Adaptive calibration system and method for assisting driving a wheelchair

Abstract

The application provides an adaptive calibration system and method for assisting driving a wheelchair, and belongs to the technical field of intelligent auxiliary equipment. The system comprises a perception module, an analysis module, a prediction module and a control module. The perception module collects multispectral data of a road surface in front of the wheelchair, and generates a high-dimensional feature vector reflecting physical properties of the ground. The analysis module maps the high-dimensional feature vector to a low-dimensional manifold space using a manifold learning algorithm, and calculates a ground friction factor in combination with a pre-set ground property knowledge base. The prediction module obtains real-time running state data of the wheelchair, constructs a dynamic trajectory, and performs stability analysis on the dynamic trajectory in combination with the ground friction factor to predict instability risks of the wheelchair. When instability risks are predicted, the control module performs online simulation on multiple compensation strategies, determines an optimal compensation operation, and controls the motor to execute the operation to maintain the running stability of the wheelchair.

Description

Technical Field

[0001] This invention relates to the field of intelligent assistive device technology, and more specifically, to an adaptive calibration system and method for assistive driving wheelchairs. Background Technology

[0002] Currently, driver assistance systems for electric wheelchairs play a crucial role in enhancing users' independent mobility. However, existing technologies still have significant limitations when dealing with complex and ever-changing road conditions. Most mainstream driver assistance systems rely on feedback data from inertial sensors or motor encoders to calibrate the wheelchair's driving status through closed-loop control. This approach is essentially a "remedial" model; the control system only detects the anomaly and compensates after the wheels have already slipped or the wheelchair's posture has noticeably swayed. This lag in control can lead to untimely control responses, or even wheelchair instability and rollover accidents, when facing slippery, icy, or soft surfaces.

[0003] In response to the complexity of road surface environments, some improvement solutions have been developed in existing technologies. For example, a related technology (such as the patent application number CN202510985299.9) discloses an intelligent wheelchair control method, which constructs a three-dimensional ground feature model by extracting historical environmental data, performs geometric morphology analysis on potholes, simulates the wheel's struggle behavior in potholes, and then constructs an optimized model for stuck-out control, thereby improving the wheelchair's ability to cope with obstacles.

[0004] However, while these methods, based on historical data modeling and scenario-specific simulations (such as pothole escape), have made progress in handling known obstacles or specific terrains, they still struggle to cope with the nonlinear abrupt changes in real-time road surface physical characteristics. For example, when a wheelchair transitions instantly from smooth ceramic tiles to high-resistance carpet, existing control logic often fails to adapt within milliseconds, easily leading to overshoot and oscillations, resulting in a bumpy ride. More importantly, current technologies still lack the ability to deeply predict the physical nature of road surfaces (such as the real-time evolution of friction coefficients and dynamic fluctuations in road bearing capacity). When facing completely unknown extreme road conditions, the safety and robustness of these systems still need further improvement. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the purpose of this invention is to provide an adaptive calibration system and method for assisted driving wheelchairs.

[0006] According to one aspect of the present invention, an adaptive calibration system for an assisted driving wheelchair includes:

[0007] The sensing module is configured to collect multispectral data of the road surface in front of the wheelchair and generate a high-dimensional feature vector reflecting the physical properties of the ground surface based on the multispectral data. The analysis module, connected to the perception module, is configured to receive the high-dimensional feature vector, map the high-dimensional feature vector to a low-dimensional manifold space using a manifold learning algorithm, and calculate the surface friction coefficient based on the mapping result of the high-dimensional feature vector in the low-dimensional manifold space and a preset surface attribute knowledge base. The prediction module is configured to acquire real-time operating status data of the wheelchair, construct a dynamic trajectory based on the real-time operating status data, and perform stability analysis on the dynamic trajectory in conjunction with the ground friction coefficient in order to predict the risk of instability of the wheelchair. The control module, connected to the prediction module and the motor of the wheelchair, is configured to perform online simulation of multiple compensation strategies based on the surface friction coefficient when the instability risk is predicted, determine the optimal compensation operation based on the online simulation results, and control the motor to execute the optimal compensation operation to maintain the operational stability of the wheelchair.

[0008] Preferably, the manifold learning algorithm is an isometric mapping algorithm.

[0009] Preferably, the prediction module performs stability analysis on the dynamic trajectory, specifically including: Calculate the instantaneous Lyapunov exponent of the dynamic trajectory, and determine the instability risk based on the instantaneous Lyapunov exponent.

[0010] Preferably, the control module determines the optimal compensation operation, specifically including: The optimal compensation operation is selected from the various compensation strategies by using a causal decision engine based on counterfactual reasoning and in accordance with the criterion of minimizing variational free energy.

[0011] Preferably, the control module controls the motor to perform the optimal compensation operation, specifically including: A high-frequency quasi-resonant component is superimposed on the control current of the motor.

[0012] Preferably, the analysis module, the prediction module, and the control module are integrated into an edge computing hub employing an in-memory computing architecture.

[0013] Preferably, it further includes: The wireless communication module is configured to upload the high-dimensional feature vector corresponding to the road surface with a matching degree lower than a preset threshold in the land surface attribute knowledge base to a cloud server, and receive the updated land surface attribute knowledge base or the model of the manifold learning algorithm from the cloud server.

[0014] Preferably, the prediction module performs stability analysis on the dynamic trajectory, specifically including: A pre-trained long short-term memory network is used, with the dynamic trajectory and the surface friction coefficient as input, to predict the instability risk.

[0015] According to another aspect of the present invention, an adaptive calibration method for an assisted driving wheelchair includes the following steps: Multispectral data of the road surface in front of the wheelchair is collected, and a high-dimensional feature vector reflecting the physical properties of the ground surface is generated based on the multispectral data. The high-dimensional feature vector is mapped to a low-dimensional manifold space using a manifold learning algorithm, and the surface friction coefficient is calculated based on the mapping result of the high-dimensional feature vector in the low-dimensional manifold space and a preset surface attribute knowledge base. The real-time operating status data of the wheelchair is acquired, a dynamic trajectory is constructed based on the real-time operating status data, and the stability of the dynamic trajectory is analyzed in combination with the ground friction coefficient to predict the risk of instability of the wheelchair. When the risk of instability is anticipated, multiple compensation strategies are simulated online based on the surface friction coefficient. The optimal compensation operation is determined based on the online simulation results, and the motor of the wheelchair is controlled to perform the optimal compensation operation to maintain the operational stability of the wheelchair.

[0016] Preferably, the step of determining the optimal compensation operation specifically includes: The optimal compensation operation is selected from the various compensation strategies by using a causal decision engine based on counterfactual reasoning and in accordance with the criterion of minimizing variational free energy.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. It has achieved a fundamental change in control mode, transforming the traditional passive "post-event remediation" control mode of driver assistance systems into an active "pre-event prediction and defense" mode. By predicting in advance, it has bought valuable time for the system to avoid risks, greatly improving the safety of wheelchairs on dangerous roads.

[0018] 2. It enhances the depth and accuracy of environmental perception. Through multispectral sensing technology, the system can penetrate macroscopic geometric appearances and directly perceive microscopic physical properties of the road surface, such as moisture content, viscoelasticity, and friction coefficient. This enables a deep understanding of the environment and provides a solid data foundation for precise control.

[0019] 3. It enhances the adaptability and stability of complex terrains. By adopting nonlinear manifold learning and causal reasoning control, it effectively solves the instability and oscillation problems of traditional linear models on unstructured and nonlinear road surfaces. Through active torque compensation, it ensures that the wheelchair can maintain dynamic balance under all terrain constraints, making the driving process smoother and more comfortable.

[0020] 4. It balances advanced algorithms with low power consumption requirements. By adopting an edge-side storage and computing architecture, it solves the contradiction between computing power and power consumption when deploying complex artificial intelligence algorithms on embedded devices. It meets the requirements of medical rehabilitation assistive devices for low power consumption and long battery life, combining advanced technology with practicality. Attached Figure Description

[0021] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A schematic diagram of the architecture of a wheelchair adaptive calibration system based on multispectral sensing and predictive control, provided for an embodiment of this application; Figure 2 A schematic flowchart illustrating the wheelchair adaptive calibration method provided in this application embodiment; Figure 3 This is a schematic diagram of the signaling interaction timing between key system modules provided in an embodiment of this application; Figure 4 This is a schematic diagram of the physical structure and sensor layout of the adaptive calibration system on a wheelchair in an embodiment of this application; Figure 5 This is a schematic diagram of the energy spectrum distribution of different surface material characteristics in the embodiments of this application; Figure 6 This is a schematic diagram illustrating the convergence characteristics of system stability in an embodiment of this application; Figure 7 A comparison diagram of the transient dynamic deviation between the adaptive calibration system provided in this application embodiment and the traditional PID control architecture when experiencing sudden road surface changes. Detailed Implementation

[0022] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0023] Example 1 This embodiment provides a wheelchair adaptive calibration system and corresponding calibration method based on multispectral sensing and predictive control. It aims to illustrate the basic architecture, core workflow and collaborative working principle between the components of the technical solution of this application, so as to achieve active and predictive stability control of wheelchairs when driving on complex road surfaces.

[0024] Please see Figure 4 This illustration shows the physical structure and sensor layout of the adaptive calibration system on a wheelchair in an embodiment of this application. In a specific application scenario, the system is mounted on an electric wheelchair body. The wheelchair body includes, but is not limited to, conventional components such as the frame, seat, battery, and drive wheels. The hardware of the adaptive calibration system provided in this application mainly includes: a multispectral photon trapping array, a high-precision dynamic observation vector unit, and an integrated control box 30.

[0025] Specifically, the multispectral photon capture array is mounted at a low front position of the wheelchair body 1, such as below the footrests, with its detection field of view aligned with and covering an area of ​​approximately one centimeter in front of the drive wheels. It should be noted that this layout ensures the system can acquire the physical properties of the new surface before the wheels contact it. The high-precision dynamics observation vector unit, specifically a six-degree-of-freedom inertial measurement unit integrating a three-axis accelerometer and a three-axis gyroscope in this embodiment, is mounted near the geometric center of the wheelchair body 1 chassis to measure the wheelchair's attitude, angular velocity, and linear acceleration in real time with high precision. The integrated control box is typically placed under or behind the wheelchair seat, housing the system's core computing and control units, and is tightly connected to the multispectral photon capture array, the high-precision dynamics observation vector unit, and the wheelchair's existing motor drive via electrical and signal interfaces.

[0026] Please see Figure 1 This is a schematic diagram of the logical architecture of the adaptive calibration system provided in the embodiments of this application. Functionally, the system can be divided into a sensing module, an analysis module, a prediction module, and a control module. It is understood that the functions of these modules can be physically derived from... Figure 4 The hardware collaboration shown is completed.

[0027] The core of the sensing module is a multispectral photon trapping array (corresponding to...) Figure 1 The multispectral sensing array S1 in the wheelchair is responsible for collecting multispectral data of the road surface in front of the wheelchair and generating high-dimensional feature vectors that reflect the physical properties of the ground surface.

[0028] The analysis module's functionality is primarily implemented on the edge computing unit within the integrated control box, corresponding to... Figure 1The module includes the Hilbert space high-dimensional feature mapping unit F2. This module is connected to the perception module to receive high-dimensional feature vectors and uses a manifold learning algorithm to map them to a low-dimensional manifold space. Then, combined with a preset surface attribute knowledge base, it calculates the surface friction coefficient of the current road surface.

[0029] The prediction module is also implemented on the edge computing unit within the integrated control box 30, corresponding to Figure 1 The system includes the Lyapunov stability checker P1 and the singularity bifurcation real-time prediction and monitoring P3. This module acquires real-time operating status data of the wheelchair provided by a high-precision dynamic observation vector unit and a motor encoder. Based on this data, it constructs the system's dynamic trajectory and, combined with the surface friction coefficient calculated by the analysis module, performs stability analysis on the future evolution trend of the dynamic trajectory, thereby predicting the risk of wheelchair instability.

[0030] The control module is also integrated into the control box and interacts directly with the wheelchair's motor drive, corresponding to... Figure 1 The system includes a counterfactual reasoning causal decision engine D1 and a magnetic field-oriented control advanced phase compensator E3. When the prediction module anticipates an instability risk, it is activated. Based on the current surface friction coefficient, it performs rapid online simulations of various virtual compensation strategies, determines an optimal compensation operation according to specific criteria, and immediately controls the motor to execute the operation to actively counteract disturbances and maintain the wheelchair's operational stability.

[0031] The following will combine Figure 2 The method flowchart shown and Figure 3 The signaling interaction timing diagram shown describes in detail the working process of the system in this embodiment. Assume a wheelchair is moving smoothly at a speed of 0.5 m / s on a dry indoor tiled floor when suddenly a wet, slippery area appears ahead due to a spilled beverage.

[0032] In step S101, the system acquires multispectral data of the road surface. During the wheelchair's movement, a multispectral photon-capturing array mounted at the front continuously detects the road surface ahead. In this embodiment, the array includes one or more tunable light sources (such as broadband light-emitting diodes or laser diode arrays) and a high-sensitivity photon detector array. The light source employs a pseudo-random phase-coded mode, emitting a series of pulsed beams with specific wavelength combinations towards the road surface. These photons interact with the road surface material, undergoing scattering and absorption, with a portion of the echo photons being captured by the detector array. The system constructs a high-dimensional feature vector matrix that reflects the microscopic physical properties of the surface (such as water content, surface roughness, viscoelastic modulus, etc.) by accurately measuring the echo delay distribution from emission to reception and the energy absorption characteristics at different wavelengths. When the wheelchair approaches a waterlogged area (corresponding to...) Figure 3At time t0, the road surface changes. The multispectral photon trapping array collects spectral echo signals that are significantly different from those of dry ceramic tiles, and quickly generates a high-dimensional feature vector representing the wet and slippery road surface (corresponding to...). Figure 3 (Spectral data is generated at time t1). Please refer to [link / reference]. Figure 5 The figure shows the characteristic energy spectrum distribution of different surface materials. It can be seen that different surfaces such as ice / water, smooth ceramic tiles and high-resistance carpets have distinct normalized reflectance intensity curves in the wavelength range of 400-2500 nanometers. This provides a theoretical basis for identifying the physical properties of road surfaces through spectral information.

[0033] Subsequently, in step S102, the system calculates the surface friction coefficient. The analysis module within the integrated control box receives the high-dimensional feature vector matrix generated in step S101. Because this vector is high-dimensional, information-redundant, and highly nonlinear, it is difficult to use directly. Therefore, the analysis module employs a manifold learning algorithm, specifically the isometric mapping algorithm in this embodiment, to reduce the dimensionality and structure the high-dimensional data. The core idea of ​​the isometric mapping algorithm is to assume that the high-dimensional observation data is an embedding of a low-dimensional manifold in a high-dimensional space. It aims to find this low-dimensional manifold while preserving the geodesic distance between points on the manifold. The algorithm first constructs a neighborhood graph for each data point, then approximates the geodesic distance by calculating the shortest path between all pairs of points in the graph. Finally, it uses multidimensional scaling techniques to reconstruct these points in the low-dimensional space, making the Euclidean distance between them equal to the original geodesic distance. Through this process, the high-dimensional, nonlinear spectral features are effectively mapped to a lower-dimensional (e.g., 2-dimensional or 3-dimensional) manifold space with Riemannian metric.

[0034] Meanwhile, a surface attribute knowledge base is pre-stored within the integrated control box. This knowledge base was pre-established through numerous experiments and includes the topological center coordinates of the spectral feature vectors of various typical landforms (such as dry cement roads, slippery tiles, ice surfaces, high-resistance carpets, and grasslands) mapped under the same manifold learning model, as well as the known physical parameters (such as the range of friction coefficients) corresponding to these landforms. The analysis module topologically aligns the projection results of the real-time acquired spectral features onto the low-dimensional manifold space with the topological centers of each landform in the knowledge base. Specifically, it calculates the geodesic distance between the current data point and each topological center in the knowledge base. Based on these distances, the system uses the following probability measure calculation formula to estimate the posterior probability distribution of the current road friction coefficient μ:

[0035] in, It is a spectral feature acquired in real time. Projected coordinates in manifold space It is related to the friction factor in the knowledge base. Related landforms The topological center coordinates, The first in the knowledge base The topological center of a landform This represents the geodesic distance on the Riemannian manifold.

[0036] The physical meaning of this formula is that the closer the current data point is to the topological center of a certain knowledge base terrain in manifold space, the higher the probability that the current road surface belongs to that terrain (and has the corresponding friction coefficient). By weighted averaging of all possible terrain types or selecting the type with the highest probability, the system ultimately calculates a specific surface friction coefficient value or its probability distribution. For the aforementioned waterlogged area, the system calculates that its friction coefficient drops sharply.

[0037] Subsequently, in step S103, the system constructs a dynamic trajectory and predicts risks. The prediction module begins functioning at the instant the wheel is about to or just after contacting the waterlogged area. It continuously acquires real-time attitude data such as angular velocity and acceleration of the wheelchair from the high-precision dynamic observation vector unit 20, and wheel speed data fed back from the motor encoder of the motor driver. These multi-dimensional real-time operating state data are fused to construct the system's real-time dynamic trajectory in phase space. The evolution of this trajectory can be described by a nonlinear dissipation equation:

[0038] in, It is the system's state vector (containing information such as attitude and velocity). It is its rate of change with respect to time. It describes the inherent dynamic characteristics of the system. It is a control input matrix. These are control commands issued by the current user or the upper-level controller, and This represents the unknown, uncertain disturbance torque caused by changes in the road surface (i.e., changes in the friction coefficient).

[0039] The core task of the prediction module is to analyze the stability of the dynamic trajectory based on the surface friction coefficient calculated in step S102, in order to predict the risk of instability. In this embodiment, the stability analysis is performed by calculating the instantaneous Lyapunov exponent of the trajectory. This is achieved through [method / method]. The Lyapunov index measures the rate of separation of two adjacent trajectories in phase space over time. Its calculation formula is:

[0040] in, The instantaneous Lyapunov exponent measures the rate of separation between two adjacent trajectories in phase space over time, and is used to predict the stability (risk of instability) of the system's dynamic trajectory; t is time. It represents the small perturbation (or initial distance) at the initial moment, which represents the distance between two adjacent trajectories at the initial moment (t=0) in phase space; The perturbation (or distance) after time t represents the distance between two adjacent trajectories in phase space after time t; ||·|| represents the distance or norm.

[0041] In practical calculations, numerical methods (such as QR decomposition) are typically used to estimate the maximum Lyapunov exponent over a finite time interval. When the system is stable, the trajectory converges. It can be negative or zero; when the system is unstable, small disturbances can be amplified, causing the trajectory to diverge. It is a positive value. Once the prediction module detects it... A shift from negative to positive indicates that the system trajectory is approaching a singular value region or bifurcation point. This implies that the system will experience significant dynamic instability within a very short time window (e.g., 5-10 milliseconds), such as wheel slippage or vehicle roll. The moment this prediction is completed corresponds to... Figure 3 t2 in the middle.

[0042] Next, the process proceeds to the judgment step S104. If the prediction module determines that there is a risk of instability (i.e. If the current driving state is stable, the process proceeds to step S105; otherwise, the system considers the current driving state to be stable, and the process returns to step S101 to continue periodic sensing and monitoring.

[0043] In the scenario described in this embodiment, since it is predicted that the water accumulation area will cause slippage, Once the value becomes positive, the system proceeds to step S105 to determine the optimal compensation tensor. The control module within the integrated control box is activated, and its internal nonlinear torque control module immediately initiates a causal inference module (corresponding to...). Figure 1 (D1 in the example). This module does not simply apply a fixed compensation rule, but rather makes an online decision based on counterfactual reasoning. It simulates multiple possible compensation strategies (e.g., "if a braking torque of size A is applied to the left wheel", "if a driving torque of size B is applied to the right wheel", etc.), and through online simulation, it quickly deduces how the wheelchair's dynamic trajectory will evolve under the combined effect of the currently predicted disturbance (based on the known decrease in the friction coefficient) and each virtual compensation strategy (i.e., causal intervention).

[0044] The selection of the optimal compensation strategy follows the criterion of minimizing variational free energy. This criterion originates from active reasoning theory and aims to make the system's future state as close as possible to its expected steady state, while minimizing control costs. Optimal Compensation Tensor Determined by the following formula:

[0045] in, It is a candidate compensation tensor (representing a specific torque compensation operation). This indicates the action of performing the intervention. It is the future state trajectory simulated under this intervention. Let E be the variational free energy function, and E denotes the expectation. By quickly solving this optimization problem, the system can determine an optimal compensation operation online and in real time, which can most effectively counteract the impending instability effects.

[0046] Finally, in step S106, the system performs compensation. Once the optimal compensation tensor is obtained... Once identified, the control module immediately translates it into specific control commands for the wheelchair motor drive (corresponding to...). Figure 3 At time t3, the compensation command is issued. In this embodiment, the compensation is not performed by drastically changing the target speed of the motor, but by a more refined low-level current vector reconstruction. Specifically, the controller, in the current loop of the original field-oriented control algorithm, according to... The magnitude and direction of the waveform are superimposed in real time with a high-frequency quasi-resonant compensation waveform of a specific frequency and amplitude. This superimposed waveform component can quickly and accurately synthesize a compensation torque at the electromagnetic level that is opposite in direction and similar in magnitude to the expected slippage or disturbance torque. This compensation torque actively and feedforward cancels out the indeterminate disturbance torque generated by the ground on the drive wheels with almost no noticeable effect on the user, thus nipping instability in the bud before it actually occurs and bringing the wheelchair's dynamic state back to a stable manifold.

[0047] After compensation is performed, the process returns to step S101, and the system enters the next control cycle to continuously perceive and calibrate the road surface.

[0048] The intended effect of this embodiment is that when the wheelchair travels to a flooded area, because the system anticipates the decrease in friction and the resulting risk of slippage, and implements precise compensation for the main driving torque, the wheelchair will hardly experience significant lateral slippage or directional deviation, and can pass through the area smoothly and safely. Please refer to [link to relevant documentation]. Figure 7This figure compares the transient dynamic deviations of the system described in this application with those of a traditional proportional-integral-derivative (PI-DE) control architecture when encountering sudden road surface changes. It can be seen that the system using the proposed solution (shown by the solid line in the figure) exhibits a low peak value (approximately 0.2 rad / s) and smooth fluctuations in its wheelchair angular velocity deviation curve; while the wheelchair with traditional PI-DE control (shown by the dashed line in the figure), when encountering the same situation, experiences severe wheel slippage and body swaying due to lag response, with a peak angular velocity deviation as high as 1.7 rad / s. Furthermore, please refer to... Figure 6 The figure shows the Lyapunov exponent of the system after it has been disturbed. The evolution trajectory over iteration steps (time) shows that the exponential value starts from a positive value (indicating an unstable trend) and, with the intervention of the control module, quickly converges to near zero in only about a few iterations, indicating that the system can rapidly recover to a stable state. These results strongly demonstrate the significant beneficial effects of the technical solution in this application on improving driving safety and stability.

[0049] Furthermore, as a preferred implementation, the complex algorithms of the analysis, prediction, and control modules in this embodiment are all deployed on an edge computing hub with an in-memory computing architecture within an integrated control box. This non-von Neumann computing architecture deeply integrates storage and computing units, greatly reducing latency and power consumption caused by data transfer. This makes it possible to run complex manifold learning and causal inference algorithms on embedded wheelchair devices with limited power consumption and size, thus balancing the system's advanced features with its practicality.

[0050] Example 2 This embodiment provides a variant to illustrate that the core technical concept of this application has good scalability and is not limited to the specific sensor model or specific manifold learning algorithm used in Embodiment 1. The main difference in this embodiment lies in the implementation of the sensing module and the analysis module.

[0051] In terms of hardware structure, the system in this embodiment is basically the same as that in Embodiment 1, but the multispectral photon trapping array is replaced with a compact hyperspectral camera. Compared to a multispectral array that only collects data from a few discrete bands, a hyperspectral camera can acquire continuous spectral information from the visible to near-infrared (e.g., 400-2500 nm) spanning hundreds of narrow bands, forming a high-dimensional hyperspectral data cube (containing two-dimensional spatial information and one-dimensional spectral information). This provides a richer data source for the fine differentiation of road surface materials.

[0052] At the software algorithm level, the analysis module replaces the isometric mapping algorithm used in Example 1 with another classic nonlinear manifold learning algorithm—the local linear embedding algorithm.

[0053] The operation of this embodiment is as follows: When the wheelchair is in motion, the hyperspectral camera continuously acquires a cube of hyperspectral data of the road surface ahead. The analysis module within the integrated control box first uses methods such as principal component analysis to perform preliminary noise reduction and dimensionality reduction on the hyperspectral data, extracting the main spectral features. Subsequently, a locally linear embedding algorithm is run. This algorithm assumes that the high-dimensional data manifold is locally linear. It first finds the k nearest neighbors for each data point, and then calculates a weight coefficient that can linearly reconstruct the data point through these neighbors. This weight coefficient describes the local geometry of the data point and its neighborhood. The key step of the algorithm is, while keeping these weight coefficients unchanged, to find a new set of coordinates in the low-dimensional space such that each point can also be linearly reconstructed by its neighbors with the same weights. In this way, the high-dimensional hyperspectral data is embedded into a low-dimensional manifold space that preserves its local topology.

[0054] The subsequent steps, including topological alignment of the projection results on the low-dimensional manifold with the surface attribute knowledge base, calculation of the surface friction factor, instability prediction based on the Lyapunov index, and control intervention through causal reasoning and torque compensation, are basically the same as the principles and methods described in Example 1.

[0055] The expected effect of this embodiment is that, due to the richer and more continuous spectral information provided by the hyperspectral camera, the system's ability to distinguish road surface materials may be further enhanced. Especially when distinguishing materials with similar physical properties but subtle differences in spectral characteristics (e.g., carpets of different materials or with varying degrees of wear), the recognition accuracy may be higher, resulting in a more accurate estimation of the friction coefficient. The successful implementation of this embodiment proves that the scope of protection for "acquiring multispectral data" and "utilizing manifold learning algorithms" in this application should not be limited to a specific sensor or algorithm, but rather encompasses all technical means capable of realizing the transformation from high-dimensional nonlinear sensing data to low-dimensional physical parameterization modeling.

[0056] Example 3 This embodiment provides another variation to illustrate that the specific implementation methods of instability prediction and moment compensation in this application are diverse. The core of this application lies in the "predictive" discovery of risks and the "proactive" compensation, rather than being limited to a specific prediction model or execution method. This embodiment mainly differs in the implementation of the prediction module and the control module.

[0057] The system hardware, as well as the structure and working principle of the sensing and analysis modules, remain consistent with those in Example 1. However, at the software level, the prediction module will replace the mathematical model prediction method based on the Lyapunov exponent with a data-driven deep learning prediction model, specifically a pre-trained long short-term memory network.

[0058] Meanwhile, the way the control module performs the compensation operation has also changed from "superimposing a quasi-resonant waveform in the control current" in Example 1 to a more direct "pulse width modulation signal of the high-frequency modulated motor driver".

[0059] The working process of this embodiment is as follows: In the prediction phase, the prediction module takes the multi-dimensional time-series data (e.g., the sequence of angular velocity, linear acceleration, motor current, and rotational speed within the past 100 milliseconds) generated by the high-precision dynamic observation vector unit and the motor encoder, as well as the current surface friction coefficient calculated by the analysis module, as input and feeds them into a pre-trained Long Short-Term Memory (LSTM) network. The LSM network is a special type of recurrent neural network, well-suited for processing and predicting long-term dependencies in time-series data. This network, through offline learning of a large amount of historical data (including data from normal driving and various instability conditions), has mastered the complex mapping pattern from sensor time series and road surface parameters to future instability states. Therefore, it can directly output a probability value of wheelchair instability within a specific future time window (e.g., 5-10 milliseconds) based on the real-time input. When this probability value exceeds a preset threshold (e.g., 90%), the prediction module determines that there is a risk of instability.

[0060] During the control phase, upon predicting the risk of instability, the causal reasoning module in the control module is also activated to perform online simulation and determine an optimal compensation tensor. This decision-making process is the same as in Example 1. However, when performing compensation, the controller no longer modifies the basic current vector. Instead, based on the magnitude and direction of the torque represented by the compensation tensor, it directly and at high frequency (e.g., at a frequency of 20 kHz) adjusts the duty cycle of the pulse width modulation signal driving the three-phase brushless DC motor. By precisely controlling the on / off times of each power switch in the inverter, the required instantaneous voltage vector can be directly synthesized at the electromagnetic level, thereby generating a precise compensation torque.

[0061] The expected effect of this embodiment is that the prediction model based on deep learning may exhibit higher sensitivity and accuracy when dealing with complex dynamic patterns that are difficult to describe by traditional mathematical models. Furthermore, directly modulating the pulse width modulation signal may provide a faster and finer torque response compared to superimposing the signal in the current loop, because it operates at a lower level of the control link. This embodiment demonstrates that the prediction stage of this application can employ various models, including but not limited to Lyapunov exponential analysis and long short-term memory networks, and the control execution stage can also employ various methods, including but not limited to current waveform superposition and pulse width modulation. All these different advanced technical implementations fall within the scope of the protection concept of this application.

[0062] Example 4 This embodiment demonstrates a system architecture with cloud collaboration and self-evolution capabilities, which aims to solve the pain points of traditional embedded systems that have fixed knowledge and cannot adapt to unknown new environments.

[0063] In this embodiment, as an optional implementation, a wireless communication module (e.g., a 4G or 5G communication module) is added to the integrated control box 30 based on the hardware of Embodiment 1, enabling it to connect to the Internet. Simultaneously, a central server is deployed in the cloud, which maintains a global, dynamically expandable "surface attribute knowledge base" and possesses significantly stronger offline computing and model training capabilities than edge devices.

[0064] The working process of this embodiment embodies an "edge-cloud" collaborative model: In daily use, the adaptive calibration system on the wheelchair end completes all sensing, analysis, prediction and control tasks completely independently and in real time, ensuring immediate response and reliable operation in a network-free environment. Its workflow is completely consistent with that of Example 1.

[0065] However, when the system's analysis module encounters an "unknown" or "fuzzy" road surface during operation—that is, when the projection of its real-time acquired spectral features onto the manifold space exceeds a preset threshold in terms of geodesic distance from all known topographic centers in the surface attribute knowledge base (i.e., the matching degree is below the threshold)—the system triggers a "new sample learning" mechanism. At this point, the system packages the original high-dimensional spectral data corresponding to this "unknown" road surface, the high-precision dynamic data at the time (attitude, velocity, etc.), and the control strategy adopted by the system afterward and the final control effect data (e.g., whether stability was successfully maintained) into a "new sample to be learned."

[0066] When the wheelchair is in a state where network connectivity is available and the machine is idle (for example, when it is back at home and connected to a Wi-Fi network), the wireless communication module in the integrated control box will upload the stored labeled "new sample" data to the central server in the cloud.

[0067] The cloud server aggregates a large amount of new sample data uploaded by wheelchair users from numerous different regions. The cloud backend can utilize more sophisticated unsupervised or semi-supervised clustering algorithms, and even introduce manual annotation, to classify and calibrate the physical parameters of this new road surface data (for example, by analyzing correlations with meteorological and geographic information data, the friction coefficient of a "slippery road surface covered with fallen leaves after rain" can be determined). Based on this massive amount of new data, the cloud server can expand the global "surface attribute knowledge base" and retrain and optimize manifold learning models (such as isometric mapping models) or prediction models (such as long short-term memory network models).

[0068] Finally, the cloud server pushes the updated and optimized incremental package of the surface attribute knowledge base or the new version of the algorithm model to all networked wheelchair terminals via the wireless communication module. After receiving the update package, the system on the wheelchair terminal will automatically update online, thereby "learning" how to identify and cope with these new road surface environments.

[0069] The intended effect of this embodiment is that the entire wheelchair system (or even a network of numerous wheelchairs) possesses the ability to continuously learn and self-evolve. Its breadth of environmental awareness and adaptability will continuously enhance with extended use and expanded usage. An unknown, slightly unstable, slippery surface covered in fallen leaves encountered today, through learning and sharing, may become a known risk item in the entire wheelchair knowledge base tomorrow, making the entire system increasingly safe and intelligent when facing new environments in the future. This demonstrates the advanced nature and scalability of this application at the system architecture level, laying the foundation for achieving higher levels of intelligent assisted driving.

[0070] The present invention also provides an adaptive calibration system for an assisted driving wheelchair. The adaptive calibration system for an assisted driving wheelchair can be implemented by executing the process steps of the adaptive calibration method for an assisted driving wheelchair. That is, those skilled in the art can understand the adaptive calibration method for an assisted driving wheelchair as a preferred embodiment of the adaptive calibration system for an assisted driving wheelchair.

[0071] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0072] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. An adaptive calibration system for an assisted driving wheelchair, characterized in that, include: The sensing module is configured to collect multispectral data of the road surface in front of the wheelchair and generate a high-dimensional feature vector reflecting the physical properties of the ground surface based on the multispectral data. The analysis module, connected to the perception module, is configured to receive the high-dimensional feature vector, map the high-dimensional feature vector to a low-dimensional manifold space using a manifold learning algorithm, and calculate the surface friction coefficient based on the mapping result of the high-dimensional feature vector in the low-dimensional manifold space and a preset surface attribute knowledge base. The prediction module is configured to acquire real-time operating status data of the wheelchair, construct a dynamic trajectory based on the real-time operating status data, and perform stability analysis on the dynamic trajectory in conjunction with the ground friction coefficient in order to predict the risk of instability of the wheelchair. The control module, connected to the prediction module and the motor of the wheelchair, is configured to perform online simulation of multiple compensation strategies based on the surface friction coefficient when the instability risk is predicted, determine the optimal compensation operation based on the online simulation results, and control the motor to execute the optimal compensation operation to maintain the operational stability of the wheelchair.

2. The system according to claim 1, characterized in that, The manifold learning algorithm is an isometric mapping algorithm.

3. The system according to claim 1, characterized in that, The prediction module performs stability analysis on the dynamic trajectory, specifically including: Calculate the instantaneous Lyapunov exponent of the dynamic trajectory, and determine the instability risk based on the instantaneous Lyapunov exponent.

4. The system according to claim 1, characterized in that, The control module determines the optimal compensation operation, specifically including: The optimal compensation operation is selected from the various compensation strategies by using a causal decision engine based on counterfactual reasoning and in accordance with the criterion of minimizing variational free energy.

5. The system according to claim 1, characterized in that, The control module controls the motor to perform the optimal compensation operation, specifically including: A high-frequency quasi-resonant component is superimposed on the control current of the motor.

6. The system according to claim 1, characterized in that, The analysis module, the prediction module, and the control module are integrated into an edge computing hub that adopts an in-memory computing architecture.

7. The system according to claim 1, characterized in that, Also includes: The wireless communication module is configured to upload the high-dimensional feature vector corresponding to the road surface with a matching degree lower than a preset threshold in the land surface attribute knowledge base to a cloud server, and receive the updated land surface attribute knowledge base or the model of the manifold learning algorithm from the cloud server.

8. The system according to claim 1, characterized in that, The prediction module performs stability analysis on the dynamic trajectory, specifically including: A pre-trained long short-term memory network is used, with the dynamic trajectory and the surface friction coefficient as input, to predict the instability risk.

9. An adaptive calibration method for an assisted driving wheelchair, characterized in that, Includes the following steps: Multispectral data of the road surface in front of the wheelchair is collected, and a high-dimensional feature vector reflecting the physical properties of the ground surface is generated based on the multispectral data. The high-dimensional feature vector is mapped to a low-dimensional manifold space using a manifold learning algorithm, and the surface friction coefficient is calculated based on the mapping result of the high-dimensional feature vector in the low-dimensional manifold space and a preset surface attribute knowledge base. The real-time operating status data of the wheelchair is acquired, a dynamic trajectory is constructed based on the real-time operating status data, and the stability of the dynamic trajectory is analyzed in combination with the ground friction coefficient to predict the risk of instability of the wheelchair. When the risk of instability is anticipated, multiple compensation strategies are simulated online based on the surface friction coefficient. The optimal compensation operation is determined based on the online simulation results, and the motor of the wheelchair is controlled to perform the optimal compensation operation to maintain the operational stability of the wheelchair.

10. The method according to claim 9, characterized in that, The step of determining the optimal compensation operation specifically includes: The optimal compensation operation is selected from the various compensation strategies by using a causal decision engine based on counterfactual reasoning and in accordance with the criterion of minimizing variational free energy.

Images