PSO-LSTM-based wheel-legged robot dynamics identification and joint torque prediction method

Abstract

The invention relates to the technical field of robots, in particular to a PSO-LSTM-based wheel-legged robot dynamics identification and joint torque prediction method. The method comprises the following steps: designing a multi-frequency Fourier excitation track considering a wheel-legged robot floating base, driving a robot to move according to the excitation track, collecting corresponding wheel-legged robot body multi-source sensor information, and constructing a neural network model training sample set; constructing a dynamic identification and moment prediction model based on a particle swarm and a long-short-term memory neural network; the excitation track serves as input, multi-element sensing information serves as output, the designed dynamics identification and torque prediction model is trained until the set precision is achieved, and the obtained model comprises accurate information of robot dynamics parameters and can be used for torque prediction in the robot movement process. According to the method, the accuracy of prediction of the dynamical model and the motor torque of the robot is greatly improved, and method and model support is provided for control of the floating-based robot.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robots, and particularly relates to a wheel-legged robot dynamics identification and joint torque prediction method based on PSO-LSTM. BACKGROUND

[0002] With the promotion of new industrialization, wheel-legged robots play an irreplaceable role in various fields. Wheel-legged robots have both wheel-type fast movement and leg-type obstacle-crossing ability, and their movement forms are more diverse. The dynamics model of the wheel-legged robot not only contains wheel driving friction and lateral slip, but also involves high nonlinear friction of leg servo joints and ground contact mechanics characteristics. This leads to the difficulty of dynamics modeling and identification far exceeding that of traditional mechanical arms.

[0003] In the prior art, the robot dynamics identification method based on neural network has been applied to a certain extent. For example, CN118848994A discloses a hydraulic mechanical arm adaptive neural network control method, which inputs the running data of the hydraulic mechanical arm, such as state data, position, speed and acceleration, into an adaptive neural network to simulate the dynamics model of the mechanical arm, predicts the dynamic response data of the mechanical arm in different states, and predicts the position coordinates of the mechanical arm at a certain time in the future. The patent application is used to predict the position of the mechanical arm, does not involve torque prediction, and does not consider the case of whole machine instability. For example, CN114378812A discloses a parallel mechanical arm predictive control method based on a discrete recurrent neural network model. The Stewart platform controls the end effector to track the preset path by adjusting the length of the six independent brakes, and the discrete recurrent neural network model is used to predict the path of the parallel mechanical arm nonlinear dynamics system. The patent application is for tracking control of the Stewart platform, does not involve torque prediction, and does not consider the case of whole machine instability. For example, CN116352724A discloses a mechanical arm dynamics identification method based on neural network torque prediction. The method inputs the position, speed and acceleration time series signals of each joint of the mechanical arm into an LSTM network, uses a mean square error (MSE) loss function for end-to-end torque prediction, and uses an offline excitation trajectory to optimize the network parameters to achieve high-precision identification of the joint torque of the fixed base mechanical arm. However, all excitation trajectory designs based on mechanical arms (such as joint position / speed / acceleration preset of the fixed base mechanical arm) assume that the base is rigid and does not move, and the whole machine balance does not need to be considered during movement. The base of the wheel-legged robot is a floating base. If the method of CN116352724A is used, the problem of overturning or attitude loss may occur due to the difficulty of maintaining the centroid projection within the support polygon, so that effective data cannot be obtained. Therefore, it is necessary to provide a wheel-legged robot dynamics identification and joint torque prediction method based on PSO-LSTM. SUMMARY

[0004] The application aims to provide a wheel-legged robot dynamics identification and joint torque prediction method based on PSO-LSTM, which applies a multi-frequency Fourier excitation trajectory to input the collected multi-source sensing data into a PSO-optimized LSTM neural network for dynamics model fitting, improves the accuracy of torque prediction, and provides data-driven support for subsequent control strategies.

[0005] To achieve the above purpose, the application adopts the following technical solutions:

[0006] A wheel-legged robot dynamics identification and joint torque prediction method based on PSO-LSTM includes the following steps:

[0007] A multi-frequency Fourier excitation trajectory considering the floating base and multiple constraints of the wheel-legged robot is designed, and the multiple constraints include body balance constraints, wheel contact constraints, leg height constraints, wheel-knee relative position constraints, wheel-knee relative velocity constraints, and trajectory acceleration constraints.

[0008] The wheel-legged robot is driven to move according to the excitation trajectory, and the corresponding multi-source sensing information of the wheel-legged robot body is collected to construct a neural network model training sample set.

[0009] A dynamics identification model is constructed, and the dynamics identification model is an LSTM network model optimized by a PSO algorithm.

[0010] The excitation trajectory in the training sample set is taken as the input, and the multi-element sensing information is taken as the output, and the dynamics identification model is trained until the set accuracy is reached.

[0011] Further, the method of designing a multi-frequency Fourier excitation trajectory considering the floating base and multiple constraints of the wheel-legged robot includes:

[0012] The height constraint of the wheel-legged robot is defined, and the height of the robot leg at the t time is z(t), which needs to satisfy:

[0013] z min ≤z(t)≤z max , wherein z min and z max are the lowest and highest heights allowed by the leg structure, respectively.

[0014] The balance constraint of the wheel-legged robot is defined, and the pitch angle is used to describe the body tilt posture, which needs to be guaranteed:

[0015] |θ(t)|≤θ max , wherein θ max is the maximum allowed tilt angle to prevent rollover or forward and backward overturning.

[0016] The wheel contact constraint of the wheel-legged robot is defined to ensure that the wheels do not leave the ground, which needs to satisfy:

[0017] f z,k (t) ≥ 0, k = 1, …, m, where f z,k (t) is the normal contact force of the kth contact point;

[0018] Mathematical expression of relative position of wheel joint center and knee joint center:

[0019]

[0020] Mathematical expression of relative velocity of wheel joint center and knee joint center:

[0021]

[0022] where z d (t) is the real-time relative distance of wheel joint center and knee joint center in vertical direction, x d (t) is the real-time relative distance of wheel joint center and knee joint center in horizontal direction, is the real-time relative velocity of wheel joint center and knee joint center in vertical direction, is the real-time relative velocity of wheel joint center and knee joint center in horizontal direction; A i is the amplitude, ω i is the frequency, ψ i is the phase, x0 is the horizontal offset, z0 is the vertical offset;

[0023] The trajectory acceleration is constrained as follows:

[0024]

[0025] where a z_max is the maximum acceleration in vertical direction, a x_max is the maximum acceleration in horizontal direction, is the absolute value of acceleration in vertical direction, is the absolute value of acceleration in horizontal direction, A zi is the vertical amplitude, A xi is the horizontal amplitude.

[0026] Further, the selection method of vertical amplitude A zi includes:

[0027] z min ≤ z d = z0 ± (A1 + A2 + … + A N ) ≤ z max

[0028] where z min represents the lower limit of z d , z max represents the upper limit of zd the upper limit of z d the relative distance in the vertical direction between the wheel joint center and the knee joint center, A1, A2, … A N is the amplitude of the N harmonic components;

[0029] The selection formula for the total amplitude in the vertical direction is:

[0030]

[0031] The horizontal amplitude A xi is selected in the same way as the vertical amplitude A zi .

[0032] Further, the selection method of the frequency ω i includes:

[0033] ω1=2πf1, ω2=2πf2, ω3=2πf3, where ω1, ω2, ω3 represent low, medium and high frequency components respectively, and satisfy f1<f2<f3;

[0034] where n2, n3∈N + , and n2, n3 are small integer multiples of f1, T 总 is the total cycle time.

[0035] Further, the method for collecting multi-source sensor information of the wheel-legged robot and constructing a training sample set includes:

[0036] Collecting robot motion and force data under multi-frequency Fourier excitation trajectory through multi-source multi-modal sensors on the wheel-legged robot, the multi-source multi-modal sensor information including joint angle, angular velocity, angular acceleration, torque and inertial measurement unit data;

[0037] Inputting the multi-frequency Fourier excitation trajectory into the wheel-legged robot control system, selecting multiple sets of different excitation parameter combinations, collecting multi-source sensor information under each excitation parameter combination, and constructing a training sample set;

[0038] The excitation parameter combination includes amplitude A i , frequency ω i and phase ψ i .

[0039] Further, the collected multi-source sensor information is subjected to forward and backward low-pass filtering processing with a Butterworth filter.

[0040] Further, the step of inputting the training sample set into the dynamics identification model for training includes:

[0041] S1, pre-process and vectorize the data of the training sample set, and then input the processed data into the LSTM network model;

[0042] S2, the LSTM layer of the LSTM network model is a three-layer stacked structure, the number of LSTM units of the first layer and the second layer is the same as the total number of wheel-leg robot joint numbers, wheel linear acceleration and inertial sensor data, and they are connected in series in the same time step, and each layer is connected in series in time sequence;

[0043] The initialization network hyperparameters of the LSTM network model are input into the PSO algorithm, the PSO algorithm performs global optimization by continuously iterating and evaluating the fitness of each particle in the preset parameter search space, obtains the best hyperparameter combination, and determines the number of units of the third layer LSTM according to the optimization result to reconstruct the corresponding network structure;

[0044] S3, the LSTM network model outputs the torque prediction value;

[0045] S4, calculate the error of the output torque prediction value, judge whether the training precision is reached, if the training precision is not reached, update the weight and bias, and output the torque prediction value again;

[0046] S5, the training is completed, and the obtained dynamics identification model is saved.

[0047] Further, in step S2, the method for the PSO algorithm to perform global search and optimization on the network hyperparameters in the optimization range comprises:

[0048] The particle is represented as: the position x of the i-th particle i is a set of hyperparameter configurations, the velocity v i indicates the update direction of the particle in the parameter space, and the root mean square error between the predicted torque after cross-validation and the real torque is taken as the fitness function:

[0049] Where, x i contains all the weights and biases of the LSTM, y t is the real torque;

[0050] For the i-th particle in the k-th iteration, the velocity update and position update of the update rule are:

[0051] Velocity update formula: Where, ω is the inertia weight, c1 and c2 are individual and social learning factors, r1 and r2 are random factors, and U(0,1) is a random factor, is the optimal position of particle i so far, g best is the global optimal position of all particles;

[0052] Position update formula:

[0053] Further, in the step S2, the input vectors of the first layer and the second layer of the LSTM network model are:

[0054]

[0055] where θ lh , θ lk , θ lw are the left hip joint angle, knee joint angle, wheel joint angle of the wheel-legged robot, respectively, are the first and second order derivatives of the joint angle, a p is the pitch angle acceleration, a y is the yaw angle acceleration, a r is the roll angle acceleration, and a x is the wheel line acceleration.

[0056] The input vectors of the first layer and the second layer are:

[0057]

[0058] where T lh , T lk , T lw are the moments of the left hip joint, knee joint and wheel joint of the wheel-legged robot, respectively.

[0059] Further, in the step S2, the hidden dimension of the third layer of the LSTM network model is set to H, and the hidden state vector of the LSTM-3 at the end of the sequence is:

[0060] The fully connected layer of the LSTM network model converts the high-dimensional time sequence feature into a 6-dimensional joint moment output through linear mapping:

[0061]

[0062] where z∈R 6 Table 6 original moment prediction values of 6 joints at the current time step;

[0063] The Dropout layer at the end of the LSTM network model can be represented as generating a mask m i ~ Bernoulli (1-p drop ), obtaining

[0064] where ⊙ represents element-wise multiplication, p drop is the dropout rate, and the output is the instantaneous moment prediction value of the 6 joints at the time step.

[0065] The technical scheme provided by the present application can include the following beneficial effects:

[0066] In the present application, the multi-frequency Fourier excitation trajectory of the wheel-legged robot considering the floating base and multiple constraints is designed, and the data collected includes inertial sensing and joint motion, force signal and other data. The PSO optimized LSTM neural network performs dynamics model fitting on the input data, taking into account the balance stability and dynamics excitation demand, ensuring the generalization ability of the method on the floating base robot. The method realizes end-to-end fitting of the overall dynamics model, improves the accuracy of torque prediction, and provides data-driven support for subsequent control strategies. BRIEF DESCRIPTION OF DRAWINGS

[0067] Fig. 1 is a schematic diagram of the relative position of the hip joint and the wheel joint of the wheel-legged robot;

[0068] Fig. 2 is a training flowchart of the PSO optimized LSTM neural network. DETAILED DESCRIPTION

[0069] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application.

[0070] The embodiments of the present application are described below in conjunction with Figs. 1-2 , a PSO-LSTM based wheel-legged robot dynamics identification and joint torque prediction method is described, which includes the following steps:

[0071] A multi-frequency Fourier excitation trajectory of the wheel-legged robot floating base and multiple constraints is designed, and the multiple constraints include body balance constraint, wheel contact constraint, leg height constraint, wheel-knee relative position constraint, wheel-knee relative velocity constraint and trajectory acceleration constraint;

[0072] The wheel-legged robot is driven to move according to the excitation trajectory, and the corresponding multi-source sensing information of the wheel-legged robot body is collected to construct a neural network model training sample set;

[0073] A dynamics identification model is constructed, and the dynamics identification model is an LSTM network model optimized by a PSO algorithm;

[0074] The excitation trajectory in the training sample set is taken as input, and the multi-element sensing information is taken as output. The dynamics identification model is trained until the set accuracy is reached;

[0075] The trained dynamic identification model is used for predicting joint torque of the wheel-legged robot.

[0076] The application is directed to a wheel-legged robot with a floating base, and a multi-frequency Fourier excitation trajectory is designed considering the floating base and multiple constraints. Data collection includes inertial sensing and joint motion / torque data, and the collected multi-source sensing data is input into a PSO optimized LSTM neural network for dynamic model fitting, taking into account balance stability and dynamic excitation requirements. In this scheme, the introduction of PSO significantly improves the efficiency of hyperparameter search and avoids the blindness of manual experience parameter tuning. Moreover, this scheme considers the generalization ability on the floating base robot, and at the same time, this method realizes the end-to-end fitting of the overall dynamic model, improves the accuracy of torque prediction, and provides data-driven support for subsequent control strategies.

[0077] Specifically, the method controls the displacement and velocity of the wheel joint center and the knee joint center in the motion plane under the premise of wheel-legged robot motion on the ground and without leaving the ground, thereby exciting the rich dynamic response of the system. Multi-source sensing information including joint angle, angular velocity, angular acceleration, driving torque, and inertial measurement unit data is collected, and after pre-processing by bidirectional low-pass filtering, it is input into the constructed PSO optimized LSTM network model to complete torque prediction. The network model performs global search and optimization on key hyperparameters such as time step, learning rate, and number of hidden neurons through the PSO algorithm, avoiding the uncertainty and limitations of manual experience parameter tuning. Through the training of the above network, an end-to-end mapping relationship between the joint motion state of the robot and the output torque is realized, which can be used to further fit the dynamic model of the entire wheel-legged robot system.

[0078] Further, the method of designing a multi-frequency Fourier excitation trajectory considering the floating base and multiple constraints of the wheel-legged robot includes:

[0079] Define the height constraint of the wheel-legged robot, let the height of the robot leg at the t-th moment be z(t), which needs to satisfy:

[0080] z min ≤z(t)≤z max

[0081] where z min ,z max are the minimum and maximum heights allowed by the leg structure, respectively.

[0082] Define the balance constraint of the wheel-legged robot, describe the body tilt attitude with the pitch angle, and need to ensure:

[0083] |θ(t)|≤θ max

[0084] where θ max is the maximum allowed inclination angle to prevent rollover or pitch-over.

[0085] Define the wheel contact constraint for the wheel-legged robot to ensure that the wheel does not leave the ground, which needs to satisfy:

[0086] f z,k (t)≥0,k=1,…,m

[0087] where f z,k (t) is the normal contact force of the kth contact point.

[0088] Mathematical expression of the relative position of the wheel joint center and the knee joint center:

[0089]

[0090] Mathematical expression of the relative velocity of the wheel joint center and the knee joint center:

[0091]

[0092] where z d (t) is the real-time relative distance of the wheel joint center and the knee joint center in the vertical direction, x d (t) is the real-time relative distance of the wheel joint center and the knee joint center in the horizontal direction, is the real-time relative velocity of the wheel joint center and the knee joint center in the vertical direction, is the real-time relative velocity of the wheel joint center and the knee joint center in the horizontal direction; A i is the amplitude, ω i is the frequency, ψ i is the phase, x0 is the horizontal offset, and z0 is the vertical offset.

[0093] The trajectory acceleration is constrained as follows:

[0094]

[0095] where a z_max is the maximum acceleration in the vertical direction, a x_max is the maximum acceleration in the horizontal direction, is the absolute value of the vertical acceleration, is the absolute value of the horizontal acceleration, A zi is the vertical amplitude, and A xi is the horizontal amplitude.

[0096] The excitation trajectory control of the traditional mechanical arm focuses on giving absolute angle / speed / acceleration time sequence for each joint (or joint combination) to excite sufficient dynamic information, without considering whether the relative displacement between joints will cause the whole machine to be unstable. Unlike the traditional mechanical arm which only plans for single joint position / speed / acceleration, the present application focuses on the relative target position and speed control and acceleration limitation between the knee joint (knee joint center) and the wheel joint (wheel joint center) in the floating base state, and ensures that the system can obtain sufficient dynamic response on both low frequency and high frequency components by applying a synthesized sinusoidal signal at different frequency bands, thereby improving the richness of the training data and the modeling accuracy.

[0097] Specifically, when designing the Fourier excitation, it is necessary to ensure that the relative displacement of the knee joint and the wheel joint in the horizontal direction (x-axis) and the vertical direction (z-axis) always satisfies that the overall centroid of the robot is within the support surface during the whole movement process, so as to ensure balance. By jointly planning x d (t), z d (t) and their derivatives, accurate control of the relative movement of the joints is realized.

[0098] The excitation method is different from the joint excitation design of the traditional fixed base mechanical arm, and can make the knee joint and the wheel joint cooperatively reach the predetermined relative position and speed under the premise that the whole machine maintains balance, so as to obtain more abundant whole body dynamic response data. In combination with a torque prediction method based on particle swarm optimization (PSO) and long short-term memory network (LSTM), the dynamic model identification is realized.

[0099] Preferably, the present application selects appropriate amplitudes A zi and A xi , frequencies ω i and phases ψ i and adds biases x0 and z0 to ensure that the vertical distance z and the horizontal distance x are within the feasible working space and ensure that the overall centroid projection of the robot is always within the support surface to maintain balance.

[0100] Specifically, the selection method of the vertical amplitude A zi includes:

[0101] z min ≤z d =z0±(A1+A2+…+A N )≤z max

[0102] wherein z min represents the lower limit of z d , z max represents the upper limit of z d , and z dThe relative distance of the wheel joint center and the knee joint center in the vertical direction, A1, A2, … A N is the amplitude of the Nth harmonic component;

[0103] The selection formula of the total amplitude in the vertical direction is:

[0104]

[0105] The horizontal amplitude A xi is selected in the same way as the vertical amplitude A zi .

[0106] Specifically, the selection method of the frequency ω i includes:

[0107] ω1=2πf1, ω2=2πf2, ω3=2πf3, where ω1, ω2, ω3 represent low, medium and high frequency components respectively, and satisfy f1

[0108] where n2, n3∈N + , and n2, n3 are small integer multiples of f1 (exactly complete an integer circle within the same period T 总 , T 总 is the total period time.

[0109] By reasonably selecting different frequency ω i components, at least spanning low frequency (slow motion) to medium-high frequency (exciting some common joint inertia, damping, a small amount of elastic effect and structural resonance), multiple frequency components of the system can be excited, which helps to more fully collect system dynamic information for neural network fitting.

[0110] By limiting n2, n3 to be small integer multiples of f1, they exactly complete an integer circle within the same period T 总 , so that all selected sine components exactly complete an integer vibration period within a certain total period T 总 , which is convenient for subsequent repeated experiments and data alignment. Finally, a suitable phase ψ i is selected to make the signal reach the expected initial value at t=0.

[0111] In an embodiment of the method, the method of collecting multi-source sensor information of the wheel-legged robot and constructing a training sample set includes:

[0112] Collecting robot motion and force data under multi-frequency Fourier excitation trajectories through multi-source multi-modal sensors on the wheel-legged robot, the multi-source multi-modal sensor information including joint angle, angular velocity, angular acceleration, torque and inertial measurement unit data;

[0113] The multi-frequency Fourier excitation trajectory is input into a wheel-legged robot control system, a plurality of groups of different excitation parameter combinations are selected, multi-source sensor information under each group of excitation parameter combinations is collected, and a training sample set is constructed;

[0114] The excitation parameter combination includes an amplitude A i , a frequency ω i and a phase ψ i .

[0115] Preferably, a plurality of groups of data are collected by repeating the above test under different excitation parameter combinations to construct a rich training sample set, so that the learning ability and generalization performance of the neural network to the system dynamics characteristics are improved.

[0116] Preferably, the multi-source sensor information collected is subjected to forward and backward low-pass filtering processing by a Butterworth filter.

[0117] Since the high-frequency range of the sampling signal usually does not contain valid information, it is necessary to filter the high-frequency noise. A Butterworth filter is used to perform forward and backward low-pass filtering processing on the joint angle data to maintain zero lag.

[0118] The Butterworth filter can eliminate phase delay through forward (Forward) + backward (Backward) bidirectional filtering, and the specific steps are as follows:

[0119] First filtering (forward): the original signal x(t) is subjected to conventional low-pass filtering to obtain a filtered signal xf(t), at this time the signal will lag;

[0120] Second filtering (backward): xf(t) is reversed (i.e., processing from the last data point forward), and is again subjected to the same low-pass filter to obtain fB(t);

[0121] Final output: fB(t) is reversed to restore the normal time sequence, at this time the signal has filtered out the high-frequency noise and compensated for the phase lag.

[0122] The Butterworth filter used in the scheme has no ripple within the cutoff frequency, is suitable for smoothing the joint angle data, and a high-order Butterworth filter can provide a steeper roll-off characteristic.

[0123] With reference to Fig. 2 , in an embodiment of the present application, the step of inputting the training sample set into the dynamics identification model for training includes:

[0124] S1, pre-process and vectorize the data of the training sample set, and then input the processed data into an LSTM network model;

[0125] S2, the LSTM layer of the LSTM network model is a three-layer stacked structure, the number of LSTM units of the first layer and the second layer is the same as the total number of wheel-leg robot joint numbers, wheel line acceleration and inertial sensor data, and they are connected in series in the same time step, and each layer is connected in series in time sequence;

[0126] The initialization network hyperparameters of the LSTM network model are input into the PSO algorithm, the PSO algorithm performs global optimization by continuously iterating and evaluating the fitness of each particle in the preset parameter search space, and the best hyperparameter combination is obtained, and the number of units of the third layer LSTM is determined according to the optimization result to reconstruct the corresponding network structure;

[0127] S3, the LSTM network model outputs a torque prediction value;

[0128] S4, the error of the output torque prediction value is calculated, and it is judged whether the training precision is reached, if the training precision is not reached, the weight and bias are updated, and the torque prediction value is output again;

[0129] S5, the training is completed, and the obtained dynamic identification model is saved.

[0130] The scheme adopts the PSO algorithm to perform global search and optimization on the key hyperparameters of the LSTM network. The value range of the preset hyperparameters is a determined optimization range. A group of particles is randomly initialized in the optimization range, that is, the particle group includes the initial position and initial speed of the initialized particles, and the position, speed and fitness of the particles are used to represent the particle characteristics, the position represents all hyperparameter combinations (such as the number of hidden layer neurons N, the time step T and the learning rate η) in the LSTM network model, the speed represents the evolution direction of each particle, and the fitness value is obtained by the fitness function, that is, each particle corresponds to a target function: and each particle represents a potential optimal solution of the above optimization problem. The scheme accurately captures the coupling dynamics from the floating base to the joint through the three-layer stacked LSTM network. The key hyperparameters are globally optimized by the PSO algorithm, which further improves the convergence speed and generalization ability of the model.

[0131] Specifically, the time step T in the value range of the hyperparameters is usually determined in combination with the sampling frequency and the dynamic characteristics of the system, and the input sequence should be able to cover the time window required for the joint to change from one significant state to another, for example, if the sampling frequency is 100 Hz and the dynamic characteristics of 0.2 s to 1 s are to be captured, T can be set to (20-100). The learning rate η is set to 1x10 -4 to 1x10 -2 , and the number of hidden layer neurons N is set to the input dimension to 5 times the input dimension.

[0132] Further, in the step S2, the method for the PSO algorithm to perform global search and optimization on the network hyperparameters in the optimization range comprises:

[0133] The position of the i-th particle is denoted as x i That is, a set of hyperparameter configurations, the velocity v i The update direction of the particle in the parameter space is denoted as x

[0134] where x i contains all the weights and biases of the LSTM, y t is the true moment;

[0135] For the i-th particle in the k-th iteration, the velocity update and position update of the update rule are:

[0136] The velocity update formula is: where ω is the inertia weight, c1, c2 are individual and social learning factors, and r1, r2 ~ U(0, 1) are random factors, is the optimal position of particle i so far, g best is the global optimal position of all particles;

[0137] The position update formula is:

[0138] In this way, multiple iterations are performed to obtain an approximate optimal solution, that is, an approximate optimal weight value that minimizes the control error of the LSTM network model.

[0139] Further, in the step S2, the input vectors of the first layer and the second layer of the LSTM network model are:

[0140]

[0141] where θ lh , θ lk , θ lw are the left hip joint angle, knee joint angle, and wheel joint angle of the wheeled-legged robot, respectively, are the first and second derivatives of the joint angle, a p is the pitch angle acceleration, a y is the yaw angle acceleration, a r is the roll angle acceleration, a x is the wheel line acceleration;

[0142] The input vectors of the first layer and the second layer are:

[0143]

[0144] where T lh , T lk , Tlw The torque of the hip joint, the knee joint and the wheel joint on the left side of the wheel-retreating robot, respectively.

[0145] Preferably, in the step S2, the hidden dimension of the third layer of the LSTM network model is set as H, and the hidden state vector of the LSTM-3 at the end of the sequence is:

[0146] The full connection layer of the LSTM network model converts the high-dimensional time sequence feature into a 6-dimensional joint torque output through linear mapping:

[0147]

[0148] Wherein, z∈R 6 Table 6 original torque prediction value of the joints at the current time step;

[0149] The Dropout layer at the end of the LSTM network model can be represented as generating a mask m i ~ Bernoulli (1-p drop , obtaining

[0150] Wherein, ⊙ represents element-wise multiplication, p drop is the dropout rate, and the output is the instantaneous torque prediction value of the 6 joints at the time step.

[0151] The performance of the torque prediction model based on the PSO plus LSTM mechanism is verified, and the root mean square error (RMS) of the model prediction accuracy is taken as the evaluation standard.

[0152] Root mean square error (RMS): In the formula, N is the number of data, τ real,i is the measured torque of joint i, is the torque predicted by the model of joint i, and ε RMS is the root mean square error.

[0153] Other configurations and operations of the wheel-legged robot dynamics identification method based on neural network torque prediction according to the embodiments of the present application are known to those skilled in the art, and will not be described in detail here.

[0154] In the description of the present specification, the description referring to the terms "embodiment", "example" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0155] While embodiments of the application have been shown and described, it is to be understood that the application is not limited to the details of the embodiments described, since the scope of the application will be defined with respect to the claims and their equivalents.

Claims

1. A PSO-LSTM-based wheel-legged robot dynamics identification and joint torque prediction method, characterized in that, The method comprises the following steps: Designing a multi-frequency Fourier excitation trajectory of a wheel-legged robot floating base and multiple constraints, the multiple constraints including body balance constraints, wheel contact constraints, leg height constraints, relative position constraints of wheel joints and knee joints, relative velocity constraints of wheel joints and knee joints, and trajectory acceleration constraints; Driving the wheel-legged robot to move according to the excitation trajectory, collecting corresponding multi-source sensor information of the body of the wheel-legged robot, and constructing a training sample set of a neural network model; Constructing a dynamics identification model, the dynamics identification model being an LSTM network model optimized by a PSO algorithm; Training the dynamics identification model with the excitation trajectory in the training sample set as input and the multi-source sensor information as output until a set accuracy is reached.

2. The method of claim 1, wherein, The method of designing a multi-frequency Fourier excitation trajectory of a wheel-legged robot floating base and multiple constraints comprises: Defining a wheel-legged robot height constraint, letting the height of the robot leg at the t-th moment be z(t), and needing to satisfy: z min ≤ z(t) ≤ z max , where z min , z max are the minimum, maximum height allowed by the leg structure, respectively. Defining a wheel-legged robot balance constraint, describing the body tilt posture with a pitch angle, and needing to ensure: | θ (t) | ≤ θ max where θ max is the maximum allowed tilt angle to prevent roll or pitch over. Defining a wheel-legged robot wheel contact constraint, ensuring that the wheels do not leave the ground, and needing to satisfy: f z,k (t) ≥ 0, k = 1,..., m, where fz, k (t) is the normal contact force at the kth contact point; Mathematical expressions of the relative positions of the centers of the wheel joints and the knee joints: Mathematical expressions of the relative velocities of the centers of the wheel joints and the knee joints: where z d (t) is the real-time relative distance of the vertical direction of the wheel joint center and the knee joint center, x d (t) is the real-time relative distance of the horizontal direction of the wheel joint center and the knee joint center, is the real-time relative speed of the vertical direction of the wheel joint center and the knee joint center, is the real-time relative speed of the horizontal direction of the wheel joint center and the knee joint center; A i is the amplitude, ω i is the frequency, ψ i is the phase, x0 is the horizontal offset, and z0 is the vertical offset; The trajectory acceleration is constrained as follows: wherein a z_max is the maximum acceleration in the vertical direction, a x_max is the maximum acceleration in the horizontal direction, is the absolute value of the acceleration in the vertical direction, is the absolute value of the acceleration in the horizontal direction, A zi is the vertical amplitude, A xi is the horizontal amplitude.

3. The method of claim 2, wherein, Vertical amplitude A zi The selection method includes: z min ≤z d =z0±(A1+A2+…+A N )≤z max wherein z min represents the lower limit of z d , z max represents the upper limit of z d , z d represents the relative distance in the vertical direction between the center of the hip joint and the center of the knee joint, A1, A2,... A N is the amplitude of the Nth harmonic component; The selection formula of the total amplitude in the vertical direction is: horizontal amplitude A xi The selection method of the vertical amplitude A zi is the same as that of the horizontal amplitude A 4. The method of claim 2, wherein, Frequency ω i The selection method includes: ω1=2πf1, ω2=2πf2, ω3=2πf3, where ω1, ω2, and ω3 represent low-frequency, medium-frequency, and high-frequency components, respectively, and satisfy f1<f2<f3; where n2, n3 e N + and n2, n3 are small integer multiples of f1, T 总 is the total cycle time.

5. The method according to any one of claims 1 to 4, characterized in that, The method of collecting multi-source sensor information of the body of the wheel-legged robot and constructing a training sample set comprises: Collecting robot movement and force data under the multi-frequency Fourier excitation trajectory through multi-source multi-modal sensors on the wheel-legged robot, the multi-source multi-modal sensor information including joint angles, angular velocities, angular accelerations, torques, and inertial measurement unit data; Inputting the multi-frequency Fourier excitation trajectory into the wheel-legged robot control system, selecting multiple different excitation parameter combinations, collecting multi-source sensor information under each excitation parameter combination, and constructing a training sample set; The excitation parameter combination includes amplitude A i , frequency ω i , and phase ψ i .

6. The method of claim 5, wherein, The collected multi-source sensor information is processed by a Butterworth filter for forward and backward low-pass filtering.

7. The method according to any one of claims 1 to 6, characterized in that, The steps of inputting the training sample set into the dynamics identification model for training comprise: S1, preprocessing and vectorizing the data of the training sample set, and then inputting the processed data into the LSTM network model; S2, the LSTM layer of the LSTM network model is a three-layer stacked structure, the number of LSTM units of the first layer and the second layer is the same as the total number of wheel-legged robot joint numbers, wheel linear accelerations, and inertial sensor data, and the same time step is connected in line, and each layer is connected in column in time sequence; The initialization network hyperparameters of the LSTM network model are input into the PSO algorithm, the PSO algorithm performs global optimization by continuously iterating and evaluating the fitness of each particle in the preset parameter search space, obtains the best hyperparameter combination, determines the number of units of the third layer LSTM with the optimization result, and reconstructs the corresponding network structure; S3, the LSTM network model outputs torque prediction values; S4, calculate the error of the output torque prediction value, judge whether the training precision is reached, if the training precision is not reached, update the weight and bias, and output the torque prediction value again; S5, the training is completed, and the dynamics identification model obtained by the training is saved.

8. The method of claim 7, wherein, In the step S2, the method for the PSO algorithm to perform global search and optimization on the network hyperparameters in the optimization range comprises: The position of the i-th particle is denoted as x i is a set of hyperparameter configurations, the velocity v i The update direction of the particle in the parameter space is denoted as d, and the root mean square error between the predicted moment and the real moment after cross-validation is used as the fitness function: where x i contains all the weights and biases of the LSTM, y t is the true torque; For the i-th particle in the k-th iteration, the velocity update and position update of the update rule are: Velocity update equation: where ω is the inertia weight, c1, c2 are the individual and social learning factors, and r1, r2 ~ U(0, 1) are random factors, is the best position of particle i so far, g best is the global best position of all particles; Position update formula:

9. The method of claim 7, wherein, In the step S2, the input vectors of the first layer and the second layer of the LSTM network model are: wherein θ lh , θ lk , θ lw are the left hip angle, the left knee angle, and the left wheel angle of the wheel-legged robot, respectively, are the first and second order derivatives of the joint angle, a p is the pitch angle acceleration, a y is the yaw angle acceleration, a r is the roll angle acceleration, and a x is the wheel line acceleration. The input vectors of the first layer and the second layer are output vectors: where T lh , T lk , T lw are the torques of the left hip, knee and wheel joint of the wheel-retreating robot, respectively.

10. The method of claim 7, wherein, In the step S2, the hidden dimension of the third layer of the LSTM network model is set as H, and the hidden state vector of the LSTM-3 at the end of the sequence is: The fully connected layer of the LSTM network model converts the high-dimensional time sequence features into 6-dimensional joint torque outputs through linear mapping: where z e R 6 Table 6 Joint original torque prediction values at current time step; The Dropout layer at the end of the LSTM network model can be represented as generating a mask m for each element i ~ Bernoulli(1 - p drop ), resulting in where denotes element-wise multiplication, p drop is the drop rate, and the output is the predicted instantaneous torque value of the 6 joints at this time step.

Images