A washout algorithm based on a data-driven parameter prediction model

Abstract

The application aims to provide a data-driven parameter prediction model-based washout algorithm, which belongs to the technical field of driving simulation. Vehicle longitudinal acceleration signals and motion platform state parameters are obtained and input into a pre-trained multi-layer perception neural network model to calculate and output a washout filter parameter combination adapted to the current working condition, thereby realizing real-time and dynamic adjustment of high-pass acceleration filter parameters and low-pass tilt coordination filter parameters according to the vehicle and motion platform state, completing online real-time optimization of the motion platform control instruction, and finally realizing optimization of the driving motion feeling simulation fidelity. The data-driven parameter prediction model breaks through the limitations of traditional washout algorithm parameters, such as fixedness and poor adaptability; the combination of offline optimization and online prediction balances real-time performance and fidelity; the active centering controller compensates the longitudinal acceleration by using the pitch angle, improves the space utilization rate, and introduces a comprehensive somatosensory error evaluation index to accurately evaluate the somatosensory simulation effect.

Description

Technical Field

[0001] This invention relates to the field of automotive driving motion simulation technology, and in particular to a washing algorithm based on a data-driven parameter prediction model. Background Technology

[0002] As an important tool for driver training, vehicle performance testing, and road safety research, the effectiveness of car driving simulators is closely related to the realism of the motion sensations experienced by the driver during use. However, due to the limitations of physical space, it is technically and economically unrealistic to completely reproduce car behavior. Therefore, how to achieve higher motion sensation realism within a limited motion space has become the core issue in the research of Motion Cueing Algorithm (MCA).

[0003] Traditional motion simulation algorithms are mainly represented by the classic washout algorithm. Its core mechanism involves selecting the desired frequency of motion signals using a combination of high-pass and low-pass filters, converting the motion signals of vehicles such as cars and airplanes into feasible displacement commands for the simulator platform. This algorithm employs a multi-channel signal processing and tilt coordination mechanism. Its main structure includes input signal preprocessing, coordinate transformation, a high-pass acceleration channel, a low-pass tilt coordination channel, and a high-pass angular velocity channel. The high-pass filter is used for transient acceleration simulation, and the low-pass filter is used for gravitational acceleration tilt compensation to adapt to the limited space of the platform. While the classic washout algorithm is simple in structure and has good real-time performance, its fixed filter parameters cannot be adjusted in real time according to changes in platform status and operating conditions, resulting in poor adaptability and limited realism in motion perception.

[0004] To overcome the limitations of classic washout algorithms, researchers have proposed several improved methods. Among them, the Model Predictive Control (MPC) motion simulation algorithm dynamically controls the output of motion commands by constructing and solving an optimization model in each control cycle, thereby improving simulation accuracy and platform space utilization efficiency. While this method improves the accuracy of motion simulation to some extent, it suffers from high computational complexity, large online optimization computation load, and insufficient real-time performance, making it difficult to apply to high-speed dynamic conditions. Furthermore, improved methods such as optimal control washout algorithms and adaptive washout algorithms also exhibit similar limitations under complex conditions.

[0005] A common problem faced by existing technical solutions is the difficulty in effectively balancing motion simulation error and real-time performance. On the one hand, improving motion realism requires complex online optimization calculations, which inevitably affects the system's real-time response capability. On the other hand, ensuring real-time performance often necessitates simplifying the algorithm structure or using fixed parameters, which reduces the accuracy and adaptability of the motion simulation. Furthermore, existing solutions generally lack efficient active centering mechanisms; when the simulation platform approaches its travel limits, a passive centering strategy is typically employed, which not only affects motion realism but also leads to low platform space utilization efficiency.

[0006] Therefore, designing a motion simulation algorithm that can guarantee real-time performance, adaptively adjust parameters under different working conditions to improve motion realism, and has an efficient centering mechanism to improve platform space utilization has become a pressing technical problem. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a washing algorithm based on a data-driven parameter prediction model. This algorithm achieves real-time prediction and adaptive adjustment of washing filter parameters by pre-training a multilayer perceptron neural network model. Specifically, it obtains the vehicle's longitudinal acceleration signal and the motion platform's state parameters as inputs to the neural network, dynamically outputs the optimal combination of high-pass acceleration filter parameters and low-pass tilt coordination filter parameters, and constructs corresponding filters based on these parameters to process the vehicle signal and generate platform longitudinal acceleration and pitch angle commands. This allows for adaptive control of the motion platform's longitudinal and pitch movements according to real-time operating conditions.

[0008] To achieve the above objectives, this invention provides a washing algorithm based on a data-driven parameter prediction model, comprising the following steps:

[0009] Step S1: Obtain the vehicle longitudinal acceleration signal and motion platform state parameters, wherein the motion platform state parameters include the platform longitudinal displacement and the platform longitudinal velocity;

[0010] Step S2: Input the vehicle longitudinal acceleration signal and the motion platform state parameters into a pre-trained multilayer perceptron neural network model, and output the washout filter parameter combination under the current working condition. The washout filter parameter combination includes high-pass acceleration filter parameters and low-pass tilt coordination filter parameters.

[0011] Step S3: Construct a high-pass acceleration filter and a low-pass tilt coordination filter based on the parameter combination of the washing filter. Process the vehicle longitudinal acceleration signal through the high-pass acceleration filter and the low-pass tilt coordination filter respectively to obtain the platform longitudinal acceleration command and the platform pitch angle command.

[0012] Step S4: Control the longitudinal movement of the motion platform according to the longitudinal acceleration command of the platform, and control the pitch movement of the motion platform according to the pitch angle command of the platform to realize the motion simulation of vehicle movement.

[0013] By employing a data-driven parameter prediction approach, dynamic adjustment and intelligent optimization of the washout algorithm parameters are achieved. Compared to traditional washout algorithms with fixed parameters, this invention can adjust filter parameters in real time based on the actual vehicle motion state and the current platform state, significantly improving the realism of motion simulation. Simultaneously, it maintains the algorithm's real-time performance, avoiding complex online optimization calculations, and providing a highly efficient and accurate motion simulation solution for driving simulators.

[0014] Furthermore, the multilayer perceptron neural network model in step S2 is pre-trained in the following manner:

[0015] Discrete sample points are selected from the influencing factor space using a non-uniform granular Latin hypercube sampling method based on sensitivity analysis. The influencing factors include vehicle longitudinal acceleration increment and platform longitudinal velocity.

[0016] For each discrete sample point, the NSGA-II algorithm is used for multi-objective optimization to obtain the corresponding Pareto optimal shuffling filter parameter combination;

[0017] The discrete sample points are used as input data, and the corresponding Pareto optimal shuffle filter parameter combinations are used as output data to train a multilayer perceptron neural network model.

[0018] A non-uniform granular Latin hypercube sampling method based on sensitivity analysis was adopted to ensure the uniform distribution and representativeness of training samples in the parameter space, thereby improving the generalization ability of the neural network model. The Pareto optimal solution set was obtained by combining the NSGA-II multi-objective optimization algorithm, providing high-quality training data for the neural network. This enabled the trained model to accurately predict the optimal parameter combination under various operating conditions, enhancing the algorithm's adaptability and robustness.

[0019] Furthermore, the sensitivity analysis is achieved in the following manner:

[0020] Determine the initial combination of influencing factors, use the particle swarm optimization algorithm for single-objective optimization, and record the optimal combination of washing filter parameters.

[0021] Each influencing factor is perturbed by 10%, while keeping other influencing factors unchanged, and then a single-objective optimization is performed again;

[0022] By comparing the output changes before and after the disturbance, the sensitivity of each influencing factor to the washing filter parameters is evaluated, and the sampling ratio of the influencing factors is set according to the sensitivity results.

[0023] Furthermore, the multi-objective optimization of the NSGA-II algorithm includes the following objective function:

[0024] Minimize the perceptual error objective function:

[0025] Objective function to maximize platform space utilization: J2 = max(|x platform |);

[0026] Where N is the total number of sampling points, f real (i) represents the real vehicle perception ratio at the i-th sampling time, f sim (i) represents the sensing force of the motion platform at the i-th sampling time, x platform This represents the longitudinal displacement of the platform.

[0027] Furthermore, the high-pass accelerometer filter in step S3 is a third-order filter, and its transfer function is:

[0028]

[0029] Among them, K hp ω is the gain coefficient. hp ζ is the natural response frequency of a second-order element. hp ω is the damping coefficient. hp1 is the cutoff frequency of the first-order element, s is the Laplace transform operator, and represents a complex variable.

[0030] Furthermore, the low-pass tilt coordinated filter in step S3 is a second-order filter, and its transfer function is:

[0031]

[0032] Where, ω lp ζ is the natural response frequency. lp is the damping coefficient.

[0033] Furthermore, step S4 also includes active centering control, which is achieved by: detecting the degree to which the longitudinal displacement of the motion platform deviates from the center position; and calculating the desired centering acceleration a based on the degree of deviation. return The centering acceleration is achieved by controlling the platform's pitch angle θ to generate the gravitational acceleration component gsinθ, which enables the platform to quickly return to the center position. s is the Laplace transform operator, representing a complex variable.

[0034] Furthermore, the pitch angular velocity in the active homing control satisfies the following constraints: in, Let θ be the pitch angular velocity. threshold The threshold for sensing the pitch angular velocity of the human body.

[0035] Furthermore, the somatosensory simulation effect in step S4 is evaluated using a human vestibular model, the transfer function of which is:

[0036]

[0037] Where, ω vest ζ represents the natural frequency of the vestibular model. vest is the damping coefficient of the vestibular model, s is the Laplace transform operator, and represents a complex variable.

[0038] Furthermore, the washing algorithm based on the data-driven parameter prediction model also includes a body sensation evaluation step:

[0039] Calculate the root mean square error (RMSE) and Pearson correlation coefficient (r) between the perceived force generated by the motion platform and the perceived force of the real vehicle.

[0040] The comprehensive perceptual error index E is calculated based on the RMSE and the Pearson correlation coefficient r. combined =α·RMSE+(1-α)·(1-r), where α is the weighting coefficient;

[0041] The effectiveness of the motion simulation is evaluated based on the comprehensive motion error index.

[0042] Compared with the prior art, the present invention has at least the following beneficial effects:

[0043] (1) This invention achieves a fundamental shift from the traditional fixed-parameter mode to an intelligent adaptive mode by constructing a washout algorithm based on a data-driven parameter prediction model. This algorithm can dynamically predict the optimal combination of washout filter parameters based on the vehicle's longitudinal acceleration signal and the real-time state of the motion platform using a pre-trained multilayer perceptron neural network model, thereby significantly improving the realism of motion simulation while ensuring real-time performance. Compared to the limitations of traditional classic washout algorithms with fixed parameters and poor adaptability to operating conditions, this invention achieves real-time parameter optimization and adjustment, effectively solving the technical problem of insufficient realism in motion simulation and providing a motion experience for driving simulators that is closer to real driving.

[0044] (2) This invention employs an innovative multi-level optimization strategy in neural network model construction and training. It identifies key influencing factors through sensitivity analysis, utilizes non-uniform granular Latin hypercube sampling to ensure the representativeness and spatial coverage of training samples, and combines the NSGA-II multi-objective optimization algorithm to simultaneously optimize two mutually constraining objective functions: minimizing haptic error and maximizing platform space utilization. This systematic optimization method not only significantly reduces offline computation and training time, but more importantly, it provides the neural network with a high-quality Pareto optimal solution set as training data. This enables the model to possess excellent generalization ability and adaptability to various driving scenarios, accurately predicting the optimal parameter combination and achieving the best balance between haptic realism and platform space utilization efficiency.

[0045] (3) In terms of filter design and control strategy, this invention adopts precise mathematical modeling of a third-order high-pass acceleration filter and a second-order low-pass tilt coordination filter, which can more accurately process acceleration signals of different frequency components, especially showing excellent performance in transient acceleration simulation and gravity compensation for continuous acceleration. More importantly, this invention introduces an active centering control mechanism, which achieves rapid centering and reset by cleverly utilizing the gravitational acceleration component generated by the platform pitch angle. Under the constraint of strictly controlling the pitch angular velocity below the human perception threshold, it not only avoids the driver perceiving additional pitch motion, but also significantly improves the utilization efficiency of the platform's longitudinal space, solving the technical problems of low efficiency and poor tactile continuity of traditional passive centering methods.

[0046] (4) This invention combines the root mean square error (RMSE) with the Pearson correlation coefficient to construct a more comprehensive and accurate haptic evaluation system. RMSE reflects the degree of difference in signal amplitude, while the Pearson correlation coefficient reflects the consistency of signal change trends. The combination of the two can simultaneously evaluate the performance of haptic simulation in both numerical accuracy and dynamic characteristics. This comprehensive evaluation method overcomes the limitations of single-index evaluation and provides a more scientific evaluation standard that is closer to the actual user experience for the performance evaluation of driving simulators. Attached Figure Description

[0047] The accompanying drawings, which form part of this specification, illustrate embodiments of the invention and, together with the specification, serve to explain the principles of the invention.

[0048] The invention will be more clearly understood with reference to the accompanying drawings and the following detailed description, wherein:

[0049] Figure 1 This is a schematic diagram of the active centering controller provided in an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram illustrating the changes in filter parameters with influencing factors provided in an embodiment of the present invention;

[0051] Figure 3 This is a heatmap showing the sensitivity of filter parameters to various influencing factors provided in this embodiment of the invention.

[0052] Figure 4 This is a schematic diagram of Latin hypercube sampling provided in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of the parameter optimization model based on NSGA-II provided in an embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram of the Parato front curve provided in an embodiment of the present invention;

[0055] Figure 7 This is a schematic diagram of an MLP structure with multiple hidden layers provided in an embodiment of the present invention;

[0056] Figure 8 This is a schematic diagram of the online verification model provided in an embodiment of the present invention;

[0057] Figure 9 This is a schematic diagram of a step acceleration signal provided in an embodiment of the present invention;

[0058] Figure 10 This is a schematic diagram of the step acceleration test condition verification provided in an embodiment of the present invention;

[0059] Figure 11 This is a schematic diagram illustrating the noise reduction performance verification provided in an embodiment of the present invention;

[0060] Figure 12 This is a schematic diagram of FTP-75 operating condition verification provided in an embodiment of the present invention. Detailed Implementation

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, and not limitations thereof. Where there is no conflict, the embodiments and technical features in the embodiments can be combined with each other. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0062] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0063] like Figure 1As shown, this invention provides a washing algorithm based on a data-driven parameter prediction model, the method comprising:

[0064] Step S1: Obtain the vehicle longitudinal acceleration signal and motion platform state parameters, wherein the motion platform state parameters include the platform longitudinal displacement and the platform longitudinal velocity;

[0065] Step S2: Input the vehicle longitudinal acceleration signal and the motion platform state parameters into a pre-trained multilayer perceptron neural network model, and output the washout filter parameter combination under the current working condition. The washout filter parameter combination includes high-pass acceleration filter parameters and low-pass tilt coordination filter parameters.

[0066] Step S3: Construct a high-pass acceleration filter and a low-pass tilt coordination filter based on the parameter combination of the washing filter. Process the vehicle longitudinal acceleration signal through the high-pass acceleration filter and the low-pass tilt coordination filter respectively to obtain the platform longitudinal acceleration command and the platform pitch angle command.

[0067] Step S4: Control the longitudinal movement of the motion platform according to the longitudinal acceleration command of the platform, and control the pitch movement of the motion platform according to the pitch angle command of the platform to realize the motion simulation of vehicle movement.

[0068] By employing a data-driven parameter prediction approach, dynamic adjustment and intelligent optimization of the washout algorithm parameters are achieved. Compared to traditional washout algorithms with fixed parameters, this invention can adjust filter parameters in real time based on the actual vehicle motion state and the current platform state, significantly improving the realism of motion simulation. Simultaneously, it maintains the algorithm's real-time performance, avoiding complex online optimization calculations, and providing a highly efficient and accurate motion simulation solution for driving simulators.

[0069] Furthermore, the multilayer perceptron neural network model in step S2 is pre-trained in the following manner:

[0070] Discrete sample points are selected from the influencing factor space using a non-uniform granular Latin hypercube sampling method based on sensitivity analysis. The influencing factors include vehicle longitudinal acceleration increment and platform longitudinal velocity.

[0071] For each discrete sample point, the NSGA-II algorithm is used for multi-objective optimization to obtain the corresponding Pareto optimal shuffling filter parameter combination;

[0072] The discrete sample points are used as input data, and the corresponding Pareto optimal shuffle filter parameter combinations are used as output data to train a multilayer perceptron neural network model.

[0073] A non-uniform granular Latin hypercube sampling method based on sensitivity analysis was adopted to ensure the uniform distribution and representativeness of training samples in the parameter space, thereby improving the generalization ability of the neural network model. The Pareto optimal solution set was obtained by combining the NSGA-II multi-objective optimization algorithm, providing high-quality training data for the neural network. This enabled the trained model to accurately predict the optimal parameter combination under various operating conditions, enhancing the algorithm's adaptability and robustness.

[0074] Furthermore, the sensitivity analysis is achieved in the following manner:

[0075] Determine the initial combination of influencing factors, use the particle swarm optimization algorithm for single-objective optimization, and record the optimal combination of washing filter parameters.

[0076] Each influencing factor is perturbed by 10%, while keeping other influencing factors unchanged, and then a single-objective optimization is performed again;

[0077] By comparing the output changes before and after the disturbance, the sensitivity of each influencing factor to the washing filter parameters is evaluated, and the sampling ratio of the influencing factors is set according to the sensitivity results.

[0078] Through systematic sensitivity analysis, the key factors most significantly affecting the parameters of the washing filter were identified, and the sampling ratio was rationally allocated, avoiding blind uniform sampling. This method significantly reduces the computational load of offline optimization, improves the efficiency of parameter space exploration, and ensures that important influencing factors are adequately sampled, providing a more effective data foundation for subsequent neural network training.

[0079] Furthermore, the multi-objective optimization of the NSGA-II algorithm includes the following objective function:

[0080] Minimize the perceptual error objective function:

[0081] Objective function to maximize platform space utilization: J2 = max(|x platform |);

[0082] Where N is the total number of sampling points, f real (i) represents the real vehicle perception ratio at the i-th sampling time, f sim (i) represents the sensing force of the motion platform at the i-th sampling time, x platform This represents the longitudinal displacement of the platform.

[0083] By simultaneously optimizing two mutually constraining objectives—motional error and platform space utilization—the optimal balance between motion realism and platform space utilization efficiency is achieved. The motional error objective function ensures a high degree of consistency between simulated motion and real perception, while the platform space utilization objective function guarantees a wider range of motion simulations within a limited physical space, maximizing the performance potential of the driving simulator.

[0084] Furthermore, the high-pass accelerometer filter in step S3 is a third-order filter, and its transfer function is:

[0085]

[0086] Among them, K hp ω is the gain coefficient. hp ζ is the natural response frequency of a second-order element. hp ω is the damping coefficient. hp1 The cutoff frequency is set to the first-order element, and a third-order high-pass filter design is employed. Compared to traditional low-order filters, this design provides more precise frequency response control and better transient acceleration simulation. By appropriately setting parameters such as gain coefficient, natural response frequency, and damping coefficient, low-frequency signal components can be effectively filtered out, highlighting the instantaneous acceleration sensation and providing the driver with a more realistic acceleration experience. This is particularly evident in dynamic conditions such as rapid acceleration and deceleration, where s is the Laplace transform operator, representing a complex variable.

[0087] Furthermore, the low-pass tilt coordinated filter in step S3 is a second-order filter, and its transfer function is:

[0088]

[0089] Where, ω lp ζ is the natural response frequency. lp The damping coefficient is represented by a second-order low-pass filter specifically designed for the tilt coordination mechanism, which can effectively extract low-frequency acceleration components and convert them into platform pitch angle control commands. By precisely controlling the natural response frequency and damping coefficient, gravity tilt compensation for continuous acceleration is achieved. Under the premise of not exceeding the physical limitations of the platform, the continuous longitudinal acceleration sensation is simulated using gravity acceleration components, significantly expanding the range of simulable acceleration. 's' is the Laplace transform operator, representing a complex variable.

[0090] Furthermore, step S4 also includes active centering control, which is implemented in the following way:

[0091] Detect the degree to which the longitudinal displacement of the motion platform deviates from the center position;

[0092] The expected homing acceleration a is calculated based on the degree of deviation. return ;

[0093] The centering acceleration is achieved by controlling the platform's pitch angle θ to generate a gravitational acceleration component gsinθ, enabling the platform to quickly return to the center position.

[0094] The active centering control mechanism intelligently utilizes the gravitational acceleration component generated by the platform's pitch angle to achieve rapid platform centering and reset, significantly improving the utilization efficiency of the platform's longitudinal space. Compared to the traditional passive centering method, active centering not only has a faster centering speed but also maintains a certain degree of haptic continuity during the centering process, avoiding interruptions in the haptic simulation and providing a better user experience for continuous, long-duration driving simulations.

[0095] Furthermore, the pitch angular velocity in the active homing control satisfies the following constraints:

[0096]

[0097] in, Let θ be the pitch angular velocity. threshold The threshold for sensing the pitch angular velocity of the human body.

[0098] By strictly controlling the pitch angular velocity below the human perception threshold, it is ensured that the driver will not perceive any additional pitch movement during active centering, thus avoiding motion illusion and discomfort. This constraint guarantees the concealment and effectiveness of the tilt coordination mechanism, making the use of the gravitational acceleration component completely transparent to the user's perception, achieving the ingenious application of technical means without affecting the user's real driving experience.

[0099] Furthermore, the somatosensory simulation effect in step S4 is evaluated using a human vestibular model, the transfer function of which is:

[0100]

[0101] Where, ω vest ζ represents the natural frequency of the vestibular model. vest is the damping coefficient of the vestibular model, s is the Laplace transform operator, representing a complex variable. Specifically, the transfer function is a result of transforming a time-domain function to the complex frequency domain, also known as a complex frequency domain function.

[0102] Introducing a vestibular model that conforms to human physiological characteristics as an evaluation standard makes the assessment of somatosensory simulation effects more scientific and objective. The vestibular model accurately reflects the human body's true perception of acceleration stimuli. Compared to simple mathematical error calculations, evaluation results based on the vestibular model are closer to the driver's subjective experience, providing more meaningful guidance for algorithm optimization and ensuring the biological rationality of somatosensory simulation.

[0103] Furthermore, it also includes a sensory evaluation step:

[0104] Calculate the root mean square error (RMSE) and Pearson correlation coefficient (r) between the perceived force generated by the motion platform and the perceived force of the real vehicle.

[0105] The comprehensive perceptual error index E is calculated based on the RMSE and the Pearson correlation coefficient r. combined =α·RMSE+(1-α)·(1-r), where α is the weighting coefficient;

[0106] The effectiveness of the motion simulation is evaluated based on the comprehensive motion error index.

[0107] This paper innovatively combines the root mean square error (RMSE) with the Pearson correlation coefficient to construct a more comprehensive and accurate haptic evaluation system. RMSE reflects the degree of difference in signal amplitude, while the Pearson correlation coefficient reflects the consistency of signal change trends. Combining the two allows for simultaneous evaluation of the haptic simulation's performance in both numerical accuracy and dynamic characteristics. This comprehensive evaluation method overcomes the limitations of single-index evaluation, providing a more scientific and user-centric evaluation standard for driving simulator performance assessment.

[0108] In summary, this invention achieves adaptive adjustment of washing algorithm parameters through a data-driven MLP parameter prediction model, solving the problems of fixed parameters and poor adaptability to operating conditions in classic washing algorithms, thereby improving the realism of the motion perception. Simultaneously, the complex optimization process (NSGA-II) is performed offline, while parameters are quickly predicted online via MLP, effectively reducing online computational complexity and addressing the issues of high computational load and insufficient real-time performance in MPC algorithms. It meets real-time requirements while maintaining high realism. Furthermore, an active centering device is designed, using pitch angle compensation for longitudinal acceleration to accelerate platform centering without affecting the motion perception, effectively improving the longitudinal space utilization within a limited physical travel distance and compensating for the lack of an efficient centering mechanism in existing solutions. In addition, the non-uniform granular Latin hypercube sampling based on sensitivity analysis ensures the generalization ability of the MLP model under a wide range of operating conditions, giving the algorithm better adaptability to operating conditions. Noise resistance testing verifies the stability and parameter prediction ability of the algorithm in the presence of noise interference.

[0109] Example 1

[0110] This embodiment provides a preferred implementation process for a washout algorithm based on a data-driven parameter prediction model:

[0111] This technical solution employs a data-driven parameter prediction model-based algorithm to optimize the dynamic simulation of the driving simulator in the longitudinal degree of freedom. The specific implementation plan includes the following four main modules and steps:

[0112] The first step is to construct a somatosensory evaluation model: The human vestibular system is the main organ for sensing motion, with the otolith organs primarily sensing linear acceleration (including gravitational acceleration) and the semicircular canals primarily sensing angular velocity. In this invention, the focus is on optimizing the somatosensory simulation of the longitudinal degree of freedom; therefore, the otolith model is primarily used to sense longitudinal acceleration. The perception of linear acceleration by the otolith organs can be approximated by a second-order high-pass filter, with the transfer function being:

[0113]

[0114] In the formula, f, These represent the input and output perceptual ratios of the otolith model, respectively, k and τ. L τ S These are the parameters for the otolith model.

[0115] Secondly, the core of the somatosensory evaluation model is to quantify the deviation between the expected somatosensory stimulus and the somatosensory stimulus generated by the actual movement of the simulator. This invention uses mean squared error (MSE) as the evaluation index.

[0116]

[0117] In the formula, N is the total number of sampling points, and t i f represents the i-th sampling time. v For a realistic car perception comparison, f p To generate a sense of force for changes in the pose of the motion platform.

[0118] The second step is the design of an active centering device: Traditional washout algorithms typically employ a passive centering strategy when the platform approaches its travel limit, which affects the realism of the motion perception. This invention proposes an active centering device, the core idea of ​​which utilizes the principle of tilt coordination. When the platform deviates from its center position during longitudinal movement, the active centering device adjusts the pitch angle of the moving platform, projecting a component of gravitational acceleration onto the longitudinal direction to generate a virtual acceleration opposite to the platform's longitudinal acceleration, thereby prompting the platform to quickly return to its center position. Figure 1 This tilt is usually below the human body's angular velocity perception threshold to avoid being noticed by the driver.

[0119] When the simulator platform changes its pitch angle θ, gravitational acceleration will produce a longitudinal component gsinθ in the platform coordinate system. This component can be used to simulate or counteract part of the longitudinal acceleration.

[0120] Compensation principle: By precisely controlling the platform's pitch angle θ, gsinθ generates a force opposite to the actual longitudinal acceleration of the platform, thereby achieving a partial simulation of longitudinal acceleration or a centering force without relying on the platform's own longitudinal motion.

[0121] Mathematical description: Let a be the desired centering acceleration of the platform in the longitudinal direction. r This acceleration component can be generated by controlling the platform's pitch angle θ.

[0122] a r =gsinθ (3)

[0123] The further the platform deviates from the center, the greater the expected acceleration to return to center. The platform's pitch angle gradually increases to ensure that the resulting pitch velocity is below the human semicircular canal's perception threshold, thus avoiding motion illusions.

[0124]

[0125] In the formula, dθ / dt is the pitch angular velocity. The threshold for sensing the pitch angular velocity of the human body.

[0126] The third step, multi-objective optimization model construction and parameter prediction: The process of establishing a filter parameter optimization model for the longitudinal-pitch channel based on the vehicle's motion characteristics is as follows:

[0127] The washout algorithm proposed in this invention employs a combination of filters in the longitudinal-pitch channel to optimize motion simulation, adapting to the limited motion travel of the simulator. When the vehicle's acceleration changes, the optimal set of washout filter parameters is obtained by combining the force signal input to the washout filter with the motion platform state parameters, and then replaced, thus achieving real-time optimization. The selection of filter parameters directly affects the motion perception effect and platform utilization; therefore, an optimization model needs to be constructed for accurate solution.

[0128] The longitudinal high-pass accelerometer is a third-order filter, and its transfer function is:

[0129]

[0130] In the formula, ω ah ξ is the natural response frequency of the second-order element of the high-pass accelerometer filter, ω0 is the cutoff frequency of the first-order element of the high-pass accelerometer filter, and ξ is the frequency of the second-order element of the high-pass accelerometer filter. ah k is the damping coefficient of the high-pass accelerometer filter. ah This is the gain coefficient, ranging between 0 and 1. These are four of the decision variables (parameters to be optimized).

[0131] The tilt-coordinated low-pass filter is second-order, and its transfer function is:

[0132]

[0133] In the formula, ω al Let ξ be the natural response frequency of the low-pass accelerometer filter.al This is the damping coefficient of the low-pass accelerometer. The above are two other parameters to be optimized.

[0134] Human perception error model:

[0135]

[0136] In the formula, L IS This is the force coordinate transformation matrix, used to transform the motion parameters obtained in the vehicle's moving coordinate system to the platform's inertial coordinate system. The matrix is ​​as follows:

[0137]

[0138] In the formula: φ, θ, ψ are the attitude angles of the motion platform, sθ is sinθ, and cθ is cosθ.

[0139] Platform longitudinal displacement model:

[0140]

[0141] Table 1 Optimization Model Expression

[0142]

[0143]

[0144] In the formula: T is the total time for single-condition parameter optimization.

[0145] The process of non-uniform granular Latin hypercube sampling based on sensitivity analysis is as follows:

[0146] Since the influencing factor space is two-dimensional and continuous, in order to balance the time efficiency of offline optimization with the accuracy of parameter prediction, it is necessary to select representative discrete sample points from the complex influencing factor space. This invention employs a non-uniform granular Latin hypercube sampling method based on sensitivity analysis to achieve this goal.

[0147] Platform displacement limitation is tentatively considered as one of the influencing factors. First, an initial set of influencing factors (vehicle acceleration increment, platform speed, and platform displacement limitation) is determined. A single-objective (F1) solution is then performed using the Particle Swarm Optimization (PSO) algorithm, and the optimal combination of output filter parameters under the current influencing factors is recorded. Subsequently, based on the initial influencing factors, a 10% perturbation is added to one parameter, while keeping the other two influencing factors unchanged. PSO is then used again to solve for the new combination of influencing factors, and the results are recorded. This process is repeated five times for each influencing factor. By comparing the output changes before and after the perturbation, the sensitivity of each influencing factor to the filter parameters is evaluated. Figure 2 , Figure 3Based on the sensitivity analysis results, the ratio of influencing factors was set to 6:3:1. Platform displacement limitation, as an influencing factor with low sensitivity, can be transformed from an influencing factor into an objective function to reduce the dimensionality of the influencing factors, thereby reducing the number of offline optimizations.

[0148] Latin hypercube sampling divides the sample space into multiple levels according to a ratio, and randomly selects sample points within each level, thereby ensuring a uniform distribution of samples throughout the parameter space. Figure 4 .

[0149] In the third step, the offline parameter optimization process using the NSGA-II algorithm is as follows:

[0150] Offline multi-objective optimization process such as Figure 5 The offline optimization of the NSGA-II algorithm obtains a Pareto optimal solution set for each typical working condition sampled by LHS, such as... Figure 6 Each solution contains a set of optimized parameters for the rinsing algorithm and the corresponding multi-objective function values. These offline optimization results establish a nonlinear mapping from influencing factors (vehicle motion parameters and platform state) to rinsing algorithm parameters (filter parameters). This data will serve as a valuable dataset for the next step of training the MLP neural network.

[0151] The third step involves constructing a multilayer perceptron (MLP) neural network parameter prediction model as follows:

[0152] To meet the needs of real-time driving simulators, this invention utilizes the powerful nonlinear mapping and generalization capabilities of MLP neural networks to construct a data-driven parameter prediction model, enabling real-time and rapid prediction of algorithm parameters, such as... Figure 7 .

[0153] First, the dataset was constructed, derived from the aforementioned NSGA-II offline optimization results. The input data consisted of a combination of influencing factors and platform displacement under typical operating conditions. The output data came from the Pareto optimal shuffling algorithm parameters obtained by NSGA-II optimization under the corresponding operating conditions. The training set to test set ratio was set to 8:2. Second, the network structure design, activation function, loss function, and optimization algorithm were selected. MLP is a feedforward neural network consisting of an input layer, one or more hidden layers, and an output layer. The number of neurons in the input layer matched the number of input parameters in the dataset. These influencing factors typically include the vehicle's longitudinal acceleration increment, platform speed, and remaining longitudinal displacement of the platform at the current moment. The number of hidden layers was determined through trial and error, using the Sigmoid function to enhance the network's nonlinear mapping capability. The number of neurons in the output layer matched the number of filter parameters, using a linear activation function. Mean squared error (MSE) was used as the loss function to measure the difference between the predicted parameters and the actual optimized parameters of the MLP model. Bayesian regularization was used as the optimization algorithm, which has good compatibility with the Sigmoid activation function used. The next step is the hyperparameter selection process. Performance is evaluated through cross-validation, and hyperparameters such as the learning rate and the number of hidden layer neurons are adjusted to obtain optimal model performance. After setting, the MLP neural network is trained. The trained and validated MLP neural network model is then deployed into the driving simulation algorithm structure. During online dynamic simulation, the MLP receives real-time inputs from factors influencing vehicle kinematic parameters and sensor data (such as current longitudinal acceleration, pitch rate, platform position, etc.). The MLP model instantly outputs the predicted optimal parameters for the current driving condition. These predicted parameters are dynamically input into the driving algorithm, thereby achieving data-driven, real-time, and adaptive dynamic simulation optimization.

[0154] Step 4: Online motion simulation testing and verification:

[0155] The washing algorithm based on a data-driven parameter prediction model proposed in this invention, after completing offline optimization and MLP model training, needs to undergo a series of online tests and verifications to comprehensively evaluate its performance, robustness, and haptic realism in actual driving simulation scenarios. The online verification model includes... Figure 8 The testing and verification process will strictly follow the preset operating conditions and use mean squared error (MSE) as the core evaluation indicator.

[0156] The step response test aims to evaluate the transient response performance of the algorithm of this invention when faced with sudden changes in vehicle motion signals. In the test, a continuous step signal that changes every few seconds is input into the driving simulator system to simulate the instantaneous acceleration or deceleration of the vehicle. Figure 9During the test, the pose changes of the motion platform were collected in real time, and the perceived force was calculated using a human vestibular model. This was then compared with the actual perceived force. Figure 10 As shown in Table 2, the MSE is calculated. A smaller MSE indicates that the algorithm can provide higher haptic realism in transient response.

[0157] Table 2 Comparison of Step Acceleration Signal Results

[0158]

[0159] Noise robustness testing is used to verify the robustness of the algorithm in the presence of noise interference. The test simulates sensor or transmission noise in a real driving environment by superimposing Gaussian white noise into the aforementioned step response test. Dynamic simulation verification is performed under noisy sensor feedback signals. Performance evaluation primarily assesses the algorithm's noise robustness by comparing the longitudinal perceptual MSE changes under noise-free conditions. Figure 11 The increase in MSE is not significant, indicating that the algorithm has good robustness, as shown in Table 3.

[0160] Table 3 Noise immunity verification results

[0161]

[0162] The FTP-75 driving cycle test aims to comprehensively evaluate the robustness of the algorithm of this invention under typical, continuous driving cycles. This test uses motion parameters obtained from the vehicle according to the FTP75 driving cycle as input. This cycle includes various driving modes commonly encountered in urban driving, such as start-up, acceleration, cruising, deceleration, and idling. The core performance evaluation metric is longitudinal motion-sensing MSE; a lower MSE value demonstrates the high robustness of the algorithm under complex conditions. Furthermore, pitch motion-sensing MSE is calculated to verify the imperceptibility of pitch angle changes, and the maximum and average displacement of the platform in the longitudinal direction are recorded to ensure that the platform does not exceed its motion limits. Figure 12 As shown.

[0163] Table 4. User Experience Evaluation under FTP-75 Operating Conditions

[0164]

[0165] In summary, this technical solution helps reduce longitudinal motion simulation efficiency (MSE) and provides a more realistic driving experience because the algorithm can better match human perception characteristics and dynamically adjust according to operating conditions. Through offline optimization and online prediction, this technical solution effectively meets the real-time requirements of driving simulators while ensuring motion realism. The active centering device helps the platform simulate more diverse longitudinal movements within a relatively small actual displacement range, which has a positive effect on improving space utilization. The algorithm of this technical solution can adaptively adjust according to real-time vehicle motion parameters and platform status, and maintain relatively stable performance in the presence of noise interference. This invention combines a complex optimization process with an intelligent prediction model, providing a novel approach to the design of motion simulation algorithms for driving simulators. Through one-time offline optimization and MLP model training, it helps to cope with diverse driving scenarios, simplifies online control logic, and helps reduce the difficulty of development and debugging, thereby promoting the overall efficiency of driving simulator system development.

[0166] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A washing algorithm based on a data-driven parameter prediction model, characterized in that, Includes the following steps: Step S1: Obtain the vehicle longitudinal acceleration signal and motion platform state parameters, wherein the motion platform state parameters include the platform longitudinal displacement and the platform longitudinal velocity; Step S2: Input the vehicle longitudinal acceleration signal and the motion platform state parameters into a pre-trained multilayer perceptron neural network model, and output the washout filter parameter combination under the current working condition. The washout filter parameter combination includes high-pass acceleration filter parameters and low-pass tilt coordination filter parameters. The multilayer perceptron neural network model is pre-trained in the following way: Discrete sample points are selected from the influencing factor space using a non-uniform granular Latin hypercube sampling method based on sensitivity analysis. The influencing factors include vehicle longitudinal acceleration increment and platform longitudinal velocity. For each discrete sample point, the NSGA-II algorithm is used for multi-objective optimization to obtain the corresponding Pareto optimal shuffling filter parameter combination; The discrete sample points are used as input data, and the corresponding Pareto optimal shuffle filter parameter combination is used as output data to train a multilayer perceptron neural network model. The multi-objective optimization of the NSGA-II algorithm includes the following objective functions: Minimize the perceptual error objective function: ; Objective function to maximize platform space utilization: ; in, The total number of sampling points. For the first The actual vehicle perception at each sampling time point is compared. For the first The motion platform sensing capability at each sampling time point For the longitudinal displacement of the platform Step S3: Construct a high-pass acceleration filter and a low-pass tilt coordination filter based on the parameter combination of the washing filter. Process the vehicle longitudinal acceleration signal through the high-pass acceleration filter and the low-pass tilt coordination filter respectively to obtain the platform longitudinal acceleration command and the platform pitch angle command. Step S4: Control the longitudinal movement of the motion platform according to the longitudinal acceleration command of the platform, and control the pitch movement of the motion platform according to the pitch angle command of the platform to realize the motion simulation of vehicle movement.

2. The washing algorithm based on a data-driven parameter prediction model according to claim 1, characterized in that, The sensitivity analysis is achieved in the following manner: Determine the initial combination of influencing factors, use the particle swarm optimization algorithm for single-objective optimization, and record the optimal combination of washing filter parameters. Each influencing factor is perturbed by 10%, while keeping other influencing factors unchanged, and then a single-objective optimization is performed again. By comparing the output changes before and after the disturbance, the sensitivity of each influencing factor to the washing filter parameters is evaluated, and the sampling ratio of the influencing factors is set according to the sensitivity results.

3. The washing algorithm based on a data-driven parameter prediction model according to claim 1, characterized in that, The high-pass accelerometer filter in step S3 is a third-order filter, and its transfer function is: in, This is the gain coefficient. The natural response frequency of a second-order element. The damping coefficient is... is the cutoff frequency of the first-order element, s is the Laplace transform operator, and represents a complex variable.

4. The washing algorithm based on a data-driven parameter prediction model according to claim 1, characterized in that, The low-pass tilt coordinated filter in step S3 is a second-order filter, and its transfer function is: in, The natural response frequency, is the damping coefficient, s is the Laplace transform operator, and represents a complex variable.

5. The washing algorithm based on a data-driven parameter prediction model according to claim 1, characterized in that, Step S4 also includes active centering control, which is implemented in the following way: Detect the degree to which the longitudinal displacement of the motion platform deviates from the center position; Calculate the expected homing acceleration based on the degree of deviation. ; By controlling the platform's pitch angle Generates gravitational acceleration component This achieves the centering acceleration, enabling the platform to quickly return to its center position.

6. The washing algorithm based on a data-driven parameter prediction model according to claim 5, characterized in that, The pitch angular velocity in the active centering control satisfies the following constraint condition: in, The pitch angular velocity, The threshold for sensing the pitch angular velocity of the human body.

7. The washing algorithm based on a data-driven parameter prediction model according to claim 1, characterized in that, The somatosensory simulation effect in step S4 is evaluated using a human vestibular model, the transfer function of which is: in, For the vestibular model, natural frequency. is the damping coefficient of the vestibular model, s is the Laplace transform operator, and represents a complex variable.

8. The washing algorithm based on a data-driven parameter prediction model according to claim 1, characterized in that, It also includes a motion assessment step: calculating the root mean square error (RMSE) and Pearson correlation coefficient between the perceived force generated by the motion platform and the perceived force of the actual vehicle. ; Based on the RMSE and the Pearson correlation coefficient Calculate the comprehensive sensory error index ,in These are the weighting coefficients; The effectiveness of the motion simulation is evaluated based on the comprehensive motion error index.

Images