Recognition method based on walking state recognition system of quadruped robot for field operation

Through the combination of ICEEMDAN with wavelet threshold noise reduction and IPO-VMD/IPO-PNN models, the signal noise problem in field operations of four-legged robots is solved, real-time recognition of the walking state of four-legged robots is realized, and the recognition accuracy and system are improved. applicability.

CN118795784BActive Publication Date: 2024-11-26JILIN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411273790.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-12
Publication Date
2024-11-26
Estimated Expiration
2044-09-12

AI Technical Summary

Technical Problem

The signal of four-legged robots contains a lot of noise in the field working environment, and it is difficult for the prior art to effectively identify their walking state, and traditional methods have problems with large signal reconstruction errors and false components.

Method used

The IMU sensor signal is processed by combining ICEEMDAN with wavelet threshold noise reduction, and signal decomposition and identification are carried out through IPO-VMD and IPO-PNN models to establish a real-time walking state recognition system.

Benefits of technology

Effectively reduce noise interference, improve signal accuracy and recognition accuracy, and realize real-time identification of the walking state of a four-legged robot, suitable for complex agricultural scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118795784B_ABST
    Figure CN118795784B_ABST
Patent Text Reader

Abstract

The identification method based on the walking state identification system of a quadruped robot for field operation belongs to the technical field of intelligent agricultural equipment. The walking state identification system of the quadruped robot of the present invention is composed of an IMU sensor, a control unit, and a computer / mobile terminal. The control unit is provided with a walking state identification program based on a quadruped robot for field operation. The walking state identification method of the quadruped robot for field operation includes an improved wavelet threshold denoising model based on ICEEMDAN, an IPO-VMD model, and an IPO-PNN model, which can realize the denoising processing and feature extraction of the attitude angle signal of the IMU sensor of the quadruped robot, and recognize the walking state of the quadruped robot through the signal feature, so as to realize the real-time identification of the walking state of the quadruped robot. The present invention can realize the online real-time identification of the operation state of intelligent agricultural machinery, and provide a new technical solution for the identification of intelligent agricultural machinery operation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent agricultural equipment, and in particular relates to a walking state recognition system and method for a field operation quadruped robot. Background Art

[0002] Quadruped robots have broad application prospects in space exploration, security inspection, patrol, rescue, etc., and therefore have attracted much attention. Compared with wheels or tracks, quadruped robots are more likely to cross artificial and natural terrains, so they have great potential for application in agricultural scenarios. The operating scenario of agricultural quadruped robots is farmland that is difficult for humans to enter and inspect, and it is impossible to observe the real-time operating status of quadruped robots in time.

[0003] The field operation environment of quadruped robots is complex, and the signals of their sensors contain a lot of noise signals, so the signals need to be preprocessed. To solve this problem, the complementary set empirical mode decomposition method based on adjustable triangular cardinality spline interpolation can be used to process the joint displacement signals of quadruped robots, effectively avoiding the mode aliasing phenomenon. However, this method has the disadvantage of large signal reconstruction error and may decompose false components, so it is not suitable to be applied to the signal denoising process. With the development of artificial intelligence, pattern recognition models based on machine learning have been widely used in various fields. Probabilistic neural networks can be used to identify typical faults of vehicle gearboxes. This method has high diagnostic efficiency, but the structural parameters of the network itself have a great influence on its performance. Without parameter optimization, the model is difficult to meet the state recognition requirements of quadruped robots in complex environments. Therefore, a real-time field operation quadruped robot walking state recognition system and method are created to effectively solve the above technical defects. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies in the prior art and provide a recognition method based on a walking state recognition system of a field operation quadruped robot, comprising the following steps:

[0005] S1 Set up the quadruped robot walking state recognition system:

[0006] The quadruped robot walking state recognition system is composed of an IMU sensor 2, a control unit 3, and a computer / mobile terminal 4, wherein the IMU sensor 2 and the control unit 3 are placed inside the quadruped robot 1, the IMU sensor 2 is electrically connected to the control unit 3, the control unit 3 is provided with a recognition program based on the walking state of the field operation quadruped robot, and the computer / mobile terminal 4 is wirelessly connected to the control unit 3;

[0007] S2 Establish a quadruped robot motion simulation model, simulate the motion state of the quadruped robot based on MATLAB / Simulink, add noise signals to simulate the environmental disturbances of the quadruped robot during field walking, simulate the normal walking of the quadruped robot, the critical stable state before falling, and the three working conditions of falling, and generate the attitude angle signal groups of the three working conditions:

[0008] ;

[0009] in: f r1 , f p1 They respectively represent the roll angle and pitch angle signals of the quadruped robot when walking normally; f r2 , f p2 They respectively represent the roll angle and pitch angle signals of the quadruped robot in the critical stable state before it falls; f r3 , f p3 They respectively represent the roll angle and pitch angle signals when the quadruped robot falls;

[0010] Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) is used to improve the wavelet threshold denoising algorithm. f r1 , f p1 , f r2 , f p2 , f r3 , f p3 Perform noise reduction, including the following steps:

[0011] S2.1 Yes f r1 , f p1 , f r2 , f p2 , f r3 , f p3 Performing ICEEMDAN includes the following steps:

[0012] S2.1.1 Add to the original signalN Sub-Gaussian white noise, construct the first group of signals with Gaussian white noise added:

[0013] ;

[0014] in: i Pick r 1. p 1. r 2. p 2. r 3. p 3; To add n The first group of signals after the group of Gaussian white noise; e 1 is the expected signal-to-noise ratio when the first group of signals is decomposed; oh n For the n Group Gaussian white noise; E 1(·) is the calculation signal oh n Operator of the first group of modal components after empirical mode decomposition; S2.1.2 Calculate the residual of the first decomposition r 1: ;

[0015] S2.1.3 Calculation of the first-order modal components IMF 1: ;

[0016] S2.1.4 Construct k ( k ≥2) Groups of signals with Gaussian white noise added: ;in: To add n After the group of Gaussian white noise k Group signal; e k For the k The expected signal-to-noise ratio when the group signal is decomposed; E k (·) is the calculation signal oh n After empirical mode decomposition k Operators for group modal components; r k-1 For the k -1 residual of decomposition;

[0017] S2.1.5 Calculate the k The residual of the decomposition r k : ;

[0018] S2.1.6 Calculate the kModal Component IMF k : ;

[0019] S2.1.7 Order k = k +1, calculate standard deviation s , determine whether the iteration termination condition is met, if so, end the iteration, otherwise repeat steps S2.1.4 to S2.1.7. The mathematical expression of the iteration termination condition is:

[0020] ;

[0021] in: IMF k+1 For the k +1 order modal component, Δ is the IMF k+1 and IMF k The standard deviation between ' is the standard deviation determined by the Cauchy convergence criterion;

[0022] S2.2 Calculate the sample entropy value of each group of modal components decomposed in step S2.1 and set the noise threshold p , if the sample entropy of the modal component is greater than or equal to p , then filter out the modal component. If the sample entropy value of the modal component is less than p , then retain the modal component, and finally get l Group modal components;

[0023] S2.3 For step S2.2 l The improved wavelet threshold denoising is performed on the group modal components, including the following steps:

[0024] S2.3.1 Set the number of wavelet decomposition layers, select the wavelet basis function, and construct the wavelet threshold function;

[0025] S2.3.2 Use the selected wavelet basis function to l The group modal components are processed by Mallat algorithm to obtain the wavelet coefficients corresponding to the wavelet decomposition layer number i ;

[0026] S2.3.3 Use improved threshold function to process wavelet coefficients;

[0027] S2.4 Reconstruct the wavelet coefficients processed in step S2.3.3 to obtain the denoised signal f r1 ' , f p1' , f r2 ' , f p2 ' , f r3 ' , f p3 ' ;

[0028] S3 pair f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' The Improve Population Optimization-Variational Mode Decomposition (IPO-VMD) includes the following steps:

[0029] S3.1 Yes f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' Perform VMD:

[0030] S3.1.1 Set the mode function after signal decomposition to m k , m k Time Series m k ( t ) is:

[0031] ;

[0032] in: A k ( t ) is the amplitude, and A k ( t )≥0; f k ( t ) is a non-monotonic decreasing phase function;

[0033] S3.1.2 Initialize the number of VMD algorithm iterations N , Lagrange multipliers l , quadratic penalty factor a , the number of decomposition levels is K ;

[0034] S3.1.3 Order n = n +1, VMD algorithm performs iterative calculation;

[0035] S3.1.4 Use k The value increases from 1 to K , m k and oh k The update method is:

[0036] ;

[0037] in: for m k Updated value; FFTD <·> is an operator that performs Fourier transform and differentiation; oh k for m k ( t )’s frequency center; for oh k Updated value; oh For t The variable obtained by Fourier transform;

[0038] S3.1.5 Yes l To update, l The update method is:

[0039] ;

[0040] in: l n+1 for l Updated value; IHT <·> is the inverse Hilbert transform;

[0041] S3.1.6 Repeat steps S3.1.3 to S3.1.5 until the termination condition is met:

[0042] ;

[0043] in: e is the convergence criterion, and e >0;

[0044] S3.2 Determining VMD parameters using the IPO algorithm a , K , including the following steps:

[0045] S3.2.1 Initialize individual positions and velocities in the IPO algorithm;

[0046] S3.2.2 Determine the fitness function of the IPO algorithm and calculate the fitness of each individual. The individual fitness function Fitness The mathematical expression is:

[0047] ;

[0048] in: PE <·> is the permutation entropy operator;

[0049] S3.2.3 Determined by fitness α , β , c Population, α , β , c The individual fitness size in the population satisfies:

[0050] ;

[0051] S3.2.4 Yes β , c Individuals in the population perform speed updates;

[0052] S3.2.5 Yes α Individuals in the population perform speed updates;

[0053] S3.2.6 Perform the first position update of the individuals in the population;

[0054] S3.2.7 Perform the second position update of the individuals in the population;

[0055] S3.2.8 Assign the individual's second position update result to the parameters of the VMD algorithm a , K, when the fitness value generated by the iterative update of the current population is less than the fitness value generated by the iterative update of the previous generation population, the individual extreme value and the group extreme value are updated, otherwise the judgment of the termination condition is entered;

[0056] S3.2.9 When the population update iteration number meets the termination condition, the update is stopped and the optimal parameters are obtained. a , K Otherwise, repeat steps S3.1 to S3.2 to continue adjusting the parameters. a , K Make updates;

[0057] S3.3 Yes f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' Repeat steps S3.1 to S3.2 to obtain the attitude angle signal VMD model with optimal parameter assignment;

[0058] S4 Establish a walking state recognition model based on Improved Population Optimization-Probabilistic Neural Networks (IPO-PNN) to perform walking state recognition:

[0059] S4.1 Use the motion simulation model of step S2 to generate Q Form an attitude angle signal group, denoise each signal in the signal group using the ICEEMDAN-improved wavelet denoising in step S2, then decompose the denoised signal using the IPO-VMD model established in step S3, and extract the permutation entropy of each signal after VMD as the signal group feature vector;

[0060] S4.2 The feature vector obtained in step S4.1 is normalized as follows:

[0061] ;

[0062] in: y Represents the normalized parameter data; y maxIndicates the maximum value of the normalized expected range; y min Indicates the minimum value of the normalized expected range; x max Indicates the maximum value in each row of parameter data; x min Indicates the minimum value in each row of parameter data; V Indicates the actual parameter value;

[0063] S4.3 Establish a 4-layer PNN topology structure, including an input layer, a hidden layer, a summation layer, and an output layer, input the data in step S4.2 into the input layer, and finally correspond to the corresponding expected output and actual output, where the expected output of the PNN is the walking state label corresponding to each group of posture angle signal groups, and initialize the structural parameters of the PNN;

[0064] S4.4 Optimize PNN using IPO algorithm;

[0065] S4.4.1 Determining the individual dimensions of the population in the IPO algorithm P d , which is calculated as:

[0066] ;

[0067] in: in is the number of neurons in the PNN input layer; h is the number of neurons in the hidden layer of PNN; sum is the number of neurons in the PNN summation layer;

[0068] S4.4.2 Determine the fitness function. The fitness function is calculated as follows:

[0069] ;

[0070] in: Y j Indicates the IPO algorithm j Individual expected output; y j Indicates the IPO algorithm j The actual output of each individual;

[0071] S4.4.3 Repeat step S3.2.1;

[0072] S4.4.4 Repeat steps S3.2.3 to S3.2.7;

[0073] S4.4.5 Every time the speed and position are updated, the weight and threshold of the PNN are updated; when the fitness value generated by the current iterative update is less than the fitness value generated by the previous iterative update, the individual extreme value update and the group extreme value update are continued, otherwise the termination condition judgment process is entered;

[0074] S4.4.6 If the number of update iterations meets the termination condition, the update is stopped and the PNN obtains the optimal weights and thresholds; otherwise, the process returns to step S4.4.4 to continue updating the weights and thresholds of the PNN;

[0075] S4.4.7 train the PNN with optimal parameter assignments;

[0076] S5 collects the attitude angle signal group of the IMU sensor (2) f r4 , f p4 ] and input it into the ICEEMDAN-improved wavelet threshold denoising algorithm in step S2. f r4 , f p4 ] to reduce noise and obtain the signal after noise reduction[ f r4 ' , f p4 ' ],Will[ f r4 ' , f p4 ' ] Input the IPO-VMD model established in step S3, obtain the signal component after VMD, extract its permutation entropy, input the extracted permutation entropy into the IPO-PNN model established in step S4, and output a label to determine the walking state of the quadruped robot (1).

[0077] Furthermore, in step S2.3.3, the threshold function is improved as follows:

[0078] ;

[0079] in: I will is the updated wavelet coefficient; m is the threshold coefficient;

[0080] m The calculation formula is:

[0081] ; Where: median(·) is the median operator; linen is the signal length.

[0082] Furthermore, in step S3.2.4, β , c The individual speeds in the population are updated as follows:

[0083] ;

[0084] in: v β(m+1) After update β The speed of individuals in the population; v γ(m+1) After update c The speed of individuals in the population; v βm For the current β The speed of individuals in the population; v γm For the current c The speed of individuals in the population; x βm For the current β The position of an individual in a population; x γm For the current c The position of an individual in a population; c 1 is the individual learning factor; c 2 is the social learning factor; r 1. r 2 is two different random numbers between [0, 1]; p βbest For now β The optimal solution for individuals in the population; p γbest For now c The optimal solution for individuals in the population; g βbest For now β The global optimal solution of the population; g γbest For now c The global optimal solution of the population; oh is the variable weight coefficient; oh max is the maximum inertia weight; oh min represents the minimum inertia weight; T c Indicates T c Iterations; T max Indicates the number of iterations.

[0085] Furthermore, in step S3.2.5, α The individual speeds in the population are updated as follows:

[0086] ;

[0087] in: v α(m+1) After update α The speed of individuals in the population; v αm For the current α The speed of individuals in the population; x αm For the current α The position of an individual in a population; c 3. c 4 is the intergroup learning factor; r 3. r 4 are two different random numbers between [0, 1]; g αbest For now α The global optimal solution of the population; F α , F β , F γ They are α , β , c The average fitness of all individuals in a population.

[0088] Furthermore, in step S3.2.6, the first position update method of the population individuals is as follows:

[0089] ;

[0090] in: j Pick α , β , c ; g jbest For population j The global optimal solution of D j for g jbest and x m The Euclidean distance between x m+1 is the position of the individual after the first position update; x m The position of the individual before the first position update; A 1. C 1 is the coefficient of two different variables; a 1 is the shrinkage factor; r 5. r 6.r 7. r 8 is 4 different random numbers between [0, 1].

[0091] Furthermore, in step S3.2.7, the second position update method of the population individuals is as follows:

[0092] ;

[0093] in: D j+1 for g jbest and x m+1 The Euclidean distance between x m+2 is the position of the individual after the first position update; A 2. C 2 is the coefficient of two different variables; F m for m The fitness value of the individual; F r is the fitness value of a random individual in the population; a 2 is different from a A shrinkage factor of 1; r 9. r 10 , r 11 , r 12 , r 13 are 5 different random numbers between [0, 1].

[0094] Furthermore, in step S3.3, the optimal parameter determination method in the attitude angle signal VMD model is:

[0095] f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' After repeating steps S3.1 to S3.2, 6 parameter groups are obtained [ a fr1 , Kfr1 ]、[ a fp1 , K fp1 ]、[ a fr2 , K fr2 ]、[ a fp2 , K fp2 ]、[ a fr3 , K fr3 ]、[ a fp3 , K fp3 ], the optimal parameters are calculated as follows:

[0096] ;

[0097] in: a rt , K rt is the optimal parameter of the VMD model of the roll angle signal; a pt , K pt are the optimal parameters of the VMD model for the pitch angle signal.

[0098] The beneficial effects of the present invention are:

[0099] The field operation environment of quadruped robots is complex, and the IMU sensor signals contain a large number of noise signals. To address this problem, the present invention proposes a method combining ICEEMDAN with wavelet threshold denoising to denoise the IMU sensor signals. A new threshold function is proposed to address the defects of hard and soft threshold functions in the wavelet threshold denoising process. The new threshold function is close to the soft threshold near the threshold and close to the hard threshold at infinity. The threshold function is optimized in a compromise manner, avoiding the defect of the hard threshold using a "one-size-fits-all" approach to process the wavelet component, resulting in the removal of signal features, and at the same time avoiding the defect of the soft threshold using a "one-size-fits-all" approach to process the wavelet component, resulting in a constant error in the signal.

[0100] The VMD method has good anti-noise interference performance. At the same time, the quadratic penalty factor and the number of signals in VMD have a great influence on the decomposition results. Aiming at the problem that the key parameters are selected based on experience, resulting in poor universality of VMD, the present invention proposes two parameters of the IPO algorithm to optimize the VMD parameters, which can obtain the best parameter combination and greatly improve the accuracy and universality of VMD.

[0101] The operation scene of the quadruped robot is a farmland where it is difficult for humans to walk, and the real-time operation status of the quadruped robot cannot be observed in time. To solve this problem, the present invention proposes a walking state recognition model based on IPO-PNN, which can realize the real-time recognition of the walking state of the quadruped robot; the proposed walking state recognition model is simple to implement, and only requires the IMU sensor to collect the attitude angle signal in real time, and perform noise reduction through the ICEEMDAN-improved wavelet threshold noise reduction method, perform IPO-VMD on the noise-reduced signal, extract the signal eigenvalue, and input it into the IPO-PNN walking state recognition model, so as to complete the real-time recognition of the walking state of the quadruped robot;

[0102] Aiming at the problem that it is difficult to obtain abnormal working condition signal samples of quadruped robots and the walking state recognition model lacks training data, the present invention establishes a quadruped robot motion simulation model based on MATLAB / Simulink, constructs a digital model of the quadruped robot under abnormal working conditions through computer simulation, thereby generating multiple groups of abnormal working condition signals, expanding the training sample volume of the walking state recognition model, and improving the reliability and recognition accuracy of the model operation;

[0103] Traditional population algorithms have the advantages of high optimization efficiency and easy implementation, but also have the disadvantages of being easy to fall into local optimality and poor convergence. In view of the defects of traditional population algorithms, the present invention proposes an IPO algorithm:

[0104] First, divide according to the fitness size α , β , c Population, in order to strengthen the information exchange between individuals within the algorithm population, and first β , c The speed of the population individuals is updated according to α , β , c The ratio of population fitness to the fitness of all individuals in the population α The speed of the population individuals is updated, which improves the information exchange ability between individuals; secondly, the individual position is updated for the first time by surrounding the prey. α , β , c The leading role of the population updates the position of each individual in the population, avoiding the individual's "greediness", and the algorithm is not easy to fall into the local optimum. At the same time, the black-winged kite's attack method is introduced in the process of surrounding the prey, and the individual position movement method is improved to expand the algorithm's optimization range; finally, the black-winged kite's migration method is quoted to update the individual position for the second time, further improving the algorithm's global optimization ability, and proposing an adaptive mutation strategy in which the mutation probability increases with the number of algorithm iterations, thereby increasing the population diversity in the later stage of the algorithm and increasing the optimization space in the later stage of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] Figure 1 This is the structural diagram of the walking state recognition system of the quadruped robot;

[0106] Figure 2 It is a flowchart of the ICEEMDAN-improved wavelet threshold and IPO-VMD signal processing model;

[0107] Figure 3 It is the flow chart of IPO-PNN model;

[0108] Figure 4 This is the waveform of the rolling angle signal of the quadruped robot;

[0109] Figure 5 This is the waveform diagram after signal noise reduction;

[0110] Figure 6 is the particle fitness curve based on the PSO algorithm;

[0111] Figure 7 is the particle fitness curve based on the IPO algorithm;

[0112] Among them: 1. Quadruped robot 2. IMU sensor 3. Control unit 4. Computer / mobile terminal. DETAILED DESCRIPTION

[0113] The present invention is described below in conjunction with the accompanying drawings:

[0114] like Figure 1 As shown, the quadruped robot walking state recognition system is composed of an IMU sensor 2, a control unit 3, and a computer / mobile terminal group 4. When the quadruped robot walking state recognition system is in operation, the IMU sensor 2 transmits the collected roll angle and pitch angle signals of the quadruped robot 1 to the control unit 3. The field operation quadruped robot walking state recognition method program in the control unit 3 determines the walking state information of the quadruped robot 1 according to the roll angle and pitch angle signals, and transmits the state information to the computer / mobile terminal 4 through the wireless connection between the control unit 3 and the computer / mobile terminal 4. The operator can monitor the walking state of the quadruped robot 1 in real time through the display interface of the computer / mobile terminal 4.

[0115] like Figure 2 , Figure 3 As shown in FIG. 1 , the steps of the method for recognizing the walking state of a quadruped robot for field operations are as follows:

[0116] S1 Set up the quadruped robot walking state recognition system:

[0117] The IMU sensor 2 and the control unit 3 are placed inside the quadruped robot 1, the IMU sensor 2 is electrically connected to the control unit 3, the control unit 3 is equipped with a recognition program based on the walking state of the field operation quadruped robot, and the computer / mobile terminal 4 is wirelessly connected to the control unit 3;

[0118] S2 Establish a quadruped robot motion simulation model, simulate the motion state of the quadruped robot based on MATLAB / Simulink, add noise signals to simulate the environmental disturbances of the quadruped robot during field walking, simulate the normal walking of the quadruped robot, the critical stable state before falling, and the three working conditions of falling, and generate the attitude angle signal groups of the three working conditions: ;

[0119] in: f r1 , f p1 They respectively represent the roll angle and pitch angle signals of the quadruped robot when walking normally; f r2 , f p2 They respectively represent the roll angle and pitch angle signals of the quadruped robot in the critical stable state before it falls; f r3 , f p3 They respectively represent the roll angle and pitch angle signals when the quadruped robot falls;

[0120] The ICEEMDAN-improved wavelet threshold denoising algorithm is used to f r1 , f p1 , f r2 , f p2 , f r3 , f p3 Perform noise reduction, including the following steps:

[0121] S2.1 Yes f r1 , f p1 , f r2 , f p2 , f r3 , f p3 Performing ICEEMDAN includes the following steps:

[0122] S2.1.1 Add to the original signal NSub-Gaussian white noise, construct the first group of signals with Gaussian white noise added:

[0123] ;

[0124] in: i Pick r 1. p 1. r 2. p 2. r 3. p 3; fn 1 is to add n The first group of signals after the group of Gaussian white noise; e 1 is the expected signal-to-noise ratio when the first group of signals is decomposed; oh n For the n Group Gaussian white noise; E 1(·) is the calculation signal oh n Operator of the first group of modal components after empirical mode decomposition;

[0125] S2.1.2 Calculate the residual of the first decomposition r 1:

[0126] ;

[0127] S2.1.3 Calculation of the first-order modal components IMF 1:

[0128] ;

[0129] S2.1.4 Construct k ( k ≥2) Groups of signals with Gaussian white noise added:

[0130] ;

[0131] in: To add n After the group of Gaussian white noise k Group signal; e k For the k The expected signal-to-noise ratio when the group signal is decomposed; E k (·) is the calculation signal oh n After empirical mode decomposition k Operators for group modal components; r k-1 For the k -1 residual of decomposition;

[0132] S2.1.5 Calculate the k The residual of the decomposition r k :

[0133] ;

[0134] S2.1.6 Calculate the k Modal Component IMF k :

[0135] ;

[0136] S2.1.7 Order k = k +1, calculate standard deviation s , determine whether the iteration termination condition is met, if so, end the iteration, otherwise repeat steps S2.1.4 to S2.1.7. The mathematical expression of the iteration termination condition is:

[0137] ;

[0138] in: IMF k+1 For the k +1 order modal component, Δ is the IMF k+1 and IMF k The standard deviation between ' is the standard deviation determined by the Cauchy convergence criterion;

[0139] S2.2 Calculate the sample entropy value of each group of modal components decomposed in step S2.1 and set the noise threshold p , if the sample entropy of the modal component is greater than or equal to p , then filter out the modal component. If the sample entropy value of the modal component is less than p , then retain the modal component, and finally get l Group modal components;

[0140] S2.3 For step S2.2 l The improved wavelet threshold denoising is performed on the group modal components, including the following steps:

[0141] S2.3.1 Set the number of wavelet decomposition layers, select the wavelet basis function, and construct the wavelet threshold function:

[0142] S2.3.2 Use the selected wavelet basis function to l The group modal components are processed by Mallat algorithm to obtain the wavelet coefficients corresponding to the wavelet decomposition layer number i ;

[0143] S2.3.3 The wavelet coefficients are processed using an improved threshold function. The improved threshold function is as follows:

[0144] ;

[0145] in: I will is the updated wavelet coefficient; m is the threshold coefficient;

[0146] m The calculation formula is:

[0147] ;

[0148] Among them: median(·) is the median operator; linen is the signal length;

[0149] S2.4 Reconstruct the wavelet coefficients processed in step S2.3.3 to obtain the denoised signal f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' ;

[0150] S3 pair f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' Perform IPO-VMD, including the following steps:

[0151] S3.1 Yes f r1 ' , f p1' , f r2 ' , f p2 ' , f r3 ' , f p3 ' Perform VMD:

[0152] S3.1.1 Set the mode function after signal decomposition to m k , m k Time Series m k ( t ) is:

[0153] ;

[0154] in: A k ( t ) is the amplitude, and A k ( t )≥0; f k ( t ) is a non-monotonic decreasing phase function;

[0155] S3.1.2 Initialize the number of VMD algorithm iterations N , Lagrange multipliers l , quadratic penalty factor a , the number of decomposition levels is K ;

[0156] S3.1.3 Order n = n +1, VMD algorithm performs iterative calculation;

[0157] S3.1.4 Use k The value increases from 1 to K , m k and oh k The update method is:

[0158] ;

[0159] in: for m k Updated value; FFTD<·> is an operator that performs Fourier transform and differentiation; oh k for m k ( t )’s frequency center; for oh k Updated value; oh For t The variable obtained by Fourier transform;

[0160] S3.1.5 Yes l To update, l The update method is:

[0161] ;

[0162] in: l n+1 for l Updated value; IHT <·> is the inverse Hilbert transform;

[0163] S3.1.6 Repeat steps S3.1.3 to S3.1.5 until the termination condition is met:

[0164] ;

[0165] in: e is the convergence criterion, and e >0;

[0166] S3.2 Determining VMD parameters using the IPO algorithm a , K , including the following steps:

[0167] S3.2.1 Initialize individual positions and velocities in the IPO algorithm;

[0168] S3.2.2 Determine the fitness function of the IPO algorithm and calculate the fitness of each individual. The individual fitness function Fitness The mathematical expression is:

[0169] ;

[0170] in: PE <·> is the permutation entropy operator;

[0171] S3.2.3 Determined by fitness α , β , c Population, α , β , cThe individual fitness size in the population satisfies:

[0172] ;

[0173] S3.2.4 Yes β , c The individuals in the population update their speed in the following way:

[0174] ;

[0175] in: v β(m+1) After update β The speed of individuals in the population; v γ(m+1) After update c The speed of individuals in the population; v βm For the current β The speed of individuals in the population; v γm For the current c The speed of individuals in the population; x βm For the current β The position of an individual in a population; x γm For the current c The position of an individual in a population; c 1 is the individual learning factor; c 2 is the social learning factor; r 1. r 2 is two different random numbers between [0, 1]; p βbest For now β The optimal solution for individuals in the population; p γbest For now c The optimal solution for individuals in the population; g βbest For now β The global optimal solution of the population; g γbest For now c The global optimal solution of the population; oh is the variable weight coefficient; oh max is the maximum inertia weight; oh min represents the minimum inertia weight; T c Indicates T c Iterations; T max Indicates the number of iterations;

[0176] S3.2.5 Yes α The individuals in the population update their speed in the following way:

[0177] ;

[0178] in: v α(m+1) After update α The speed of individuals in the population; v αm For the current α The speed of individuals in the population; x αm For the current α The position of an individual in a population; c 3. c 4 is the intergroup learning factor; r 3. r 4 are two different random numbers between [0, 1]; g αbest For now α The global optimal solution of the population; F α , F β , F γ They are α , β , c The average fitness of all individuals in the population;

[0179] S3.2.6 Perform the first position update of the population individuals in the following manner:

[0180] ;

[0181] in: j Pick α , β , c ; g jbest For population j The global optimal solution of D j for g jbest and x m The Euclidean distance between x m+1 is the position of the individual after the first position update; x m The position of the individual before the first position update; A 1. C1 is the coefficient of two different variables; a 1 is the shrinkage factor; r 5. r 6. r 7. r 8 is 4 different random numbers between [0, 1];

[0182] S3.2.7 Perform the second position update of the individuals in the population. The update method is as follows:

[0183] ;

[0184] in: D j+1 for g jbest and x m+1 The Euclidean distance between x m+2 is the position of the individual after the first position update; A 2. C 2 is the coefficient of two different variables; F m for m The fitness value of the individual; F r is the fitness value of a random individual in the population; a 2 is different from a A shrinkage factor of 1; r 9. r 10 , r 11 , r 12 , r 13 are 5 different random numbers between [0, 1].

[0185] S3.2.8 Assign the individual's second position update result to the parameters of the VMD algorithm a , K , when the fitness value generated by the iterative update of the current population is less than the fitness value generated by the iterative update of the previous generation population, the individual extreme value and the group extreme value are updated, otherwise the judgment of the termination condition is entered;

[0186] S3.2.9 When the population update iteration number meets the termination condition, the update is stopped and the optimal parameters are obtained. a , K Otherwise, repeat steps S3.1 to S3.2 to continue adjusting the parameters. a , K Make updates;

[0187] S3.3 Yes fr1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' Repeat steps S3.1 to S3.2 to obtain the attitude angle signal VMD model with optimal parameter assignment. The optimal parameter determination method is:

[0188] f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' After repeating steps S3.1 to S3.2, 6 parameter groups are obtained [ a fr1 , K fr1 ]、[ a fp1 , K fp1 ]、[ a fr2 , K fr2 ]、[ a fp2 , K fp2 ]、[ a fr3 , K fr3 ]、[ a fp3 , K fp3 ], the optimal parameters are calculated as follows:

[0189] ;

[0190] in: a rt ,K rt is the optimal parameter of the VMD model of the roll angle signal; a pt , K pt is the optimal parameter of the pitch angle signal VMD model; S4 establishes a walking state recognition model based on IPO-PNN to perform walking state recognition:

[0191] S4.1 Use the motion simulation model of step S2 to generate Q Form an attitude angle signal group, denoise each signal in the signal group using the ICEEMDAN-improved wavelet denoising in step S2, then decompose the denoised signal using the IPO-VMD model established in step S3, and extract the permutation entropy of each signal after VMD as the signal group feature vector;

[0192] S4.2 The feature vector obtained in step S4.1 is normalized as follows:

[0193] ;

[0194] in: y Represents the normalized parameter data; y max Indicates the maximum value of the normalized expected range; y min Indicates the minimum value of the normalized expected range; x max Indicates the maximum value in each row of parameter data; x min Indicates the minimum value in each row of parameter data; V Indicates the actual parameter value;

[0195] S4.3 Establish a 4-layer PNN topology structure, including an input layer, a hidden layer, a summation layer, and an output layer, input the data in step S4.2 into the input layer, and finally correspond to the corresponding expected output and actual output, where the expected output of the PNN is the walking state label corresponding to each group of posture angle signal groups, and initialize the structural parameters of the PNN;

[0196] Among them, the walking state label settings are shown in Table 1:

[0197] Table 1

[0198]

[0199] S4.4 Optimize PNN using IPO algorithm;

[0200] S4.4.1 Determining the individual dimensions of the population in the IPO algorithm Pd , which is calculated as:

[0201] ;

[0202] in: in is the number of neurons in the PNN input layer; h is the number of neurons in the hidden layer of PNN; sum is the number of neurons in the PNN summation layer;

[0203] S4.4.2 Determine the fitness function. The fitness function is calculated as follows:

[0204] ;

[0205] in: Y j Indicates the IPO algorithm j Individual expected output; y j Indicates the IPO algorithm j The actual output of each individual;

[0206] S4.4.3 Repeat step S3.2.1;

[0207] S4.4.4 Repeat steps S3.2.3 to S3.2.7;

[0208] S4.4.5 Every time the speed and position are updated, the weight and threshold of the PNN are updated; when the fitness value generated by the current iterative update is less than the fitness value generated by the previous iterative update, the individual extreme value update and the group extreme value update are continued, otherwise the termination condition judgment process is entered;

[0209] S4.4.6 If the number of update iterations meets the termination condition, the update is stopped and the PNN obtains the optimal weights and thresholds; otherwise, the process returns to step S4.4.4 to continue updating the weights and thresholds of the PNN;

[0210] S4.4.7 train the PNN with optimal parameter assignments;

[0211] S5 collects the attitude angle signal group of IMU sensor 2 [ f r4 , f p4 ] and input it into the ICEEMDAN-improved wavelet threshold denoising algorithm in step S2. f r4 , f p4 ] to reduce noise and obtain the signal after noise reduction[ f r4 ', f p4 ' ],Will[ f r4 ' , f p4 ' ] Input the IPO-VMD model established in step S3, obtain the signal component after VMD, extract its permutation entropy, input the extracted permutation entropy into the IPO-PNN model established in step S4, and output a label to determine the walking state of the quadruped robot 1.

[0212] The ICEEMDAN-improved wavelet threshold denoising method of the present invention has excellent performance. The specific experimental data and analysis are as follows: Figure 4 , Figure 5 As shown, after the signal is subjected to ICEEMDAN-improved wavelet threshold denoising, the number of "burrs" in the signal waveform is significantly reduced, which proves the feasibility of the ICEEMDAN-improved wavelet threshold denoising method of the present invention.

[0213] In the traditional method, the PSO algorithm is generally used to optimize the PNN to establish the walking state recognition model. Compared with this method, the performance of the IPO-PNN in terms of particle fitness is more superior. The specific experimental data and analysis are as follows: Figure 6 to 7 As shown, on the basis that the number of algorithm termination iterations is 300, when the PNN is optimized based on the PSO algorithm, the final fitness of the particles of the particle swarm is 0.1644, while when the present invention optimizes the PNN based on the IPO algorithm, the final fitness of the particles of the particle swarm is 0.1559. It can be seen that the walking state recognition model established by the present invention has better performance.

Claims

1. A recognition method based on a walking state recognition system of a quadruped robot for field operations, characterized in that: The following steps are involved: S1 Set up the quadruped robot walking state recognition system: The quadruped robot walking state recognition system is composed of an IMU sensor (2), a control unit (3), and a computer / mobile terminal (4), wherein the IMU sensor (2) and the control unit (3) are placed inside the quadruped robot (1), the IMU sensor (2) and the control unit (3) are electrically connected, the control unit (3) is provided with a recognition program based on the walking state of the field operation quadruped robot, and the computer / mobile terminal (4) is wirelessly connected to the control unit (3); S2 Establish a quadruped robot motion simulation model, simulate the motion state of the quadruped robot based on MATLAB / Simulink, add noise signals to simulate the environmental disturbances of the quadruped robot during field walking, simulate the normal walking of the quadruped robot, the critical stable state before falling, and the three working conditions of falling, and generate the attitude angle signal groups of the three working conditions: ; in: f r1 , f p1 They respectively represent the roll angle and pitch angle signals of the quadruped robot when walking normally; f r2 , f p2 They respectively represent the roll angle and pitch angle signals of the quadruped robot in the critical stable state before it falls; f r3 , f p3 They respectively represent the roll angle and pitch angle signals when the quadruped robot falls; The ICEEMDAN-improved wavelet threshold denoising algorithm is used to f r1 , f p1 , f r2 , f p2 , f r3 , f p3 Perform noise reduction, including the following steps: S2.1 Yes f r1 , f p1 , f r2 , f p2 , f r3 , f p3 Performing ICEEMDAN includes the following steps: S2.1.1 Add to the original signal N Sub-Gaussian white noise, construct the first group of signals with Gaussian white noise added: ; in: i Pick r 1. p 1. r 2. p 2. r 3. p 3; To add n The first group of signals after the group of Gaussian white noise; e 1 is the expected signal-to-noise ratio when the first group of signals is decomposed; ω n For the n Group Gaussian white noise; E 1(·) is the calculation signal ω n Operator of the first group of modal components after empirical mode decomposition; S2.1.2 Calculate the residual of the first decomposition r 1: ; S2.1.3 Calculation of the first-order modal component IMF 1: ; S2.1.4 Construct k ( k ≥2) Groups of signals with Gaussian white noise added: ; in: To add n After the group of Gaussian white noise k Group signal; e k For the k The expected signal-to-noise ratio when the group signal is decomposed; E k (·) is the calculation signal ω n After empirical mode decomposition k Operators for group modal components; r k-1 For the k -1 residual of decomposition; S2.1.5 Calculate the k The residual of the decomposition r k : ; S2.1.6 Calculate the k Modal Component IMF k : ; S2.1.7 Order k = k +1, calculate standard deviation σ , determine whether the iteration termination condition is met, if so, end the iteration, otherwise repeat steps S2.1.4 to S2.1.

7. The mathematical expression of the iteration termination condition is: ; in: IMF k+1 For the k +1 order modal component, Δ is the IMF k+1 and IMF k The standard deviation between ' is the standard deviation determined by the Cauchy convergence criterion; S2.2 Calculate the sample entropy value of each group of modal components decomposed in step S2.1 and set the noise threshold p , if the sample entropy of the modal component is greater than or equal to p , then filter out the modal component. If the sample entropy value of the modal component is less than p , then retain the modal component, and finally get l Group modal components; S2.3 For step S2.2 l The improved wavelet threshold denoising is performed on the group modal components, including the following steps: S2.3.1 Set the number of wavelet decomposition layers, select the wavelet basis function, and construct the wavelet threshold function; S2.3.2 Use the selected wavelet basis function to l The group modal components are processed by Mallat algorithm to obtain the wavelet coefficients corresponding to the wavelet decomposition layer number θ ; S2.3.3 Use improved threshold function to process wavelet coefficients; S2.4 Reconstruct the wavelet coefficients processed in step S2.3.3 to obtain the denoised signal f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' ; S3 pair f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' Perform IPO-VMD, including the following steps: S3.1 Yes f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' Perform VMD: S3.1.1 Set the mode function after signal decomposition to m k , m k Time Series m k ( t ) is: ; in: A k ( t ) is the amplitude, and A k ( t )≥0; φ k ( t ) is a non-monotonic decreasing phase function; S3.1.2 Initialize the number of VMD algorithm iterations N , Lagrange multipliers λ , quadratic penalty factor a , the number of decomposition levels is K ; S3.1.3 Order n = n +1, VMD algorithm performs iterative calculation; S3.1.4 Use k The value increases from 1 to K , m k and ω k The update method is: ; in: for m k Updated value; FFTD <·> is an operator that performs Fourier transform and differentiation; ω k for m k ( t )’s frequency center; for ω k Updated value; ω For t The variable obtained by Fourier transform; S3.1.5 Yes λ To update, λ The update method is: ; in: λ n+1 for λ Updated value; IHT <·> is the inverse Hilbert transform; S3.1.6 Repeat steps S3.1.3 to S3.1.5 until the termination condition is met: ; in: ε is the convergence criterion, and ε >0; S3.2 Determining VMD parameters using the IPO algorithm a , K , including the following steps: S3.2.1 Initialize individual positions and velocities in the IPO algorithm; S3.2.2 Determine the fitness function of the IPO algorithm and calculate the fitness of each individual. The individual fitness function Fitness The mathematical expression is: ; in: PE <·> is the permutation entropy operator; S3.2.3 Determined by fitness α , β , γ Population, α , β , γ The individual fitness size in the population satisfies: ; S3.2.4 Yes β , γ Individuals in the population perform speed updates; S3.2.5 Yes α Individuals in the population perform speed updates; S3.2.6 Perform the first position update of the individuals in the population; S3.2.7 Perform the second position update of the individuals in the population; S3.2.8 Assign the individual's second position update result to the parameters of the VMD algorithm a , K , when the fitness value generated by the iterative update of the current population is less than the fitness value generated by the iterative update of the previous generation population, the individual extreme value and the group extreme value are updated, otherwise the judgment of the termination condition is entered; S3.2.9 When the population update iteration number meets the termination condition, the update stops and the optimal parameters are obtained. a , K Otherwise, repeat steps S3.1 to S3.2 to continue adjusting the parameters. a , K Make updates; S3.3 Yes f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' Repeat steps S3.1 to S3.2 to obtain the attitude angle signal VMD model with optimal parameter assignment; S4 Establish a walking state recognition model based on IPO-PNN to perform walking state recognition: S4.1 Use the motion simulation model of step S2 to generate Q Form an attitude angle signal group, denoise each signal in the signal group using the ICEEMDAN-improved wavelet denoising in step S2, then decompose the denoised signal using the IPO-VMD model established in step S3, and extract the permutation entropy of each signal after VMD as the signal group feature vector; S4.2 The feature vector obtained in step S4.1 is normalized as follows: ; in: y Represents the normalized parameter data; y max Indicates the maximum value of the normalized expected range; y min Indicates the minimum value of the normalized expected range; x max Indicates the maximum value in each row of parameter data; x min Indicates the minimum value in each row of parameter data; V Indicates the actual parameter value; S4.3 Establish a 4-layer PNN topology structure, including an input layer, a hidden layer, a summation layer, and an output layer, input the data in step S4.2 into the input layer, and finally correspond to the corresponding expected output and actual output, where the expected output of the PNN is the walking state label corresponding to each group of posture angle signal groups, and initialize the structural parameters of the PNN; S4.4 Optimize PNN using IPO algorithm; S4.4.1 Determining the individual dimensions of the population in the IPO algorithm P d , which is calculated as: ; in: in is the number of neurons in the PNN input layer; h is the number of neurons in the hidden layer of PNN; sum is the number of neurons in the PNN summation layer; S4.4.2 Determine the fitness function. The fitness function is calculated as follows: ; in: Y j Indicates the IPO algorithm j Individual expected output; y j Indicates the IPO algorithm j The actual output of each individual; S4.4.3 Repeat step S3.2.1; S4.4.4 Repeat steps S3.2.3 to S3.2.7; S4.4.5 Every time the speed and position are updated, the weight and threshold of the PNN are updated; when the fitness value generated by the current iterative update is less than the fitness value generated by the previous iterative update, the individual extreme value update and the group extreme value update are continued, otherwise the termination condition judgment process is entered; S4.4.6 If the number of update iterations meets the termination condition, the update is stopped and the PNN obtains the optimal weights and thresholds; otherwise, the process returns to step S4.4.4 to continue updating the weights and thresholds of the PNN; S4.4.7 train the PNN with optimal parameter assignments; S5 collects the attitude angle signal group of the IMU sensor (2) f r4 , f p4 ] and input it into the ICEEMDAN-improved wavelet threshold denoising algorithm in step S2. f r4 , f p4 ] to perform noise reduction and obtain the signal group after noise reduction [ f r4 ' , f p4 ' ],Will[ f r4 ' , f p4 ' ] Input the IPO-VMD model established in step S3, obtain the signal component after VMD, extract its permutation entropy, input the extracted permutation entropy into the IPO-PNN model established in step S4, and output a label to determine the walking state of the quadruped robot (1).

2. The recognition method based on the walking state recognition system of the field operation quadruped robot according to claim 1 is characterized in that: In step S2.3.3, the threshold function is improved as follows: ; in: θ' is the updated wavelet coefficient; m is the threshold coefficient; m The calculation formula is: ; Among them: median(·) is the median operator; len is the signal length.

3. The recognition method based on the walking state recognition system of a field operation quadruped robot according to claim 1 is characterized in that: In step S3.2.4, β , γ The individual speeds in the population are updated as follows: ; in: v β(m+1) After update β The speed of individuals in the population; v γ(m+1) After update γ The speed of individuals in the population; v βm For the current β The speed of individuals in the population; v γm For the current γ The speed of individuals in the population; x βm For the current β The position of an individual in a population; x γm For the current γ The position of an individual in a population; c 1 is the individual learning factor; c 2 is the social learning factor; r 1. r 2 is two different random numbers between [0, 1]; p βbest For now β The optimal solution for individuals in the population; p γbest For now γ The optimal solution for individuals in the population; g βbest For now β The global optimal solution of the population; g γbest For now γ The global optimal solution of the population; ω' is the variable weight coefficient; ω max is the maximum inertia weight; ω min represents the minimum inertia weight; T c Indicates T c Iterations; T max Indicates the number of iterations.

4. The recognition method based on the walking state recognition system of the field operation quadruped robot according to claim 1 is characterized in that: In step S3.2.5, α The individual speeds in the population are updated as follows: ; in: v α(m+1) After update α The speed of individuals in the population; v αm For the current α The speed of individuals in the population; x αm For the current α The position of an individual in a population; c 3. c 4 is the intergroup learning factor; r 3. r 4 are two different random numbers between [0, 1]; g αbest For now α The global optimal solution of the population; F α , F β , F γ They are α , β , γ The average fitness of all individuals in a population.

5. The identification method based on the walking state identification system of the field operation quadruped robot according to claim 1 is characterized in that: In step S3.2.6, the first position update method of the population individuals is as follows: ; in: j Pick α , β , γ ; g jbest For population j The global optimal solution of D j for g jbest and x m The Euclidean distance between x m+1 is the position of the individual after the first position update; x m The position of the individual before the first position update; A 1. C 1 is the coefficient of two different variables; a 1 is the shrinkage factor; r 5. r 6. r 7. r 8 is 4 different random numbers between [0, 1].

6. The identification method based on the walking state identification system of the field operation quadruped robot according to claim 1 is characterized in that: In step S3.2.7, the second position update method of the population individuals is as follows: ; in: D j+1 for g jbest and x m+1 The Euclidean distance between x m+2 is the position of the individual after the first position update; A 2. C 2 is the coefficient of two different variables; F m for m The fitness value of the individual; F r is the fitness value of a random individual in the population; a 2 is different from a A shrinkage factor of 1; r 9. r 10 , r 11 , r 12 , r 13 are 5 different random numbers between [0, 1].

7. The identification method based on the walking state identification system of the field operation quadruped robot according to claim 1 is characterized in that: In step S3.3, the optimal parameter determination method in the attitude angle signal VMD model is: f r1 ' , f p1 ' , f r2 ' , f p2 ' , f r3 ' , f p3 ' After repeating steps S3.1 to S3.2, 6 parameter groups are obtained [ a fr1 , K fr1 ]、[ a fp1 , K fp1 ]、[ a fr2 , K fr2 ]、[ a fp2 , K fp2 ]、[ a fr3 , K fr3 ]、[ a fp3 , K fp3 ], the optimal parameters are calculated as follows: ; in: a rt , K rt is the optimal parameter of the VMD model of the roll angle signal; a pt , K pt is the optimal parameter of the VMD model for the pitch angle signal.