Uncertain reasoning learning control method for self-balancing of unmanned bicycle
Through the uncertainty inference learning control method, the probability dynamic model is learned from interactive empirical data, which solves the problem of high model accuracy requirements in lateral balance control of unmanned bicycles, and achieves higher robustness and generalization ability.
Patent Information
- Application Number
- CN202411697069.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-25
- Publication Date
- 2025-04-22
AI Technical Summary
The existing lateral balance control method of unmanned bicycles relies on accurate dynamic models, making it difficult to effectively deal with factors such as vehicle ground contact, external environmental interference and vehicle body state uncertainty, resulting in poor performance and poor robustness of control strategies.
The uncertainty reasoning learning control method is adopted, and the control strategy is automatically adjusted to cope with uncertainty interference by learning probability dynamic models from interactive empirical data.
It improves the robustness and generalization ability of unmanned bicycle balance control, reduces the impact of model deviation on the control strategy, simplifies the controller design process, and improves the dynamic performance of the system.
Smart Images
Figure SMS_13 
Figure SMS_14 
Figure SMS_15
Abstract
Description
Technical Field
[0001] The invention relates to the field of balance control of unmanned bicycles, and in particular to an uncertainty reasoning learning control method for lateral balance of unmanned bicycles. Background Art
[0002] Unmanned bicycles have the advantages of being light, flexible, environmentally friendly and energy-saving. They can play an important role in lightweight and short-distance logistics scenarios and are a potential intelligent transportation tool. The lateral tilt balance is the key problem that needs to be solved first to realize the functions of unmanned bicycles. At present, the common control methods for this problem are: linear quadratic optimal regulator (LQR), active disturbance rejection control (ADRC), partial feedback linearization control (PFLC), fault-tolerant sliding mode control (FTSMC), reduced-order sliding mode control (ROSMC), non-singular terminal sliding mode control (NSTSMC), fuzzy control (FC) and model predictive control (MPC), etc. These methods usually rely on bicycle dynamics models to realize control strategy calculations, so they have high requirements for model accuracy. However, the dynamics model established by physical principles often requires some idealized assumptions, such as pure rolling of wheels, flat ground, accurate sensor detection, etc. It is difficult to accurately characterize factors such as contact between the vehicle and the ground, external environmental interference, road condition changes, and uncertainty of the vehicle state, so there will be a large difference between it and the actual system. When these uncertain factors appear, the above-mentioned model-based controller may be difficult to maintain the original performance, or even unsuitable for use. Manually re-tuning the controller parameters or changing the control structure is a common solution to match the uncertainty disturbance, but such a solution is usually complicated, inefficient, and time-consuming. Therefore, designing a control method that is independent of the prior dynamic model of the unmanned bicycle and can evolve autonomous learning based on empirical data to cope with uncertainty disturbances, improve the generalization ability of the system, and has important practical value for the application research of unmanned bicycles.
[0003] Uncertainty reasoning learning is an important branch in the field of artificial intelligence. Its basic principle is to start from uncertain evidence, use uncertainty representation and reasoning and other technical means to learn from real experience data, and finally come up with a reasonable or basically reasonable conclusion with uncertainty. The core of uncertainty reasoning learning is model representation and reasoning method. The former represents uncertainty by establishing a mathematical model, and the latter condenses logical conclusions by designing uncertainty reasoning. The key point of uncertainty reasoning learning control is to give the state of the research object a certain probability distribution, combine the state evolution data to establish a probabilistic dynamics model to characterize the uncertainty dynamics of the system, and integrate the probabilistic dynamics model into planning and decision-making. Uncertainty reasoning learning control has been successfully applied to robots, drones, underwater vehicles and other fields. Summary of the invention
[0004] In view of the shortcomings of the prior art, the present invention provides a self-balancing control method for an unmanned bicycle based on uncertainty reasoning learning control. The overall technical solution of the unmanned bicycle self-balancing uncertainty reasoning learning control method that can solve the above technical problems is as follows.
[0005] The unmanned bicycle experimental prototype and its structural diagram are attached. Figure 1 Attached Figure 1 In the figure, B, H, F, and R represent the frame, front fork, front wheel, and rear wheel respectively; M1, M2, M3, and M4 are the center of mass positions of the corresponding components respectively.
[0006] Description of motion variables: ν is the forward velocity of the vehicle; is the lateral tilt angle (tilt velocity) of the vehicle body; is the heading angle (tilt velocity) of the vehicle body; is the handlebar turning angle (inclination velocity); τ δ Input torque for the control of the handlebars.
[0007] Assuming that the unmanned bicycle moves on the ground, its balance problem can be described as inputting the appropriate handlebar control torque τ δ , turn the handlebar to the appropriate position, and correct the lateral tilt angle of the vehicle body through dynamic coupling Make it tilt Always in balance nearby.
[0008] Select a parameterized control strategy π(θ) to represent the control torque τ of the handlebar δ , where θ is the policy parameter; the state of the vehicle is Remember J π (θ) is the total cost of the state sequence.
[0009] The inference learning algorithm of the control strategy is as follows:
[0010] ST1: Initialize control strategy π(θ) (handlebar torque τ δ )’s policy parameter θ.
[0011] ST2: Loop LOOP1 :
[0012] ST2-1: Change strategy π(θ)(handlebar torque τ δ ) is applied to unmanned bicycles.
[0013] ST2-2: Collect vehicle status and handlebar torque τ δ Generate empirical datasets.
[0014] ST2-3: Learning the Probabilistic Dynamic Model (PDM) of Autonomous Bicycles from Empirical Datasets.
[0015] ST2-4: Loop LOOP2 :
[0016] ST2-4-1: Use PDM to make N steps of the bicycle Inference prediction.
[0017] ST2-4-2: Calculate the total cost J of the N-step prediction state π (θ).
[0018] ST2-4-3: Gradient search optimization control strategy π(θ)(handlebar torque τ δ ).
[0019] ST2-5: End the loop until the search converges LOOP2 .
[0020] ST2-6: Update control strategy π(θ * )(Handlebar torque τ δ ).
[0021] ST3: End the loop until the search converges. LOOP1 .
[0022] ST4: Solidify the control strategy π(θ) and the unmanned bicycle achieves self-balancing movement.
[0023] In the above-mentioned inference learning control method, ST2-3 is used to learn the probabilistic dynamic model (PDM) of the unmanned bicycle. Its technical elements are as follows:
[0024] Assume that the system state vector Control input u t =τ δ Assume that the vehicle state of the unmanned bicycle system at the next moment is the same as the vehicle state x at the current moment t , handlebar control amount u t According to the Markov decision process, its dynamic model can be characterized as:
[0025] x t+1 =f(x t ,u t ) (1)
[0026] Where f(·) represents a nonlinear transformation function
[0027] Assume that the state x at time t t It obeys the normal distribution, that is: Among them, μ t is the mean, σ tis the variance. Through the Gaussian process regression method (GPR), combined with the interactive data of bicycle motion, the probabilistic dynamic model (PDM) of the system is obtained as follows:
[0028]
[0029] In the above inference learning control method, ST2-4-2 is used to calculate the total cost J of the N-step predicted state π (θ), its technical elements are as follows:
[0030] Using the PDM obtained from ST2-3, the state of the unmanned bicycle system is predicted in the long term, and the state x0(θ),...,x is recursively calculated. N The probability distribution of (θ) is P(x0(θ)),...,P(x N (θ)).
[0031] Select state x t The cost function is a quadratic function:
[0032]
[0033] Among them, x target =(0,0,0,0) T is the target state vector of bicycle balance control, Λ is a 4×4 diagonal matrix, that is: Λ=diag([L1,L2,L3,L4]), L i (i=1,2,3,4) is the weight of the corresponding dimension. Take c(x t ), and calculate the total cost J of the N-step predicted state according to the following formula π (θ):
[0034]
[0035] In the above inference learning control method, ST2-4-3 is used to solve the optimal strategy parameter θ of the control strategy. * , its technical elements are as follows:
[0036] Construct the following optimization problem to minimize the cost function J π (θ), optimization parameter θ:
[0037]
[0038] Using the conjugate gradient method to solve equation (5), we can get the optimal control strategy parameter θ * , that is, the optimal control strategy π(θ * ).
[0039] Beneficial effects of the present invention:
[0040] 1. The uncertainty reasoning learning control method of the present invention does not need to establish a priori dynamics model of the unmanned bicycle through complicated physical principles, thus simplifying the controller design process.
[0041] 2. The uncertainty reasoning learning control method of the present invention learns the probabilistic dynamic model from the interactive experience data, fully considers the uncertainty of the unmanned bicycle during operation, reduces the impact of model deviation on the control strategy, and improves the robustness and generalization ability of the system. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 Unmanned bicycle prototype and its structural diagram
[0043] Figure 2 Unmanned bicycle prototype control system structure diagram
[0044] Figure 3 Continuous video of unmanned bicycle learning to walk on cement road
[0045] Figure 4 Evolution curve of lateral tilt angle at different learning stages
[0046] Figure 5 Evolution curve of lateral tilt angle velocity at different learning stages
[0047] Figure 6 Handlebar angle evolution curve at different learning stages
[0048] Figure 7 Evolution curve of handlebar angular velocity at different learning stages
[0049] Figure 8 Evolution curve of handlebar control torque at different learning stages
[0050] Fig. 9 Video of unmanned bicycles running on the lawn road
[0051] Fig.10 The lateral inclination angle of a driverless bicycle on a lawn
[0052] Fig.11 The lateral tilt speed of an unmanned bicycle on a lawn
[0053] Fig.12 The handlebar corner of a driverless bicycle on a lawn road
[0054] Fig.13 The handlebar turning speed of a driverless bicycle on a lawn road
[0055] Fig.14 Handlebar control torque of unmanned bicycle running on grass road DETAILED DESCRIPTION
[0056] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention is further described in detail in conjunction with the following specific embodiments, but the present invention is not limited to the following embodiments.
[0057] The unmanned bicycle used in this embodiment is shown in the attached Figure 1 The measurement and control system is attached. Figure 2 .
[0058] The unmanned bicycle prototype is equipped with a single-chip microcomputer system, an electronic control drive unit, a Zigbee module, an on-board power supply system, etc. The incremental encoder and the absolute encoder measure the speed of the bicycle rear wheel, the handlebar angle, etc.; the MPU6050 gyroscope measures the posture of the vehicle body; the STM32 single-chip microcomputer is responsible for receiving the vehicle body status information fed back by each sensor, and servo-controlling the handlebar motor and the wheel motor (the servo cycle is 5ms); the Zigbee module is used to send the status information and control quantity to the mobile terminal (the communication cycle is 50ms); the current sensor collects the driving current of the motor, and the bridge drives the motor according to the PWM signal of the single-chip microcomputer.
[0059] In order to make the embodiment description as simple and understandable as possible, the following linear strategy is selected As a parameterized initial control strategy:
[0060]
[0061] Where:
[0062] is the expanded state vector, and θ=(θ1,θ2,θ3,θ4,θ5) is the control strategy parameter vector to be optimized.
[0063] Consider the handlebar drive motor torque τ of the unmanned bicycle δ Output capacity limitation, in this embodiment The following limiting processing is performed:
[0064]
[0065] Where: u max =τ δmax / 8, τ δmax It is the upper limit of the handlebar motor output torque.
[0066] In this embodiment, in order to reduce the experimental hardware loss caused by the unmanned bicycle falling due to improper strategy at the initial stage and improve the learning efficiency, numerical simulation calculation and manual rough adjustment of strategy parameters are first performed to make the system slightly balanced. Although the effect is ideal, the parameters obtained by rough adjustment can still be used as the initial strategy parameters of reasoning learning, specifically:
[0067] θ=(0.201,0.004,1.303,-0.003,-0.000217) (8)
[0068] Set the maximum torque τ of the handlebar motor output δmax =10(N·m), the inference prediction step length N=40. In addition, when the absolute value of the vehicle body roll angle Or when the absolute value of the handlebar angle |δ|>1rad, straighten the vehicle body and end this round of learning.
[0069] On a cement road, the unmanned bicycle is driven to walk and perform inference learning of balance control, and motion data (x t ,u t ).
[0070] From the data curve of the implementation column (see Appendix Figure 4 ~Attached Figure 8 ) It can be seen that:
[0071] (1) At the initial moment (T = 0), the lateral inclination angle of the unmanned bicycle is The fluctuation of and the swing of the handlebar angle δ are both large, indicating that the system has obvious vibration and poor dynamic performance.
[0072] (2) After two rounds of learning (T = 2), within the first 5 seconds of interaction, the lateral tilt angle of the vehicle The oscillation amplitude is in the range of [-0.020,0.030]rad; after 5s, the lateral inclination angle of the vehicle body The oscillation amplitude decreases and gradually approaches the equilibrium point, and finally oscillates within the range of [-0.010, 0.010] rad.
[0073] (3) After 4 rounds of learning (T = 4), the vehicle body lateral inclination angle The small amplitude oscillation converges in the range of [-0.004, 0.002] rad, and the swing amplitude of the handlebar angle δ is stable in the range of [-0.050, 0.020] rad, indicating that the unmanned bicycle achieves a good straight-line balanced walking.
[0074] In addition, from the attached Figure 3 It can also be seen that after four times of interactive learning, the unmanned bicycle system uses the roadblocks distributed in a straight line and labeled with numbers (0, 1, ..., 9, with a roadblock interval of about 2 meters) as a reference for the travel route, and achieves better self-balancing linear motion.
[0075] It can be seen that the uncertain reasoning learning balance control method provided by the present invention can enable the unmanned bicycle to achieve better balance control learning, and with the deepening of iterative learning, the dynamic performance of the unmanned bicycle's balanced walking is significantly improved.
[0076] In order to examine the robustness and generalization ability of the uncertain inference learning control results, this embodiment applies the control strategy learned on the cement road surface to the lawn road surface (which has different damping, friction, softness, etc. from the cement road surface) and conducts an unmanned bicycle balance control test. The results are shown in the attached figure. Fig. 9 ~Attached Fig.14 shown.
[0077] From the attached Fig.10 ~Attached Fig.14 It can be seen that the balance effect of the unmanned bicycle on the lawn road is consistent with that on the cement road. This shows that even if the driving road surface is changed from a cement road to a lawn road, the uncertainty reasoning learning balance controller of the present invention still maintains good dynamic performance and shows strong robustness and generalization ability.
[0078] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, and therefore the present invention is not limited to the above specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.
Claims
1. An uncertainty reasoning learning control method for self-balancing of an unmanned bicycle, including the following specific algorithm steps: Step 1: Initialize the control strategy π(θ) (handlebar torque τ δ )’s policy parameter θ. Step 2: Loop LOOP1 : Step 2-1: Set the policy π(θ)(handlebar torque τ δ ) is applied to unmanned bicycles. Step 2-2: Collect vehicle status and handlebar control torque τ δ , generate empirical interaction datasets. Step 2-3: Learn the autonomous bicycle probabilistic dynamic model (PDM) from the empirical dataset. Step 2-4: Loop LOOP2 : Step 2-4-1: Use PDM to make N steps of the bicycle Inference prediction. Step 2-4-2: Calculate the total cost J of the N-step prediction state π (θ). Step 2-4-3: Gradient search optimization control strategy π(θ)(handlebar torque τ δ ). Step 2-5: End the loop until the search converges LOOP2 . Step 2-6: Update the control strategy π(θ * )(Handlebar torque τ δ ). Step 3: End the loop until the search converges LOOP1 . Step 4: Solidify the control strategy π(θ) and the unmanned bicycle can achieve self-balancing motion.
2. The vehicle body state according to claim 1 Features: is the lateral tilt angle (tilt velocity) of the vehicle body; is the turning angle of the handlebar (inclination speed).
3. The total cost J of step 2-4-2 according to claim 1 π (θ), characterized in that: and 4. The state x according to claim 3 t The cost c(xt(θ)) is characterized by: Λ is a 4×4 diagonal matrix, Λ=diag([L1,L2,L3,L4]), L i (i=1,2,3,4) is the weight of the corresponding dimension, and x target =(0,0,0,0) T is the target state vector for bicycle balance control.
5. The gradient search optimization control strategy π(θ) of step 2-4-3 according to claim 1, characterized in that: Construct the following optimization problem: The optimal control strategy parameter θ can be obtained by solving it using the conjugate gradient method strategy search method. * , that is, the optimal control strategy π(θ * ).