This action might not be possible to undo. Are you sure you want to continue?
A COMPARATIVE STUDY OF MODELING AND IDENTIFICATION OF THE PNEUMATIC ARTIFICIAL MUSCLE (PAM) MANIPULATOR BASED ON RECURRENT NEURAL NETWORKS
Kyoung Kwan AHN1 Ho Pham Huy ANH2 Nguyen The KIET3 School of Mechanical and Automotive Engineering, University of Ulsan, Ulsan, Korea
Electrical and Electronic Faculty, HCM City University of Technology, HCMCity, Viet Nam
The paper deals with the PAM manipulator modeling and identification based on recurrent neural networks. The first time, the most powerful types of neural network-based nonlinear autoregressive models, namely, NNARMAX, NNOE and NNARX models will be applied comparatively of the PAM manipulator identification. Furthermore, the evaluation of different nonlinear Neural Network AutoRegressive models of PAM manipulator with various hidden layer nodes is completely discussed. On this basis the merit and demerit features of each identified model of the highly nonlinear PAM manipulator have been analyzed and compared. The results show that the nonlinear NNARX model yields more performance and higher accuracy than the other nonlinear NNARMAX and NNOE model schemes. These results can be applied to model and identify not only of the PAM manipulator but also of other nonlinear and time-varied parametric industrial systems. Keywords: pneumatic artificial muscle (PAM), PAM manipulator, recurrent neural networks, nonlinear NNARMAX model, nonlinear NNARX model, nonlinear NNOE model, modeling and identification.
1. INTRODUCTION. PAM actuators have been used in various precision robotic tasks as well as in human exoskeletons for enhancement of strength and mobility . Due to their highly nonlinear and time-varying parameter nature, PAM manipulator modeling always presents a challenging problem that has been approached via many methodologies. Recently, in , Ahn et al. applied Modified GA for optimizing parameters of pseudo-linear ARX model of the PAM manipulator. In , authors identified successfully PAM manipulator based on nonlinear neural NNARX model. In this paper, auto-regressive recurrent neural networks are applied to model and identify the PAM manipulator system. The 1link PAM manipulator test system was designed. The contribution focuses on carrying experimental modeling and analyzing from gathered results the advantages of PAM manipulator identification using neural networks with different auto-regressive structures and different hidden layer nodes as well. Modeling results of the complex dynamic systems such as
PAM manipulator show that the newly proposed nonlinear model presented in this study can be applied in online control with better dynamic property, strong robustness and suitable for the identification of various plants, including linear and nonlinear process without regard greatly changing external environments. This paper is organized as follows: Section 2 gives a description of basis characteristics of pneumatic artificial muscle. In section 3, the procedure of modeling and identification of the PAM manipulator based on auto-regressive recurrent neural network is presented. Section 4 shows the hardware set-up for 1-link PAM manipulator identification. Section 5 presents the results of modeling and identification of the 1-link PAM manipulator based on recurrent neural network. Section 6 is the conclusion. BASIS CHARACTERISTICS OF PNEUMATIC ARTIFICIAL MUSCLE. The basic mechanics characteristic of PAM can be depicted with the equation below: (2.1) F = P[a (1 − h) 2 − b] 2.
International Symposium on Electrical & Electronics Engineering 2007 - Oct 24, 25 2007 - HCM City, Vietnam -98-
.. h is contraction rate. which used regressor: X (k ) = [ y (k − 1).εu(k −1). q denotes the shift operator. ˆ y = QT X (k ) (3. u (k − 1). u (k − 1)... the NARX model is called a series parallel model.. the ARMAX model (F=D=1).... i=1. C (q ) B(q ) (3.…na. prediction errors (associated with the C polynomial) ˆ • eu (k − i ) = y (k − i ) − yu k − i . Nonlinear Output Error Models.. This makes the identification of these models difficult. NONLINEAR AUTO-REGRESSIVE NEURAL NETWORKS-BASED PAM MANIPULATOR IDENTIFICATION Based on these regressors. F is nonlinear with h. This highly nonlinear feature is shown in Fig.. P and h are three critical variables to determine the characteristics of artificial muscle PAM. The NARMAX and NOE models are recurrent models. Nonlinear Auto-Regressive Moving Average with eXogenous input models.. simulation errors (associated with the D polynomial) ( ) Fig. Description of Nonlinear Autoregressive Model The nomenclature of nonlinear dynamical model is based on the terminology used to categorize linear input-output models... which used regressor as ˆ ˆ X (k ) = [ y(k − 1). Following this nomenclature of linear models. When P is constant. the curves tell us that artificial muscle is substantially a time-varying and nonlinear system.1 illustrates the ε-F relationships of artificial muscle... 3. different types of model structures can be constructed. A(q) is a polynomial in q-1. while the NOE is referred to as a parallel model.. due to the friction and wire resistance between rubber tubes and mesh shell. Fig. 25 2007 . y(k − na ).…nb. simulated outputs from past u(k) (associated with the F polynomial) ˆ • e(k − i ) = y (k − i ) − y (k − i ).... Vietnam -99- . This model can be given in a “pseudo-linear” regression form International Symposium on Electrical & Electronics Engineering 2007 . time-invariant system can always be represented by a NARX model . discrete. because they use the estimated output that constitutes a feedback..1) as the Box-Jenkins (BJ) model (A=1).1 extracted from  representing the relationship the contraction rate h and the exerting force F of a pneumatic artificial muscle (PAM)... a and b are constants related to the structure of artificial muscle PAM. (k −nb). i=1.u(k −ne)] y u ε On the soft computing and system identification of nonlinear system. Some special cases of (3. its parameters are easy to be estimated. after making some moderate assumptions. In this figure. i. u (k − nb )] NNOE. it has been proved that any nonlinear. Nonlinear Auto-Regressive with eXogenous input models.HCM City. 1: h -F relationships of artificial muscle (extracted from ) Equation (1) indicates that F. Meanwhile.1) e(k ) u (k ) + A(q ) y (k ) = F (q ) D(q ) where.. measured process outputs (associated with the A polynomial) ˆ • y (k − i ).. P is the internal pressure (bar). control signals (associated with the B polynomial) • y(k-i). this paper tries to apply typical recurrent neural networks models for identifying the PAM system. To present a dynamic system..e the components of X(k) can be given by: • u(k-i). For instance.Track 1: AUTOMATIC CONTROL where F is the contraction force...2) where the regressors.Oct 24. Because the NARX model structure is non-recursive. (k −na).. u (k − nb )] NNARMAX. in the following the linear empirical model structures that can be summarized by the general family . A. where X(k) =[y(k −1). y(k − na ).u(k −1). the output-error (OE) model (A=C=D=1) and the ARX model (F=C=D=1). it is natural to construct similar nonlinear neural model as: NNARX. In order to overcome the disadvantage of nonlinear and lag characteristics of artificial muscle. artificial muscle has an obvious characteristic of lag.
F): q q m ˆ yi (w. Vietnam -100- . H40-8-3600ZO) and fed back to the computer through an 32-bit counter board (COMPUTING MEASUREMENT. and 6c represent experimental diagram for modeling and identifying neural network NNARX model.) Table 1 presents the configuration of the hardware set-up installed from Fig.5 Experiment system for the PAM manipulator Figure 4: Diagram for obtaining PRBS training data of the PAM manipulator It needs to obtain the experimental data which describe the underlying intrinsic features of the 1-axes PAM manipulator. through a D/A board (ADVANTECH.Oct 24. Consequently.6a. is detected by a rotary encoder (METRONIX. The lists of experimental hardware No. A block diagram of the experimental apparatus is shown in Fig. u(t).4.5.Track 1: AUTOMATIC CONTROL B Description of MLPNN The Multilayer-Perceptron (or MLP) network is probably the most-often considered member of the neural network family.5 as to model and identify the auto-regressive neural networks model of the PAM manipulator.7 GHz) which sends the PRBS voltage signal u(t) to control the proportional valve (FESTO.) F1(. 2. HARDWARE SET-UP OF MODELING AND IDENTIFICATION OF THE PAM MANIPULATOR An experiment system is illustrated in Fig. The main reason for this is its ability to model simple as well as very complex functional relationships. Joint angle. Table 1. are a set of inputs. International Symposium on Electrical & Electronics Engineering 2007 . NNARMAX model and NNOE model of the PAM manipulator. Fig.W) = F∑ ijhj (w) +W0 = Fi ∑ ij. f j ∑wjl zl + wj0 +W0 i i W i W j =1 j =1 l =1 (3. PCI QUAD-4 card) which changes digital pulse signals to joint angle value y(t). or the training data as they are usually called.) y1 w10 1 Fig. 6b.3) The weights (specified by the vector θ. Fig. and corresponding desired outputs. θ [deg]. 4. The hardware includes a PC (Pentium 1. 2 hidden units (also called “nodes” or “neurons”). PCI 1720 card). and 2 outputs is shown in Fig. the joint of the PAM manipulator will be rotated. The pneumatic line is conducted under the pressure of 4[bar] and the software is coded in C program language. This has been proven through a large number of practical applications. 2: Structure of feed-forward MLP-network The class of MLP-networks considered in this paper is furthermore confined to those having only one hidden layer and only hyperbolic tangent and linear activation functions (f. 1 2 3 4 5 Name Proportional valve Pneumatic artificial muscle D/A board Counter board Rotary encoder Model name MPYE-5-1/8HF-710 B MAS-10-N-220-AA-MCFK PCI 1720 PCI QUAD-4 H40-8-3600ZO Company FESTO FESTO ADVANTECH COMPUTING MEASUREMENT METRONIX w22 w12 w21 w20 w11 w10 F2(. and they are determined from a set of examples through the process called training.) y2 f1(. y(t).HCM City. A fully connected two layer feed-forward MLP-network with 3 inputs. z3 z2 z1 w13 w22 w12 w21 w11 w20 1 w23 f2(. or alternatively by the matrices w and W) are the adjustable parameters of the network. 25 2007 . The examples. The rotating torque is generated by the pneumatic pressure difference supplied from air-compressor between the antagonistic PAM. MPYE-5-1/8HF-710B).
2. 6c: Modeling and identifying neural NNOE model of the PAM manipulator Fig. the convergence of Fitness values and the performance in comparison the real PAM manipulator output and predictive NN model output are showed in followed Fig.Oct 24. The results concern the structure of each NN Model. 25 2007 . Vietnam -101- . This experimental PRBS input-output data is used for training and estimating the resulting ARX model. • The hidden layer of all three NN models composes 5 nodes. Number of hidden nodes = 5) 30 PAM manipulator response NNARMAX model [2. PRBS input and Joint Angle output from (0–20)[s] will be used for training while PRBS input and Joint Angle output from (20– 40)[s] will be used for validation purpose.HCM City. nb=2) for 3 studied NN Models (NNARX. NNARMAX and NNOE. • The training algorithm uses LevenbergMarquardt method.(y-yhat) 0 5 10 15 20 Time [s] 25 30 35 40 10 Er r .8-9-10: a) NNARMAX Model: e(t-2) e(t-1) u(t-2) yhat(t) Fig. 8: PAM manipulator NNARMAX Model International Symposium on Electrical & Electronics Engineering 2007 . 7: Training data obtained from the prototype PAM manipulator.[d ge ] ro er e 5 0 -5 -10 Fig.Track 1: AUTOMATIC CONTROL 5. NNARMAX and NNOE). two following tests would be realized: 1) Identification the PAM manipulator based on recurrent Neural Network Model with the same recursive configuration (na=2. • The Tangent Hyperbolic function for hidden nodes and linear function for output nodes.2. Number of hidden nodes = 5) 2 10 10 Criterion of fitness 1 10 0 10 -1 10 -2 0 2 4 6 8 10 Iteration 12 14 16 18 20 2nd Link PAM manipulator Output and nonlinear NNARMAX Model Output (NNARMAX [2. 6b: Modeling and identifying neural NNARMAX model of the PAM manipulator Fig. IDENTIFICATION & MODELING NNARX MODEL of THE PROTOTY PE PA M MANIPULATOR P B in u -[ ] R S pt v 5. RESULTS OF PAM MANIPULATOR MODELING AND IDENTIFICATION BASED ON AUTO-REGRESSIVE NEURAL NETWORKS MODELS In comparison the performance of three typical recurrent Neural Network Models: NNARX.5 0 30 5 10 15 20 25 30 35 40 10 0 20 10 J in A g [d g o t n le e ] -10 0 -20 -10 -30 -20 0 200 400 600 800 1000 1200 time (samples) 1400 1600 1800 2000 -30 Error of 2nd Link PAM manipulator Output and predictive NNARMAX Model Output .2] response 20 O tp t r s o s .5 5 u(t-1) y(t-2) y(t-1) FITNESS CONVERGENCE OF NONLINEAR NNARMAX MODEL OF THE 2nd LINK 2-AXES PAM MANIPULATOR (NNARMAX[2.[d ge ] u u epne er e 4. • The number of iterations in training process is 100.2.2] .2] . 6a: Modeling and identifying neural NNARX model of the PAM manipulator Fig. 0 200 400 600 800 1000 1200 time (samples) 1400 1600 1800 2000 Fig.7 presents the PRBS input applied to the tested 2-axes PAM manipulator and the responding joint angle output collected from it.
(y-yhat) E r . Number of hidden nodes = 10) 30 PAM manipulator respons e NNARMAX model [2.2] .(y-y hat) 0.(y -yhat) e 2 1 0 -1 0 2 4 6 8 10 time .2 Er r.Track 1: AUTOMATIC CONTROL b) NNOE Model: u(t-2) u(t-1) yhat(t) yhat(t-2) yhat(t-1) 10 1 FITNESS CONVERGENCE OF NONLINEAR NNOE MODEL OF THE 2nd Link 2-AXES PAM MANIPULATOR (NNOE[2. Number of hidden nodes = 10) 3 10 -30 0 200 400 600 800 1000 1200 time (samples) 1400 1600 1800 2000 10 2 Error of 2nd Link PAM manipulator Output and predictive NNOE Model Output .[d g e u u sp n e re ] 10 0 yhat(t) u(t-1) y(t-2) -10 y(t-1) -20 FITNESS CONVERGENCE OF NONLINEAR NNARMAX MODEL OF THE 2nd LINK 2-AXES PAM MANIPULATOR (NNARMAX[2.2] . Number of hidden nodes = 5) 10 Criterion of fitness 0 10 -1 10 -2 10 -3 0 10 20 30 40 50 Iteration 60 70 80 90 100 2) Modeling and Identification the PAM manipulator based on recurrent Neural Network Model with the same typical configuration (na=2. 10: PAM manipulator NNARX Model results Fig.2] respons e 20 yhat(t-1) 2nd Link PAM manipulat or Output and nonlinear NNOE Model Output (NNOE [2. 25 2007 . Number of hidden nodes = 5) 30 PAM manipulator response NNOE model [2.2] response 20 0 Ot u r s os -[ er e u t epne dg ] p e 10 -10 0 -20 -10 -30 -20 0 200 400 600 800 1000 1200 time (samples) 1400 1600 1800 2000 -30 Error of 2nd Link PAM manipulator Output and predictive NNARX Model Output .2] . the convergence of Fitness values and the performance in comparison the real PAM manipulator output and predictive NN Model output are showed in following Fig. It keeps all testing features like the 1st TEST except the hidden layer of all three NN models composes 10 nodes.2] . The results concern the structure of each NN Model.[d g e rro e re ] Criterion of fitness 2 0 -2 0 200 400 600 800 1000 1200 time (samples) 1400 1600 1800 2000 10 1 10 0 10 -1 10 -2 Fig. Number of hidden nodes = 10) Otp t r s o s -[d ge ] u u e p ne e r e 10 30 PAM manipulator response NNOE model [2. Vietnam -102- .[d ge ] ro er e 0.2] response 20 Otp t r s o s .11-12-13: a) NNARMAX Model: e(t-2) e(t-1) u(t-2) 2nd Link PAM manipulator Output and nonlinear NNOE Model Output (NNOE [2. 11 PAM manipulator NNARMAX Model results 10 -2 10 -3 b) NNOE Model: 0 10 20 30 40 50 Iteration 60 70 80 90 100 u(t-2) u(t-1) yhat(t) yhat(t-2) 2nd Link PAM manipulator Output and nonlinear NNARX Model Output (NNARX [2.HCM City. 9: PAM manipulator NNOE Model result 10 -3 0 10 20 30 40 50 Iteration 60 70 80 90 100 c) NNARX Model: u(t-2) u(t-1) yhat(t) y(t-2) 2nd Link PAM manipulator Output and nonlinear NNARMAX Model Output (NNARMAX [2.2] .[s ec] 12 14 16 18 20 Fig.(y-yhat) 20 Criterion of fitness 10 1 1 0 -1 -2 0 2 4 6 8 10 time .[d g e ro e re ] 0 2 4 6 8 10 12 14 16 18 Error of 2nd Link PAM manipulator Output and predictive NNARMAX Model Output . Number of hidden nodes = 5) -20 10 2 -30 2 Er r .2 0 200 400 600 800 1000 1200 time (samples) 1400 1600 1800 2000 E o -[ e r e r r dg ] r e 0 2 4 6 8 10 12 14 16 18 20 Error of 2nd Link PAM manipulat or Output and predict iv NNOE Model Output . 12: PAM manipulator NNOE Model results International Symposium on Electrical & Electronics Engineering 2007 . NNARMAX and NNOE). nb=2) for 3 typical NN Models (NNARX.1 0 -0.1 -0.2.[s ec ] 12 14 16 18 20 10 0 10 -1 Fig.2] response 20 O tp t re o se .2] .2.[d ge ] u u epne er e 10 0 y(t-1) -10 10 3 FITNESS CONVERGENCE OF NONLINEAR NNARX MODEL OF THE 2nd LINK 2-AXES PAM MANIPULATOR (NNARX[2.2] . Number of hidden nodes = 5) 30 PAM manipulator res ponse NNARX model [2.2.Oct 24.
1691-1724. and Fuzzy Logic Models.Track 1: AUTOMATIC CONTROL c) NNARX Model: u(t-2) 6.[degree] 0. and H. Ljung. “Measurement and Modeling of Mc-Kibben pneumatic artificial muscle.HCM City. 12(1). nonlinear NNARX model always yields more performance and higher accuracy than the other nonlinear NNARMAX and NNOE model schemes.[degree] 10 0 -10 -20 -30 0 200 400 600 800 1000 1200 time (samples) 1400 1600 1800 2000 Error of 2nd Link PAM manipulator Output and predictive NNARX Model Output . (2007). Automatic. the model-based identification is discussed with respect to neural networks as universal approximator yielding the possibility of modeling nonlinear autoregressive models with the exogenous variable. K.H. pp. 25 2007 .1 0 -0. Second. and H. Korea. the newly proposed auto-regressive neural networks models presented in this study can be applied in adaptive online control.H. Anh. Furthermore. FESTO Pneumatic Artificial Muscle (PAM) Manual. (1996). “System modeling and identification of the two-link pneumatic artificial muscle (PAM) manipulator optimized with genetic algorithm”. 6. Hannaford. CONCLUSIONS In the paper. First. Neural Networks. 2.” J. Anh. all 3 auto-regressive neural networks models. “Learning and Soft Computing: Support Vector Machines. Sjoberg. Vietnam -103- . namely. (2006). Consequently. in IMECHE Journal. predictive control with better dynamic property.” IEEE Trans. (2001). P.2] . London. (Accepted). 90~102. 5. Third. increasing hidden layer nodes from 5 to 10 will improve the performance of neural NNARMAX and NNOE model.P. strong robustness.. “Nonlinear Black-box Modeling in System Identification: a unified overview. 2004-2005.2 0 200 400 600 800 1000 1200 time (samples) 1400 1600 1800 2000 REFERENCES 1.2] . Acknowledgment This research was supported by the BK21 (Republic of Korea) u(t-1) yhat(t) y(t-2) y(t-1) 10 3 FITNESS CONVERGENCE OF NONLINEAR NNARX MODEL OF THE 2nd LINK 2-AXES PAM MANIPULATOR (NNARX[2. Fig. Number of hidden nodes = 10) 10 2 Criterion of fitness 10 1 10 0 10 -1 10 -2 10 -3 0 10 20 30 40 50 Iteration 60 70 80 90 100 2nd Link PAM manipulator Output and nonlinear NNARX Model Output (NNARX [2. J. Ahn. on Robotics and Automation. pp.P. important remarks can be derived. Zhang.2 Error . (1995). Conf. have overcome successfully the nonlinear characteristic of the prototype PAM manipulator system and resulting recurrent neural networks models would surely enhance the control performance of PAM manipulator.(y-yhat) 0. International Symposium on Electrical & Electronics Engineering 2007 . 3. in Proceedings 2006 IEEEICASE Int. NNARX-NNARMAX and NNOE demonstrate their superb performance in comparison with conventional MLPNN model and linear ARX model as well. 3. Ahn. K. Busan. Number of hidden nodes = 10) 30 PAM manipulator response NNARX model [2. 4. Chou and B. K. Q. Neural networks utilized in modeling and identification of PAM manipulator in this paper. K.2] response 20 Output response .Oct 24. K. such models prove quite suitable for the identification of various plants. due to the extraordinary capacity in learning nonlinear characteristic.1 -0.” The MIT Press and Cambridge. and L. These results can be applied to model and identify not only of the PAM manipulator but also of other nonlinear and time-varied parametric industrial systems. 4744-4749. including linear and nonlinear process without regard greatly changing external environments. pp. 13 PAM manipulator NNARX Model results From previous resulted figures. Vojislav. “A new approach of modeling and identification of the pneumatic artificial muscle (PAM) manipulator based on neural network”. C. except NNARX model with hidden layer 5 nodes proved better.