Intelligent wheelchair dispatching method and system

Abstract

The present application relates to the technical field of intelligent wheelchair scheduling, and particularly relates to an intelligent wheelchair scheduling method and system. The method comprises the following steps: extracting obstacle avoidance failure decisions in historical scheduling tasks, and generating clustering logic of decision failure through clustering analysis; analyzing the motion trajectory of the sudden obstacle based on the clustering logic to deduce the deviation, and then performing sudden risk prediction lag analysis to obtain lag regression data; using the lag regression data to perform pre-deviation correction memory learning of wheelchair scheduling braking, iteratively optimizing the decision through reinforcement learning, and finally outputting the optimized decision data. The present application optimizes the intelligent wheelchair scheduling technology to make the intelligent wheelchair scheduling technology more perfect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent wheelchair scheduling, and in particular to an intelligent wheelchair scheduling method and system. BACKGROUND

[0002] As an auxiliary moving device, the intelligent wheelchair still has many deficiencies in autonomous scheduling and obstacle avoidance in complex environments. On the one hand, the wheelchair will encounter dynamic obstacles such as pedestrians, pets or sudden moving objects in actual operation. The obstacle avoidance strategy based on static environment or rules in the past is difficult to respond to these changes in real time, thereby leading to obstacle avoidance failure. However, a conventional intelligent wheelchair scheduling method often lacks deep analysis and learning of historical task data, and cannot summarize the error-prone decision-making mode. The autonomous decision-making ability of the intelligent wheelchair is weak, thereby causing the problem of large obstacle avoidance error. SUMMARY

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

[0004] To achieve the above-mentioned purpose, an intelligent wheelchair scheduling method, the method comprising the following steps:

[0005] Step S1: extracting historical scheduling tasks from a database of the intelligent wheelchair, and then extracting obstacle avoidance failure decisions of the intelligent wheelchair in the historical scheduling tasks; performing cluster analysis on the obstacle avoidance failure decisions to generate decision failure cluster logic;

[0006] Step S2: performing decision sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis based on the decision failure cluster logic to obtain decision deduction deviation; performing decision sudden risk prediction lag analysis based on the decision deduction deviation to obtain sudden risk prediction lag regression data;

[0007] Step S3: performing wheelchair scheduling braking pre-deviation correction memory learning according to the sudden risk prediction lag regression data to obtain deviation correction memory learning data; performing reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

[0008] The present application also provides an intelligent wheelchair scheduling system for executing the intelligent wheelchair scheduling method as described above, the intelligent wheelchair scheduling system comprising:

[0009] An obstacle avoidance failure decision extraction module is configured to extract historical scheduling tasks from a database of the intelligent wheelchair, and then extract obstacle avoidance failure decisions of the intelligent wheelchair in the historical scheduling tasks; perform cluster analysis on the obstacle avoidance failure decisions to generate decision failure cluster logic;

[0010] The obstacle avoidance failure decision analysis module is configured to perform sudden obstacle wheelchair avoidance movement trajectory deduction deviation analysis based on the decision failure clustering logic to obtain decision deduction deviation; and perform sudden risk prediction lag analysis based on the decision deduction deviation to obtain sudden risk prediction lag regression data.

[0011] The memory learning module is configured to perform pre-deviation correction memory learning of the wheelchair scheduling braking based on the sudden risk prediction lag regression data to obtain deviation correction memory learning data; and perform reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

[0012] The beneficial effects of the present application are that by extracting historical scheduling tasks from the database of the intelligent wheelchair, especially focusing on the obstacle avoidance failure decisions in history, a deep analysis basis can be provided for subsequent scheduling optimization. Cluster analysis of obstacle avoidance failure decisions helps to reveal the commonalities and differences of wheelchair obstacle avoidance failure under different circumstances, thereby generating cluster logic of decision failure. This process enables the system to identify which decision 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 the decision failures that have occurred, improving the decision accuracy and adaptability of the intelligent wheelchair. Based on the decision failure cluster logic, the motion trajectory deviation analysis of the sudden obstacle is carried out, which helps to predict and analyze the dynamic changes of the sudden obstacle and its influence on the decision of the intelligent wheelchair. By analyzing the deviation of the motion trajectory in detail, the system can identify how the current decision pattern is disturbed by the sudden obstacle under specific circumstances, and further understand the specific source of the deviation. This provides important data support for risk prediction, especially the sudden risk prediction lag data obtained through lag analysis, which can help the system to identify potential safety hazards in advance, reduce the reaction delay of the wheelchair to sudden obstacles, and enhance its environmental adaptability. Through the combination of pre-deviation correction memory learning and reinforcement learning, the decision optimization ability of the intelligent wheelchair is significantly improved. First, pre-deviation correction memory learning enables the system to correct the sudden risk prediction lag data and convert these corrected data into memory for future decision reference. In this way, the wheelchair can gradually accumulate past experience to avoid the same deviation in similar situations. Second, reinforcement learning iteration further optimizes the decision path in multiple scheduling tasks, enabling the intelligent wheelchair to continuously adjust its obstacle avoidance and scheduling strategies through a feedback mechanism. After multiple training and adjustments, the final output of the decision optimization data can significantly improve the autonomous scheduling ability, reaction speed and safety of the wheelchair in complex environments. Overall, through intelligent learning and optimization, the response ability of the intelligent wheelchair in dynamic environments is greatly enhanced, improving the user experience and safety guarantee. Therefore, the present application is an optimization of the traditional intelligent wheelchair scheduling method, solving the problem that the traditional intelligent wheelchair scheduling method often lacks deep analysis and learning of historical task data, and cannot summarize the error-prone decision patterns, improving the autonomous decision-making ability of the intelligent wheelchair and reducing the obstacle avoidance error. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A step flowchart for an intelligent wheelchair scheduling method;

[0014] Figure 2 A detailed implementation step flowchart for step S2 in the Figure 1 intelligent wheelchair scheduling method.

[0015] Figure 3 For Figure 1 Detailed implementation step flow diagram of step S3. DETAILED DESCRIPTION

[0016] Please refer to Figures 1 to 3 An intelligent wheelchair scheduling method, the method comprises the following steps:

[0017] Step S1: extracting historical scheduling tasks from the database of the intelligent wheelchair, and then extracting the obstacle avoidance failure decision of the intelligent wheelchair in the historical scheduling tasks; clustering analysis is performed on the obstacle avoidance failure decision to generate decision failure clustering logic;

[0018] Step S2: based on the decision failure clustering logic, the sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis of the decision is carried out, and the decision deduction deviation is obtained; based on the decision deduction deviation, the sudden risk prediction lag analysis of the decision is carried out to obtain sudden risk prediction lag regression data;

[0019] Step S3: according to the sudden risk prediction lag regression data, the pre-deviation correction memory learning of wheelchair scheduling braking is carried out, and the deviation correction memory learning data is obtained; the deviation correction memory learning data is subjected to reinforcement learning iteration, and the decision optimization data is output.

[0020] In the embodiment of the application, reference Figure 1 The is a step flow diagram of an intelligent wheelchair scheduling method of the application, in this example, the intelligent wheelchair scheduling method comprises the following steps:

[0021] Step S1: extracting historical scheduling tasks from the database of the intelligent wheelchair, and then extracting the obstacle avoidance failure decision of the intelligent wheelchair in the historical scheduling tasks; clustering analysis is performed on the obstacle avoidance failure decision to generate decision failure clustering logic;

[0022] In the embodiment of the present application, historical scheduling task data is extracted from the database of the intelligent wheelchair, which records the timestamp of task execution, wheelchair position coordinates, obstacle detection information, obstacle avoidance state marker and task completion in the form of a relational data table, from which obstacle avoidance tasks with failure markers are screened to obtain a historical obstacle avoidance failure data set. Then, the historical obstacle avoidance failure data set is structured and processed, each failure record is decomposed into basic elements such as real-time position information of the wheelchair, speed of the wheelchair, steering angle of the wheelchair, relative position of the obstacle, speed and acceleration of the obstacle, and an obstacle avoidance failure decision sequence is formed after processing. The obstacle avoidance failure decision sequence is classified by clustering method, and 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 calculated first, and the failure decisions are divided into several categories according to the similarity, each category representing a typical failure mode. After clustering, a decision failure clustering logic is formed, which records the hierarchical relationship between failure types in a tree structure and is stored in a logical index table for subsequent step calling.

[0023] In another embodiment, stored historical scheduling task data is called from the database of the intelligent wheelchair, which contains task records executed by the wheelchair in different scenarios, including complete path trajectories and control instruction data of obstacle avoidance success and obstacle avoidance failure. The scheduling tasks of obstacle avoidance failure are extracted in time sequence through task indexing, and the decision instructions and environmental sensor parameter data corresponding to the obstacle avoidance failure are separated from these task records. The environmental parameters include laser radar scanning point cloud data, ultrasonic distance measurement data and infrared detection results. These obstacle avoidance failure decision data are processed by data vectorization according to the type, position, relative speed of the obstacle and the steering angle of the wheelchair, each failure decision is converted into a multi-dimensional feature vector composed of obstacle features and decision parameters, and a clustering method combining Euclidean distance and cosine similarity is used to analyze the similarity of these vectors. By setting the clustering radius threshold value such as 0.15 and the minimum sample number threshold value such as 10, clustering clusters are divided, and a decision failure clustering logic is formed, which represents the repeated obstacle avoidance failure mode of the wheelchair under similar obstacle dynamic conditions.

[0024] Step S2: Based on the decision failure clustering logic, the deviation analysis of the decision of the sudden obstacle avoidance movement trajectory of the wheelchair is carried out to obtain the decision deviation; based on the decision deviation, the sudden risk prediction lag analysis is carried out to obtain the sudden risk prediction lag regression data;

[0025] In the embodiment of the present application, after obtaining the decision failure clustering logic, first, the dynamic parameters of the obstacles matching the clustering results are called from the database, including the obstacle position coordinates, speed vector, acceleration vector and motion direction angle. The historical scene is restored through these parameters to form an obstacle dynamic scene. The scene records the trajectory points of the obstacles and the relative position of the wheelchair in a continuous time sequence. Then, the decision failure clustering logic is called in the obstacle dynamic scene to compare and analyze the obstacle avoidance behavior of the wheelchair. The time sequence deviation data between the wheelchair and the obstacle motion trajectory in the obstacle avoidance failure process is extracted, including the lateral displacement difference, longitudinal displacement difference and angle deviation of the wheelchair and the obstacle at each time point. The deviation data is weighted and accumulated to form trajectory deduction deviation. The sudden risk prediction lag analysis is performed through the trajectory deduction deviation. In the analysis process, first, the time delay between the decision trigger time of the wheelchair in the obstacle avoidance action and the motion state change time of the obstacle is calculated. The time delay is compared with the failure types in the clustering logic to obtain the typical lag interval under each failure mode. Then, the lag intervals extracted in multiple scenes are regressed to generate sudden risk prediction lag regression data, which is stored in the form of a corresponding relationship table of time delay and obstacle motion intensity for the next step call.

[0026] In another embodiment, based on the obtained decision failure clustering logic, an obstacle dynamic scene is reconstructed in an experimental environment. The scene consists of two types of obstacles: one is a crowd with a speed of 0.8 m / s to 1.2 m / s, and the other is a small moving object with a speed of 0.5 m / s to 1.5 m / s. The obstacle motion trajectory sequence is constructed using the speed, direction and acceleration of the obstacle, and the obstacle avoidance path is simulated using the sensor data of the wheelchair. For each dynamic scene, the avoidance failure state generated based on the decision failure clustering logic is extracted. The actual avoidance trajectory generated by the wheelchair is differentiated from the ideal avoidance trajectory using the time sequence trajectory deviation comparison method. The cumulative deviation of the trajectory in position, speed and turning angle is recorded to obtain the decision deduction deviation data. On this basis, the prediction delay caused by the sudden motion state of the obstacle is calculated. The lag situation within the prediction period of 100 ms to 200 ms is extracted through time window analysis, and the difference between the time point of the actual obstacle avoidance action and the time point in the prediction trajectory is compared to form the sudden risk prediction lag data. The sudden risk prediction lag regression data is obtained by fitting the lag data using linear regression

[0027] Step S3: Pre-deviation correction memory learning of wheelchair scheduling braking is performed according to the sudden risk prediction lag regression data to obtain deviation correction memory learning data. The deviation correction memory learning data is iteratively learned to output decision optimization data.

[0028] In the embodiment of the present application, after obtaining the burst risk prediction lag regression data, first, the data is normalized to scale the time delay and obstacle movement intensity data into the interval [0, 1] to form the burst risk prediction lag normalized data, then the decision deduction deviation is convolved, the convolution kernel length is set to 5, and the convolution method is one-dimensional sliding window operation to extract the continuity feature of the trajectory deviation with time, and the decision convolution deviation is obtained. On this basis, the burst risk prediction lag normalized data and the decision convolution deviation are jointly input into the deviation correction memory learning module. The module obtains the wheelchair braking response delay, maximum braking acceleration, steering angle change rate and avoidance trajectory offset and other decision deviation parameters through the deviation parameter extraction method. Then, the time series regression method is used to analyze the evolution law of the deviation parameters to form the deviation parameter evolution law sequence. The sequence is aggregated and analyzed with the lag normalized data to obtain the correction amount clustering memory data. Further, the correction amount clustering memory data is supplemented and trained by the obstacle burst trajectory learning data to form the burst behavior correction training data. After the correction training is completed, the deviation correction memory learning data is output. Finally, the deviation correction memory learning data is subjected to reinforcement learning iteration. In the iteration process, the correction data is used as a reward signal to optimize the action selection process. The output result after multiple iterations is the decision optimization data, which records the optimal obstacle avoidance trajectory correction scheme and braking trigger advance amount, and is used as the input of subsequent scheduling control of the intelligent wheelchair.

[0029] Preferably, step S1 comprises the following steps:

[0030] Extracting historical scheduling tasks from the database of the intelligent wheelchair, and extracting an obstacle avoidance failure data set from the historical scheduling tasks;

[0031] Random sample selection is performed on the obstacle avoidance failure data set to output an obstacle avoidance failure sample, and then the obstacle avoidance failure decision of the intelligent wheelchair in the sample is extracted;

[0032] Performing logical analysis on the obstacle avoidance failure decision to obtain a decision failure logic;

[0033] Performing clustering analysis on the decision failure logic to generate a decision failure clustering logic.

[0034] In the embodiment of the application, historical scheduling task data is extracted from the database of the intelligent wheelchair. The database adopts a relational data table structure. Each task record contains a task number, an execution time, an execution location, a wheelchair running state, obstacle information collected by a sensor, a wheelchair speed, a wheelchair steering angle, a braking state, and a task completion state. The database capacity is 20,000 historical records. Each record is sampled and stored at a time interval of 100 ms. First, the historical scheduling task data is screened. Only records containing obstacle detection information and marked as obstacle avoidance failure are retained, thereby obtaining an obstacle avoidance failure dataset. The dataset contains multi-dimensional parameters such as real-time position (unit: m), speed (unit: m / s), acceleration (unit: m / s2), steering angle (unit: °), obstacle position (unit: m), obstacle speed (unit: m / s), and obstacle acceleration (unit: m / s2) of the wheelchair in different environments. Then, random sample selection is performed on the obstacle avoidance failure dataset. A fixed-length random sampling method is adopted. Samples are extracted from the obstacle avoidance failure dataset at a proportion of 5%. The number of samples is not less than 1,000. Each sample retains time series data for 10 s to ensure data integrity. The obstacle avoidance failure samples are output and stored in a structured manner. Then, obstacle avoidance failure decision information of the intelligent wheelchair is extracted from the obstacle avoidance failure samples. The decision information is obtained by reading the wheelchair control instruction sequence. The instruction sequence contains a braking trigger time, a braking intensity (unit: m / s2), a steering angle change rate (unit: ° / s), and an acceleration / deceleration threshold (unit: m / s2). The decision information is arranged in time sequence to form an obstacle avoidance failure decision.

[0035] After obtaining the obstacle avoidance failure decision, logical analysis is performed on the obstacle avoidance failure decision. The logical analysis adopts a state transition-based time sequence analysis method. First, a state transition matrix is constructed. The state (speed, steering angle, acceleration, and obstacle relative position) of the wheelchair at each sampling time is taken as a row of the matrix, and the state at the next time is taken as a column of the matrix. The state transition law is obtained by calculating the transition probability of the matrix. Then, the causal relationship between each decision instruction and the dynamic parameters of the obstacle is analyzed. The braking trigger time and the relative distance of the obstacle, the steering angle change rate and the direction change of the obstacle speed are calculated for correlation coefficient. The Pearson correlation coefficient method is used to obtain the correlation values between variables. Variables with a correlation value greater than 0.7 are taken as high-correlation decision features, thereby forming a decision failure logic. The decision failure logic is stored in the form of a multi-dimensional array, recording the state transition law and the high-correlation feature parameters of each obstacle avoidance failure decision.

[0036] After the decision failure logic extraction is completed, the decision failure logic is subjected to cluster analysis, the cluster analysis adopts a hierarchical cluster algorithm based on Euclidean distance, first, a feature vector is calculated for each piece of decision failure logic, the feature vector contains brake trigger time delay (unit: s), steering angle rate of change (unit: ° / s), brake intensity (unit m / s²), obstacle speed direction deviation (unit °), then the Euclidean distance between all feature vectors is calculated, a distance matrix is constructed, and according to the distance matrix, step-by-step merging is performed from the smallest distance, a cluster tree diagram is formed, the distance threshold is set to 0.5, and when the Euclidean distance is less than 0.5, merging is performed into the same class, and finally, multiple decision failure modes are obtained, each class corresponds to a typical failure scene, the cluster result is stored in the form of a logic index table, the feature parameter range and decision behavior sequence of each class of failure mode are recorded in the table, and the index table is used as decision failure cluster logic for calling in subsequent steps.

[0037] Preferably, step S2 comprises the following steps:

[0038] Obstacle dynamic parameters are extracted from the obstacle avoidance failure samples, and a scene restoration is performed to obtain an obstacle dynamic scene, wherein the obstacles include crowds and moving objects.

[0039] Based on the decision failure cluster logic, a sudden obstacle wheelchair avoidance movement trajectory derivation deviation analysis is performed on the obstacle dynamic scene, to obtain a decision derivation deviation;

[0040] Based on the decision derivation deviation and the obstacle dynamic scene, a sudden risk prediction lag analysis is performed to obtain sudden risk prediction lag data;

[0041] The sudden risk prediction lag data is subjected to regression analysis, and sudden risk prediction lag regression data is output.

[0042] As an example of the present application, reference is made to Figure 2 In this example, step S2 comprises:

[0043] Step S201: Obstacle dynamic parameters are extracted from the obstacle avoidance failure samples, and a scene restoration is performed to obtain an obstacle dynamic scene, wherein the obstacles include crowds and moving objects.

[0044] In the embodiment of the application, the obstacle dynamic parameters are extracted from the obtained obstacle avoidance failure samples, the obstacle dynamic parameters include three-dimensional coordinate position (unit: m) of the obstacle at each time sampling point, velocity vector (unit: m / s), acceleration vector (unit: m / s2), motion direction angle (unit: °), obstacle type identification (crowd or moving object) and relative distance (unit: m) between the obstacle and the intelligent wheelchair, the data sampling frequency is 100 Hz, each sample contains 10 s of continuous obstacle motion data, the obstacle motion trajectory is reconstructed according to the time sequence, the interpolation reconstruction algorithm is used to perform linear interpolation completion on the lost time points, the obstacle dynamic scene is formed, the obstacle dynamic scene is stored in the form of a matrix, the row represents the time point and represents the parameter value, the number of crowd type obstacles is greater than 20, the parameters of each individual are recorded independently, the speed range of the moving object type obstacle 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, the continuous trajectory of the obstacle in the plane coordinate system is generated through the reconstructed matrix data and is time-synchronized and aligned with the wheelchair historical trajectory, and the obstacle dynamic scene is obtained.

[0045] Step S202: performing sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis on the obstacle dynamic scene based on the decision failure clustering logic, and obtaining decision deduction deviation;

[0046] 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.

[0047] In another embodiment, the specific embodiment of the sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis based on the decision-making of the obstacle dynamic scene by the decision failure clustering logic is that, first, the clustering logic is called in the restored obstacle dynamic scene, each obstacle dynamic state is input into the clustering logic for matching, a trajectory pattern highly similar to the historical avoidance failure is extracted, and then the intelligent wheelchair is avoided in the trajectory pattern to deduce the path. The deduction process is based on the actual control parameters of the wheelchair, wherein the maximum turning angle of the wheelchair is ±30°, the maximum acceleration is 2.8 m / s², the maximum deceleration is 3.2 m / s², the sampling period is 50 ms, the avoidance trajectory obtained by deduction is compared with the ideal safe avoidance trajectory by time difference, the position deviation, velocity deviation and turning angle deviation are calculated, and the deviation values are integrated and accumulated in the entire time window. Finally, the decision deduction deviation is output, which reflects the degree of insufficient avoidance of the wheelchair under the condition of sudden obstacles. The acquisition of the "ideal safe avoidance trajectory" in the implementation is that, on the basis that the obstacle dynamic scene has been completely restored, the feasible path space is strictly constrained according to the structural limit parameters of the wheelchair (maximum turning angle ±30°, maximum acceleration 2.8 m / s², maximum deceleration 3.2 m / s², minimum safe distance 0.8 m), and then the current and predicted motion state of the obstacle is input into the path space. The trajectory generation algorithm based on the combination of geometric path search and time optimal control (for example, A* search in the grid space and minimum acceleration and deceleration scheduling on the time axis) is used to calculate the obstacle avoidance path that meets the minimum safe distance and can pass within the given response time. Then, the path is sampled at 50 ms to obtain a continuous time sequence that meets the spatial avoidance constraint and the wheelchair dynamics constraint. This time sequence is the ideal safe avoidance trajectory.

[0048] Step S203: sudden risk prediction lag analysis based on decision deduction deviation and obstacle dynamic scene for decision-making to obtain sudden risk prediction lag data;

[0049] In the embodiment of the present application, based on the obtained decision deduction deviation and the obstacle dynamic scene, the sudden risk prediction lag analysis of decision making is first to calculate the time delay (unit: s) between the triggering time of each decision instruction of the wheelchair in the obstacle avoidance process and the time point of the change of the speed and acceleration of the obstacle, to establish a corresponding table of the time delay, the current braking intensity (unit: m / s2) of the wheelchair and the steering angle change rate (unit: ° / s), and then to calculate the distribution density (unit: pieces / m2) of the obstacle in each time slice. For the crowd obstacle, a two-dimensional grid density calculation method is used to divide the scene into 0.5m*0.5m grids to count the number of people in each grid to obtain a density distribution map. The sudden motion state of the obstacle is analyzed, including the speed sudden increase value (unit: m / s), the direction sudden change value (unit: °) and the acceleration sudden change value (unit: m / s2). These sudden motion states are vectorized to form an obstacle motion vector sequence. Through the time stamp in the real-time data stream, the wheelchair decision deduction deviation and the obstacle motion vector sequence are time-aligned to calculate the real-time prediction resource demand, and the maximum decision delay time sequence of the wheelchair in the obstacle avoidance process is extracted according to the real-time prediction resource demand. The sudden risk prediction lag data is obtained by corresponding the time sequence and the decision deduction deviation matrix. The data is stored in the form of a multi-dimensional array for regression analysis.

[0050] Step S204: regression analysis of the sudden risk prediction lag data, outputting sudden risk prediction lag regression data.

[0051] In the embodiment of the present application, the sudden risk prediction lag data is subjected to regression analysis. The regression analysis adopts a multiple linear regression method to take the time delay (unit: s) as the dependent variable, and takes the sudden increase value of the obstacle speed (unit: m / s), the sudden change value of the obstacle direction (unit: °), the sudden change value of the obstacle acceleration (unit: m / s2), the braking intensity of the wheelchair (unit: m / s2) and the steering angle change rate (unit: ° / s) as the independent variables to establish a regression equation to calculate the regression coefficients of the independent variables. The regression coefficients are solved by the least square method, and the iteration number is set to 100 times and the convergence threshold is set to In order to ensure the calculation accuracy, the regression training is performed on all sample data to obtain the sudden risk prediction lag regression data. The data records the time delay characteristics under each type of sudden scene in the form of a parameter weight matrix and is input to the subsequent steps.

[0052] Preferably, the sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis of decision making comprises the following steps:

[0053] According to the decision failure clustering logic, the intelligent wheelchair obstacle avoidance failure state of the sudden obstacle in the obstacle dynamic scene is extracted;

[0054] Extract the back-and-forth cycle trajectory in the obstacle avoidance failure state, analyze the trajectory range nonlinear relationship in the back-and-forth cycle trajectory; wherein the back-and-forth cycle trajectory refers to an invalid motion mode of the intelligent wheelchair in a local space, that is, the motion direction path swings back and forth in a short time; wherein the trajectory range nonlinear relationship refers to the degree of change and irregularity of the activity range of the wheelchair over time in the obstacle avoidance failure process, by analyzing the back-and-forth cycle trajectory, a curve of the wheelchair swinging back and forth in the motion direction path can be constructed;

[0055] From the trajectory range nonlinear relationship, analyze the intermittent start-stop speed change relationship of the intelligent wheelchair; in the trajectory range nonlinear relationship and the speed change relationship, the disorder degree of the wheelchair steering angle is coupled;

[0056] According to the trajectory range nonlinear relationship, the speed change relationship and the disorder degree of the wheelchair steering angle, the avoidance action trajectory time deviation accumulation fitting of the intelligent wheelchair in the obstacle dynamic scene is carried out, and the avoidance trajectory deviation accumulation data is obtained.

[0057] Based on the avoidance trajectory deviation accumulation data, the sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis of decision is carried out, and the decision deduction deviation is obtained.

[0058] In the embodiment of the application, according to the decision failure clustering logic, the intelligent wheelchair obstacle avoidance failure state of the sudden obstacle in the obstacle dynamic scene is extracted, first, the clustering center parameters are read from the decision failure clustering logic database, including the key characteristic values such as the brake trigger delay time value of 0.85 seconds, the steering angle change rate of 15.3° / s, and the obstacle speed direction deviation of 28.7°, then the wheelchair position coordinate sequence and the obstacle position coordinate sequence are extracted in the obstacle dynamic scene data according to the 100ms sampling interval, the Euclidean distance change curve between the two is calculated, when the distance is less than the safety threshold 1.2m and the wheelchair does not execute effective obstacle avoidance action, it is marked as an obstacle avoidance failure point, all obstacle avoidance failure points in the 10s time window are continuously extracted to form an obstacle avoidance failure state sequence, which includes the position coordinates, speed value, acceleration value and steering angle value of the wheelchair in the failure decision state, and the position coordinates, speed value and acceleration value of the obstacle at the corresponding moment, these parameters are organized into a structured data table, the row represents the time point and the column represents the parameter value, so as to obtain the intelligent wheelchair obstacle avoidance failure state.

[0059] The back-and-forth circulating trajectory in the obstacle avoidance failure state is extracted, the trajectory range nonlinear relationship is analyzed in the back-and-forth circulating trajectory, the wheelchair position coordinates in the obstacle avoidance failure state sequence are analyzed in time sequence, a sliding window method is used to set the window length to 2s and the step length to 0.5s, the number of changes of the wheelchair movement direction is calculated in each window, and when the number of changes of the direction is more than 4 times, it is determined that it is a back-and-forth circulating trajectory, all back-and-forth circulating trajectory segments are extracted and merged into a continuous sequence, the Fourier transform is performed on the merged trajectory sequence to extract the frequency spectrum characteristics, the amplitude and phase of the main frequency component are calculated, the amplitude of the main frequency component represents the cycle intensity, and the phase represents the cycle starting point, the trajectory points are projected onto a two-dimensional plane to calculate the envelope curve of the trajectory, the envelope curve is fitted by using a least square method to obtain a trajectory range function, the function represents the change relationship of the wheelchair activity range with time in the obstacle avoidance process, the second derivative of the trajectory range function is calculated to obtain the nonlinear change rate of the trajectory range, and when the nonlinear change rate is greater than 2.5, it indicates that the trajectory range sharply expands, and when the nonlinear change rate is less than 2.5, it indicates that the trajectory range sharply shrinks, and the nonlinear change rate sequence is corresponded to the time stamp to form trajectory range nonlinear relationship data.

[0060] The intermittent start-stop speed change relationship of the intelligent wheelchair is analyzed from the trajectory range nonlinear relationship. First, the wheelchair speed time sequence is differentiated to obtain an acceleration sequence, the acceleration threshold is set to ±0.8 m / s², when the absolute value of the acceleration is greater than the threshold, it is marked as a start or stop point, all start-stop points are continuously extracted to form a start-stop sequence, the time interval between adjacent start-stop points is calculated to obtain a start-stop period, the distribution characteristics of the start-stop period are analyzed, including a mean value of 1.75s, a standard deviation of 0.43s, a maximum value of 2.8s, and a minimum value of 0.9s, the start-stop point time stamp is time-aligned with the trajectory range nonlinear relationship data table, the trajectory range nonlinear change rate at the moment when the start-stop point appears is calculated, the correlation between the start-stop frequency and the trajectory range change rate is calculated by a Pearson correlation coefficient method, the correlation coefficient is 0.78, indicating that the two are in a strong correlation relationship, and a speed change relationship matrix is constructed, the matrix row represents a time point list, represents a speed value, an acceleration value, a start-stop mark and a corresponding trajectory range nonlinear change rate.

[0061] The wheelchair steering angle time sequence is extracted from the back-and-forth cycle trajectory of the obstacle avoidance failure state, the steering angle sequence is segmented according to the sampling point interval of 0.01s, each segment contains 50 continuous sampling points, the local autocorrelation coefficient of each segment steering angle data is calculated, the autocorrelation coefficient reflects the continuity and regularity of the steering angle change with time, the segment with autocorrelation coefficient less than 0.6 is marked as an unordered fluctuation segment, at the same time, the first order difference of the wheelchair speed sequence in the same trajectory is obtained to get the instantaneous acceleration, the standard deviation of the acceleration change sequence is calculated, the segment with standard deviation higher than 0.15m / s² is coupled with the unordered segment, the steering angle, the nonlinear fitting coefficient of the trajectory range and the speed change amplitude are coupled by time alignment method, and the three are integrated into a three-dimensional feature vector, the first dimension of the feature vector is the nonlinear coefficient of the trajectory range, the second dimension is the speed change amplitude, and the third dimension is the steering angle unordered index, wherein the steering angle unordered index is the value of the proportion of the marked segment to the whole trajectory segment, the three-dimensional feature vector is linearly combined by principal component analysis to obtain the coupled steering angle unordered degree coefficient, which is used to represent the irregular fluctuation degree of the steering angle of the wheelchair in the current trajectory segment, and is corresponding to the trajectory range and the speed change relationship matrix, which is used for subsequent avoidance action trajectory time sequence deviation accumulation fitting input.

[0062] According to the trajectory range nonlinear relationship, the speed change relationship and the unordered degree of the wheelchair steering angle, the avoidance action trajectory time sequence deviation accumulation fitting of the intelligent wheelchair in the obstacle dynamic scene is carried out, and the avoidance trajectory deviation accumulation data is obtained. First, a multiple regression model is constructed, the nonlinear change rate of the trajectory range, the speed change rate and the steering angle unordered degree are used as independent variables, and the relative position deviation between the wheelchair and the obstacle is used as the dependent variable. The least square method is used to solve the regression coefficient, and the iteration is calculated for 100 times until convergence. The regression coefficients are 0.42, 0.35 and 0.23, respectively, indicating the contribution weight 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, and the sum of squares of the residual is 0.087, indicating that the fitting accuracy is high. The residual sequence is accumulated according to the time sequence to obtain the cumulative residual curve, which represents the cumulative effect of the avoidance trajectory deviation with time. The cumulative residual curve is smoothed by using the exponential smoothing method, and the smoothing coefficient is set to 0.15. The avoidance trajectory deviation accumulation data is obtained, which is stored in the form of time sequence and contains three components of horizontal cumulative deviation, longitudinal cumulative deviation and angle cumulative deviation.

[0063] The sudden obstacle wheelchair avoidance motion trajectory deviation analysis based on the avoidance trajectory deviation cumulative data for decision making obtains a decision-making deviation. First, the sudden motion characteristics of the obstacle are extracted from the obstacle dynamic scene, including the velocity mutation variable, the direction mutation variable, and the acceleration mutation variable. The sliding difference method is used to calculate the parameter change rate of adjacent time points. When the change rate exceeds the threshold value, it is marked as a sudden point. The velocity mutation threshold is 0.6 m / s, the direction mutation threshold is 25°, and the acceleration mutation threshold is 1.2 m / s². The sudden motion characteristics of the obstacle are time-aligned with the avoidance trajectory deviation cumulative data, and the time delay between them is calculated. The cross-correlation function method is used to determine the time offset corresponding to the maximum correlation. The time offset is in the range of 0.3s to 1.2s, and the average value is 0.75s. The time offset and the avoidance trajectory deviation cumulative data are combined to construct a decision-making deviation matrix. The matrix row represents the time point list representing the lateral deviation value, the longitudinal deviation value, the angle deviation value, and the corresponding time delay value. The main deviation mode is extracted by singular value decomposition method for matrix dimension reduction processing. The singular values with a contribution rate greater than 95% and the corresponding feature vectors are retained to obtain a decision-making deviation feature space. The feature space describes the decision-making deviation distribution law of the wheelchair when facing sudden obstacles.

[0064] Preferably, wherein the avoidance action trajectory timing deviation accumulation fitting of the intelligent wheelchair comprises the following steps:

[0065] The effective avoidance action of the intelligent wheelchair is analyzed through the obstacle dynamic scene;

[0066] The sawtooth trajectory data is obtained by sawtooth trajectory analysis on the trajectory range nonlinear relationship. The trajectory curvature deviation trend is obtained by curvature mutation trend deviation analysis on the sawtooth trajectory data;

[0067] The acceleration / deceleration change slope in the speed change relationship is analyzed;

[0068] The steering angle and speed coupling disorder analysis of the wheelchair steering angle is performed according to the trajectory curvature deviation trend and the acceleration / deceleration change slope, and the steering angle disorder coupling data is obtained;

[0069] The avoidance multidimensional deviation is obtained by avoidance multidimensional deviation random process decomposition of the effective avoidance action according to the trajectory curvature deviation trend, the acceleration / deceleration change slope, and the steering angle disorder coupling data; wherein the multidimensional includes time dimension and space dimension;

[0070] The avoidance action trajectory timing deviation accumulation fitting of the intelligent wheelchair is performed based on the multidimensional avoidance deviation, and the avoidance trajectory deviation cumulative data is obtained.

[0071] In the embodiment of the present application, the effective avoidance action of the intelligent wheelchair in the obstacle dynamic scene is analyzed. Firstly, the wheelchair position coordinate sequence and the obstacle position coordinate sequence are extracted from the obstacle dynamic scene data, the motion trajectory within 10s is recorded, the relative distance change curve between the wheelchair and the obstacle is calculated, the safety distance threshold is set to 1.5m, the avoidance requirement is triggered when the relative distance is less than the safety threshold, the control instruction sequence of the wheelchair in the avoidance process is extracted, including the speed adjustment instruction, the steering angle adjustment instruction and the brake instruction, the instruction sampling interval is 50ms, the time window analysis is performed on the control instruction sequence, the window length is set to 1s, the step is 0.2s, the included angle between the wheelchair position change vector and the obstacle position change vector is calculated in each window, the effective avoidance action is determined when the included angle is greater than 45° and the relative distance increases, all effective avoidance action segments are extracted and combined into a continuous sequence, the avoidance action sequence is characterized, the avoidance amplitude, the avoidance duration and the safety margin after avoidance are calculated, the avoidance amplitude is defined as the maximum lateral displacement of the wheelchair deviating from the original trajectory, the numerical range is between 0.8m and 2.3m, the avoidance duration is defined as the time interval from the avoidance to the normal driving, the numerical range is between 1.2s and 3.5s, the safety margin is defined as the minimum distance between the wheelchair and the obstacle during the avoidance process, the numerical range is between 0.6m and 1.8m, the avoidance action feature vector is constructed, the vector dimension is 15, including the avoidance amplitude, the avoidance duration, the safety margin, the wheelchair speed at the avoidance starting time, the wheelchair acceleration, the wheelchair steering angle, the obstacle relative position, the obstacle relative speed, the obstacle relative acceleration and other parameters, the principal component analysis method is used for dimension reduction processing of the feature vector, the principal components with a contribution rate greater than 95% are retained, the main mode of the avoidance action is obtained, the main mode of the avoidance action is associated with the obstacle motion feature, the cross-correlation function between the two is calculated, the time delay corresponding to the maximum correlation is determined, the time delay range is between 0.2s and 0.9s, the average value is 0.55s, the effective avoidance action data is constructed, the characteristic parameters of the avoidance action, the corresponding obstacle motion feature and the time delay value are recorded, and the basic data is provided for subsequent trajectory deviation analysis.

[0072] The sawtooth trajectory data is obtained by analyzing the nonlinear relationship of the trajectory range. First, the wheelchair position coordinate sequence is extracted from the nonlinear relationship data of the trajectory range, and the motion trajectory within 10s is recorded. The trajectory is projected onto a two-dimensional plane, the directional angle change between adjacent trajectory points is calculated, the directional angle change threshold is set to ±30°, and when the absolute value of the directional angle change is greater than the threshold and the sign alternately changes, it is marked as a sawtooth trajectory point. All sawtooth trajectory points are continuously extracted to form a sawtooth trajectory segment, and the geometric characteristics of the sawtooth trajectory segment are calculated, including sawtooth amplitude, sawtooth period and sawtooth symmetry. The sawtooth amplitude is defined as the lateral displacement between adjacent turning points, the numerical range is between 0.3m and 1.2m, the sawtooth period is defined as the time interval between adjacent turning points in the same direction, the numerical range is between 0.8s and 2.2s, and the sawtooth symmetry is defined as the ratio of left and right turning amplitudes, the numerical range is between 0.65 and 1.35. The wavelet transform method is used for time-frequency analysis of the sawtooth trajectory, Morlet wavelet is selected as the base function, the scale parameter range is set to 1 to 64, the wavelet coefficient matrix is calculated, the peak position and peak intensity of the wavelet energy spectrum are extracted, the peak position represents the main period of the sawtooth, and the peak intensity represents the significance of the sawtooth. The sawtooth trajectory feature vector is constructed, the vector dimension is 12, and the parameters include sawtooth amplitude, sawtooth period, sawtooth symmetry, wavelet energy spectrum peak position and peak intensity. The clustering analysis method is used for classification of the sawtooth trajectory, the K-means algorithm is used, the number of categories is set to 3, which corresponds to weak sawtooth type, medium sawtooth type and strong sawtooth type respectively, the center vector and sample distribution of each category are calculated, the proportion of weak sawtooth type is 25%, the proportion of medium sawtooth type is 45%, and the proportion of strong sawtooth type is 30%. The sawtooth trajectory data table is constructed, the characteristic parameters, category labels and corresponding time stamps of the sawtooth trajectory are recorded, and the input data is provided for subsequent curvature mutation trend analysis.

[0073] The curvature mutation trend offset analysis is performed on the sawtooth trajectory data to obtain the trajectory curvature offset trend. First, the wheelchair position coordinate sequence is extracted from the sawtooth trajectory data table, and the motion trajectory within 10s is recorded. The local curvature of the trajectory is calculated, and the curvature calculation formula is wherein , y', x", y" represent the first and second order derivatives of x and y coordinates respectively, the derivatives are calculated by central difference method with a step size of 3 sampling points, a sliding window analysis is performed on the curvature sequence, the window length is set to 0.5 s and the step size is 0.1 s, the mean, standard deviation and rate of change of curvature are calculated in each window, the mean curvature range is between -2.5 and 2.5, the standard deviation range is between 0.3 and 1.8, the rate of change range is between -3.0 and 3.0, the curvature mutation threshold is set to ±1.5, and the curvature mutation point is marked when the absolute value of the rate of change of curvature is greater than the threshold, all curvature mutation points are extracted to form a mutation sequence, the time interval and curvature change amplitude between adjacent mutation points are calculated, the time interval range is between 0.3 s and 1.5 s, the curvature change amplitude range is between 1.2 and 4.5, the linear regression method is used to analyze the time trend of the curvature mutation sequence, the regression equation is , wherein represents the curvature at time t, and b are regression coefficients, the regression coefficients are solved by least squares method, and the iteration is calculated for 50 times until convergence, the regression coefficients The value of a ranges between -0.8 and 0.8, the value of b ranges between -1.5 and 1.5, The value greater than 0 indicates that the curvature increases, The value less than 0 indicates that the curvature decreases, the residual sequence of the regression equation is calculated, the residual standard deviation range is between 0.4 and 1.2, the residual standard deviation greater than 0.8 indicates that the curvature changes are unstable, the autoregressive moving average model is used to analyze the time correlation of the curvature sequence, the model order is set to ARMA(2, 1), the model parameters and prediction errors are calculated, the curvature mutation trend feature vector is constructed, the vector dimension is 10, including the mean, standard deviation, rate of change, mutation frequency, regression coefficient, residual standard deviation, ARMA model parameters, etc., the principal component analysis method is used to extract the main change mode, the principal components with a contribution rate greater than 90% are retained, and the trajectory curvature deviation trend data is obtained, which is stored in the form of time sequence and contains the curvature value, curvature change rate, curvature mutation mark and trend prediction value, providing input data for subsequent avoidance trajectory deviation analysis.

[0074] The acceleration / deceleration change slope in the speed change relationship is analyzed. First, the wheelchair speed time sequence is extracted from the speed change relationship, the speed change in the next 10s is recorded, the sliding window analysis is performed on the speed sequence, the window length is set to 0.8s, the step length is 0.2s, the linear regression equation of the speed in each window is calculated, the equation form is v(t)=kt+c, wherein v(t) represents the speed at time t, k and c are the regression coefficients, the regression coefficients are solved by the least square method, and the regression coefficients k are iteratively calculated for 30 times until convergence. The regression coefficient k is the speed change slope, which represents the average value of acceleration. The value of k is greater than 0, indicating an acceleration process. The value of k is less than 0, indicating a deceleration process. The absolute value of k represents the intensity of acceleration / deceleration, and the numerical range is between 0.1 m / s² and 2.5 m / s². The goodness of fit R² of the regression equation is calculated. The R² value ranges between 0.75 and 0.98. The R² value is greater than 0.9, indicating that the speed change presents good linear characteristics. The speed sequence is differentiated to obtain the acceleration sequence. The central difference method is adopted, and the difference step length is 5 sampling points. The statistical characteristics of the acceleration sequence are calculated, including mean, standard deviation, skewness and kurtosis. The mean range is between-1.5 m / s² and 1.5 m / s². The standard deviation range is between 0.2 m / s² and 1.0 m / s². The skewness range is between-1.2 and 1.2. The kurtosis range is between 2.0 and 5.0. The acceleration threshold is set to ±0.8 m / s². When the absolute value of acceleration is greater than the threshold, it is marked as a significant acceleration / deceleration point. 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. The time interval ranges between 0.4s and 1.8s. The acceleration change amplitude ranges between 0.6 m / s² and 2.2 m / s². The piecewise linear fitting method is used to model the acceleration sequence. The sequence is divided into multiple linear segments. The slope of each segment represents the acceleration value, i.e. the acceleration change rate. The acceleration value ranges between-3.0 m / s² and 3.0 m / s². An acceleration / deceleration feature vector is constructed, and the vector dimension is 14, including speed change slope, goodness of fit, acceleration statistical characteristics, acceleration / deceleration frequency, jerk value and other parameters. The clustering analysis method is used to classify the acceleration / deceleration mode. The K-means algorithm is used, and the number of categories is set to 4, corresponding to stable type, slow change type, rapid change type and mixed type respectively. The center vector and sample distribution of each category are calculated. The stable type accounts for 20%, the slow change type accounts for 35%, the rapid change type accounts for 25%, and the mixed type accounts for 20%. An acceleration / deceleration change slope data table is constructed, which records the characteristic parameters, category labels and corresponding time stamps of acceleration / deceleration, providing input data for subsequent avoidance trajectory deviation analysis.

[0075] The curvature value sequence and the curvature change rate sequence are extracted from the trajectory curvature deviation trend data table, the speed change slope sequence and the acceleration sequence are extracted from the acceleration / deceleration change slope, the parameter changes within 10 seconds are recorded, these sequences are time-aligned with the wheelchair steering angle sequence, a multivariate time sequence matrix is constructed, the matrix dimension is 1000x5, the row represents the time point, and the column represents the curvature value, the curvature change rate, the speed change slope, the acceleration value and the steering angle value respectively, the cross-correlation function between each variable is calculated, the time delay corresponding to the maximum correlation is determined, the time delay between the steering angle and the curvature value is 0.15s, the time delay between the steering angle and the curvature change rate is 0.08s, the time delay between the steering angle and the speed change slope is 0.22s, and the time delay between the steering angle and the acceleration value is 0.18s, the independent influence of each variable on the steering angle is evaluated by using the partial correlation analysis method, the partial correlation coefficient is calculated, the partial correlation coefficient of the curvature value is 0.72, the partial correlation coefficient of the curvature change rate is 0.65, the partial correlation coefficient of the speed change slope is 0.58, and the partial correlation coefficient of the acceleration value is 0.63, the phase diagram of the steering angle and other variables is constructed, the phase difference and the phase synchronization index are calculated, the phase difference ranges between and , the phase synchronization index ranges between 0.4 and 0.9, the phase synchronization index less than 0.6 indicates that there is a significant phase disorder, the Granger causality test method is used to analyze the causal relationship between the variables, the lag order is set to 5, the F statistic and the p value are calculated, the F statistic greater than 4.0 and the p value less than 0.05 indicate that there is a significant causal relationship, the vector autoregression model is constructed, the model order is VAR(3), the model parameter matrix is estimated, the prediction error and the information criterion of the model are calculated, the model with the minimum AIC information criterion is selected as the optimal model, the impulse response function of the model is calculated, the dynamic influence of the impact of one variable on other variables is analyzed, the steering angle disorder index is constructed, the index calculation formula is wherein , , respectively represent the actual steering angle, the angular velocity and the angular acceleration, , , represent the model predicted value, , , represent the weight coefficient, the weight coefficient is set to 0.5, 0.3 and 0.2 respectively, the disorder index ranges between 0 and 10, and the index greater than 5 indicates that there is a serious steering disorder, the steering angle disorder coupling data is constructed, and the steering angle disorder index, the phase difference, the phase synchronization index, the Granger causality test result and the corresponding time stamp are recorded.

[0076] The multi-dimensional deviation random process of the effective avoidance action is decomposed according to the trajectory curvature deviation trend, the change slope of acceleration / deceleration and the steering angle misalignment coupling data to obtain the multi-dimensional avoidance deviation. First, the avoidance action feature vector is extracted from the effective avoidance action, the curvature feature vector is extracted from the trajectory curvature deviation trend, the speed feature vector is extracted from the change slope of acceleration / deceleration, and the misalignment feature vector is extracted from the steering angle misalignment coupling data table. The feature vectors are combined into a comprehensive feature matrix with a dimension of 1000*40. The row represents the time point, and the column represents each feature parameter. The feature matrix is standardized to have a zero mean and a unit variance. The principal component analysis method is used to reduce the dimension of the standardized feature matrix, and the principal components with a contribution rate greater than 95% are retained to obtain the reduced feature matrix with a dimension of 1000*12. The avoidance deviation random process model is constructed. The Gaussian process regression method is used, the kernel function is selected as the radial basis function, the kernel width parameter is set to 0.8, and the noise variance parameter is set to 0.05. The Gaussian process model is trained, and the iteration is calculated for 200 times until convergence. The log-likelihood value and the prediction variance of the model are calculated. If the log-likelihood value is greater than -500, the model fitting is good. The avoidance deviation random process is decomposed into time dimension and space dimension components. The time dimension component represents the time sequence deviation of the avoidance action, including avoidance start time deviation, avoidance duration deviation and avoidance recovery time deviation. The space dimension component represents the spatial deviation of the avoidance action, including avoidance amplitude deviation, avoidance trajectory deviation and avoidance endpoint deviation. The singular spectrum analysis method is used to decompose the time dimension component. The embedding window length is set to 50 sampling points. The first three singular values and the corresponding singular vectors are extracted. The time dimension deviation sequence is reconstructed. The wavelet decomposition method is used for multi-scale analysis of the space dimension component. The Daubechies wavelet is selected, and the decomposition level is 4. The wavelet coefficients at each scale are calculated. The space dimension deviation sequence is reconstructed. The time dimension deviation and the space dimension deviation are combined into a multi-dimensional avoidance deviation data table. The time deviation parameters, the space deviation parameters and the corresponding time stamps and space coordinates are recorded. The input data is provided for subsequent trajectory time sequence deviation accumulation fitting.

[0077] The avoidance action trajectory time sequence deviation accumulation fitting of the intelligent wheelchair is performed based on the multi-dimensional avoidance deviation to obtain the avoidance trajectory deviation accumulation data. First, the time deviation sequence and the space deviation sequence are extracted from the multi-dimensional avoidance deviation data. The deviation changes within 10s are recorded. The time deviation sequence is accumulated and summed to obtain the time deviation accumulation curve. The space deviation sequence is vector accumulated to obtain the space deviation accumulation curve. The exponential smoothing method is used to smooth the accumulation curve. The smoothing coefficient is set to 0.12 to reduce the influence of random fluctuations. The deviation accumulation model is constructed, and the model form is: wherein represents the accumulation deviation at time . denotes the initial bias, denotes the instantaneous bias at time t-i, and denotes the weight coefficient and the decay coefficient, denotes the time delay, the model parameters are estimated by least squares method, the iteration is calculated for 150 times until convergence, the weight coefficient has a value range of 0.6 to 1.2, and the decay coefficient has a value range of 0.05 to 0.25, the fitting error and the prediction error of the calculation model are calculated, the root mean square value of the fitting error is less than 0.15, which indicates that the model fitting precision is high, the trend of the bias accumulation curve is extracted by using the adaptive filtering method, the filter type is Kalman filter, and the smooth bias accumulation trend curve is obtained after filtering, and a space-time joint bias accumulation model is constructed, and the model form is wherein denotes the joint cumulative bias at time t and space position s, denotes the pure time bias component, denotes the pure space bias component, denotes the space-time interaction bias component, the model parameters are estimated by using tensor decomposition method, the joint bias data is organized as a three-dimensional tensor, the tensor dimension is 100x100x3, which respectively represents time point, space point and bias component, the Tucker decomposition method is used, the core tensor rank is set to (5, 5, 2), the iteration is calculated for 200 times until convergence, the joint bias accumulation field is reconstructed, the reconstruction error and the explained variance ratio are calculated, and the explained variance ratio is greater than 90%, which indicates that the decomposition effect is good, and finally the avoidance trajectory bias accumulation data is obtained, the data is stored in the form of multi-dimensional array, including time dimension cumulative bias, space dimension cumulative bias and space-time joint cumulative bias.

[0078] Preferably, the decision-making sudden risk pre-judgment lag analysis includes the following steps:

[0079] Obtain the hardware configuration parameters of the intelligent wheelchair, wherein the hardware includes CPU, GPU and memory;

[0080] Calculate the distribution density of the obstacles in the dynamic scene of the obstacles, and analyze the sudden motion state of the obstacles; wherein the sudden motion state includes velocity, direction, and acceleration;

[0081] Obtain the real-time prediction resource demand by performing obstacle motion vector prediction derivation on the sudden motion state of the obstacles, and then performing real-time data processing resource demand amount evaluation;

[0082] Quantify the calculation bottleneck of the intelligent wheelchair by calculating the real-time prediction resource demand according to the hardware configuration parameters, and obtain the calculation bottleneck quantization data of the intelligent wheelchair;

[0083] Extract the time sequence with the maximum decision delay based on decision deduction deviation;

[0084] According to the calculation bottleneck quantization data and the time sequence with the maximum decision delay, the burst risk pre-judgment lag analysis is performed to obtain burst risk pre-judgment lag regression data.

[0085] In the embodiment of the application, the hardware configuration parameters of the intelligent wheelchair are obtained, wherein the hardware includes CPU, GPU and memory. First, the hardware configuration information of the intelligent wheelchair control unit is read through the system information interface. The CPU model, core number, main frequency and cache size are obtained by using a standard system call function. The specific CPU parameters read include the model, the core number is 4 cores and 8 threads, the basic frequency is 1.6 GHz, the maximum turbo frequency is 3.9 GHz, the three-level cache is 6 MB, and the instruction set supports AVX2. Then, the GPU configuration information is read, including the GPU model, the CUDA core number is 256, the GPU frequency is 1.3 GHz, the video memory capacity is 8 GB, the computing power is 5.3 TFLOPS, and the support is CUDA 10.2 and TensorRT 7.1 acceleration library. Then, the memory configuration information is obtained, including the memory capacity of 16 GB, the memory type of LPDDR4, the memory frequency of 2133 MHz, and the memory bandwidth of 34.1 GB / s in dual-channel mode. The storage device information is also read, including the main storage of 256 GB SSD with a read speed of 550 MB / s and a write speed of 520 MB / s, and the auxiliary storage of 64 GB eMMC with a read speed of 280 MB / s and a write speed of 250 MB / s. The hardware resource usage state is collected in real time through the system monitoring interface, including CPU utilization, GPU utilization, memory occupancy and storage I / O load, with a sampling frequency of 10 Hz. The resource usage fluctuation within 30 seconds is recorded. The average usage, peak usage and standard deviation of each resource are calculated. The CPU average utilization is 45%, the peak is 78%, and the standard deviation is 12%. The GPU average utilization is 35%, the peak is 65%, and the standard deviation is 15%. The memory average occupancy is 42%, the peak is 60%, and the standard deviation is 8%. The hardware resource configuration parameter table is constructed, and the specification parameters and resource usage state of each hardware component are recorded to provide basic data for subsequent load evaluation.

[0086] The distribution density of obstacles in the dynamic obstacle scene is calculated, and the sudden motion state of the obstacles is analyzed. First, the obstacle position coordinate sequence is extracted from the dynamic obstacle scene data, the obstacle distribution in the next 10s is recorded, the scene space is divided into 0.5m×0.5m grid units, the number of obstacles in each grid unit at each time point is calculated, and a time-varying density distribution matrix is constructed. The matrix has dimensions of 20×20×1000, representing the number of x-direction grid, y-direction grid and time point, respectively. The global average density is 0.08 / , the maximum local density is 1.5 / , which appears in the center area of the scene. The kernel density estimation method is used to generate the continuous density distribution function, the Gaussian kernel is selected as the kernel function, and the bandwidth parameter is set to 0.75m. The density gradient field is calculated, the gradient direction points to the direction of the fastest density increase, and the gradient amplitude represents the density change rate. When the gradient amplitude is greater than 0.5, it indicates that there is a significant density change area. The time difference of the obstacle position sequence is calculated to obtain the velocity vector of the obstacle. The velocity vector contains two components: size and direction. The average value of the velocity size is 1.2m / s, the maximum value is 3.5m / s, and the standard deviation is 0.65m / s. The velocity direction is distributed in the range of 0 to 360°, mainly concentrated in the directions of 45°, 135°, 225° and 315°. The time difference of the velocity vector is calculated to obtain the acceleration vector of the obstacle. The average value of the acceleration vector is 0.8m / s², the maximum value is 2.7m / s², and the standard deviation is 0.55m / s². The sudden motion threshold is set, the velocity mutation threshold is 1.5m / s², the direction mutation threshold is 45°, and the acceleration mutation threshold is 2.0m / s². When the parameter changes exceed the threshold, it is marked as a sudden motion point. The sudden motion state sequence is formed by extracting all the sudden motion points. The frequency and spatial distribution characteristics of the sudden motion are calculated. The sudden motion frequency is 0.8 times / s, mainly distributed in the edge area of the scene. The obstacle distribution density and sudden motion state data table is constructed, which records the space-time density distribution, velocity distribution, direction distribution, acceleration distribution and sudden motion label.

[0087] The sudden motion state of the obstacle is derived for obstacle motion vector prediction, and then real-time data processing resource demand is evaluated to obtain real-time prediction resource demand. First, the historical trajectory data of the obstacle is extracted from the obstacle distribution density and the sudden motion state, and the motion trajectory within 10s is recorded. A state vector is constructed for each obstacle, which includes position coordinates, velocity vector, acceleration vector and motion direction angle. Kalman filtering algorithm is used for trajectory prediction. The prediction parameters are set based on the uniform acceleration motion model, the position prediction error is controlled within 0.05m, the velocity prediction error is controlled within 0.15m / s, and the acceleration prediction error is controlled within 0.25m / s². The trajectory of the obstacle within the next 5s is predicted, the prediction step is 100ms, and 50 prediction points are generated. The uncertainty of the predicted trajectory is analyzed, and the 95% confidence interval is calculated. The position prediction error increases with the increase of the prediction time. The average error after 1s is 0.3m, the average error after 3s is 0.8m, and the average error after 5s is 1.5m. The sudden motion point is processed, that is, the particle filtering algorithm is used, the number of particles is set to 1000, the prediction model parameters are adjusted through the historical sudden intensity, the multimodal distribution of the predicted trajectory is calculated, the three most probable trajectories and their probabilities are extracted, the calculation complexity of real-time prediction is evaluated, the time complexity and space complexity of the algorithm are analyzed, the number of floating point operations for single prediction is calculated, Kalman filtering is about 5000 times, and particle filtering is about 500000 times. According to the number of obstacles and the sudden motion frequency in the scene, the overall calculation load is estimated. The average load is 20 million floating point operations per second, and the peak load is 100 million floating point operations per second. Considering the data transmission and storage overhead, the memory bandwidth requirement is evaluated as 200MB per second, and the storage space requirement is 20MB per second. A real-time prediction resource demand data table is constructed, which records the calculation load, memory bandwidth, storage space and their relationship with the number of obstacles and the sudden frequency, providing a basis for subsequent calculation bottleneck quantification.

[0088] The real-time predicted resource demand is quantified according to the hardware configuration parameters to obtain the calculation bottleneck quantization data of the intelligent wheelchair. First, the 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 proportion of each hardware component. The CPU load proportion is the actual calculation load divided by the CPU computing power. When the number of obstacles is 20 and the burst frequency is 0.8 times per second, the CPU load proportion is 35.7%. When the number of obstacles increases to 50 and the burst frequency is 2.0 times per second, the CPU load proportion rises to 89.3%. The GPU load proportion is the actual calculation load divided by the GPU computing power. When the parallel prediction algorithm is executed, the GPU load proportion is 1.9%. The memory bandwidth load proportion is the actual bandwidth demand divided by the memory bandwidth. When the data transmission rate is 200 MB / s, the memory bandwidth load proportion is 0.6%. A load proportion matrix is constructed, with different combinations of obstacle number and burst frequency as rows and the load proportions of CPU, GPU and memory as columns. A bottleneck identification algorithm is used to analyze the load proportion matrix to 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 per second. The GPU saturation critical point is far from being reached. The memory bandwidth saturation critical point is a data transmission rate of 34 GB / s, which is much higher than the current demand. It is determined that CPU is the main bottleneck component. The CPU bottleneck coefficient is calculated. The bottleneck coefficient is defined as the ratio of the load proportion to the critical load proportion. A bottleneck coefficient greater than 0.8 indicates that it is close to the bottleneck state. Through multiple experimental data fitting, the relationship between the bottleneck coefficient and the number of obstacles and the burst frequency is fitted, with fitting parameters of 0.012, 0.085 and 0.007. The fitting error is less than 5%. The fitted relationship is used to predict the bottleneck coefficient in different scenarios to generate a bottleneck coefficient surface graph. The horizontal axis represents the number of obstacles, the vertical axis represents the burst frequency, and the surface height represents the bottleneck coefficient. The bottleneck coefficient contour line is extracted. The contour line with a bottleneck coefficient of 0.8 represents the safe operation boundary. The calculation bottleneck quantization data of the intelligent wheelchair is constructed, recording the bottleneck coefficient, load proportion and safety margin under different scene parameters.

[0089] The time sequence with the maximum decision delay is extracted based on decision deduction deviation. First, the timestamp sequence and the corresponding deviation value sequence are extracted from the decision deduction deviation. The threshold detection is performed on the deviation value sequence. The deviation threshold is set as the horizontal deviation 0.3 m, the vertical deviation 0.3 m and the angle deviation 15°. When any deviation component exceeds the threshold, it is marked as a decision deduction time point. The time interval between adjacent time points is calculated. The average value of the interval is 0.85 s, and the standard deviation is 0.32 s. The first-order difference is performed on the time interval sequence to construct the difference sequence matrix. The K-means clustering algorithm is used to divide the difference sequence into a small delay class, a medium delay class and a large delay class. The average value of the time interval difference of the small delay class is 0.15 s, the average value of the medium delay class is 0.45 s, and the average value of the large delay class is 0.85 s. The secondary screening is performed on the large delay class samples. The screening threshold is set as the horizontal deviation difference 0.2 m. When the deviation difference exceeds the threshold, the sample is retained. The time sequence with the maximum decision delay is extracted from all samples that meet the conditions. The average delay time of the sequence is 1.05 s, and the maximum delay time is 1.85 s. The role is to find the key time period of the wheelchair in the obstacle avoidance process, and to provide an accurate target time window for the lag analysis of the sudden risk prediction.

[0090] The burst risk pre-judgment lag analysis according to the calculation bottleneck quantization data and the time sequence with the maximum decision delay is performed to obtain burst risk pre-judgment lag regression data. First, the bottleneck coefficient sequence is extracted from the calculation bottleneck quantization data, the delay time sequence is extracted from the time sequence data table with the maximum decision delay, the two sequences are matched according to the scene parameters, a joint data matrix is constructed, the matrix row represents different scene parameter combinations, and the list represents the bottleneck coefficient, the delay time, the deviation value and the obstacle burst intensity. A multivariate regression analysis method is used to establish a lag pre-judgment relationship. The relationship between the pre-judgment lag time and the obstacle quantity, the burst frequency and the bottleneck coefficient is obtained through data fitting. The fitting parameters include a constant term 0.35s, an obstacle quantity coefficient 0.008s / individual, a burst frequency coefficient 0.12, a bottleneck coefficient coefficient 0.45s / unit bottleneck coefficient, and three interaction term coefficients 0.003, 0.006 and 0.08. The goodness of fit is 0.85, which means that the fitting relationship explains 85% of the variation of the lag time. The root mean square value of the prediction error is 0.15s. The cross-validation of the fitting relationship is performed by using a 10-fold cross-validation method. The root mean square value of the validation error is 0.18s, which is slightly higher than the training error, indicating that the fitting relationship has good generalization ability. A lag pre-judgment response surface is constructed. The horizontal axis is the obstacle quantity, the vertical axis is the burst frequency, and the surface height represents the pre-judgment lag time. Multiple response surfaces are generated under different bottleneck coefficients to analyze the variation of the lag time with the scene parameters. When the bottleneck coefficient increases from 0.3 to 0.8, the average lag time increases from 0.55s to 0.95s, with an increase of 72.7%. The gradient field of the lag time is calculated. The gradient direction points to the parameter combination with the fastest increase of the lag time, and the gradient amplitude represents the sensitivity of the lag time. The significance of the fitting parameters is tested by calculating the t-statistic and the p-value. The p-value of all parameters is less than 0.05, indicating that the parameters have statistical significance. The burst risk pre-judgment lag regression data is constructed.

[0091] Preferably, step S3 comprises the following steps:

[0092] The decision evolution deviation is convolved to obtain a decision convolution deviation. The burst risk pre-judgment lag regression data is normalized to obtain burst risk pre-judgment lag normalized data.

[0093] The wheelchair scheduling braking pre-deviation correction memory learning is performed on the decision convolution deviation according to the burst risk pre-judgment lag normalized data to obtain deviation correction memory learning data.

[0094] The deviation correction memory learning data is iterated by reinforcement learning to output decision optimization data.

[0095] As an example of the present application, referring to Figure 3 In this example, step S3 comprises:

[0096] Step S301: Convolution processing is performed on the decision deduction deviation to obtain a decision convolution deviation; and normalization processing is performed on the burst risk prediction lag regression data to obtain burst risk prediction lag normalized data.

[0097] In the embodiment of the application, the decision deduction deviation is subjected to convolution processing to obtain a decision convolution deviation, and the burst risk prediction lag regression data is subjected to normalization processing to obtain burst risk prediction lag normalized data. Firstly, a deviation sequence is extracted from the decision deduction deviation, and the change of the deviation within 10 seconds is recorded. A one-dimensional convolution kernel is designed, the kernel length is 5, and the kernel weight is [0.1, 0.2, 0.4, 0.2, 0.1]. Convolution operation is performed on the deviation sequence, a sliding window method is adopted, the step length is 1, the boundary is subjected to mirror filling processing, the convolution calculation formula is that the current output is equal to the inner product of the input sequence and the convolution kernel, the local features and the time continuity of the deviation sequence are extracted through the convolution operation, the decision convolution deviation sequence is obtained, and the maximum and minimum value normalization processing is performed on the burst risk prediction lag regression data, so that the numerical range is scaled to the interval [0, 1]. The normalization formula is (x-min) / (max-min), wherein x is the original value, min and max are respectively the minimum value and the maximum value of the data, the burst risk prediction lag normalized data is obtained after processing, the data dimension remains unchanged, the numerical range is unified to the interval [0, 1], and subsequent processing is facilitated.

[0098] Step S302: Pre-deviation correction memory learning of wheelchair scheduling braking is performed on the decision convolution deviation according to the burst risk prediction lag normalized data to obtain deviation correction memory learning data.

[0099] In the embodiment of the application, the pre-deviation correction memory learning of the wheelchair scheduling braking is performed according to the burst risk prediction lag normalized data, deviation correction memory learning data is obtained, a memory learning network is first constructed, the network structure is a three-layer feedforward neural network, the number of input layer neurons is 15, corresponding to the combined features of the decision convolution deviation and the burst risk prediction lag normalized data (wherein the burst risk prediction lag normalized data includes the prediction delay of the appearance time, position, speed and acceleration of the obstacle in the decision made by the intelligent wheelchair in the dynamic environment, and the wheelchair state such as the current speed, acceleration, steering angle and braking response delay), the number of hidden layer neurons is 30, the activation function is ReLU, and the number of output layer neurons is 8, corresponding to the corrected braking parameters, including the braking advance, the braking intensity, the steering angle and the steering rate, the network is trained by using the stochastic gradient descent method, the initial value of the learning rate is set to 0.01, the exponential decay strategy is used, the decay is 0.9 times of the original value every 1000 steps, the batch size is set to 64, the training iteration number is 10000 times, the loss function uses the mean square error, the memory mechanism is introduced in the training process, the historical samples are stored in the experience replay buffer, the buffer size is 1000, 128 historical samples are randomly sampled each time to mix with the current batch samples, the memory capacity and the generalization ability of the network are enhanced, and the deviation correction memory learning data is obtained after the training is completed.

[0100] Step S303: performing reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

[0101] In the embodiment of the application, the reinforcement learning iteration is performed on the deviation correction memory learning data to output the decision optimization data, a reinforcement learning environment is first constructed, the state space is the wheelchair position, speed, steering angle and surrounding obstacle distribution, the action space is the adjustment amount of the braking intensity and the steering angle, the reward function is designed based on the safety distance, the smoothness and the target achievement degree, the deep Q network algorithm is used to realize the reinforcement learning, the Q network structure is a four-layer fully connected network, the input layer dimension is the state space dimension, the output layer dimension is the action space dimension, the number of hidden layer neurons is 128 and 64 respectively, the activation function is ReLU, the ε-greedy strategy is used for exploration and utilization balance, the initial ε value is 0.9, gradually decays 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, the training iteration number is 50000 times, the average reward and the success rate are recorded every 1000 iterations, the training is stopped when the average reward of 5 consecutive evaluations is less than 1%, and finally the decision optimization data is output, including the optimal braking time, the braking intensity, the steering angle and the adjustment strategy thereof.

[0102] Preferably, the wheelchair scheduling bias correction memory learning of the brake includes the following steps:

[0103] The decision convolution bias is subjected to bias parameter extraction to obtain decision bias parameters, wherein the parameters include wheelchair brake response delay, maximum brake acceleration, steering angle and avoidance trajectory;

[0104] The decision bias parameters are subjected to evolution law regression analysis to generate bias parameter evolution law;

[0105] Based on the bias parameter evolution law and the decision lag normalized data, bias correction safety margin approximation clustering is performed, and memory learning is performed, so as to obtain correction amount clustering memory data; wherein the bias correction safety margin refers to the evolution law of the key parameters such as brake response delay, maximum brake acceleration, steering angle and avoidance trajectory of the wheelchair, and by real-time correction and optimization of these parameters, it is ensured that the wheelchair can maintain sufficient safety distance and reaction time in various scenes, so as to quantify the dynamic safety buffer reserved by the wheelchair when facing sudden obstacles;

[0106] According to the correction amount clustering memory data and the sudden risk prediction lag normalized data, the wheelchair scheduling brake bias correction memory learning is performed to obtain bias correction memory learning data.

[0107] In the embodiment of the application, the decision convolution bias is subjected to bias parameter extraction to obtain decision bias parameters, first, the key feature points are extracted from the decision convolution bias, the feature point extraction threshold is set to 1.5 times the standard deviation of the bias value, when the bias value exceeds the threshold, it is marked as a feature point, the time distribution of the feature point is calculated to obtain the brake response delay parameter, the brake response delay is defined as the time interval from the appearance of the obstacle to the start of the brake, the average value is 0.72s, and the standard deviation is 0.18s, the acceleration change is calculated by the second order difference of the bias sequence, the maximum brake acceleration parameter is extracted, the average value of the maximum brake acceleration is 1.85m / s², and the range is between 1.2 and 2.5 m / s², the steering angle change is extracted from the bias, the steering angle is defined as the included angle between the forward direction of the wheelchair and the target direction, the average maximum steering angle is 28°, and the standard deviation is 7.5°, the trajectory deviation is calculated by the integral of the position bias, the avoidance trajectory parameter is extracted, including the maximum lateral offset 1.2m, the trajectory length 2.8m, and the trajectory curvature 0.35 , a decision bias parameter vector is constructed, the vector dimension is 12, and the statistical characteristics of the above four types of parameters are included, which provides input data for subsequent evolution law analysis.

[0108] The evolutionary law regression analysis is performed on the decision bias parameters to generate a bias parameter evolution law. First, the decision bias parameters are arranged in time sequence to construct a time series matrix, with the matrix dimension being 1000x12, the row representing the time point, and the column representing each bias parameter. Trend analysis is performed on the time series, and the long-term trend is extracted using the moving average method with a window length of 50 sampling points. The trend slope of each parameter is calculated, with the brake response delay trend slope being 0.002 s, the maximum brake acceleration trend slope being -0.05 m / s², the steering angle trend slope being 0.15° / s, and the maximum lateral offset trend slope of the avoidance trajectory being 0.01 m / s. Periodic analysis is performed on the time series using the autocorrelation function method, and the autocorrelation coefficients of each parameter are calculated. The main period of the brake response delay is 1.8 s, the main period of the maximum brake acceleration is 2.2 s, the main period of the steering angle is 1.5 s, and the main period of the avoidance trajectory is 2.5 s. The polynomial regression method is used to fit the parameter evolution curve, with the polynomial order being 3. The goodness of fit R² is 0.82, 0.78, 0.85, and 0.76, respectively. The bias parameter evolution law is constructed, and the trend slope, periodic characteristics, and regression coefficients of each parameter are recorded to provide a law basis for subsequent safety margin clustering.

[0109] Based on the bias parameter evolution law and the decision lag normalized data, the bias correction safety margin is approximately clustered, and memory learning is performed to obtain the correction amount clustering memory data. First, the bias parameter evolution law and the decision lag normalized data are fused to construct a fusion feature vector, with the vector dimension being 20 and containing the evolution law parameters and the lag normalized data. Principal component analysis is performed on the fusion features to extract the main feature axes, and the principal components with a contribution rate greater than 95% are retained. The dimension of the reduced features is 8. The safety margin is calculated, and the safety margin is defined as the ratio of the brake response delay correction amount to the obstacle speed, with the unit being m / s². The average safety margin is 0.45 m / s², and the standard deviation is 0.12 m / s². The fuzzy C-means clustering algorithm is used to cluster the safety margin, with the cluster number being set to 5, the fuzziness parameter being 2.0, the iteration number being 100, and the convergence threshold being 0.001. The cluster centers and sample membership degrees are calculated, with the cluster centers being 0.25, 0.35, 0.45, 0.55, and 0.65 m / s², respectively. A memory learning network is constructed, with the network structure being a four-layer feedforward neural network, the input layer dimension being 8, the hidden layer dimensions being 16 and 12, respectively, and the output layer dimension being 5, corresponding to the membership degrees of the five clusters. The training sample number is 1000, the learning rate is 0.005, the batch size is 32, the training iteration number is 5000, and the loss function is cross-entropy. After training, the correction amount clustering memory data is obtained, including the cluster centers, sample distribution, and network parameters.

[0110] The wheelchair scheduling deviation correction memory learning is performed according to the correction amount clustering memory data and the sudden risk prediction lag normalized data, deviation correction memory learning data is obtained, first, the clustering center and the weight matrix are extracted from the correction amount clustering memory data, the lag time and the obstacle motion parameter are extracted from the sudden risk prediction lag normalized data (wherein the sudden risk prediction lag normalized data includes the prediction delay of the appearance time, position, speed and acceleration of the obstacle in the decision made by the intelligent wheelchair in the dynamic environment, and the wheelchair state such as the current speed, acceleration, steering angle and brake response delay), a joint feature vector is constructed, the vector dimension is 25, including the clustering feature and the lag feature, a memory learning algorithm is designed, a radial basis function network structure is adopted, the input dimension is 25, the number of hidden layer nodes is 40, the output dimension is 8, corresponding to the mean and standard deviation of four types of correction parameters, the radial basis function selects the Gaussian kernel, the kernel width parameter is adaptively set, which is inversely proportional to the sample density, the number of training samples is 1500, the training set and the validation set are divided by the cross-validation method, the proportion is 8:2, the gradient descent method is used for training, the learning rate is 0.008, the momentum factor is 0.6, the batch size is 50, the number of training iterations is 6000, the loss function is mean square error, a memory decay mechanism is introduced in the training process, the weight of the long-term sample gradually decreases, the decay coefficient is 0.95, at the same time, a sample importance weighting strategy is adopted, the weight of the sudden scene sample is increased by 50%, after the training is completed, the network is pruned and optimized, the connections with the weight absolute value less than 0.05 are removed, the network parameter quantity is reduced by 35% after pruning, a deviation correction memory learning data table is constructed, the network parameters, correction parameter distribution characteristics and scene adaptation mapping relationship are recorded in the table, the correction parameters include the brake advance mean 0.65s, the standard deviation 0.18s, the brake force mean 1.5m / s², the standard deviation 0.4m / s², the steering angle mean 22°, the standard deviation 6.5°, and the steering rate mean 25° / s, the standard deviation 8° / s, which provide basic data for subsequent reinforcement learning iteration.

[0111] The application further provides an intelligent wheelchair scheduling system for executing the intelligent wheelchair scheduling method as described above, which comprises:

[0112] The obstacle avoidance failure decision extraction module is used for extracting historical scheduling tasks from a database of the intelligent wheelchair, and then extracting the obstacle avoidance failure decisions of the intelligent wheelchair in the historical scheduling tasks; and performing clustering analysis on the obstacle avoidance failure decisions to generate decision failure clustering logic.

[0113] The obstacle avoidance failure decision analysis module is used for performing motion trajectory deduction deviation analysis on the sudden obstacles based on the decision failure clustering logic to obtain decision deduction deviation; and performing sudden risk prediction lag analysis based on the decision deduction deviation to obtain sudden risk prediction lag regression data.

[0114] The memory learning module is used for pre-deviation correction memory learning of wheelchair scheduling braking according to the burst risk prediction lag normalized data, obtains deviation correction memory learning data; and performs reinforcement learning iteration on the deviation correction memory learning data, and outputs decision optimization data.

[0115] The above description is merely a specific implementation of the present application, which enables those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A smart wheelchair dispatching method, characterized in that, The method comprises the following steps: Step S1: extracting historical scheduling tasks from the database of the intelligent wheelchair, and then extracting the obstacle avoidance failure decision of the intelligent wheelchair in the historical scheduling tasks; Performing cluster analysis on the obstacle avoidance failure decision to generate a decision failure cluster logic; Step S2 comprises: Extracting the dynamic parameters of the obstacles from the obstacle avoidance failure samples and restoring the scene to obtain an obstacle dynamic scene, wherein the obstacles include crowds and moving objects; Performing decision sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis on the obstacle dynamic scene based on the decision failure cluster logic to obtain a decision deduction deviation; wherein the decision sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis comprises: According to the decision failure cluster logic, extracting the intelligent wheelchair obstacle avoidance failure state of the sudden obstacle in the obstacle dynamic scene; Extracting the back-and-forth cycle trajectory in the obstacle avoidance failure state, and analyzing the trajectory range nonlinear relationship in the back-and-forth cycle trajectory; From the trajectory range nonlinear relationship, analyzing the intermittent start-stop speed change relationship of the intelligent wheelchair; in the trajectory range nonlinear relationship and the speed change relationship, coupling the disorder degree of the wheelchair steering angle; According to the trajectory range nonlinear relationship, the speed change relationship and the disorder degree of the wheelchair steering angle, performing intelligent wheelchair avoidance action trajectory time deviation accumulation fitting on the obstacle dynamic scene to obtain avoidance trajectory deviation accumulation data; Based on the avoidance trajectory deviation accumulation data, performing decision sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis to obtain a decision deduction deviation; Based on the decision deduction deviation and the obstacle dynamic scene, performing sudden risk prediction lag analysis to obtain sudden risk prediction lag data; wherein the sudden risk prediction lag analysis comprises: Obtaining the hardware configuration parameters of the intelligent wheelchair, wherein the hardware includes CPU, GPU and memory; Calculating the distribution density of the obstacles in the obstacle dynamic scene and analyzing the sudden motion state of the obstacles; wherein the sudden motion state includes speed, direction and acceleration; Performing obstacle motion vector prediction derivation on the sudden motion state of the obstacles, and then performing real-time data processing resource demand amount evaluation to obtain real-time prediction resource demand amount; According to the hardware configuration parameters, performing calculation load bottleneck quantification on the real-time prediction resource demand amount to obtain calculation bottleneck quantification data of the intelligent wheelchair; Based on the decision deduction deviation, extracting the time sequence with the longest decision delay; the time sequence with the longest decision delay refers to: extracting the time stamp sequence and the corresponding deviation value sequence from the decision deduction deviation, performing threshold detection on the deviation value sequence, and finding out the key time period in which the wheelchair responds the slowest in the obstacle avoidance process; According to the burst risk pre-judgment lag analysis of the decision-making of the calculation bottleneck quantization data and the time sequence with the maximum decision delay, burst risk pre-judgment lag data is obtained; the burst risk pre-judgment lag analysis of the decision-making is that: the bottleneck coefficient sequence is extracted from the calculation bottleneck quantization data, the delay time sequence is extracted from the time sequence with the maximum decision delay, the joint data matrix is constructed, the matrix row represents different scene parameter combinations, the list represents the bottleneck coefficient, the delay time, the deviation value and the obstacle burst intensity, the lag pre-judgment relationship is established, and the relationship between the pre-judgment lag time and the obstacle quantity, the burst frequency and the bottleneck coefficient is obtained; Regression analysis is performed on the burst risk pre-judgment lag data, and burst risk pre-judgment lag regression data is output; Step S3: normalizing the burst risk pre-judgment lag regression data to obtain burst risk pre-judgment lag normalized data; pre-position deviation correction memory learning of wheelchair scheduling braking is performed according to the burst risk pre-judgment lag normalized data, and deviation correction memory learning data is obtained; the deviation correction memory learning data is iterated by reinforcement learning, and decision optimization data is output.

2. The intelligent wheelchair dispatching method of claim 1, wherein, Step S1 includes the following steps: Extracting historical scheduling tasks from the database of the intelligent wheelchair, and extracting an obstacle avoidance failure data set from the historical scheduling tasks; Random sample selection is performed on the obstacle avoidance failure data set, and an obstacle avoidance failure sample is output, and then the obstacle avoidance failure decision of the intelligent wheelchair in the sample is extracted; Logical analysis is performed on the obstacle avoidance failure decision to obtain a decision failure logic; Cluster analysis is performed on the decision failure logic to generate a decision failure cluster logic.

3. The intelligent wheelchair dispatching method of claim 1, wherein, Wherein, the trajectory time sequence deviation accumulation fitting of the avoidance action of the intelligent wheelchair includes the following steps: Analyzing the effective avoidance action of the intelligent wheelchair through the dynamic scene of the obstacle; Sawtooth trajectory analysis is performed on the non-linear relationship of the trajectory range to obtain sawtooth trajectory data; curvature mutation trend offset analysis is performed on the sawtooth trajectory data to obtain trajectory curvature offset trend; Analyzing the change slope of acceleration / deceleration in the speed change relationship; Coupling disorder analysis of the steering angle and the speed is performed on the disordered degree of the wheelchair steering angle according to the trajectory curvature offset trend and the change slope of acceleration / deceleration to obtain steering angle disorder coupling data; Multi-dimensional avoidance deviation is obtained by performing avoidance multi-dimensional deviation random process decomposition on the effective avoidance action according to the trajectory curvature offset trend, the change slope of acceleration / deceleration and the steering angle disorder coupling data; wherein, the multi-dimension includes time dimension and space dimension; Based on the multi-dimensional avoidance deviation, the trajectory time sequence deviation accumulation fitting of the avoidance action of the intelligent wheelchair is performed to obtain avoidance trajectory deviation accumulation data.

4. The intelligent wheelchair dispatching method of claim 1, wherein, Step S3 includes the following steps: Convolution processing is performed on the decision deduction deviation to obtain decision convolution deviation; normalization processing is performed on the burst risk pre-judgment lag regression data to obtain burst risk pre-judgment lag normalized data; Pre-position deviation correction memory learning of wheelchair scheduling braking is performed on the decision convolution deviation according to the burst risk pre-judgment lag normalized data, and deviation correction memory learning data is obtained; Reinforcement learning iteration is performed on the deviation correction memory learning data, and decision optimization data is output.

5. The intelligent wheelchair dispatching method of claim 4, wherein, The pre-position deviation correction memory learning of the wheelchair scheduling braking includes the following steps: The decision convolution deviation is subjected to deviation parameter extraction to obtain a decision deviation parameter; wherein the parameter includes wheelchair brake response delay, maximum brake acceleration, steering angle and avoidance trajectory; The decision deviation parameter is subjected to evolution law regression analysis to generate a deviation parameter evolution law; Based on the deviation parameter evolution law and the decision lag normalized data, deviation correction safety margin approximation clustering is performed, and memory learning is performed to obtain correction amount clustering memory data; wherein the deviation correction safety margin refers to the evolution law of key parameters such as wheelchair brake response delay, maximum brake acceleration, steering angle and avoidance trajectory, which is obtained by real-time correction and optimization of these parameters, to ensure that the wheelchair can maintain sufficient safety distance and reaction time in various scenes, to quantify the dynamic safety buffer reserved by the wheelchair when facing sudden obstacles; According to the correction 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.

6. An intelligent wheelchair dispatch system characterized by, The intelligent wheelchair scheduling system for performing the intelligent wheelchair scheduling method of claim 1 comprises: An obstacle avoidance failure decision extraction module is configured to extract historical scheduling tasks from a database of the intelligent wheelchair, and then extract the obstacle avoidance failure decisions of the intelligent wheelchair in the historical scheduling tasks; the obstacle avoidance failure decisions are subjected to clustering analysis to generate decision failure clustering logic; An obstacle avoidance failure decision analysis module is configured to perform sudden obstacle wheelchair avoidance motion trajectory deduction deviation analysis based on the decision failure clustering logic to obtain decision deduction deviation; and sudden risk prediction lag analysis based on the decision deduction deviation to obtain sudden risk prediction lag regression data; A memory learning module is configured to perform pre-position deviation correction memory learning of wheelchair scheduling braking based on the sudden risk prediction lag normalized data to obtain deviation correction memory learning data; and perform reinforcement learning iteration on the deviation correction memory learning data to output decision optimization data.

Images