the tire rolling direction and the force offset direction by the time step. The slip distance recording process is started when the force offset angle θ of the stuck wheel is greater than 10°. The total recording time is set to 2 seconds, the number of sample points is collected to 100, and the slip distance recorded for each frame of data is calculated. The slip velocity values ​​were fitted with a continuous function using fifth-order gradient interpolation, and the third-order derivative was used to solve the rate-of-change curve to obtain the slip distance rate of change value. This rate of change value is expressed in meters per square second. A rate-of-change-adhesion comparison matrix was established based on the residual adhesion level corresponding to each frame. Each element in the matrix corresponds to the degree of adhesion loss and the degree of sudden change in slip trend under the current slip state. These rate-of-change values ​​for all sample wheels were standardized using a maximum sideslip velocity conversion rate of 0.6 m / s as the normalization factor. The normalized slip distance rate of change was finally stored and sorted as a set of key stability change indicators for subsequent power vector scheduling analysis.

[0118] In the slip lateral adhesion loss data obtained in step S2463, each set of data corresponds to the remaining lateral adhesion capacity and slip angle change information of the stuck wheel within a specific time period. First, a slip distance integral function is established based on the adhesion loss sequence. The time segment interval is set to a fixed sliding window of 0.02 seconds. The lateral slip displacement value of the stuck wheel is recorded in each sliding window. This displacement is obtained by multiplying the lateral velocity component calculated from the angle between the tire rolling direction and the force offset direction by the time step. The slip distance recording process is started when the force offset angle θ of the stuck wheel is greater than 10°. The total recording time is set to 2 seconds, the number of sample points is collected to 100, and the slip distance recorded for each frame of data is calculated. The slip velocity values ​​are fitted with a continuous function using fifth-order gradient interpolation, and the third-order derivative is used to solve the rate of change curve to obtain the slip distance rate of change value. The unit of this rate of change value is m / s². Combined with the residual adhesion level corresponding to each frame, a rate-of-change-adhesion comparison matrix is ​​established. Each element in the matrix corresponds to the degree of adhesion loss and the degree of sudden change in the slip trend under the current slip state. Such rate of change values ​​of all sample wheel groups are normalized. The normalization factor is based on the maximum side slip speed conversion rate of 0.6m / s. Finally, the normalized slip distance rate of change is stored and sorted as a set of key stability change indicator data for subsequent power vector scheduling analysis. Based on the torque regulation input formed by the slip lateral adhesion loss data in step S2463 and the slip distance rate of change data in step S2464, the friction coefficient reduction and slip trend change level of the current terrain section of the left and right wheels are extracted respectively. The friction coefficient reduction is determined by The difference between the residual adhesion in the current slip direction is calculated, and the low adhesion area threshold is set to 0.15. When the residual adhesion of a certain wheel group is lower than this value, the required compensation torque is realized by increasing the torque of the opposite wheel to achieve the differential suppression strategy. The torque scheduling distribution ratio between the two sides of the wheel group is constructed according to the symmetrical difference of the slip change rate. By constructing the differential weight coefficient α, α is calculated as the difference in slip change rate divided by the average value of the change rate on both sides. If the slip change rate on the left is 0.18 and that on the right is 0.09, then α=1, indicating a complete deviation to the right. The system needs to dispatch the total torque to the right wheel, and the torque ratio For example, a 1:2 distribution is set. In the actual drive system, the maximum single-wheel drive output is set to 18N·m, and the total dispatch output under the current vehicle load level is 24N·m. The left wheel output is 8N·m and the right wheel output is 16N·m. This distribution method will combine the current road slope direction vector, wheelbase length, slip direction, and center of gravity offset angle to construct a multi-dimensional torque vector distribution structure, ultimately forming the torque vector distribution data for both wheels. Its structure fields include left wheel output torque, right wheel output torque, inter-wheel differential ratio, main slip direction angle, torque change rate, and output time label. Using the torque vector distribution data of the two wheels obtained in step S2465 as the differential control input, the acceleration input and energy output curves of the two wheel groups on both sides are fine-tuned in real time by setting the dynamic kinetic energy compensation threshold. First, a perturbation test is performed on the output torque in the direction of the stuck wheel. The response acceleration growth value of the wheel group on this side under the residual adhesion state is evaluated by simulating a torque input of +1N·m. If the acceleration increase is lower than the set sensitivity threshold of 0.12m / s², the current wheel group is considered to be in a low coupling state, and the system performs side slip-related energy transfer control, that is, partially switching the remaining driving force of the left wheel to the right wheel to strengthen its traction side adhesion zone. The transfer ratio is calculated according to the incremental slip change rate of the stuck wheel and the residual adhesion of the other wheel. The inverse of the ability is weighted and normalized, and on this basis, a dynamic kinetic energy differential function based on the time-varying slip vector direction is constructed. The function input includes the time label, the output torque of the left and right wheels, the adhesion loss level, and the rate of change of the slip angle and torque ratio. The output is the corrected differential vector output curve, which is used to update the torque command in the wheelchair drive control logic. The specific control signal is applied to the electric drive controller through PWM modulation, and the control cycle is 20ms. In this control strategy, all kinetic energy distribution ratios and correction directions are deduced and verified by the regression values ​​of the preset experimental database. Finally, the kinetic energy differential control data that meets the mechanical stability conditions is output in each stuck scenario, which is used as the input basis for the next step of stuck escape control optimization.

[0119] Step S3 includes the following steps:

[0120] Step S31: performing feature selection processing on the wheel jamming and breaking free behavior learning data to obtain jamming and breaking free behavior feature selection data;

[0121] Step S32: constructing a model based on the selected data of the stuck-and-release behavior characteristics based on the deep Q network algorithm to obtain a stuck-and-release control model;

[0122] Step S33: Optimizing the training feature importance parameters based on the stuck-and-release control model to obtain a stuck-and-release control optimization model;

[0123] Step S34: Send the stuck and escape control optimization model to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

[0124] As an example of the present invention, refer to Figure 3 As shown, in this example, step S3 includes:

[0125] Step S31: performing feature selection processing on the wheel jamming and breaking free behavior learning data to obtain jamming and breaking free behavior feature selection data;

[0126] In the embodiment of the present invention, the acquired wheel stuck and break-free behavior learning data includes original behavior sequence data of multiple dimensions such as the force state, center of gravity change, slip distance, friction coefficient, steering angle change, stuck duration, and break-free acceleration response experienced by the wheel under various ground pothole forms. In order to extract behavioral features with decision-making value, a method based on mutual information sorting combined with principal component analysis is used for feature selection processing. First, all the original behavior data are sliced ​​and segmented according to each stuck-break-free process. The single sequence frame length is set to 2 seconds, the sampling interval is 0.02 seconds, and the stuck behavior samples containing 100 frames of data are obtained. Then, each segment is sliced ​​and segmented. The dynamic variables in the segment data are uniformly normalized to the range of [0,1] for subsequent processing. The mutual information entropy calculation method is then used to solve the information gain value between each behavioral feature variable and the label state (whether the vehicle successfully escaped). The variables with the top ten information gain values ​​are selected to form a preliminary feature candidate set. The variables include the friction force change rate, center of gravity offset amplitude, stuck wheel drive torque input increment, lateral slip distance, and vehicle body posture angle change rate. The candidate set features are then processed by principal component analysis, and the top five principal component dimensions with a cumulative variance contribution rate of more than 95% are extracted as the final stuck escape behavior feature selection data for deep reinforcement learning input.

[0127] Step S32: constructing a model based on the selected data of the stuck-and-release behavior characteristics based on the deep Q network algorithm to obtain a stuck-and-release control model;

[0128] In an embodiment of the present invention, the stuck-and-breakaway behavior feature data extracted in step S31 is selected as the state input of reinforcement learning, and a policy value estimation method based on a deep Q network (DQN) is used to construct a model. The state space is composed of five-dimensional continuous variables, including the normalized center of gravity offset amplitude, the friction coefficient level, the slip speed, the stuck duration progress, and the torque application rate. The length of each state vector is 5, corresponding to a frame of stuck-and-breakaway behavior scene. Five discrete behaviors are set in the action space, namely low-speed direct drive, high-speed direct drive, two-wheel reverse torque output, unilateral power drive, and direction disturbance input. The ε-greedy strategy is used for training behavior selection. The initial exploration rate is set to 0.9, which decays to 0.1 every 1000 rounds. The network structure adopts a three-layer fully connected neural network with 5 neurons in the input layer, 64 and 128 neurons in the hidden layers, and 5 neurons in the output layer. The activation function adopts ReLU and the loss function adopts Huber Loss, the optimizer is Adam, the learning rate is set to 0.001, the number of training rounds is 30,000, the experience replay batch size is 32, and the target network is used for Q-value update delay stabilization processing. The training output is the Q-value function estimation table corresponding to the five types of control strategies in each state. Finally, the stuck-free control model structure and parameter matrix are obtained for subsequent optimization step calls.

[0129] Step S33: Optimizing the training feature importance parameters based on the stuck-and-release control model to obtain a stuck-and-release control optimization model;

[0130] In an embodiment of the present invention, based on the preliminary stuck-and-breakaway control model obtained in step S32, a training feature importance parameter optimization process is carried out to improve the control strategy's ability to respond stably to actual operating disturbances. This process first collects the model's operating response results on a standard test set and a random operating condition set, including response parameters such as the stuck duration, output torque consumption, total wheelchair posture deviation, and slip distance peak value under each round of control output. After normalizing the response parameters, a first-order partial derivative sensitivity analysis method is used to calculate the sensitivity value of each input feature's influence on the response index under the current model parameters. A list of disturbance weights for each feature's contribution to the objective function is obtained. A feature importance ranking is then constructed based on the disturbance weights, and a combined positive and negative disturbance experiment is performed on the top three features. Each set of disturbance experiments is repeated 2000 times. By comparing the control performance changes of the model before and after the disturbance, the parameters are adaptively adjusted with a fine-grained learning rate. Finally, the feature input distribution is retrained and the parameters are iterated until convergence. The stuck-and-breakaway control optimization model is output. The optimized model has the ability to enhance the strategy convergence speed and improve the response accuracy under complex terrain stuck conditions.

[0131] Step S34: Send the stuck and escape control optimization model to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

[0132] In an embodiment of the present invention, the stuck-and-release control optimization model obtained in step S33 includes a network weight matrix, a state-action value mapping function, an action strategy table, and a state update mechanism. The model parameters are organized in a standard tensor format, and each parameter is structurally normalized according to the input feature standardization specification. The model is uploaded to the embedded control core of the wheelchair control system through a wireless data communication module. The control core adopts a 32-bit floating-point processing architecture, an operating frequency of 200MHz, and a periodic scheduling time window set to 50ms. After the model is deployed, the stuck behavior state data uploaded in real time from the sensor array is read in a time-triggered manner. The sensor data is vectorized according to the input normalization rule and then enters the control optimization model. The model outputs a Q value table based on the current state, selects the multi-wheel output torque vector parameters and the wheelchair center of gravity adjustment direction vector corresponding to the optimal action strategy, and applies the control vector to the wheel drive motor through the PWM control channel to execute the entire process of the stuck-and-release control strategy.

[0133] Step S32 includes the following steps:

[0134] Step S321: performing state feature encoding processing on the selected data of the stuck and escape behavior features, constructing a continuous state space representation, and obtaining an escape behavior state feature set;

[0135] Step S322: performing deep Q network value parameter training on the discrete feature set of the breakaway behavior state based on the deep Q network algorithm to obtain the breakaway behavior state value parameter;

[0136] Step S323: constructing a model based on the breaking free behavior state value parameter to obtain a stuck breaking free control model.

[0137] In an embodiment of the present invention, after completing feature selection of the stuck-and-break-free behavior learning data, in order to realize the state space construction in the reinforcement learning process, it is necessary to perform state feature encoding processing on the feature selected data. The state encoding strategy adopted is based on the normalized continuous variable vector representation method. The specific operation process is to sequentially set the selected five-dimensional behavior feature variables as the five dimensions of the input vector, including the stuck duration, the wheelchair center of gravity offset amplitude, the driving wheel instantaneous output torque, the slip direction offset angle, and the ground friction coefficient change rate. The stuck duration ranges from 0 to 3 seconds, the wheelchair center of gravity offset amplitude ranges from 0 to 150 mm, and the driving wheel instantaneous output torque ranges from 0 to 15 N. m, the slip direction offset angle range is -30 to 30°, and the friction coefficient change rate range is -0.4 to 0.4. All data are interval-normalized in the preprocessing stage and mapped to the [0,1] interval to form a continuous state vector. At the same time, to maintain the temporal continuity in the state space, a sliding window mechanism is used to perform average filtering on the multi-frame state vectors within each control cycle. The window size is set to five frames and the step size is one frame, thereby constructing a temporally smoothed state feature set of the breakaway behavior. This state feature set, as the input state space representation of the deep Q network, has continuity, normalization, and anti-disturbance capabilities, and can meet the requirements of subsequent network training for stable input of the state feature dimension. After obtaining the state feature set of the escape behavior, in order to train the state value function parameters, the temporal difference learning method based on the deep Q network is used for network training. The network architecture adopted is a three-layer fully connected structure. The input layer receives a state vector with a dimension of 5, the number of hidden layer nodes is set to 64 and 128 respectively, and the output dimension of the output layer is 5, corresponding to five types of control actions, including linear drive, direction disturbance, single-wheel high-frequency oscillation drive, torque step input and short-term brake assist action. The training data is an experience data set after the state-action-reward triplet is reorganized. Each data contains the current state vector, the current action number, the instant escape state score obtained, the next state vector and whether it is terminated. Stop marking, during the training process, the experience replay pool mechanism is used to evenly sample the data, the batch size is set to 32, the training rounds are set to 25000 rounds, the loss function is a mixture of mean square error and Huber loss function, and the target network delayed synchronization strategy is used to improve stability. The target network is updated every 100 rounds during training, the learning rate is set to 0.0008, the optimizer is Adam, the ε-greedy strategy is used to select actions, the initial exploration rate is set to 0.95, and it decreases by 0.00003 each round until it reaches 0.05. After the training is completed, the corresponding value function parameters for each state-action combination are obtained, that is, the state value parameters of the escape behavior, which are used in the subsequent model strategy construction process.Based on the training completed in step S322, the value function Q value of each state vector under different action outputs has been obtained. To achieve the final construction of the stuck-free control model, it is necessary to construct a mapping function based on the state value function parameters and establish an action selection strategy mechanism. The specific operation is to first build a state-action Q value mapping table. This table is output in real time through the trained neural network forward reasoning method. After receiving the state input, the Q value distribution corresponding to all actions can be instantly calculated. Then, a hybrid control method based on the maximum Q value selection strategy and the soft maximum strategy is introduced. When the Q value difference is greater than 0.3, the maximum Q value strategy is used to directly select the corresponding action. When the Q value difference is less than 0. 3. The Softmax function is used to perform weighted normalization on the Q value, and then the action sampling decision is made according to the weighted probability. The strategy mechanism regulates the decision stability by setting a dual-threshold structure to ensure the flexibility of the control strategy under uncertain conditions. Finally, the state input dimension, network structure parameters, Q value output format, action strategy function, weight update mechanism, target network structure, loss function expression and other contents are uniformly encapsulated to form a stuck-free control model. This model can realize the real-time computing capability of outputting a set of targeted torque control parameters and control action numbers after inputting a continuous state vector, and meet the actual control cycle requirements deployed on the intelligent wheelchair control terminal.

[0138] Step S33 includes the following steps:

[0139] Step S331: collecting operation response parameters based on the sticking and breaking free control model, thereby obtaining the sticking and breaking free operation response parameters;

[0140] Step S332: performing parameter sensitivity disturbance analysis on the stuck-and-break-free operation response parameters to obtain parameter sensitivity disturbance data;

[0141] Step S332: assigning feature importance weights to the stuck-and-break-free operation response parameters according to the parameter sensitivity disturbance data to obtain the response parameter feature importance weights;

[0142] Step S333: Optimizing the feature importance parameters of the stuck-and-release control model according to the feature importance weights of the response parameters to obtain an optimized stuck-and-release control model.

[0143] In an embodiment of the present invention, after the construction of the jam-and-break free control model is completed, in order to analyze the feedback behavior of the model during actual control execution and collect related dynamic response parameters, high-frequency data acquisition and processing is required for the model's online control results. The acquisition content includes four types of operational response parameters, namely, the real-time output torque value of the wheelchair after the control command is issued, the wheelchair slip angle change rate, the wheelchair body posture offset angle, and the time required for successful actual break free. The sampling frequency is set to 50 Hz, and the data acquisition window period is set to 10 seconds. Multiple independent experimental scenarios are used to control the wheelchair to travel to potholes of different depths, different friction coefficients, and different slopes, and artificially impose initial jamming conditions. After each round of experiments, the four types of response data collected are sorted into time series and processed by least squares interpolation to ensure data continuity. The jam-and-break free operational response parameters are obtained. This parameter set includes both physical and mechanical quantities and system behavior response indicators, and can be directly used as the input basis for training feature optimization. After obtaining the stuck-breakaway response parameters, a parameter sensitivity perturbation analysis is performed to clarify the degree of influence of each response parameter on the control model output. The perturbation analysis strategy adopted is a combined one-dimensional increment method and Latin hypercube sampling method. First, each item in the stuck-breakaway response parameter set is selected as the perturbed variable. While keeping other variables fixed, a perturbation increment ranging from -10% to +10% is applied to the target variable with a step size of 2%. The changing trends of the model output results after the perturbation are recorded. The breakaway success time and control energy consumption are used as model output references. Subsequently, the Latin hypercube sampling method is introduced to perform multiple joint perturbation sampling on the entire parameter space. The number of sampling groups is set to 500, each group containing a random combination of four response parameters within the perturbation range. For each perturbation combination, a model control simulation is performed and the output change is recorded. The sensitivity of each response parameter to the control performance is calculated based on the output fluctuation amplitude, thus forming parameter sensitivity perturbation data. This data is expressed as a function of the difference between the perturbation input and output, which facilitates the subsequent feature importance weighting operation.After obtaining the parameter sensitivity disturbance data, the normalized gradient analysis method is used to assign feature weights to the disturbance data. The sensitivity weight of each response parameter is quantified as a percentage in the form of a score. The maximum disturbance impact output amplitude is defined as a weight of 100%, and the remaining parameters are mapped proportionally. Subsequently, the weights of the original input feature dimensions in the control model are redistributed according to the sensitivity weight. The feature channel weighting mechanism is used to set the adjustment factor of the network connection strength of each feature input channel according to the sensitivity weight. The factor setting range is 0.5 to 1.5. The channel amplification factor is increased to 1 for channels with higher sensitivity weights. .5, the channel factor of those with lower sensitivity weights is compressed to 0.5. At the same time, L2 regularization constraints are applied to the response gradients of neurons in the hidden layer to high-weight feature channels to improve the stability of high-weight features in the weight update process, and a frozen gating layer is set to prevent non-critical feature channels from being abnormally activated in the later stages of training. This optimization mechanism does not involve modification of the control logic, but only makes structural adjustments to the training feature input structure and channel activation path. After iterative training and updating of multiple experimental scenarios, the training feature importance parameter optimization process of the stuck-free control model can be completed, and finally the stuck-free control optimization model is obtained.

[0144] The present invention also provides an intelligent wheelchair control system for executing the intelligent wheelchair control method described above, the intelligent wheelchair control system comprising:

[0145] A historical monitoring model extraction module is used to extract historical environmental monitoring data on the wheelchair's path from the memory of the electronic monitoring device carried by the wheelchair; based on the historical environmental monitoring data, a three-dimensional ground feature model of the path is constructed;

[0146] The breakaway behavior learning module is configured to analyze the geometric morphology of potholes on a three-dimensional ground feature model to obtain geometric morphology data of the potholes; simulate wheel jam scenarios based on the geometric morphology data of the potholes to obtain wheel jam scenario simulation data; and simulate learning of wheel jam breakaway behavior based on the wheel jam scenario simulation data using the geometric morphology data of the potholes to obtain wheel jam breakaway behavior learning data.

[0147] The escape control model construction module is used to construct a model for the wheel jamming and escape behavior learning data, and then optimize the training feature importance parameters to obtain the jamming and escape control optimization model; the jamming and escape control optimization model is sent to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

[0148] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for controlling an intelligent wheelchair, characterized in that: The following steps are involved: Step S1: extracting historical environmental monitoring data on the wheelchair's path from the memory of the electronic monitoring device carried by the wheelchair; and constructing a three-dimensional ground feature model of the path based on the historical environmental monitoring data; Step S2: performing ground pothole geometry analysis on the three-dimensional ground feature model to obtain ground pothole geometry data; Perform wheel jam scenario simulation based on ground pothole geometry data to obtain wheel jam scenario simulation data; Perform wheel jamming and release behavior simulation learning on wheel jamming scenario simulation data based on ground pothole geometry data to obtain wheel jamming and release behavior learning data; Step S3: constructing a model for the wheel jamming and breaking free behavior learning data, and then optimizing the training feature importance parameters to obtain a jamming and breaking free control optimization model; Sending the stuck-and-breakaway control optimization model to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method; Step S2 includes the following steps: Step S21: performing ground pothole geometry analysis on the three-dimensional ground feature model to obtain ground pothole geometry data; Step S22: Obtain the size and structure of the wheelchair; Step S23: performing wheel jam scenario simulation based on the ground pothole geometry data, thereby obtaining wheel jam scenario simulation data; Step S24: performing wheel jamming and release behavior simulation learning on the wheel jamming scenario simulation data according to the ground pothole geometry data and the size and structure of the wheelchair to obtain wheel jamming and release behavior learning data; Step S24 includes the following steps: Step S241: Evaluate the pothole friction characteristics of the ground pothole geometry data to generate pothole friction characteristic data; Step S242: Deducing the load-force deflection of the stuck wheel based on the wheel jam scenario simulation data based on the size and structure of the wheelchair, to obtain the load-force deflection data of the stuck wheel; Step S243: extracting the depth, width, and inclination of the potholes in the ground; performing a wheel steering angle force fulcrum analysis based on the depth, width, inclination, and friction characteristic data of the potholes to obtain steering angle force fulcrum data; Step S244: performing wheelchair center of gravity fluctuation deflection distribution analysis on the steering angle force fulcrum data based on the stuck wheel load force deflection data to obtain wheelchair center of gravity fluctuation deflection distribution data; Step S245: Based on the wheelchair center of gravity fluctuation and deflection distribution data and the steering angle force fulcrum data, the multi-wheel drive output torque is controlled based on the pothole friction characteristic data and the depth, width, and pothole inclination of the ground pothole geometry data to obtain the multi-wheel drive output torque; and the breakaway acceleration kinetic energy of the stuck wheel is extracted from the multi-wheel drive output torque. Step S246: performing sideslip-related kinetic energy differential control on the pothole friction characteristic data and the depth, width, and pothole inclination in the ground pothole geometry data based on the breakaway acceleration kinetic energy, thereby obtaining kinetic energy differential control data; Step S247: performing wheel jamming and breaking-free behavior simulation learning based on the multi-wheel drive output torque and kinetic energy differential control data to obtain wheel jamming and breaking-free behavior learning data.

2. The intelligent wheelchair control method according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: extracting historical environmental monitoring data on the wheelchair's running path from the memory of the electronic monitoring device carried on the wheelchair; Step S12: Analyzing the path ground characteristics of the historical environmental monitoring data to obtain the historical path ground characteristics; Step S13: Marking the spatial coordinates of the historical path ground features to obtain the historical path coordinate ground features; Step S14: constructing a three-dimensional ground feature model of the path based on the ground features of the historical path coordinates.

3. The intelligent wheelchair control method according to claim 1, characterized in that: Step S244 includes the following steps: Perform spatial vector decomposition on the load force deflection data of the stuck wheel to obtain the three-axis load offset vector of the stuck wheel; According to the three-axis load offset vector of the stuck wheel, the transient reaction torque between the stuck wheel and the support contact surface is deduced from the steering angle force support data to obtain the transient reaction torque between the contact surfaces; The instantaneous reaction torque between the contact surfaces is solved by calculating the angular acceleration and the mass center line acceleration between the stuck wheel steering angles to obtain the instantaneous center of gravity offset vector; Performing center of gravity offset trajectory iterative processing on the instantaneous center of gravity offset vector to obtain center of gravity offset trajectory iterative data; The wheelchair center of gravity fluctuation deviation distribution is analyzed based on the center of gravity offset trajectory iterative data to obtain the wheelchair center of gravity fluctuation deviation distribution data.

4. The intelligent wheelchair control method according to claim 1, characterized in that: Step S246 includes the following steps: The non-constant acceleration variance is calculated for the breakaway acceleration kinetic energy to obtain the non-constant acceleration variance; The depth, width and inclination of the potholes in the ground are analyzed based on the non-constant acceleration variance to identify the lateral component of the steering angle, and obtain the lateral component of the steering angle increment data. Based on the non-constant acceleration variance and the lateral component force increment data of the steering angle, the slip lateral adhesion loss data is quantified for the friction characteristic data of the pothole morphology to obtain the slip lateral adhesion loss data; Perform slip distance change rate analysis based on slip lateral adhesion loss data to generate slip distance change rate; Performing torque vector distribution on both wheels according to the slip lateral adhesion loss data and the slip distance change rate to obtain torque vector distribution data on both wheels; The sideslip-related kinetic energy differential control is performed based on the torque vector distribution data of the wheels on both sides, thereby obtaining the kinetic energy differential control data.

5. The intelligent wheelchair control method according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: performing feature selection processing on the wheel jamming and breaking free behavior learning data to obtain jamming and breaking free behavior feature selection data; Step S32: constructing a model based on the selected data of the stuck-and-release behavior characteristics based on the deep Q network algorithm to obtain a stuck-and-release control model; Step S33: Optimizing the training feature importance parameters based on the stuck-and-release control model to obtain a stuck-and-release control optimization model; Step S34: Send the stuck and escape control optimization model to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

6. The intelligent wheelchair control method according to claim 5, characterized in that: Step S32 includes the following steps: Step S321: performing state feature encoding processing on the selected data of the stuck and escape behavior features, constructing a continuous state space representation, and obtaining an escape behavior state feature set; Step S322: performing deep Q network value parameter training on the discrete feature set of the breakaway behavior state based on the deep Q network algorithm to obtain the breakaway behavior state value parameter; Step S323: constructing a model based on the breaking free behavior state value parameter to obtain a stuck breaking free control model.

7. The intelligent wheelchair control method according to claim 5, characterized in that: Step S33 includes the following steps: Step S331: collecting operation response parameters based on the sticking and breaking free control model, thereby obtaining the sticking and breaking free operation response parameters; Step S332: performing parameter sensitivity disturbance analysis on the stuck-and-break-free operation response parameters to obtain parameter sensitivity disturbance data; Step S332: assigning feature importance weights to the stuck-and-break-free operation response parameters according to the parameter sensitivity disturbance data to obtain the response parameter feature importance weights; Step S333: Optimizing the feature importance parameters of the stuck-and-release control model according to the feature importance weights of the response parameters to obtain an optimized stuck-and-release control model.

8. An intelligent wheelchair control system, characterized in that: For executing the intelligent wheelchair control method according to claim 1, the intelligent wheelchair control system comprises: A historical monitoring model extraction module is used to extract historical environmental monitoring data on the wheelchair's path from the memory of the electronic monitoring device carried by the wheelchair; based on the historical environmental monitoring data, a three-dimensional ground feature model of the path is constructed; The breakaway behavior learning module is configured to analyze the geometric morphology of potholes on a three-dimensional ground feature model to obtain geometric morphology data of the potholes; simulate wheel jam scenarios based on the geometric morphology data of the potholes to obtain wheel jam scenario simulation data; and simulate learning of wheel jam breakaway behavior based on the wheel jam scenario simulation data using the geometric morphology data of the potholes to obtain wheel jam breakaway behavior learning data. The escape control model construction module is used to construct a model for the wheel jamming and escape behavior learning data, and then optimize the training feature importance parameters to obtain the jamming and escape control optimization model; the jamming and escape control optimization model is sent to the intelligent wheelchair control terminal to execute the intelligent wheelchair control method.

Images