Intelligent wheelchair scheduling method and system

Abstract

The invention relates to the technical field of intelligent wheelchair scheduling, in particular to an intelligent wheelchair scheduling method and system. The method comprises the following steps: extracting an obstacle avoidance failure decision in a historical scheduling task, and generating a clustering logic of decision failure through clustering analysis; analyzing the motion trail deduction deviation of the sudden obstacle based on clustering logic, and further performing sudden risk pre-judgment lag analysis to obtain lag regression data; and lagging regression data is utilized to carry out front deviation correction memory learning of wheelchair scheduling braking, decision is optimized through reinforcement learning iteration, and finally optimized decision data is output. According to the method, the intelligent wheelchair scheduling technology is optimized, so that the intelligent wheelchair scheduling technology is more perfect.

Description

Technical Field

[0001] This invention relates to the field of intelligent wheelchair scheduling technology, and in particular to an intelligent wheelchair scheduling method and system. Background Technology

[0002] As an assistive mobile device, intelligent wheelchairs still have many shortcomings in their autonomous scheduling and obstacle avoidance capabilities in complex environments. On the one hand, wheelchairs encounter dynamic obstacles in actual operation, such as pedestrians, pets, or suddenly moving objects. Traditional obstacle avoidance strategies based on static environments or rules are unable to cope with these changes in real time, leading to obstacle avoidance failures. Furthermore, traditional intelligent wheelchair scheduling methods often lack in-depth analysis and learning of historical task data, failing to identify error-prone decision-making patterns. This results in weak autonomous decision-making capabilities for intelligent wheelchairs, leading to large obstacle avoidance errors. Summary of the Invention

[0003] Therefore, it is necessary to provide an intelligent wheelchair scheduling method and system to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, an intelligent wheelchair scheduling method is provided, the method comprising the following steps: Step S1: Extract historical scheduling tasks from the database of smart wheelchairs, and then extract obstacle avoidance failure decisions of smart wheelchairs from the historical scheduling tasks; perform cluster analysis on the obstacle avoidance failure decisions to generate decision failure clustering logic; Step S2: Based on the clustering logic of decision failure, perform deviation analysis of the wheelchair avoidance trajectory in the event of a sudden obstacle to obtain the decision inference deviation; based on the decision inference deviation, perform lag analysis of the sudden risk prediction to obtain the lag regression data of the sudden risk prediction. Step S3: Based on the delayed normalized data of the sudden risk prediction, perform pre-deviation correction memory learning for wheelchair dispatching and braking to obtain deviation correction memory learning data; perform reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

[0005] The present invention also provides an intelligent wheelchair scheduling system for executing the intelligent wheelchair scheduling method described above, the intelligent wheelchair scheduling system comprising: The obstacle avoidance failure decision extraction module is used to extract historical scheduling tasks from the database of intelligent wheelchairs, and then extract the obstacle avoidance failure decisions of intelligent wheelchairs from the historical scheduling tasks; perform cluster analysis on the obstacle avoidance failure decisions to generate decision failure clustering logic; The obstacle avoidance failure decision analysis module is used to analyze the deviation of the wheelchair avoidance trajectory in sudden obstacles based on the clustering logic of decision failure, and to obtain the decision inference deviation; based on the decision inference deviation, the module analyzes the lag of sudden risk prediction to obtain the lag regression data of sudden risk prediction. The memory learning module is used to perform pre-deviation correction memory learning based on the delayed normalized data of sudden risk prediction for wheelchair dispatching and braking, and obtain deviation correction memory learning data; it then performs reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

[0006] The beneficial effects of this invention lie in its ability to provide a profound analytical foundation for subsequent scheduling optimization by extracting historical scheduling tasks from the database of intelligent wheelchairs, particularly focusing on historical obstacle avoidance failure decisions. Cluster analysis of obstacle avoidance failure decisions helps reveal the commonalities and differences in wheelchair obstacle avoidance failures under different circumstances, thereby generating a clustering logic for decision failures. This process allows the system to identify which decision-making patterns frequently fail in practical applications and whether these failures are related to specific environmental factors or task types. Through cluster analysis, the system can provide strong guidance for future scheduling decisions and avoid repeating past decision failures, improving the decision-making accuracy and adaptability of the intelligent wheelchair. Based on the decision failure clustering logic, deviation analysis of the motion trajectory projection of sudden obstacles helps predict and analyze the dynamic changes of sudden obstacles and their impact on intelligent wheelchair decision-making. Through detailed analysis of the deviation in motion trajectory projection, the system can identify how the current decision-making pattern is interfered with by sudden obstacles in specific situations and further understand the specific sources of projection deviations. This provides crucial data support for risk prediction, especially the lag regression data obtained through lag analysis for sudden risk prediction. This data helps the system identify safety hazards in advance, reduce the wheelchair's reaction delay to sudden obstacles, and enhance its environmental adaptability. The combination of pre-bias correction memory learning and reinforcement learning significantly improves the intelligent wheelchair's decision-making optimization capabilities. First, pre-bias correction memory learning allows the system to correct lag regression data for sudden risk prediction and convert this corrected data into memory for future decision-making reference. In this way, the wheelchair can gradually accumulate past experience, thus avoiding the recurrence of the same biases in similar situations. Second, reinforcement learning iteration further drives the system to optimize decision paths in multiple scheduling tasks, enabling the intelligent wheelchair to continuously adjust its obstacle avoidance and scheduling strategies through feedback mechanisms. After multiple training and adjustments, the final output decision optimization data significantly improves the wheelchair's autonomous scheduling ability, reaction speed, and safety in complex environments. Overall, through intelligent learning and optimization, the intelligent wheelchair's ability to cope with dynamic environments is greatly enhanced, improving user experience and safety. Therefore, this invention is an optimization of a traditional intelligent wheelchair scheduling method. It solves the problem that traditional intelligent wheelchair scheduling methods often lack in-depth analysis and learning of historical task data and cannot summarize error-prone decision-making patterns. This improves the autonomous decision-making ability of intelligent wheelchairs and reduces obstacle avoidance errors. Attached Figure Description

[0007] Figure 1 A flowchart illustrating the steps of an intelligent wheelchair scheduling method; Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2. Figure 3 for Figure 1 A detailed flowchart illustrating the implementation steps of step S3. Detailed Implementation

[0008] Please see Figures 1 to 3 A method for intelligent wheelchair scheduling, the method comprising the following steps: Step S1: Extract historical scheduling tasks from the database of smart wheelchairs, and then extract obstacle avoidance failure decisions of smart wheelchairs from the historical scheduling tasks; perform cluster analysis on the obstacle avoidance failure decisions to generate decision failure clustering logic; Step S2: Based on the clustering logic of decision failure, perform deviation analysis of the wheelchair avoidance trajectory in the event of a sudden obstacle to obtain the decision inference deviation; based on the decision inference deviation, perform lag analysis of the sudden risk prediction to obtain the lag regression data of the sudden risk prediction. Step S3: Based on the delayed normalized data of the sudden risk prediction, perform pre-deviation correction memory learning for wheelchair dispatching and braking to obtain deviation correction memory learning data; perform reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

[0009] In this embodiment of the invention, reference Figure 1 The above is a flowchart illustrating the steps of an intelligent wheelchair scheduling method according to the present invention. In this example, the intelligent wheelchair scheduling method includes the following steps: Step S1: Extract historical scheduling tasks from the database of smart wheelchairs, and then extract obstacle avoidance failure decisions of smart wheelchairs from the historical scheduling tasks; perform cluster analysis on the obstacle avoidance failure decisions to generate decision failure clustering logic; In this embodiment of the invention, historical scheduling task data is extracted from the database of the intelligent wheelchair. This database records the timestamp of task execution, wheelchair position coordinates, obstacle detection information, obstacle avoidance status markers, and task completion status in the form of a relational data table. Obstacle avoidance tasks with failure markers are filtered out to obtain a historical obstacle avoidance failure data set. Subsequently, the historical obstacle avoidance failure data set is structured, and each failure record is decomposed into basic elements such as the wheelchair's real-time position information, wheelchair speed, wheelchair turning angle, relative position of the obstacle, obstacle speed, and acceleration. After processing, an obstacle avoidance failure decision sequence is formed. The obstacle avoidance failure decision sequence is classified by a clustering method. The clustering method used is hierarchical clustering analysis based on Euclidean distance. In the clustering process, the similarity between each obstacle avoidance failure sequence is first calculated. Based on the similarity, the failure decisions are divided into several categories, each representing a typical failure mode. After the clustering is completed, a decision failure clustering logic is formed. This logic records the hierarchical relationship between failure types in a tree structure and stores it in a logical index table for subsequent steps.

[0010] In another embodiment, historical scheduling task data stored in the smart wheelchair's database is retrieved. This database contains task records executed by the wheelchair in different scenarios, including complete path trajectories and control command data for successful and failed obstacle avoidance. The scheduling tasks for failed obstacle avoidance are extracted according to the time sequence using a task index. Then, the decision commands and environmental sensor parameter data corresponding to the obstacle avoidance failures are separated from these task records. The environmental parameters include lidar scanning point cloud data, ultrasonic distance measurement data, and infrared detection results. These obstacle avoidance failure decision data are vectorized according to the type of obstacle, its location, relative speed, and the wheelchair's turning angle. Each failure decision is transformed into a multi-dimensional feature vector composed of obstacle features and decision parameters. A clustering method combining Euclidean distance and cosine similarity is used to analyze the similarity of these vectors. Clusters are formed by setting a cluster radius threshold (e.g., 0.15) and a minimum sample size threshold (e.g., 10), thus forming a decision failure clustering logic. This logic represents the repeated obstacle avoidance failure patterns of the wheelchair under similar obstacle dynamic conditions.

[0011] Step S2: Based on the clustering logic of decision failure, perform deviation analysis of the wheelchair avoidance trajectory in the event of a sudden obstacle to obtain the decision inference deviation; based on the decision inference deviation, perform lag analysis of the sudden risk prediction to obtain the lag regression data of the sudden risk prediction. In this embodiment of the invention, after obtaining the decision failure clustering logic, the system first retrieves obstacle dynamic parameters matching the clustering results from the database. These parameters include obstacle position coordinates, velocity vectors, acceleration vectors, and motion direction angles. The historical scene is then reconstructed using these parameters to form an obstacle dynamic scene. The scene records the obstacle trajectory points and the wheelchair's relative position in a continuous time series. Subsequently, the decision failure clustering logic is invoked within the obstacle dynamic scene to compare and analyze the wheelchair's obstacle avoidance behavior. The system extracts the temporal deviation data between the wheelchair and obstacle trajectories during obstacle avoidance failures. This deviation data includes the lateral position of the wheelchair and obstacle at each time point. The deviations, longitudinal displacement difference, and angular deviation are weighted and accumulated to form the trajectory extrapolation deviation. This trajectory extrapolation deviation is used for the lag analysis of sudden risk prediction in decision-making. In the analysis process, the time delay between the decision trigger moment when the wheelchair performs obstacle avoidance actions and the moment when the obstacle's motion state changes is first calculated. This time delay is compared with the failure types in the clustering logic to obtain the typical lag interval for each failure mode. Then, regression calculation is performed on the lag intervals extracted from multiple scenarios to generate sudden risk prediction lag regression data. This data is stored in the form of a correspondence table between time delay and obstacle motion intensity for use in the next step.

[0012] In another embodiment, based on the obtained decision failure clustering logic, a dynamic obstacle scenario is reconstructed in an experimental environment. This scenario consists of two types of obstacles: one is a group of people with speeds between 0.8 m / s and 1.2 m / s, and the other is small moving objects with speeds between 0.5 m / s and 1.5 m / s. The speed, direction, and acceleration of the obstacles are used to construct a sequence of obstacle trajectories. Then, combined with sensor data from the wheelchair, an obstacle avoidance path is simulated. For each dynamic scenario, the obstacle avoidance failure state is extracted based on the decision failure clustering logic, and the temporal trajectory deviation is utilized. The comparison method differs the actual obstacle avoidance trajectory generated by the wheelchair with the ideal obstacle avoidance trajectory, recording the cumulative deviations in position, velocity, and steering angle to obtain decision-making deviation data. Based on this, the prediction delay caused by the sudden movement of the obstacle is calculated. The lag within the prediction period of 100ms to 200ms is extracted through time window analysis, and the difference between the actual obstacle avoidance action and the time point in the predicted trajectory is compared to form sudden risk prediction lag data. Linear regression is then used to fit the lag data to obtain sudden risk prediction lag regression data. Step S3: Based on the delayed normalized data of the sudden risk prediction, perform pre-deviation correction memory learning for wheelchair dispatching and braking to obtain deviation correction memory learning data; perform reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

[0013] In this embodiment of the invention, after obtaining the lag regression data for sudden risk prediction, the data is first normalized to scale the time delay and obstacle motion intensity data to the interval [0,1], forming lag normalized data for sudden risk prediction. Then, the decision inference bias is convolved with a kernel length of 5 and a one-dimensional sliding window operation to extract the continuity features of trajectory deviation over time, thus obtaining the decision convolution bias. Based on this, the lag normalized data for sudden risk prediction and the decision convolution bias are input together into the bias correction memory learning module. This module obtains decision bias parameters such as the wheelchair's braking response delay, maximum braking acceleration, steering angle change rate, and avoidance trajectory offset through bias parameter extraction. Next, the evolution of the deviation parameters is analyzed using time series regression to form a sequence of deviation parameter evolution. This sequence is then aggregated with lag normalized data to obtain correction amount clustering memory data. Further, obstacle sudden trajectory learning data is used to supplement the correction amount clustering memory data for training, forming sudden behavior correction training data. After completing the correction training, deviation correction memory learning data is output. Finally, reinforcement learning iteration is performed on the deviation correction memory learning data. During the iteration process, the correction data is used as a reward signal to optimize the action selection process. The results after multiple iterations are the decision optimization data, which records the optimal obstacle avoidance trajectory correction scheme and braking trigger advance, serving as input for subsequent scheduling and control of the intelligent wheelchair.

[0014] Preferably, step S1 includes the following steps: Extract historical scheduling tasks from the database of smart wheelchairs, and extract obstacle avoidance failure datasets from the historical scheduling tasks; Random samples are selected from the obstacle avoidance failure dataset, and the obstacle avoidance failure samples are output. Then, the obstacle avoidance failure decisions of the smart wheelchair in the samples are extracted. Logical analysis is performed on the obstacle avoidance failure decision to obtain the decision failure logic; Cluster analysis is performed on the decision failure logic to generate decision failure cluster logic.

[0015] In this embodiment of the invention, historical scheduling task data is extracted from the database of the intelligent wheelchair. The database adopts a relational data table structure. Each task record includes task number, execution time, execution location, wheelchair operating status, obstacle information collected by sensors, wheelchair speed, wheelchair steering angle, braking status, and task completion status. The database capacity is 20,000 historical records. Each record is sampled and stored at 100 ms intervals. First, the historical scheduling task data is filtered, retaining only records that contain obstacle detection information and are marked as obstacle avoidance failures, thus obtaining an obstacle avoidance failure dataset. This dataset contains multi-dimensional parameters such as the wheelchair's real-time position (in meters), speed (in meters / s), acceleration (in meters / s²), steering angle (in degrees), obstacle position (in meters), obstacle speed (in meters / s), and obstacle acceleration (in meters / s²) under different environments. Then, random samples are selected from the obstacle avoidance failure dataset using a fixed-length random sampling method. Samples are drawn from the obstacle avoidance failure dataset at a rate of 5%, with a sample size of no less than 1,000 records. Each sample retains 10 consecutive records. The system uses time-series data to ensure data integrity, outputs obstacle avoidance failure samples and stores them in a structured manner, and then extracts obstacle avoidance failure decision information of the smart wheelchair from the obstacle avoidance failure samples. This decision information is obtained by reading the wheelchair control command sequence, which includes braking trigger time, braking intensity (unit m / s²), steering angle change rate (unit ° / s), and acceleration / deceleration threshold (unit m / s²). These decision information are arranged in chronological order to form the obstacle avoidance failure decision.

[0016] After obtaining the obstacle avoidance failure decision, a logical analysis is performed on the decision. The logical analysis adopts a time-series analysis method based on state transition. First, a state transition matrix is ​​constructed, with the wheelchair's state (velocity, steering angle, acceleration, and relative position to the obstacle) at each sampling moment as the rows of the matrix and the state at the next moment as the columns. The state transition rules are obtained by calculating the transition probabilities of the matrix. Then, the causal relationship between each decision command and the obstacle's dynamic parameters is analyzed. The correlation coefficients between braking trigger time and relative distance to the obstacle, and between steering angle change rate and obstacle speed direction change are calculated. The Pearson correlation coefficient method is used to obtain the correlation values ​​between each variable, and the variable pairs with correlation values ​​greater than 0.7 are regarded as high-correlation decision features, which constitute the decision failure logic. The decision failure logic is stored in the form of a multi-dimensional array, recording the state transition rules and high-correlation feature parameters of each obstacle avoidance failure decision.

[0017] After extracting the decision failure logic, cluster analysis is performed on the logic using a hierarchical clustering algorithm based on Euclidean distance. First, a feature vector is calculated for each decision failure logic. This feature vector includes braking trigger time delay (in seconds), steering angle change rate (in degrees / s), braking intensity (in m / s²), and obstacle speed direction deviation (in degrees). Then, the Euclidean distance between all feature vectors is calculated to construct a distance matrix. Based on the distance matrix, the nodes are gradually merged starting from the minimum distance to form a clustering dendrogram. A distance threshold of 0.5 is set; nodes with a Euclidean distance less than 0.5 are merged into the same class, resulting in multiple decision failure modes, each corresponding to a typical failure scenario. The clustering results are stored in a logical index table, which records the feature parameter range and decision behavior sequence for each failure mode. This index table serves as the decision failure clustering logic for subsequent steps.

[0018] Preferably, step S2 includes the following steps: The obstacle dynamic parameters are extracted from the obstacle avoidance failure samples and the scene is reconstructed to obtain the obstacle dynamic scene, wherein the obstacles include crowds and moving objects; Based on the clustering logic of decision failure, the deviation analysis of the motion trajectory extrapolation of wheelchair avoidance of sudden obstacles in dynamic obstacle scenarios is conducted to obtain the decision extrapolation deviation. Lag analysis of sudden risk prediction based on decision-making inference bias and dynamic obstacle scenarios, in order to obtain sudden risk prediction lag data; The regression analysis is used to analyze the lagged data of the sudden risk prediction and output the lagged regression data of the sudden risk prediction.

[0019] As an example of the present invention, reference is made to... Figure 2 As shown, step S2 in this example includes: Step S201: Extract obstacle dynamic parameters from obstacle avoidance failure samples and reconstruct the scene to obtain an obstacle dynamic scene, wherein the obstacles include crowds and moving objects; In this embodiment of the invention, obstacle dynamic parameters are extracted from the obtained obstacle avoidance failure samples. These parameters include the obstacle's three-dimensional coordinates (in meters), velocity vector (in meters per second), acceleration vector (in meters per second²), motion direction angle (in degrees), obstacle type identifier (crowd or moving object), and relative distance (in meters) between the obstacle and the intelligent wheelchair at each time sampling point. The data sampling frequency is 100 Hz, and each sample contains 10 consecutive seconds of obstacle motion data. These obstacle dynamic parameters are used to reconstruct the obstacle trajectory in chronological order. An interpolation reconstruction algorithm is used to linearly interpolate and fill in the missing time points, forming an obstacle dynamic scene. The obstacle dynamic scene is stored in matrix form, with rows representing time points and columns representing parameter values. The number of crowd-type obstacles is greater than 20, and parameters are recorded independently for each individual. The velocity range for moving object obstacles is between 0.5 m / s and 3.0 m / s, and the acceleration range is between 0 m / s² and 1.5 m / s². Between m / s², the continuous trajectory of the obstacle in the planar coordinate system is generated by the reconstructed matrix data and synchronized with the historical trajectory of the wheelchair in time to obtain the dynamic scene of the obstacle.

[0020] Step S202: Based on the decision failure clustering logic, analyze the deviation of the wheelchair avoidance motion trajectory in the dynamic obstacle scene to make decisions, and obtain the decision inference deviation; In this embodiment of the invention, a deviation analysis of the trajectory deduction for wheelchair avoidance of sudden obstacles in a dynamic obstacle scene is performed based on decision failure clustering logic. The decision deduction deviation is obtained by first reading the cluster centers and feature parameters from the decision failure clustering logic. The cluster centers include key features such as braking trigger delay time, steering angle change rate, and obstacle speed direction deviation. There are five main clusters, each corresponding to different types of obstacle avoidance failure modes. The cluster features are matched with the dynamic obstacle scene, and a nearest neighbor search algorithm is used to calculate the feature distance between each spatiotemporal point in the scene and the cluster center. Points with a feature distance less than a threshold of 0.6 are marked as potential risk points. All potential risk points are extracted to form a risk distribution map, which is represented as a heatmap. The color from blue to red indicates risk from low to high. A time-series analysis is performed on the risk distribution map using a sliding window method with a window length of 2 seconds and a step size of 0. Within each 5-second window, the mean, standard deviation, and rate of change of the risk value are calculated. The mean risk ranges from 0.2 to 0.8, the standard deviation from 0.1 to 0.3, and the rate of change from -0.2 to 0.4 per second. Based on the risk distribution map, a wheelchair avoidance trajectory projection model is constructed using the Monte Carlo method, generating 1000 avoidance trajectories. Each trajectory includes the wheelchair's position, velocity, acceleration, and turning angle sequence within the next 5 seconds. Trajectory generation considers the wheelchair's kinematic constraints: the maximum speed is limited to 2.0 m / s, the maximum acceleration to 1.5 m / s², and the maximum turning angle rate to 30° / s. The generated avoidance trajectories are evaluated, and the safety margin, smoothness, and target achievement rate of each trajectory are calculated. The safety margin is defined as the minimum distance between the trajectory and the obstacle, with an average of 0.85 m. Smoothness is defined as the standard deviation of the trajectory curvature, with an average of 0.28. The target achievement rate is defined as the ratio of the distance between the trajectory endpoint and the target point, with an average value of 0.75. The top 10 trajectories with the highest comprehensive scores are selected as the optimal avoidance trajectories. The deviation between the optimal trajectory and the actual trajectory is calculated. The deviation includes position deviation, velocity deviation, acceleration deviation, and steering angle deviation. The average position deviation is 0.42m, the average velocity deviation is 0.35m / s, the average acceleration deviation is 0.48m / s², and the average steering angle deviation is 12°. The decision-making inference deviation is constructed to provide input data for subsequent prediction of sudden risks.

[0021] In another embodiment, a specific implementation of the deviation analysis of the sudden obstacle avoidance trajectory extrapolation for wheelchairs in decision-making dynamic scenarios based on decision failure clustering logic is as follows: First, in the reconstructed obstacle dynamic scenario, clustering logic is invoked, and the dynamic state of each obstacle is input into the clustering logic for matching. Trajectory patterns highly similar to historical obstacle avoidance failures are extracted. Then, the avoidance path is extrapolated for the intelligent wheelchair under this trajectory pattern. The extrapolation process is based on the actual control parameters of the wheelchair, where the maximum steering angle of the wheelchair is ±30°, the maximum acceleration is 2.8m / s², the maximum deceleration is 3.2m / s², and the sampling period is 50ms. The extrapolated avoidance trajectory is compared with the ideal safe avoidance trajectory using time-series difference, and the position deviation, velocity deviation, and steering angle deviation are calculated. The deviation values ​​are then integrated and accumulated over the entire time-series window, and finally, the decision extrapolation deviation is output. This deviation reflects the degree of inadequacy of the wheelchair's avoidance under sudden obstacle conditions. The acquisition of the "ideal safe avoidance trajectory" is achieved by first strictly constraining the feasible path space according to the structural limit parameters of the wheelchair (maximum steering angle ±30°, maximum acceleration 2.8m / s², maximum deceleration 3.2m / s², minimum safe distance 0.8m) after the obstacle dynamic scene has been fully reconstructed. Then, within this path space, the current and predicted motion states of the obstacle are used as inputs, and a trajectory generation algorithm based on the combination of geometric path search and time-optimal control (e.g., performing A* search in the gridded space and minimum acceleration / deceleration scheduling on the time axis) is used to calculate an obstacle avoidance path that meets the requirement of not being less than the minimum safe distance throughout the entire process and can be passed within a given response time. Then, the path is sampled at 50ms intervals and smoothed by interpolation to obtain a continuous time series that simultaneously meets the spatial avoidance constraints and wheelchair dynamic constraints. This time series is the ideal safe avoidance trajectory.

[0022] Step S203: Analyze the lag in sudden risk prediction based on decision-making inference bias and obstacle dynamic scenarios to obtain lag data for sudden risk prediction; In this embodiment of the invention, a lag analysis of sudden risk prediction is performed based on the obtained decision-making inference deviation and the dynamic obstacle scene. First, the time delay (in seconds) between the trigger time of each decision command of the wheelchair during obstacle avoidance and the time point of change of the obstacle's velocity and acceleration is calculated. A correspondence table is established between this time delay and the wheelchair's current braking intensity (in m / s²) and steering angle change rate (in ° / s). Then, the distribution density of obstacles in each time slice (in units / m²) is calculated. For obstacles involving crowds, a two-dimensional grid density calculation method is used to divide the scene into 0.5m × 0.5m grids and count the number of people in each grid. The density distribution map is used to analyze the sudden motion states of obstacles, including sudden increases in velocity (unit: m / s), sudden changes in direction (unit: °), and sudden changes in acceleration (unit: m / s²). These sudden motion states are vectorized to form an obstacle motion vector sequence. The real-time predicted resource demand is calculated by aligning the wheelchair decision-making deviation with the obstacle motion vector sequence using timestamps in the real-time data stream. Based on this, the time series of the maximum decision delay during obstacle avoidance is extracted. This time series is mapped to the decision-making deviation matrix to obtain the lag data for sudden risk prediction. This data is stored in the form of a multidimensional array for regression analysis.

[0023] Step S204: Perform regression analysis on the lagged data of the sudden risk prediction and output the lagged regression data of the sudden risk prediction.

[0024] In this embodiment of the invention, regression analysis is performed on the delayed data of sudden risk prediction. The regression analysis uses a multiple linear regression method with time delay (in seconds) as the dependent variable and obstacle velocity abrupt change (in m / s), obstacle direction abrupt change (in degrees), obstacle acceleration abrupt change (in m / s²), wheelchair braking intensity (in m / s²), and steering angle change rate (in degrees / s) as independent variables to establish a regression equation and calculate the regression coefficients of each independent variable. The regression coefficients are solved by the least squares method, with the number of iterations set to 100 and the convergence threshold set to [missing value]. To ensure computational accuracy, regression training was performed on all sample data to obtain lagged regression data for predicting sudden risks. This data, in the form of a parameter weight matrix, records the time delay characteristics of each type of sudden scenario and serves as input for subsequent steps.

[0025] Preferably, the deviation analysis of the wheelchair avoidance trajectory prediction for sudden obstacles includes the following steps: Based on the decision failure clustering logic, extract the obstacle avoidance failure status of the intelligent wheelchair in the dynamic obstacle scene when there is a sudden obstacle. Extract the back-and-forth loop trajectory in the obstacle avoidance failure state, and analyze the nonlinear relationship of the trajectory range in the back-and-forth loop trajectory; wherein the back-and-forth loop trajectory refers to an invalid motion mode in which the intelligent wheelchair gets stuck in a local space, that is, the movement direction path swings back and forth in a short period of time; wherein the nonlinear relationship of the trajectory range refers to the degree and irregularity of the wheelchair's range of motion changing over time during the obstacle avoidance failure process. By analyzing the back-and-forth loop trajectory, the curve of the wheelchair swinging back and forth in the movement direction path can be constructed. The intermittent start-stop speed variation relationship of the intelligent wheelchair is analyzed from the nonlinear relationship of the trajectory range; the degree of disorder of the wheelchair steering angle is coupled in the nonlinear relationship of the trajectory range and the speed variation relationship. Based on the nonlinear relationship of the trajectory range, the relationship of speed change, and the disorder of the wheelchair turning angle, the timing deviation of the obstacle avoidance action trajectory of the intelligent wheelchair is cumulatively fitted in the obstacle dynamic scene to obtain the cumulative deviation data of the obstacle avoidance trajectory. Analysis of the deviation in wheelchair avoidance trajectory simulation for sudden obstacles based on accumulated data of avoidance trajectory deviation yields the decision simulation deviation.

[0026] In this embodiment of the invention, based on the decision failure clustering logic, the obstacle avoidance failure state of the intelligent wheelchair in the obstacle dynamic scene is extracted. First, the cluster center parameters are read from the decision failure clustering logic database, including key feature values ​​such as braking trigger delay time of 0.85 seconds, steering angle change rate of 15.3° / s, and obstacle velocity direction deviation of 28.7°. Then, the wheelchair position coordinate sequence and obstacle position coordinate sequence are extracted from the obstacle dynamic scene data at a sampling interval of 100ms. The Euclidean distance change curve between the two is calculated. When the distance is less than the safety threshold of 1.2m and the wheelchair does not perform an effective obstacle avoidance action, it is marked as an obstacle avoidance failure point. All obstacle avoidance failure points within a 10s time window are continuously extracted to form an obstacle avoidance failure state sequence. This sequence includes the position coordinates, velocity value, acceleration value, and steering angle value of the wheelchair in the failure decision state, as well as the position coordinates, velocity value, and acceleration value of the obstacle at the corresponding time. These parameters are organized into a structured data table, with rows representing time points and columns representing the parameter values, thereby obtaining the obstacle avoidance failure state of the intelligent wheelchair.

[0027] The process involves extracting back-and-forth loop trajectories from obstacle avoidance failure states and analyzing the nonlinear relationships within these trajectories. This is achieved through time-series analysis of the wheelchair position coordinates in the obstacle avoidance failure state sequence. A sliding window method is used, with a window length of 2 seconds and a step size of 0.5 seconds. Within each window, the number of changes in the wheelchair's direction of motion is calculated. When the number of direction changes exceeds four, it is considered a back-and-forth loop trajectory. All back-and-forth loop trajectory segments are extracted and merged into a continuous sequence. Fourier transform is performed on the merged trajectory sequence to extract spectral features, calculating the amplitude and phase of the dominant frequency component. The amplitude of the dominant frequency component represents the loop intensity, and the phase represents the loop start point. The trajectory points are projected onto a two-dimensional plane to calculate the trajectory envelope curve. The least squares method is used to fit the envelope curve to obtain the trajectory range function, which represents the relationship between the wheelchair's activity range during obstacle avoidance and time. The nonlinear rate of change of the trajectory range is obtained by calculating the second derivative of the trajectory range function. A nonlinear rate of change greater than 2.5 indicates a rapid expansion of the trajectory range, while a rate less than 2.5 indicates a rapid contraction of the trajectory range. The nonlinear rate of change sequence is correlated with timestamps to form nonlinear relationship data for the trajectory range.

[0028] This study analyzes the intermittent start-stop speed variation relationship of an intelligent wheelchair from the perspective of trajectory range nonlinearity. First, the wheelchair speed time series is differentiated to obtain the acceleration sequence. An acceleration threshold of ±0.8 m / s² is set. When the absolute value of acceleration exceeds the threshold, it is marked as a start or stop point. All start and stop points are continuously extracted to form a start-stop sequence. The time interval between adjacent start and stop points is calculated to obtain the start-stop cycle. The distribution characteristics of the start-stop cycle are analyzed, including a mean of 1.75 s, a standard deviation of 0.43 s, a maximum value of 2.8 s, and a minimum value of 0.9 s. The start-stop point timestamps are aligned with the trajectory range nonlinearity relationship data table. The trajectory range nonlinearity change rate at the time of start and stop point occurrence is calculated. The correlation between the start-stop frequency and the trajectory range change rate is calculated using the Pearson correlation coefficient method. A correlation coefficient of 0.78 indicates a strong correlation. A speed variation relationship matrix is ​​constructed, where rows represent time points and columns represent speed values, acceleration values, start / stop markers, and the corresponding trajectory range nonlinearity change rate.

[0029] The wheelchair steering angle time series was extracted from the loop trajectory of the obstacle avoidance failure state. The steering angle series was segmented according to the sampling point interval of 0.01s, with each segment containing 50 consecutive sampling points. The local autocorrelation coefficient was calculated for each segment of steering angle data. The autocorrelation coefficient reflects the continuity and regularity of the steering angle change over time. Segments with an autocorrelation coefficient lower than 0.6 were marked as disordered fluctuation segments. At the same time, the instantaneous acceleration was obtained by first-order difference of the wheelchair velocity sequence in the same trajectory segment. The standard deviation of the acceleration change sequence was calculated, and segments with a standard deviation higher than 0.15m / s² were associated with disordered fluctuation segments. The steering angle was then aligned using a time alignment method. The nonlinear fitting coefficients of the sequence segment and trajectory range, as well as the velocity change amplitude, are coupled and synthesized into a three-dimensional feature vector. The first dimension of the feature vector is the trajectory range nonlinear coefficient, the second dimension is the velocity change amplitude, and the third dimension is the steering angle disorder index, where the steering angle disorder index is the proportion of the marked segment to the entire trajectory segment. The three-dimensional feature vector is linearly combined through principal component analysis to obtain the coupled steering angle disorder coefficient. This coefficient is used to characterize the degree of irregular fluctuation of the wheelchair's steering angle in the current trajectory segment and corresponds to the trajectory range and velocity change relationship matrix. It is used as the input for the subsequent avoidance action trajectory time sequence deviation accumulation fitting.

[0030] Based on the nonlinear relationship of trajectory range, the relationship of velocity change, and the disorder of wheelchair turning angle, the time-series deviation of the obstacle avoidance action trajectory of the intelligent wheelchair in a dynamic obstacle scenario is cumulatively fitted to obtain the cumulative deviation data of the avoidance trajectory. First, a multivariate regression model is constructed, with the nonlinear change rate of trajectory range, the change rate of velocity, and the disorder of turning angle as independent variables, and the relative position deviation between the wheelchair and the obstacle as the dependent variable. The least squares method is used to solve the regression coefficients, and the calculation is iterated 100 times until convergence. The regression coefficients are 0.42, 0.35, and 0.23, respectively, representing the contribution weights of the three factors to the avoidance trajectory deviation. Then, the residual between the predicted deviation value and the actual deviation value is calculated at each time point. The sum of squared residuals of 0.087 indicates high fitting accuracy. The residual sequence is accumulated in time order to obtain the cumulative residual curve, which represents the cumulative effect of the avoidance trajectory deviation over time. The cumulative residual curve is smoothed using the exponential smoothing method with a smoothing coefficient of 0.15 to obtain the cumulative deviation data of the avoidance trajectory. This data is stored in time series form and includes three components: lateral cumulative deviation, longitudinal cumulative deviation, and angular cumulative deviation.

[0031] This paper analyzes the deviation of wheelchair avoidance trajectory projection for sudden obstacles based on accumulated avoidance trajectory deviation data. The decision projection deviation is obtained by first extracting sudden motion features of the obstacle from the dynamic obstacle scene, including velocity, direction, and acceleration abrupt changes. The sliding difference method is used to calculate the rate of change of parameters at adjacent time points. A sudden point is marked when the rate of change exceeds a threshold: 0.6 m / s for velocity abrupt change, 25° for direction abrupt change, and 1.2 m / s² for acceleration abrupt change. The sudden motion features of the obstacle are time-aligned with the accumulated avoidance trajectory deviation data, and the time delay between them is calculated. A cross-sectional analysis is then performed. The correlation function method is used to determine the time offset corresponding to the maximum correlation. The time offset ranges from 0.3s to 1.2s, with an average of 0.75s. The time offset is combined with the cumulative data of avoidance trajectory deviation to construct a decision inference deviation matrix. The rows of the matrix represent time points, and the columns represent the lateral deviation value, longitudinal deviation value, angular deviation value, and corresponding time delay value. The matrix is ​​dimensionality reduced by the singular value decomposition method to extract the main deviation patterns. Singular values ​​with a contribution rate greater than 95% and their corresponding eigenvectors are retained to obtain the decision inference deviation feature space. This feature space describes the distribution law of decision deviation when wheelchairs face sudden obstacles.

[0032] Preferably, the cumulative fitting of the timing deviation of the obstacle avoidance trajectory of the intelligent wheelchair includes the following steps: The effective obstacle avoidance actions of intelligent wheelchairs are analyzed through dynamic obstacle scene analysis. A sawtooth trajectory analysis is performed on the nonlinear relationship of the trajectory range to obtain sawtooth trajectory data; a curvature abrupt change trend offset analysis is performed on the sawtooth trajectory data to obtain the trajectory curvature offset trend; Analyze the slope of acceleration / deceleration in the velocity change relationship; Based on the trajectory curvature deviation trend and the slope of acceleration / deceleration, the degree of disorder of the wheelchair steering angle is analyzed to obtain the steering angle mismatch coupling data. Based on the trajectory curvature offset trend, acceleration / deceleration slope change, and steering angle misalignment coupling data, the effective avoidance action is decomposed into a multi-dimensional avoidance deviation stochastic process to obtain the multi-dimensional avoidance deviation; the multi-dimensional deviation includes the time dimension and the spatial dimension. Based on multi-dimensional avoidance deviation, the temporal deviation of the avoidance action trajectory of the intelligent wheelchair is cumulatively fitted to obtain the cumulative data of the avoidance trajectory deviation.

[0033] In this embodiment of the invention, the effective avoidance actions of an intelligent wheelchair are analyzed through obstacle dynamic scene analysis. First, the wheelchair position coordinate sequence and obstacle position coordinate sequence are extracted from the obstacle dynamic scene data. The movement trajectory is recorded for 10 consecutive seconds. The relative distance change curve between the wheelchair and the obstacle is calculated. A safe distance threshold of 1.5m is set. When the relative distance is less than the safe threshold, an avoidance requirement is triggered. The control command sequence of the wheelchair during the avoidance process is extracted, including speed adjustment command, steering angle adjustment command, and braking command. The command sampling interval is 50ms. Time window analysis is performed on the control command sequence. The window length is set to 1s and the step size is 0.2s. The angle between the wheelchair position change vector and the obstacle position change vector is calculated in each window. When the angle is greater than 45° and the relative distance increases, it is determined to be an effective avoidance action. All effective avoidance action segments are extracted and merged into a continuous sequence. Feature extraction is performed on the avoidance action sequence to calculate the avoidance amplitude, avoidance duration, and safety margin after avoidance. The avoidance amplitude is defined as the maximum lateral displacement of the wheelchair from the original trajectory, with a value range of 0.8m to 2m. Within a 3m range, the avoidance duration is defined as the time interval from the start of avoidance to the resumption of normal driving, ranging from 1.2s to 3.5s. The safety margin is defined as the minimum distance between the wheelchair and the obstacle during the avoidance process, ranging from 0.6m to 1.8m. A feature vector of avoidance action is constructed with a dimension of 15, including parameters such as avoidance amplitude, avoidance duration, safety margin, wheelchair speed at the start of avoidance, wheelchair acceleration, wheelchair turning angle, relative position of the obstacle, relative speed of the obstacle, and relative acceleration of the obstacle. Principal component analysis is used to reduce the dimensionality of the feature vector, retaining principal components with a contribution rate greater than 95%, thus obtaining the main pattern of avoidance action. The main pattern of avoidance action is correlated with the obstacle motion characteristics, and the cross-correlation function between the two is calculated to determine the time delay corresponding to the maximum correlation. The time delay ranges from 0.2s to 0.9s, with an average of 0.55s. Effective avoidance action data is constructed, recording the feature parameters of the avoidance action, the corresponding obstacle motion characteristics, and the time delay value, providing basic data for subsequent trajectory deviation analysis.

[0034] A sawtooth trajectory analysis was performed on the nonlinear relationship of the trajectory range to obtain sawtooth trajectory data. First, the wheelchair position coordinate sequence was extracted from the nonlinear relationship data of the trajectory range, and the motion trajectory within 10 consecutive seconds was recorded. The trajectory was projected onto a two-dimensional plane, and the change in direction angle between adjacent trajectory points was calculated. The threshold for the change in direction angle was set to ±30°. When the absolute value of the change in direction angle is greater than the threshold and the sign changes alternately, it is marked as a sawtooth trajectory point. All sawtooth trajectory points were continuously extracted to form a sawtooth trajectory segment. The geometric characteristics of the sawtooth trajectory segment were calculated, including sawtooth amplitude, sawtooth period, and sawtooth symmetry. Sawtooth amplitude is defined as the lateral displacement between adjacent turning points, with a value range between 0.3m and 1.2m. Sawtooth period is defined as the time interval between adjacent turning points in the same direction, with a value range between 0.8s and 2.2s. Sawtooth symmetry is defined as the ratio of left and right turning amplitudes, with a value range between 0.65 and 1.35. Wavelet transform was used. A time-frequency analysis was performed on the zigzag trajectory using Morlet wavelets as the basis function, with the scale parameter ranging from 1 to 64. The wavelet coefficient matrix was calculated, and the peak positions and intensities of the wavelet energy spectrum were extracted. The peak position represents the main period of the zigzag, and the peak intensity represents the saliency of the zigzag. A feature vector of the zigzag trajectory was constructed with a dimension of 12, including parameters such as zigzag amplitude, zigzag period, zigzag symmetry, peak position of the wavelet energy spectrum, and peak intensity. Cluster analysis was used to classify the zigzag trajectory using the K-means algorithm, with 3 categories corresponding to weak, moderate, and strong zigzag patterns. The center vector and sample distribution of each category were calculated, with weak zigzag patterns accounting for 25%, moderate zigzag patterns for 45%, and strong zigzag patterns for 30%. A zigzag trajectory data table was constructed, recording the feature parameters, category labels, and corresponding timestamps of the zigzag trajectory, providing input data for subsequent curvature change trend analysis.

[0035] A curvature abrupt change trend shift analysis was performed on the sawtooth trajectory data to obtain the trajectory curvature shift trend. First, the wheelchair position coordinate sequence was extracted from the sawtooth trajectory data table, and the motion trajectory within 10 consecutive seconds was recorded. The local curvature of the trajectory was calculated using the following formula: ,in Let y', x'', and y'' represent the first and second derivatives of the x and y coordinates, respectively. The derivatives are calculated using the central difference method with a difference step size of 3 sampling points. A sliding window analysis is performed on the curvature sequence, with a window length of 0.5 s and a step size of 0.1 s. Within each window, the mean, standard deviation, and rate of change of curvature are calculated. The mean curvature ranges from -2.5 to 2.5, the standard deviation from 0.3 to 1.8, and the rate of change from -3.0 to 3.0. A curvature mutation threshold of ±1.5 is set; when the absolute value of the rate of change of curvature exceeds the threshold, it is marked as a curvature mutation point. All curvature mutation points are extracted to form a mutation sequence. The time interval and curvature change amplitude between adjacent mutation points are calculated, with the time interval ranging from 0.3 to 1.5 seconds and the curvature change amplitude ranging from 1.2 to 4.5. Linear regression is used to analyze the time trend of the curvature mutation sequence. The regression equation is... ,in Represents the curvature at time t. Let and be the regression coefficients. The regression coefficients are solved using the least squares method, iterated 50 times until convergence. The value of is between -0.8 and 0.8, and the value of b is between -1.5 and 1.5. A value greater than 0 indicates that the curvature is increasing. A value less than 0 indicates a decreasing curvature trend. The residual sequence of the regression equation is calculated, with the residual standard deviation ranging from 0.4 to 1.2. A residual standard deviation greater than 0.8 indicates unstable curvature changes. An autoregressive moving average model is used to analyze the time correlation of the curvature sequence, with the model order set to ARMA(2,1). Model parameters and prediction errors are calculated, and a curvature abrupt change trend feature vector is constructed. The vector dimension is 10 and includes the curvature mean, standard deviation, rate of change, abrupt change frequency, regression coefficient, residual standard deviation, ARMA model parameters, etc. Principal component analysis is used to extract the main change patterns, retaining principal components with a contribution rate greater than 90%, thus obtaining trajectory curvature offset trend data. This data is stored in time series form, including curvature values, curvature change rate, curvature abrupt change markers, and trend prediction values, providing input data for subsequent avoidance trajectory deviation analysis.

[0036] To analyze the acceleration / deceleration slope in the velocity change relationship, the wheelchair velocity time series is first extracted from the velocity change relationship, recording the velocity changes over a continuous 10 seconds. A sliding window analysis is then performed on the velocity series, with a window length of 0.8 seconds and a step size of 0.2 seconds. Within each window, a linear regression equation for the velocity is calculated, in the form v(t) = kt + c, where v(t) represents the velocity at time t, and k and c are regression coefficients. The regression coefficients are solved using the least squares method, iterated 30 times until convergence. The regression coefficient k is the velocity change slope, representing the average acceleration. The value of k is used, where a value greater than 0 indicates acceleration, and a value less than 0 indicates deceleration. The absolute value of k represents the intensity of acceleration / deceleration, ranging from 0.1 m / s² to 2.5 m / s². The goodness of fit R² of the regression equation is calculated, with an R² value ranging from 0.75 to 0.98. An R² value greater than 0.9 indicates good linearity in velocity change. The acceleration sequence is obtained by differentiating the velocity sequence using the central difference method with a difference step size of 5 sampling points. The statistical characteristics of the acceleration sequence are calculated, including mean, standard deviation, skewness, and kurtosis, with the mean ranging from -1.5. The acceleration values ​​range from 0.2 m / s² to 1.0 m / s², with a standard deviation between -1.2 and 1.2, and a kurtosis between 2.0 and 5.0. An acceleration threshold of ±0.8 m / s² is set. Points with absolute acceleration values ​​greater than the threshold are marked as significant acceleration / deceleration points. All significant acceleration / deceleration points are extracted to form an acceleration / deceleration sequence. The time interval and acceleration change amplitude between adjacent acceleration / deceleration points are calculated, with the time interval ranging from 0.4 s to 1.8 s and the acceleration change amplitude ranging from 0.6 m / s² to 2.2 m / s². A piecewise linear fitting method is used to model the acceleration sequence, dividing it into multiple linear segments. The slope of each segment represents the acceleration value, i.e., the rate of change of acceleration, ranging from -3.0 m / s² to 3.0 m / s². Between m / s², acceleration / deceleration feature vectors are constructed with a dimension of 14, including parameters such as velocity change slope, goodness of fit, acceleration statistical features, acceleration / deceleration frequency, and jerk value. Cluster analysis is used to classify acceleration / deceleration patterns using the K-means algorithm, with 4 categories corresponding to stationary, gradually changing, rapidly changing, and mixed types. The center vector and sample distribution of each category are calculated, with stationary type accounting for 20%, gradually changing type for 35%, rapidly changing type for 25%, and mixed type for 20%. An acceleration / deceleration change slope data table is constructed, recording the acceleration / deceleration feature parameters, category labels, and corresponding timestamps, providing input data for subsequent avoidance trajectory deviation analysis.

[0037] Curvature value and rate of change sequences were extracted from the trajectory curvature offset trend data table. Velocity change slope and acceleration sequences were extracted from the acceleration / deceleration change slope. Parameter changes were recorded over 10 consecutive seconds. These sequences were time-aligned with the wheelchair steering angle sequence to construct a multivariate time series matrix. The matrix has dimensions of 1000×5, with rows representing time points and columns representing curvature values, rate of change of curvature, velocity change slope, acceleration values, and steering angle values, respectively. The cross-correlation function between the variables was calculated, and the time delay corresponding to the maximum correlation was determined. The time delay between the steering angle and the curvature value was calculated as follows: The time delays between the time of curvature change and the time of velocity change slope are 0.15s, 0.08s, 0.22s, and 0.18s, respectively. Partial correlation analysis is used to assess the independent influence of each variable on the steering angle. Partial correlation coefficients are calculated: 0.72 for curvature, 0.65 for the rate of curvature change, 0.58 for the velocity change slope, and 0.63 for the acceleration. A phase diagram of the steering angle and other variables is constructed, and the phase difference and phase synchronization index are calculated. The phase difference range is within... to Between 0.4 and 0.9, the phase synchronization index is between 0.4 and 0.9. A phase synchronization index less than 0.6 indicates significant phase misalignment. The Granger causality test method is used to analyze the causal relationship between variables. The lag order is set to 5. The F statistic and p value are calculated. An F statistic greater than 4.0 and a p value less than 0.05 indicate significant causal relationship. A vector autoregression model is constructed with a model order of VAR(3). The model parameter matrix is ​​estimated, and the prediction error and information criterion of the model are calculated. The model with the smallest AIC information criterion is selected as the optimal model. The impulse response function of the model is calculated, and the dynamic impact of the shock of one variable on other variables is analyzed. The steering angle misalignment index is constructed. The index calculation formula is as follows: ,in , , These represent the actual steering angle, angular velocity, and angular acceleration, respectively. , , This represents the model's predicted value. , , The weighting coefficients are set to 0.5, 0.3, and 0.2 respectively. The misalignment index ranges from 0 to 10. An index greater than 5 indicates severe steering misalignment. Steering angle misalignment coupling data is constructed, and the steering angle misalignment index, phase difference, phase synchronization index, Granger causality test results, and corresponding timestamps are recorded.

[0038] Based on the trajectory curvature offset trend, acceleration / deceleration slope change, and steering angle misalignment coupling data, a multi-dimensional deviation stochastic process decomposition of the effective avoidance action is performed to obtain the multi-dimensional avoidance deviation. First, avoidance action feature vectors are extracted from the effective avoidance action; curvature feature vectors are extracted from the trajectory curvature offset trend; velocity feature vectors are extracted from the acceleration / deceleration slope change; and misalignment feature vectors are extracted from the steering angle misalignment coupling data table. These feature vectors are merged into a comprehensive feature matrix with dimensions of 1000×40, where rows represent time points and columns represent feature parameters. The feature matrix is ​​standardized to ensure each feature has zero mean and unit variance. Principal component analysis is used to reduce the dimensionality of the standardized feature matrix, retaining principal components with a contribution rate greater than 95%, resulting in a dimensionality-reduced feature matrix with dimensions of 1000×12. A stochastic process model of the avoidance deviation is constructed using Gaussian process regression with a radial basis function kernel, a kernel width parameter of 0.8, and a noise variance parameter of 0.05. The Gaussian process model is trained and iterated 200 times until convergence. The log-likelihood and prediction variance of the model are calculated. A log-likelihood greater than -500 indicates a good model fit. The avoidance deviation stochastic process is decomposed into two components: time and space. The time component represents the temporal deviation of the avoidance action, including the avoidance initiation time deviation, avoidance duration deviation, and avoidance recovery time deviation. The space component represents the spatial deviation of the avoidance action, including the avoidance amplitude deviation, avoidance trajectory deviation, and avoidance endpoint deviation. Singular spectrum analysis is used to decompose the time component. The embedding window length is set to 50 sampling points. The first three singular values ​​and their corresponding singular vectors are extracted to reconstruct the time dimension deviation sequence. Wavelet decomposition is used to perform multi-scale analysis on the space component. Daubechies wavelets are selected, and the decomposition level is four. Wavelet coefficients at each scale are calculated to reconstruct the space dimension deviation sequence. The time dimension deviation and the space dimension deviation are combined into a multi-dimensional avoidance deviation data table. The table records the time deviation parameters, space deviation parameters, and corresponding timestamps and spatial coordinates, providing input data for subsequent trajectory temporal deviation cumulative fitting.

[0039] Based on multi-dimensional avoidance deviation, a time-series deviation accumulation fitting method is used for the avoidance motion trajectory of an intelligent wheelchair to obtain accumulated avoidance trajectory deviation data. First, time deviation sequences and spatial deviation sequences are extracted from the multi-dimensional avoidance deviation data, and deviation changes within a continuous 10 seconds are recorded. The time deviation sequences are accumulated and summed to obtain the time deviation accumulation curve. The spatial deviation sequences are vector-accumulated to obtain the spatial deviation accumulation curve. Exponential smoothing is used to smooth the accumulation curves, with a smoothing coefficient set to 0.12 to reduce the influence of random fluctuations. A deviation accumulation model is constructed, with the following form: ,in Indicates time Cumulative deviation at the location, Indicates the initial deviation. This represents the instantaneous deviation at time ti. and Indicates the weighting coefficient and the attenuation coefficient. The time delay is represented by the least squares method to estimate the model parameters, which are iteratively calculated 150 times until convergence. The weighting coefficients are... The value ranges from 0.6 to 1.2, and the attenuation coefficient is... The value ranges from 0.05 to 0.25. The fitting error and prediction error of the model are calculated. A root mean square value of fitting error less than 0.15 indicates high model fitting accuracy. An adaptive filtering method is used to extract the trend of the cumulative deviation curve. The filter type is a Kalman filter. After filtering, a smooth cumulative deviation trend curve is obtained. A spatiotemporal joint cumulative deviation model is constructed, and the model form is as follows: ,in This represents the joint cumulative deviation at time t and spatial location s. Represents the pure time deviation component. Represents the pure spatial deviation component. To represent the spatiotemporal interaction deviation components, tensor decomposition is used to estimate model parameters. The joint deviation data is organized into a three-dimensional tensor with dimensions of 100×100×3, representing time points, spatial points, and deviation components, respectively. The Tucker decomposition method is used, with the kernel tensor rank set to (5,5,2). The calculation is iterated 200 times until convergence, and the joint deviation cumulative field is reconstructed. The reconstruction error and the proportion of explained variance are calculated. An explained variance proportion greater than 90% indicates a good decomposition effect. Finally, the cumulative deviation data of the avoidance trajectory is obtained. The data is stored in the form of a multidimensional array, including the cumulative deviation in the time dimension, the cumulative deviation in the spatial dimension, and the spatiotemporal joint cumulative deviation.

[0040] Preferably, the delayed analysis of sudden risk prediction in decision-making includes the following steps: Obtain the hardware configuration parameters of the intelligent wheelchair, wherein the hardware includes CPU, GPU and memory; Calculate the distribution density of obstacles in a dynamic obstacle scene and analyze the sudden motion state of the obstacles; wherein the sudden motion state includes velocity, direction, and acceleration; The obstacle motion vector is predicted and derived for the sudden motion state of the obstacle, and then the resource demand is evaluated in real time to obtain the real-time predicted resource demand. The computational bottleneck quantification data for the intelligent wheelchair is obtained by calculating the real-time predicted resource demand based on the hardware configuration parameters. Extract the time series with the largest decision delay based on decision inference bias; Based on the computational bottleneck quantification data and the time series with the largest decision delay, a lag analysis of sudden risk prediction is performed to obtain sudden risk prediction lag regression data.

[0041] In this embodiment of the invention, the hardware configuration parameters of the intelligent wheelchair are obtained. The hardware includes a CPU, GPU, and memory. First, the hardware configuration information of the intelligent wheelchair control unit is read through the system information interface. Standard system call functions are used to obtain the CPU model, number of cores, clock speed, and cache size. Specifically, the CPU parameters read include the model, 4 cores and 8 threads, a base frequency of 1.6GHz, a maximum turbo frequency of 3.9GHz, 6MB of L3 cache, and AVX2 instruction set support. Next, the GPU configuration information is read, including the GPU model, 256 CUDA cores, a GPU frequency of 1.3GHz, 8GB of video memory, a computing power of 5.3 TFLOPS, and support for CUDA 10.2 and TensorRT 7.1 acceleration libraries. Then, the memory configuration information is obtained, including a memory capacity of 16GB, LPDDR4 memory type, a memory frequency of 2133MHz, a memory bandwidth of 34.1GB / s, and dual-channel mode. Simultaneously, the storage device information is read, including 256GB of main storage. The SSD has a read speed of 550MB / s and a write speed of 520MB / s. The secondary storage is 64GB eMMC with a read speed of 280MB / s and a write speed of 250MB / s. Hardware resource usage status, including CPU utilization, GPU utilization, memory usage, and storage I / O load, is collected in real time through the system monitoring interface at a sampling frequency of 10Hz. Resource usage fluctuations are recorded over 30 consecutive seconds, and the average utilization, peak utilization, and standard deviation of each resource are calculated. The average CPU utilization is 45%, the peak is 78%, and the standard deviation is 12%. The average GPU utilization is 35%, the peak is 65%, and the standard deviation is 15%. The average memory usage is 42%, the peak is 60%, and the standard deviation is 8%. A hardware resource configuration parameter table is constructed, which records the specifications and resource usage status of each hardware component, providing basic data for subsequent computing load assessment.

[0042] The distribution density of obstacles in a dynamic obstacle scene is calculated, and the sudden movement states of obstacles are analyzed. First, the obstacle position coordinate sequence is extracted from the obstacle dynamic scene data, and the obstacle distribution over a continuous 10 seconds is recorded. The scene space is divided into 0.5m × 0.5m grid cells. The number of obstacles in each grid cell is calculated at each time point, and a time-varying density distribution matrix is ​​constructed. The matrix has dimensions of 20 × 20 × 1000, representing the number of grid cells in the x-direction, the number of grid cells in the y-direction, and the number of time points, respectively. The global average density is calculated to be 0.08 obstacles per second. The maximum local density is 1.5 per 100 cells. Appearing in the central region of the scene, a continuous density distribution function is generated using the kernel density estimation method. A Gaussian kernel is chosen as the kernel function, and the bandwidth parameter is set to 0.75m. The density gradient field is calculated, with the gradient direction pointing towards the direction of the fastest density increase. The gradient magnitude represents the rate of density change; a gradient magnitude greater than 0.5 indicates the existence of a significant density change region. The obstacle position sequence is time-differencing to calculate the obstacle's velocity vector. The velocity vector contains two components: magnitude and direction. The average velocity magnitude is 1.2 m / s, the maximum is 3.5 m / s, and the standard deviation is 0.65 m / s. The velocity direction is distributed within the range of 0 to 360°, mainly concentrated in four directions: 45°, 135°, 225°, and 315°. The velocity vector is time-differencing to calculate the obstacle's acceleration vector. The average acceleration vector is 0.8 m / s², the maximum is 2.7 m / s², and the standard deviation is 0.55 m / s². Sudden motion thresholds are set: a velocity abrupt change threshold of 1.5 m / s², a direction abrupt change threshold of 45°, and an acceleration abrupt change threshold of 2.0. When the parameter change exceeds the threshold, it is marked as a sudden motion point. All sudden motion points are extracted to form a sudden motion state sequence. The frequency and spatial distribution characteristics of sudden motion are calculated. The frequency of sudden motion is 0.8 times / s, mainly distributed in the edge area of ​​the scene. An obstacle distribution density and sudden motion state data table are constructed. The table records the spatiotemporal density distribution, velocity distribution, direction distribution, acceleration distribution and sudden motion markers.

[0043] The method derives obstacle motion vector prediction for sudden obstacle movement states, followed by real-time data processing resource requirement assessment to obtain real-time predicted resource requirements. First, historical trajectory data of obstacles is extracted from obstacle distribution density and sudden movement states, recording the trajectory over a continuous 10 seconds. A state vector is constructed for each obstacle, containing position coordinates, velocity vector, acceleration vector, and motion direction angle. A Kalman filter algorithm is used for trajectory prediction, with prediction parameters set based on a uniform acceleration motion model. The position prediction error is controlled within 0.05m, the velocity prediction error within 0.15m / s, and the acceleration prediction error within 0.25m / s². The obstacle trajectory is predicted for the next 5 seconds with a prediction step size of 100ms, generating 50 prediction points. Uncertainty analysis is performed on the predicted trajectory, calculating a 95% confidence interval. The position prediction error increases with prediction time, with an average error of 0.3m after 1 second and 0.8m after 3 seconds. The average error after 5 seconds is 1.5m. Sudden movement points are processed using a particle filtering algorithm with 1000 particles. Prediction model parameters are adjusted based on historical sudden movement intensities. The multimodal distribution of the predicted trajectory is calculated, and the three most likely trajectories and their probabilities are extracted. The computational complexity of real-time prediction is evaluated, and the algorithm's time and space complexity are analyzed. The number of floating-point operations per prediction is calculated: approximately 5000 for Kalman filtering and approximately 500,000 for particle filtering. Based on the number of obstacles and the frequency of sudden movements in the scene, the overall computational load is estimated, with an average load of 20 million floating-point operations per second and a peak load of 100 million floating-point operations per second. Considering data transmission and storage overhead, the memory bandwidth requirement is estimated at 200MB per second, and the storage space requirement at 20MB per second. A real-time prediction resource requirement data table is constructed, recording the computational load, memory bandwidth, storage space, and their relationship with the number of obstacles and the frequency of sudden movements, providing a basis for subsequent computational bottleneck quantification.

[0044] Based on hardware configuration parameters, the computational load bottleneck of real-time predicted resource demand is quantified to obtain computational bottleneck quantification data for the intelligent wheelchair. First, performance indicators of CPU, GPU, and memory are extracted from the hardware configuration parameter table, including CPU computing power of 56 GFLOPS, GPU computing power of 5.3 TFLOPS, and memory bandwidth of 34.1 GB / s. These performance indicators are compared and analyzed with the real-time predicted resource demand to calculate the load ratio of each hardware component. The CPU load ratio is the actual computational load divided by the CPU computing power. When the number of obstacles is 20 and the burst frequency is 0.8 times / second, the CPU load ratio is 35.7%. When the number of obstacles increases to 50 and the burst frequency is 2.0 times / second, the CPU load ratio rises to 89.3%. The GPU load ratio is the actual computational load divided by the GPU computing power. When executing the parallelized prediction algorithm, the GPU load ratio is 1.9%. The memory bandwidth load ratio is the actual bandwidth demand divided by the memory bandwidth. When the data transfer rate is 200 MB / s, the memory bandwidth load ratio is 0.6%. A load ratio matrix is ​​constructed, where the rows represent different numbers of obstacles and... The burst frequency combination, column-based to represent the load ratio of CPU, GPU, and memory, uses a bottleneck identification algorithm to analyze the load ratio matrix and calculate the critical point at which each hardware component reaches saturation. The CPU saturation critical point is 56 obstacles and a burst frequency of 2.2 times / second. The GPU saturation critical point is far from being reached. The memory bandwidth saturation critical point is a data transfer rate of 34GB / s, which is far higher than the current demand. The CPU is identified as the main bottleneck component. The CPU bottleneck coefficient is calculated, defined as the ratio of the load ratio to the critical load ratio. A bottleneck coefficient greater than 0.8 indicates that the CPU is approaching a bottleneck. The relationship between bottleneck coefficient and the number of obstacles and sudden frequency was obtained by fitting multiple sets of experimental data. The fitting parameters were 0.012, 0.085 and 0.007, respectively, with a fitting error of less than 5%. The bottleneck coefficient under different scenarios was predicted using the fitting relationship, and a bottleneck coefficient surface plot was generated. The horizontal axis represents the number of obstacles, the vertical axis represents the sudden frequency, and the surface height represents the bottleneck coefficient. The bottleneck coefficient contour lines were extracted. The contour line with a bottleneck coefficient of 0.8 represents the safe operation boundary. The computational bottleneck quantification data of the intelligent wheelchair was constructed, and the bottleneck coefficient, load ratio and safety margin under different scenario parameters were recorded.

[0045] Based on the decision-making deviation, the time series with the largest decision delay is extracted. First, a timestamp sequence and the corresponding deviation value sequence are extracted from the decision-making deviation. Threshold detection is performed on the deviation value sequence, with thresholds set as 0.3m for horizontal deviation, 0.3m for vertical deviation, and 15° for angular deviation. When any deviation component exceeds the threshold, it is marked as a decision-making time point. The time interval between adjacent time points is calculated, with an average interval of 0.85s and a standard deviation of 0.32s. First-order differencing is performed on the time interval sequence to construct a difference sequence matrix. The K-means clustering algorithm is used to classify the difference sequence into small-delay, medium-delay, and... The time interval difference is divided into equal delay and large delay categories. The average difference is 0.15s for the small delay category, 0.45s for the medium delay category, and 0.85s for the large delay category. The large delay category samples are further screened, with a screening threshold of 0.2m for the lateral deviation difference. When the deviation difference exceeds the threshold, the sample is retained. All samples that meet the conditions are extracted to form the time series with the largest decision delay. The average delay time of this series is calculated to be 1.05s, and the maximum delay time is 1.85s. Its purpose is to find the critical time period when the wheelchair responds the slowest during obstacle avoidance, and to provide an accurate target time window for the lag analysis of sudden risk prediction.

[0046] Based on the calculated bottleneck quantification data and the time series with the largest decision delay, a lag analysis of the sudden risk prediction is performed to obtain lag regression data for sudden risk prediction. First, the bottleneck coefficient sequence is extracted from the calculated bottleneck quantification data, and the delay time series is extracted from the time series data table with the largest decision delay. The two sequences are matched according to scenario parameters to construct a joint data matrix. The matrix rows represent different combinations of scenario parameters, and the columns represent the bottleneck coefficient, delay time, deviation value, and obstacle sudden intensity. A multiple regression analysis method is used to establish the lag prediction relationship. Through data fitting, the relationship between the prediction lag time and the number of obstacles, sudden frequency, and bottleneck coefficient is obtained. The fitting parameters include a constant term of 0.35s, an obstacle quantity coefficient of 0.008 seconds / obstacle, a sudden frequency coefficient of 0.12, a bottleneck coefficient of 0.45 seconds / unit, and three interaction term coefficients of 0.003, 0.006, and 0.08, respectively. The goodness of fit is 0.85, indicating a good fit relationship. 85% of the lag time variation was explained, and the root mean square error of the prediction was calculated to be 0.15 s. Cross-validation of the fitted relationship was performed using a 10-fold cross-validation method, and the root mean square error of the validation was 0.18 seconds, slightly higher than the training error, indicating that the fitted relationship has good generalization ability. A lag prediction response surface was constructed, with the horizontal axis representing the number of obstacles, the vertical axis representing the burst frequency, and the surface height representing the prediction lag time. Multiple response surfaces were generated under different bottleneck coefficients, and the variation of lag time with scene parameters was analyzed. When the bottleneck coefficient increased from 0.3 to 0.8, the average lag time increased from 0.55 s to 0.95 s, an increase of 72.7%. The gradient field of the lag time was calculated, with the gradient direction pointing to the parameter combination with the fastest increase in lag time. The gradient magnitude represents the sensitivity of lag time. The significance of the fitted parameters was tested, and the t-statistic and p-value were calculated. The p-values ​​of all parameters were less than 0.05, indicating that the parameters are statistically significant. Lag regression data for burst risk prediction was constructed.

[0047] Preferably, step S3 includes the following steps: The decision extrapolation bias is processed by convolution to obtain the decision convolution bias; the lag regression data of the sudden risk prediction is normalized to obtain the lag normalized data of the sudden risk prediction. Based on the delayed normalized data of sudden risk prediction, the decision convolution bias is used to perform pre-bias correction memory learning for wheelchair scheduling braking, and bias correction memory learning data is obtained. The bias-correcting memory learning data is subjected to reinforcement learning iterations to output decision optimization data.

[0048] As an example of the present invention, reference is made to... Figure 3 As shown, step S3 in this example includes: Step S301: Perform convolution processing on the decision inference bias to obtain the decision convolution bias; perform normalization processing on the lag regression data of sudden risk prediction to obtain the lag normalized data of sudden risk prediction. In this embodiment of the invention, the decision-making inference bias is processed by convolution to obtain the decision convolution bias, and the lag regression data of sudden risk prediction is normalized to obtain the lag normalized data of sudden risk prediction. First, the bias sequence is extracted from the decision-making inference bias, and the bias changes within 10 consecutive seconds are recorded. A one-dimensional convolution kernel is designed with a kernel length of 5 and kernel weights of [0.1, 0.2, 0.4, 0.2, ...]. [0.1] Convolution operation is performed on the deviation sequence using a sliding window method with a stride of 1 and mirror filling of the boundaries. The convolution calculation formula is that the current output is equal to the inner product of the input sequence and the convolution kernel. Local features and temporal continuity of the deviation sequence are extracted through convolution operation to obtain the decision convolution deviation sequence. At the same time, the lag regression data for sudden risk prediction is normalized to the maximum and minimum values, scaling the numerical range to the [0,1] interval. The normalization formula is (x-min) / (max-min), where x is the original value, and min and max are the minimum and maximum values ​​of the data, respectively. After processing, the lag normalized data for sudden risk prediction is obtained. The data dimension remains unchanged, and the numerical range is unified to the [0,1] interval, which is convenient for subsequent processing.

[0049] Step S302: Based on the delayed normalized data of sudden risk prediction, perform pre-bias correction memory learning on the decision convolution bias to obtain bias correction memory learning data. In this embodiment of the invention, a pre-bias correction memory learning method is used to correct the wheelchair scheduling and braking based on the sudden risk prediction lag normalized data to obtain bias correction memory learning data. First, a memory learning network is constructed. The network structure is a three-layer feedforward neural network with 15 neurons in the input layer, corresponding to the combined features of the decision convolutional bias and the sudden risk prediction lag normalized data (wherein the sudden risk prediction lag normalized data includes the prediction delay of the time, position, speed, and acceleration of obstacles in the decision made by the intelligent wheelchair in a dynamic environment, as well as the wheelchair's own state such as current speed, acceleration, steering angle, and braking response delay). The hidden layer has 30 neurons, and the activation function is R. The eLU has 8 neurons in its output layer, corresponding to the corrected braking parameters, including braking advance, braking intensity, steering angle, and steering rate. The network is trained using stochastic gradient descent with an initial learning rate of 0.01, employing an exponential decay strategy that reduces the learning rate to 0.9 times its original value every 1000 steps. The batch size is set to 64, and the number of training iterations is 10,000. The loss function is mean squared error. A memory mechanism is introduced during training, using an experience replay buffer to store historical samples. The buffer size is 1000, and 128 historical samples are randomly sampled and mixed with the current batch samples in each training iteration to enhance the network's memory and generalization capabilities. After training, bias-corrected memory learning data is obtained.

[0050] Step S303: Perform reinforcement learning iterations on the bias correction memory learning data and output decision optimization data.

[0051] In this embodiment of the invention, reinforcement learning iterations are performed on bias correction memory learning data to output decision optimization data. First, a reinforcement learning environment is constructed. The state space includes the wheelchair position, speed, steering angle, and distribution of surrounding obstacles. The action space includes the adjustment amounts of braking force and steering angle. The reward function is designed based on safe distance, stability, and goal achievement. A deep Q-network algorithm is used to implement reinforcement learning. The Q-network structure is a four-layer fully connected network. The input layer dimension is the state space dimension, and the output layer dimension is the action space dimension. The number of hidden layer neurons is 128 and 64, respectively, and the activation function is ReLU. An ε-greedy strategy is adopted to balance exploration and exploitation. The initial ε value is 0.9, which is gradually reduced to 0.1. The discount factor γ is set to 0.95. The target network update frequency is 100 steps, the experience replay buffer size is 10000, the training batch size is 32, the learning rate is 0.001, and the number of training iterations is 50000. An evaluation is performed every 1000 iterations, and the average reward and success rate are recorded. Training stops when the average reward improvement of 5 consecutive evaluations is less than 1%. The final output is decision optimization data, including the optimal braking timing, braking force, steering angle and its adjustment strategy.

[0052] Preferably, the pre-deflection correction memory learning for wheelchair scheduling braking includes the following steps: The decision convolution bias is extracted to obtain the decision bias parameters; wherein the parameters include the wheelchair's braking response delay, maximum braking acceleration, steering angle and avoidance trajectory; Evolutionary regression analysis is performed on the decision deviation parameters to generate the evolutionary laws of the deviation parameters; Based on the evolution law of deviation parameters and decision lag normalized data, an approximate clustering of deviation correction safety margin is performed, and memory learning is carried out to obtain correction amount clustering memory data. The deviation correction safety margin refers to the dynamic safety buffer amount reserved by the wheelchair when facing sudden obstacles by comprehensively considering the evolution law of key parameters such as braking response delay, maximum braking acceleration, steering angle and avoidance trajectory of the wheelchair. By correcting and optimizing these parameters in real time, it is ensured that the wheelchair can maintain sufficient safety distance and reaction time in various scenarios. Based on the corrected amount clustering memory data and the delayed normalized data of sudden risk prediction, wheelchair scheduling braking deviation correction memory learning is performed to obtain deviation correction memory learning data.

[0053] In this embodiment of the invention, deviation parameters are extracted from the decision convolution deviation to obtain decision deviation parameters. First, key feature points are extracted from the decision convolution deviation, and the feature point extraction threshold is set to 1.5 times the standard deviation of the deviation value. When the deviation value exceeds the threshold, it is marked as a feature point. The time distribution of the feature points is calculated to obtain the braking response delay parameter. The braking response delay is defined as the time interval from the appearance of the obstacle to the start of wheelchair braking, with an average value of 0.72s and a standard deviation of 0.18s. The acceleration change is calculated by second-order difference of the deviation sequence to extract the maximum braking acceleration parameter. The average maximum braking acceleration is 1.85m / s², ranging from 1.2 to 2.5 m / s². The steering angle change is extracted from the deviation. The steering angle is defined as the angle between the wheelchair's forward direction and the target direction, with an average maximum steering angle of 28° and a standard deviation of 7.5°. The trajectory deviation is calculated by integrating the position deviation to extract the avoidance trajectory parameters, including a maximum lateral offset of 1.2m, a trajectory length of 2.8m, and a trajectory curvature of 0.35. A decision bias parameter vector with a dimension of 12 is constructed, which contains the statistical characteristics of the above four types of parameters, providing input data for subsequent evolutionary pattern analysis.

[0054] Evolutionary regression analysis was performed on the decision deviation parameters to generate their evolutionary patterns. First, the decision deviation parameters were arranged chronologically to construct a time series matrix with dimensions of 1000×12, where rows represent time points and columns represent individual deviation parameters. Trend analysis was performed on the time series, using the moving average method to extract the long-term trend. The window length was set to 50 sampling points. The trend slopes of each parameter were calculated: braking response delay trend slope was 0.002s, maximum braking acceleration trend slope was -0.05m / s², steering angle trend slope was 0.15° / s, and the maximum lateral offset trend slope of the avoidance trajectory was... At a speed of 0.01 m / s, periodic analysis was performed on the time series. The autocorrelation function method was used to calculate the autocorrelation coefficients of each parameter. The principal period of braking response delay was 1.8 s, the principal period of maximum braking acceleration was 2.2 s, the principal period of steering angle was 1.5 s, and the principal period of avoidance trajectory was 2.5 s. The parameter evolution curves were fitted using a polynomial regression method with a polynomial order of 3. The goodness of fit R² were 0.82, 0.78, 0.85, and 0.76, respectively. The evolution law of deviation parameters was constructed, and the trend slope, periodic characteristics, and regression coefficients of each parameter were recorded to provide a basis for subsequent safety margin clustering.

[0055] Based on the evolution law of deviation parameters and the normalized decision lag data, an approximate clustering of the deviation correction safety margin is performed, followed by memory learning to obtain clustered memory data of the correction amount. First, the evolution law of deviation parameters and the normalized decision lag data are fused to construct a fused feature vector with a dimension of 20, containing both the evolution law parameters and the normalized lag data. Principal component analysis is performed on the fused features to extract the main feature axes, retaining the principal components with a contribution rate greater than 95%. The dimensionality of the reduced feature vector is 8. The safety margin is calculated as the ratio of the correction amount of braking response delay to the obstacle speed, in m / s². The average safety margin is 0.45 m / s², and the standard deviation is 0.12. The safety margin metric was clustered using a fuzzy C-means clustering algorithm with a density of m / s². The number of clusters was set to 5, the fuzziness parameter to 2.0, the number of iterations to 100, and the convergence threshold to 0.001. The cluster centers and sample membership degrees were calculated, with cluster centers at 0.25, 0.35, 0.45, 0.55, and 0.65 m / s². A memory learning network was constructed, consisting of a four-layer feedforward neural network with an input layer dimension of 8, hidden layers dimensions of 16 and 12, and an output layer dimension of 5, corresponding to the membership degrees of the five clusters. The training sample size was 1000, the learning rate was 0.005, the batch size was 32, the training iterations were 5000, and the loss function was cross-entropy. After training, the corrected cluster memory data was obtained, including cluster centers, sample distribution, and network parameters.

[0056] Based on the corrected amount clustering memory data and the sudden risk prediction lag normalized data, wheelchair scheduling braking deviation correction memory learning is performed to obtain deviation correction memory learning data. First, cluster centers and weight matrices are extracted from the corrected amount clustering memory data. Lag time and obstacle motion parameters are extracted from the sudden risk prediction lag normalized data (wherein the sudden risk prediction lag normalized data includes the prediction delay of obstacle appearance time, position, speed, and acceleration in the decision made by the intelligent wheelchair in the dynamic environment, as well as the wheelchair's own state such as current speed, acceleration, steering angle, and braking response delay). A joint feature vector with a dimension of 25 is constructed, containing clustering features and lag features. A memory learning algorithm is designed, adopting a radial basis function network structure with an input dimension of 25, 40 hidden layer nodes, and an output dimension of 8, corresponding to the mean and standard deviation of the four types of correction parameters. A Gaussian kernel is selected as the radial basis function, and the kernel width parameter is adaptive. The training set should be set to be inversely proportional to the sample density, with 1500 training samples. Cross-validation should be used to divide the training and validation sets into an 8:2 ratio. Gradient descent should be used for training, with a learning rate of 0.008, a momentum factor of 0.6, a batch size of 50, and 6000 training iterations. The loss function should be mean squared error. A memory decay mechanism should be introduced during training, gradually reducing the weight of samples in the later stages with a decay coefficient of 0.95. A sample importance weighting strategy should also be used, increasing the weight of samples in sudden scenarios by 50%. After training, the network should be pruned to remove connections with absolute weights less than 0.05, reducing the number of network parameters by 35%. A bias-corrected memory learning data table should be constructed, recording the distribution characteristics of network parameters and corrected parameters, as well as the scene adaptation mapping relationship. Corrected parameters include a braking advance of 0.65s (mean) and a standard deviation of 0.18s, and a braking force of 1.5m / s² (mean) and a standard deviation of 0.4. The mean steering angle is 22° and the standard deviation is 6.5°, and the mean steering speed is 25° / s and the standard deviation is 8° / s, providing basic data for subsequent reinforcement learning iterations.

[0057] The present invention also provides an intelligent wheelchair scheduling system for executing the intelligent wheelchair scheduling method described above, the intelligent wheelchair scheduling system comprising: The obstacle avoidance failure decision extraction module is used to extract historical scheduling tasks from the database of intelligent wheelchairs, and then extract the obstacle avoidance failure decisions of intelligent wheelchairs from the historical scheduling tasks; perform cluster analysis on the obstacle avoidance failure decisions to generate decision failure clustering logic; The obstacle avoidance failure decision analysis module is used to analyze the deviation of the motion trajectory prediction of sudden obstacles based on the decision failure clustering logic to obtain the decision prediction deviation; and to perform sudden risk prediction lag analysis based on the decision prediction deviation to obtain sudden risk prediction lag regression data. The memory learning module is used to perform pre-deviation correction memory learning based on the delayed normalized data of sudden risk prediction for wheelchair dispatching and braking, and obtain deviation correction memory learning data; it then performs reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

[0058] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the 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 invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for intelligent wheelchair scheduling, characterized in that, Includes the following steps: Step S1: Extract historical scheduling tasks from the database of smart wheelchairs, and then extract obstacle avoidance failure decisions of smart wheelchairs from the historical scheduling tasks. Cluster analysis is performed on the obstacle avoidance failure decisions to generate decision failure clustering logic; Step S2: Based on the decision failure clustering logic, perform a deviation analysis of the wheelchair avoidance trajectory in the sudden obstacle to obtain the decision inference deviation; Lag analysis of sudden risk prediction based on decision-making inference bias, to obtain sudden risk prediction lag regression data; wherein, step S2 includes: The obstacle dynamic parameters are extracted from the obstacle avoidance failure samples and the scene is reconstructed to obtain the obstacle dynamic scene, wherein the obstacles include crowds and moving objects; Based on the clustering logic of decision failure, the deviation analysis of the motion trajectory extrapolation of wheelchair avoidance of sudden obstacles in dynamic obstacle scenarios is conducted to obtain the decision extrapolation deviation. A delayed analysis of sudden risk prediction based on decision-making inference bias and dynamic obstacle scenarios is used to obtain delayed sudden risk prediction data; wherein, the delayed sudden risk prediction analysis includes: Obtain the hardware configuration parameters of the intelligent wheelchair, wherein the hardware includes CPU, GPU and memory; Calculate the distribution density of obstacles in a dynamic obstacle scene and analyze the sudden motion state of the obstacles; wherein the sudden motion state includes velocity, direction, and acceleration; The obstacle motion vector is predicted and derived for the sudden motion state of the obstacle, and then the resource demand is evaluated in real time to obtain the real-time predicted resource demand. The computational bottleneck quantification data for the intelligent wheelchair is obtained by calculating the real-time predicted resource demand based on the hardware configuration parameters. Extract the time series with the largest decision delay based on decision inference bias; Based on the computational bottleneck quantification data and the time series with the largest decision delay, a lag analysis of sudden risk prediction is performed to obtain sudden risk prediction lag regression data. The regression analysis is used to predict the lagged data of sudden risks, and the output is the lagged regression data of sudden risk prediction. Step S3: Based on the delayed normalized data of the sudden risk prediction, perform pre-deviation correction memory learning for wheelchair dispatching and braking to obtain deviation correction memory learning data; perform reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

2. The intelligent wheelchair scheduling method according to claim 1, characterized in that, Step S1 includes the following steps: Extract historical scheduling tasks from the database of smart wheelchairs, and extract obstacle avoidance failure datasets from the historical scheduling tasks; Random samples are selected from the obstacle avoidance failure dataset, and the obstacle avoidance failure samples are output. Then, the obstacle avoidance failure decisions of the smart wheelchair in the samples are extracted. Logical analysis is performed on the obstacle avoidance failure decision to obtain the decision failure logic; Cluster analysis is performed on the decision failure logic to generate decision failure cluster logic.

3. The intelligent wheelchair scheduling method according to claim 1, characterized in that, The deviation analysis of wheelchair avoidance trajectory projection for sudden obstacles includes the following steps: Based on the decision failure clustering logic, extract the obstacle avoidance failure status of the intelligent wheelchair in the dynamic obstacle scene when there is a sudden obstacle. Extract the back-and-forth loop trajectory in the obstacle avoidance failure state, and analyze the nonlinear relationship of the trajectory range in the back-and-forth loop trajectory; wherein the back-and-forth loop trajectory refers to an invalid motion mode in which the intelligent wheelchair gets stuck in a local space, that is, the movement direction path swings back and forth in a short period of time; wherein the nonlinear relationship of the trajectory range refers to the degree and irregularity of the wheelchair's range of motion changing over time during the obstacle avoidance failure process. By analyzing the back-and-forth loop trajectory, the curve of the wheelchair swinging back and forth in the movement direction path can be constructed. The intermittent start-stop speed variation relationship of the intelligent wheelchair is analyzed from the nonlinear relationship of the trajectory range; the degree of disorder of the wheelchair steering angle is coupled in the nonlinear relationship of the trajectory range and the speed variation relationship. Based on the nonlinear relationship of the trajectory range, the relationship of speed change, and the disorder of the wheelchair turning angle, the timing deviation of the obstacle avoidance action trajectory of the intelligent wheelchair is cumulatively fitted in the obstacle dynamic scene to obtain the cumulative deviation data of the obstacle avoidance trajectory. Analysis of the deviation in wheelchair avoidance trajectory simulation for sudden obstacles based on accumulated data of avoidance trajectory deviation yields the decision simulation deviation.

4. The intelligent wheelchair scheduling method according to claim 3, characterized in that, The cumulative fitting of the temporal deviation of the obstacle avoidance trajectory of the intelligent wheelchair includes the following steps: The effective obstacle avoidance actions of intelligent wheelchairs are analyzed through dynamic obstacle scene analysis. A sawtooth trajectory analysis is performed on the nonlinear relationship of the trajectory range to obtain sawtooth trajectory data; a curvature abrupt change trend offset analysis is performed on the sawtooth trajectory data to obtain the trajectory curvature offset trend; Analyze the slope of acceleration / deceleration in the velocity change relationship; Based on the trajectory curvature deviation trend and the slope of acceleration / deceleration, the degree of disorder of the wheelchair steering angle is analyzed to obtain the steering angle mismatch coupling data. Based on the trajectory curvature offset trend, acceleration / deceleration slope change, and steering angle misalignment coupling data, the effective avoidance action is decomposed into a multi-dimensional avoidance deviation stochastic process to obtain the multi-dimensional avoidance deviation; the multi-dimensional deviation includes the time dimension and the spatial dimension. Based on multi-dimensional avoidance deviation, the temporal deviation of the avoidance action trajectory of the intelligent wheelchair is cumulatively fitted to obtain the cumulative data of the avoidance trajectory deviation.

5. The intelligent wheelchair scheduling method according to claim 1, characterized in that, Step S3 includes the following steps: The decision extrapolation bias is processed by convolution to obtain the decision convolution bias; the lag regression data of the sudden risk prediction is normalized to obtain the lag normalized data of the sudden risk prediction. Based on the delayed normalized data of sudden risk prediction, the decision convolution bias is used to perform pre-bias correction memory learning for wheelchair scheduling braking, and bias correction memory learning data is obtained. The bias-correcting memory learning data is subjected to reinforcement learning iterations to output decision optimization data.

6. The intelligent wheelchair scheduling method according to claim 5, characterized in that, The pre-deflection correction memory learning for wheelchair shunting braking includes the following steps: The decision convolution bias is extracted to obtain the decision bias parameters; wherein the parameters include the wheelchair's braking response delay, maximum braking acceleration, steering angle and avoidance trajectory; Evolutionary regression analysis is performed on the decision deviation parameters to generate the evolutionary laws of the deviation parameters; Based on the evolution law of deviation parameters and decision lag normalized data, an approximate clustering of deviation correction safety margin is performed, and memory learning is carried out to obtain correction amount clustering memory data. The deviation correction safety margin refers to the dynamic safety buffer amount reserved by the wheelchair when facing sudden obstacles by comprehensively considering the evolution law of key parameters such as braking response delay, maximum braking acceleration, steering angle and avoidance trajectory of the wheelchair. By correcting and optimizing these parameters in real time, it is ensured that the wheelchair can maintain sufficient safety distance and reaction time in various scenarios. Based on the corrected amount clustering memory data and the delayed normalized data of sudden risk prediction, wheelchair scheduling braking deviation correction memory learning is performed to obtain deviation correction memory learning data.

7. An intelligent wheelchair dispatching system, characterized in that, For performing the intelligent wheelchair scheduling method as described in claim 1, the intelligent wheelchair scheduling system includes: The obstacle avoidance failure decision extraction module is used to extract historical scheduling tasks from the database of intelligent wheelchairs, and then extract the obstacle avoidance failure decisions of intelligent wheelchairs from the historical scheduling tasks; perform cluster analysis on the obstacle avoidance failure decisions to generate decision failure clustering logic; The obstacle avoidance failure decision analysis module is used to analyze the deviation of the wheelchair avoidance trajectory in sudden obstacles based on the clustering logic of decision failure, and to obtain the decision inference deviation; based on the decision inference deviation, the module analyzes the lag of sudden risk prediction to obtain the lag regression data of sudden risk prediction. The memory learning module is used to perform pre-deviation correction memory learning based on the delayed normalized data of sudden risk prediction for wheelchair dispatching and braking, and obtain deviation correction memory learning data; it then performs reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

Images