- Email: [email protected]
Design and Simulation of ANFIS Controller for Increasing the Accuracy of Leaf Spring Test Bench
ScienceDirect ScienceDirect Procedia Computer Science 00 (2019) 000–000
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com Procedia Computer Science 00 (2019) 000–000
ScienceDirect
www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia
Procedia Computer Science 158 (2019) 169–176
3rd World Conference on Technology, Innovation and Entrepreneurship (WOCTINE) 3rd World Conference on Technology, Innovation and Entrepreneurship (WOCTINE) Design and Simulation of ANFIS Controller for Increasing the Design and Simulation of Leaf ANFIS Controller for Increasing the Accuracy of Spring Test Bench
Accuracy of Leaf Spring Test Bench * Elif Üstünışık , Ahmet Kırlı Elif Üstünışık , Ahmet Kırlı
* Technical University, Istanbul 34000, Turkey Department of Mechatronics Engineering Yıldız
Abstract
Department of Mechatronics Engineering Yıldız Technical University, Istanbul 34000, Turkey
Artificial Intelligence (AI) has been in use in several research fields and industries, including automotive. Intelligent control is a Abstract control technique that use different AI approaches like genetic algorithm, machine learning, neural networks and fuzzy logic. In Artificial Intelligence (AI) has been in use in several research fields and industries, including automotive. Intelligent control is a this study, Adaptive Neuro-Fuzzy Inference System (ANFIS) is used for the correlation of the experiment and simulation results control technique that use different AI approaches like genetic algorithm, machine learning, neural networks and fuzzy logic. In of a 5 degrees of freedom (DOF) servo-hydraulic leaf spring test bench. A multi-body simulation (MBS) software named Simpack this study, Adaptive Neuro-Fuzzy Inference System (ANFIS) is used for the correlation of the experiment and simulation results is used to model the actual leaf spring test bench in the simulation environment. However, the results of the simulations do not of a 5 degrees of freedom (DOF) servo-hydraulic leaf spring test bench. A multi-body simulation (MBS) software named Simpack fully correlate with the results of experiments for the same scenarios. Simpack is a dynamic analyses software and is not a control is used to model the actual leaf spring test bench in the simulation environment. However, the results of the simulations do not design tool, but it has an interface that exchanges data with Matlab simultaneously. Therefore, Matlab/Simulink, with its powerful fully correlate with the results of experiments for the same scenarios. Simpack is a dynamic analyses software and is not a control controller design toolboxes has been used for co-simulation with Simpack. ANFIS toolbox has been used to improve the simulation design tool, but it has an interface that exchanges data with Matlab simultaneously. Therefore, Matlab/Simulink, with its powerful model within the Simpack. In this study, the power of the libraries of both Simpack and Matlab/Simulink are combined to perform controller design toolboxes has been used for co-simulation with Simpack. ANFIS toolbox has been used to improve the simulation better simulations. First, MBS model of test bench is improved by using the experimental data. Second, Sugeno type ANFIS with model within the Simpack. In this study, the power of the libraries of both Simpack and Matlab/Simulink are combined to perform grid partitioning is designed by training different experimental datasets. The objective of the training is to evaluate the piston forces better simulations. First, MBS model of test bench is improved by using the experimental data. Second, Sugeno type ANFIS with that correspond to the actual displacement. Lastly, the trained ANFIS model is implemented to the MBS model and co-simulations grid partitioning is designed by training different experimental datasets. The objective of the training is to evaluate the piston forces are performed. The results showed that the simulation results were better suited to experimental data when ANFIS was used. Piston that correspond to the actual displacement. Lastly, the trained ANFIS model is implemented to the MBS model and co-simulations performances are improved by 88,5%, 74,3%, 73,7% for piston 1, 2 and 3 respectively. are performed. The results showed that the simulation results were better suited to experimental data when ANFIS was used. Piston performances are improved by 88,5%, 74,3%, 73,7% for piston 1, 2 and 3 respectively. © 2019 The Author(s). Published by Elsevier B.V. © 2019 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and © 2019 The Author(s). Published by Elsevier B.V. Entrepreneurship Entrepreneurship Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Entrepreneurship Keywords: Co-simulation, ANFIS, Leaf spring, Control, Correlation * Corresponding author. E-mail address: [email protected] Keywords: Co-simulation, ANFIS, Leaf spring, Control, Correlation * Corresponding author. E-mail address: [email protected] 1877-0509 © 2019 The Author(s). Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Entrepreneurship 1877-0509 © 2019 The Author(s). Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Entrepreneurship
1877-0509 © 2019 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Entrepreneurship 10.1016/j.procs.2019.09.040
170 2
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
1. Introduction Advanced ride comfort and handling stability are two main expectations from the vehicle suspension systems which absorb the vibrations and impacts from the road surface. They protect the vehicle components from damage and provide a more comfortable ride for the drivers. Leaf springs are one of the most preferred components of the suspension systems. They are easy to maintain, cost effective and likely to be used to reduce vehicle weight [1-2]. Therefore, analyzing the characteristics of leaf springs with high precision is a necessity [3]. The desire for the high ride comfort of passengers has been increasing over the years. For this reason, automotive manufacturers are investing more in research and development projects for new suspension types [4-6]. Different intelligent control methods like neural network, fuzzy logic have been used for that purpose [7-8]. In this paper, an Adaptive Neuro-Fuzzy Inference System is used to improve the accuracy of the MBS model of leaf spring test bench which is a complex servo hydraulic system. This test bench has a complicated coupling relationship and nonlinear feature. It is used to find the forces acting on the leaf springs when certain inputs are applied and to investigate the effects of these forces on the durability of the leaf springs. Simpack as a powerful multi-body simulation software is used for modeling the test bench. Experiments are conducted on the actual leaf spring test bench and a large amount of dataset is gathered by using data acquisition system. However, the simulation results and actual system results differ for the same inputs. The aim of this study is to solve this correlation problem. There are different approaches that have been used in literature. Hao et al. [9] designed a leaf spring test bench to conduct different types of tests like permanent deformation, stiffness and fatigue. They used PID (Proportion Integration Differentiation) controller to increase the precision of these experiments. Tagawa et al. [10] correlated their test results analytically by formulating the forcedeflection characteristics of leaf springs. Tank et al. [11] correlated CAE (Computer Aided Engineering) and Rig test results for stress and stiffness by using FEA (Finite Element Analysis) approach. Ghuku et al. [12] used image processing technique to obtain more precise results. In this study, leaf springs are tested under different load conditions on the servo-hydraulic test bench. Virtual prototype of the leaf spring test bench created in the Simpack environment is improved with the data acquired from the experiments. Nevertheless, the results of the multi-body simulations did not fully correlate with the experiment results. This phenomenon has been investigated by the authors and an offline FLC is used to improve the results in the previous studies [3]. That solution was unique to the current situation and it required feedback signal. In this study, in addition to fuzzy logic, artificial neural networks have also become a part of the solution process to obtain a more generic result which is independent of both feedback and case. The main contribution of this paper is, by using an intelligent control method, ANFIS, to find a generic solution for the correlation problem of this highly nonlinear test system. For this purpose, the power of Simpack in modelling and the power of Matlab in controls are combined. Co-simulations are performed to obtain more accurate results by using their superior sides. The remainder of the paper is organized as follows: ANFIS architecture is described in Section 2. Design and implementation of ANFIS is explained in Section 3. Related leaf spring test bench is identified in Section 4. Simulation results and discussions are presented in Section 5. Conclusions derived from this study are provided in Section 6. 2. Methodology 2.1. ANFIS Architecture ANFIS is a widely applied AI that combines the advantages of both Neural Networks (NN) and Fuzzy Logic (FL). It is generally used for complex and nonlinear systems in various fields. Denai et al. [13] used ANFIS approach to control neuromuscular system with large uncertainties and highly nonlinearity. Garcia et al. [14] designed an ANFIS based energy management system which consists of battery, renewable energy sources and hydrogen. Kurnaz et al. [15] controlled the autonomous behaviour of unmanned aerial vehicle by using ANFIS methodology. ANFIS architecture was first proposed by Jang [16] in 1993 as shown in Figure 1 where circle represents a fixed node and square represents an adaptive node. It uses the NN learning algorithm to generate a Tagaki-Sugeno type Fuzzy Inference System that approaches a nonlinear system with a variety of linear systems. Fuzzy rules and
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
171 3
membership functions (MFs) are obtained by training the system using experimental data sets [17]. In order to determine the parameters of the adaptive system, back propagation or hybrid learning methods are used in the learning process.
Fig. 1. ANFIS Architecture
ANFIS has Multi Input Single Output (MISO) structure and it consists of five layers. In Figure 1, there are two inputs and one output where A1 and A2 are MFs of input x and B1 and B2 are MFs of input y. First layer represents the MFs of each input. In the second layer, AND operation is performed with the outputs of the Layer 1 and rules are formed as follows; • •
Rule 1: If x is 𝐴𝐴𝐴𝐴1 and y is 𝐵𝐵𝐵𝐵1 , then f1 = p1x + q1y + r1,
Rule 2: if x is 𝐴𝐴𝐴𝐴2 and y is 𝐵𝐵𝐵𝐵2 , then f2 = p2x + q2y + r2
where p1, q1, r1 and p2, q2, r2 are the parameters of the output functions. Outputs of each node are called as firing strength of a rule. In the third layer, firing strengths are normalized. The output of the fourth layer is the first order Sugeno function where all the nodes are adaptive. These nodes find the consequent parameters. The last layer is the defuzzification layer which calculates the crisp output of the ANFIS. 2.2. MANFIS Architecture There are different ways to obtain multi-output by using ANFIS structure. One of them is using Multi-output Adaptive Neuro-Fuzzy Inference System (MANFIS) which combines as many ANFIS as desired. This problem requires three outputs. Thus, three ANFIS are integrated into the simulation model to generate the MANFIS architecture [18], as shown in Figure 2. ANFIS
F1
ANFIS
F2
ANFIS
F3
X
Y
Fig. 2. MANFIS Structure
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
172 4
3. Leaf Spring Test Bench SAE (Society of Automotive Engineers) is a globally active professional association and standards organization. This international organization offers the recommended applications and technical standards for the construction, design, and testing of vehicle components. Leaf springs can be tested under SAE standards as well as national testing standards. Permanent deformation, fatigue life estimation, durability and strength tests are some of the tests conducted on leaf springs. The most efficient results are obtained from full vehicle tests because multiaxial forces (vertical, lateral, longitudinal) act on the components when they interact with each other. However, component-based tests are highly preferred because they are faster, easier and more cost-effective than these tests [19]. On the other hand, virtual tests allow designers to test products without physical prototypes. This method not only shortens the testing cycle but also considerably reduces the cost. Fatigue lifetime is defined as the number of deflection cycles that will withstand a leaf spring without failure [20]. According to design rules, maximum-minimum test load, the desired number of cycles and the frequency are specified. Tests are conducted until failure occurs. If the tests reach or exceed the specified cycles, it means that the leaf spring has passed the required fatigue test. In this study, tests are carried out on a component-based test system which is designed to estimate the fatigue lifetime of leaf springs. First, signals are applied to the test bench to observe the movements of the individual axes. In addition to that, the signals that represents different road conditions were applied to the system.These challenging roads simulates the desired fatigue life for the durability of the vehicle. It is important to select the amplitude and frequency ranges to better demonstrate the rough road durability during the tests [21]. Data sets are gathered for different road scenarios by data acquisition system. The MBS model of the test bench is trained by using these experimental data sets during the correlation process. 4. Design and Implementation of ANFIS For this study, various tests have been performed by applying different road data at a frequency of 0.5 Hz on the actual test bench. Leaf springs are exposed to different loads during the experiments and large amounts of data are collected. 10% of these data is used to train the system and 90% is used to evaluate the results. A hybrid learning algorithm is used for training the ANFIS. In this hybrid algorithm, every epoch is a combination of a forward and backward pass. Hybrid algorithm uses least squares fitting method in the forward pass and gradient descent method in the backward pass. During the trial and error procedure, different numbers and types of MFs are integrated into the system for each piston of the test bench. It is decided that five to eight MFs are enough for this system and Gaussmf and Gauss2mf type MFs give the minimum error. The co-simulation model is designed as shown in Figure 3. As mentioned in the previous sections, the test system has five servo-hydraulic pistons. Analysis from the simulation model showed that the three pistons were not functioning properly. Therefore, intelligent control is designed and performed for these three pistons. Piston1 Disp
ANFIS
Piston1 Force
Piston2 Disp
ANFIS
Piston2 Force
Piston3 Disp
ANFIS
Piston3 Force
SIMAT
Piston1 Disp
To Workspace
Piston2 Disp
To Workspace
Piston3 Disp
To Workspace
Fig. 3. Block Diagram of the Co-simulation Model
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
173 5
Virtual prototype of test bench is extracted into the SIMAT block in order to connect Matlab and Simpack. During the simulation, Matlab and Simpack are in charge of the control loop and the dynamics of the mechanical system, respectively. The connection between two programs are secured via Ethernet Network using a TCP/IP communication. SIMAT is an interface that can only be actuated when torque and force are given as an input. [22] For that reason, inputs of the SIMAT are pistons forces. It is trained to find the corresponding forces of the displacement inputs. As the hybrid learning algorithm repeats inside the ANFIS, it improves the control parameters to get the required control performance. 5. Simulation Results and Discussions In this study, each ANFIS model is trained for the relevant piston. All ANFIS models are combined as MANFIS structure to control all the pistons simultaneously. The results of experiment on the actual test system, Simpack simulation and co-simulation for Piston 1, 2, 3 are shown in Figure 4, 5 and 6 respectively. All the graphs are normalized. The goodness of fit method is used to find how well co-simulation results fit the experiment results. Normalized Root Mean Square Error (NRMSE) method is used to evaluate the improvement. Root Mean Square Error (RMSE) measures the error between predicted values of the model and the actual values observed for n different predictions and times t by using Eq. 1 [23].
RMSE =
∑
n t =1
($ y t − yt ) 2 n
(1)
Normalization of RMSE facilitates the comparison of data sets. It is expressed in Eq. 2.
NRMSE =
RMSE ymax − ymin
(2)
It is calculated that the results are improved by 88,54% for Piston 1, 74,38% for Piston 2 and 73,78% for Piston 3.
Fig. 4. Comparison of ANFIS (Co-simulation)& Experiment& Simpack Simulation Results for Piston 1
174 6
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
Fig. 5. Comparison of ANFIS(Co-simulation)& Experiment& Simpack Simulation Results for Piston 2
Fig. 6. Comparison of ANFIS(Co-simulation)& Experiment& Simpack Simulation Results for Piston 3
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
175 7
6. Conclusion This paper is focused on solving the correlation problem between experiments and simulations results of leaf spring test bench by integrating an intelligent controller, ANFIS. Virtual prototype of the 5-DOF test bench is created in Simpack environment. After the simulations performed, different displacements were observed when the same experimental force inputs were applied by the pistons. It is realized that three pistons of the test bench were not working properly. Although several improvements were made on the MBS model, a controller was still needed. The Simpack does not include a control design tool, however MATLAB / Simulink with its comprehensive and powerful control library, has a co-simulation interface, the SIMAT, that can simultaneously exchanges data with Simpack. Therefore, SIMAT made it possible to take advantage of the superior features of both software for this problem. Then the test data collected from the actual system is used to train adaptive neuro fuzzy system for each piston within MATLAB. The MANFIS structure is integrated to the Simpack simulations where a multi-output system is required. The results of the co-simulations demonstrate that ANFIS considerably improved the simulation results and minimized the error. Piston performances are improved 88,54% for Piston 1, 74,38% for Piston 2 and 73,78% for Piston 3. 7. Acknowledgment We would like to express our appreciations to Y. Emre Erginsoy and İ. Oğuz Er for their valuable and constructive suggestions about this research work. References [1]
P. Kothari, and A. Patel (2014) “A Review Paper on Design & Analysis of Leaf Spring”, International Journal of Engineering Research
[2]
P. Solanki, and A. K. Kaviti (2015) “Study and Review on the Analysis of Leaf Spring”, International Journal for Scientific Research
& Technology 3(3). & Development 3(9). [3]
E. Üstünışık, A. Kırlı, İ. O. Er, Y. E. Erginsoy (2019) “Improvement of the Simulation Model for High Accuracy Leaf Spring Test
[4]
A. Kırlı (2015) “Design and Application of an Active Suspension System on a 6 DOF Half Vehicle Model” IEEE XXV International
[5]
B. Ekici (2005) “Multi Response Optimisation in a Three Link Leaf Spring Model”, International Journal of Vehicle Design 38(4):
[6]
D. Liu, H. Chen, R. Jiang, W. Liu (2010) “Study of Ride Comfort of Active Suspension Based on Model Reference Neural Network
Bench by Implementing Fuzzy Logic” Intelligent and Fuzzy Systems Conference. Conference on Automation Technologies (ICAT), Sarajevo, 2015, pp. 1-6. 326-346. Control System”, Sixth International Conference on Natural Computation 4:1860-1864. [7]
N. Al-Holou, T. Lahdhiri, D. Joo, J. Weaver, and F. Al-Abbas (2010) “Sliding Mode Neural Network Inference Fuzzy Logic Control
[8]
N. M. Ghazaly, A. O. Moaaz (2014) “The Future Development and Analysis of Vehicle Active Suspension System”, IOSR Journal of
for Active Suspension Systems”, IEEE Trans. Fuzzy Syst. 10 (2): 234–246. Mechanical and Civil Engineering 11(5):19-25. [9]
K. Hao, W. Xu, B. Li, X. Fen (2014) “Design of Test bench for Automotive Leaf Spring”, Advanced Materials Research: 1718-1722.
[10] Y.Tagawa, S. Komatsu, and M. Nagai (1995) `Suspension Control System by Model Matching Method', JSAE Review. [11] R. Tank, S. Kurna (2015) “Investigation of Stresses and Deflection in Multi Stage Leaf Spring of Heavy Duty Vehicle by FEM and Its Experimental Verification”, SAE Technical Paper. [12] S. Ghuku, K. N. Saha (2016) “Design Development and Performance Analysis of Leaf Spring Testing Set up in Elastic Domain”, Journal of the Association of Engineers 86. [13] M. A. Denai, F. Palis, A. Zeghbib (2004) “Anfis Based Modelling and Control of Nonlinear Systems:A Tutorial”, IEEE International Conference of Systems, Man and Cybernetics. [14] P. García, C. A. García, L. M. Fernández, F. Llorens, F. Jurado (2014) “ANFIS Based Control of a Grid-Connected Hybrid System Integrating Renewable Energies, Hydrogen and Batteries”, IEEE Transactions on Industrial Informatics 10(2): 1107 – 1117. [15] S. Kurnaz, O. Çetin, O. Kaynak (2010) “Adaptive Neuro-fuzzy Inference System Based Autonomous Flight Control of Unmanned Air Vehicles”, Expert Systems with Applications 37: 1229-1234.
8
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
176
[16] T.O.S. Hanafy (2014) “A New Algorithm to Model Highly Nonlinear System based CANFIS”, International Journal of Computer Applications 94(17). [17] J.-S. R. Jang (1993) “ANFIS: Adaptive Network Based Fuzzy Inference System,” IEEE Trans. Syst. Man Cybern 23(3): 665–685. [18] A. Hajian, H. Zomorrodian, P. Styles (2011) “Depth Estimation of Cavities from Microgravity Data Through Multi Adaptive Neuro Fuzzy Interference System”, Near Surface 2011-The 17th European Meeting of Environmental and Engineering Geophysics. [19] M. Duruş, L. Kırkayak, A. Ceyhan, K. Kozan (2015) “Fatigue Life Prediction of Z Type Leaf Spring and New Approach to Verification Method”, 3rd International Conference on Material and Component Performance under Variable Amplitude Loading: 143-150. [20] S. Mouleeswaran, S. Vijayarangan (2006) “Static Analysis and Fatigue Life Prediction of Steel and Composite Leaf Spring for Light Passenger Vehicles”, Journal of Scientific & Industrial Research 66: 128-134. [21] P. Jayakumar, J. Alanoly and R. Johnson (2005) “Three Link Leaf Spring Model for Road Loads”, Society of Automotive Engineers Warrendale, PA, SAE Paper. [22] Simpack Homepage, http://www.simpack.com [23] S. Oba, M. Sato, I. Takemasa, M. Monden, K. Matsubara, S. Ishii (2003) “A Bayesian Missing Value Estimation Method for Gene Expression Profile Data”, Bioinformatics 19(6): 2088-2096.
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com Procedia Computer Science 00 (2019) 000–000
ScienceDirect
www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia
Procedia Computer Science 158 (2019) 169–176
3rd World Conference on Technology, Innovation and Entrepreneurship (WOCTINE) 3rd World Conference on Technology, Innovation and Entrepreneurship (WOCTINE) Design and Simulation of ANFIS Controller for Increasing the Design and Simulation of Leaf ANFIS Controller for Increasing the Accuracy of Spring Test Bench
Accuracy of Leaf Spring Test Bench * Elif Üstünışık , Ahmet Kırlı Elif Üstünışık , Ahmet Kırlı
* Technical University, Istanbul 34000, Turkey Department of Mechatronics Engineering Yıldız
Abstract
Department of Mechatronics Engineering Yıldız Technical University, Istanbul 34000, Turkey
Artificial Intelligence (AI) has been in use in several research fields and industries, including automotive. Intelligent control is a Abstract control technique that use different AI approaches like genetic algorithm, machine learning, neural networks and fuzzy logic. In Artificial Intelligence (AI) has been in use in several research fields and industries, including automotive. Intelligent control is a this study, Adaptive Neuro-Fuzzy Inference System (ANFIS) is used for the correlation of the experiment and simulation results control technique that use different AI approaches like genetic algorithm, machine learning, neural networks and fuzzy logic. In of a 5 degrees of freedom (DOF) servo-hydraulic leaf spring test bench. A multi-body simulation (MBS) software named Simpack this study, Adaptive Neuro-Fuzzy Inference System (ANFIS) is used for the correlation of the experiment and simulation results is used to model the actual leaf spring test bench in the simulation environment. However, the results of the simulations do not of a 5 degrees of freedom (DOF) servo-hydraulic leaf spring test bench. A multi-body simulation (MBS) software named Simpack fully correlate with the results of experiments for the same scenarios. Simpack is a dynamic analyses software and is not a control is used to model the actual leaf spring test bench in the simulation environment. However, the results of the simulations do not design tool, but it has an interface that exchanges data with Matlab simultaneously. Therefore, Matlab/Simulink, with its powerful fully correlate with the results of experiments for the same scenarios. Simpack is a dynamic analyses software and is not a control controller design toolboxes has been used for co-simulation with Simpack. ANFIS toolbox has been used to improve the simulation design tool, but it has an interface that exchanges data with Matlab simultaneously. Therefore, Matlab/Simulink, with its powerful model within the Simpack. In this study, the power of the libraries of both Simpack and Matlab/Simulink are combined to perform controller design toolboxes has been used for co-simulation with Simpack. ANFIS toolbox has been used to improve the simulation better simulations. First, MBS model of test bench is improved by using the experimental data. Second, Sugeno type ANFIS with model within the Simpack. In this study, the power of the libraries of both Simpack and Matlab/Simulink are combined to perform grid partitioning is designed by training different experimental datasets. The objective of the training is to evaluate the piston forces better simulations. First, MBS model of test bench is improved by using the experimental data. Second, Sugeno type ANFIS with that correspond to the actual displacement. Lastly, the trained ANFIS model is implemented to the MBS model and co-simulations grid partitioning is designed by training different experimental datasets. The objective of the training is to evaluate the piston forces are performed. The results showed that the simulation results were better suited to experimental data when ANFIS was used. Piston that correspond to the actual displacement. Lastly, the trained ANFIS model is implemented to the MBS model and co-simulations performances are improved by 88,5%, 74,3%, 73,7% for piston 1, 2 and 3 respectively. are performed. The results showed that the simulation results were better suited to experimental data when ANFIS was used. Piston performances are improved by 88,5%, 74,3%, 73,7% for piston 1, 2 and 3 respectively. © 2019 The Author(s). Published by Elsevier B.V. © 2019 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and © 2019 The Author(s). Published by Elsevier B.V. Entrepreneurship Entrepreneurship Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Entrepreneurship Keywords: Co-simulation, ANFIS, Leaf spring, Control, Correlation * Corresponding author. E-mail address: [email protected] Keywords: Co-simulation, ANFIS, Leaf spring, Control, Correlation * Corresponding author. E-mail address: [email protected] 1877-0509 © 2019 The Author(s). Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Entrepreneurship 1877-0509 © 2019 The Author(s). Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Entrepreneurship
1877-0509 © 2019 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 3rd World Conference on Technology, Innovation and Entrepreneurship 10.1016/j.procs.2019.09.040
170 2
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
1. Introduction Advanced ride comfort and handling stability are two main expectations from the vehicle suspension systems which absorb the vibrations and impacts from the road surface. They protect the vehicle components from damage and provide a more comfortable ride for the drivers. Leaf springs are one of the most preferred components of the suspension systems. They are easy to maintain, cost effective and likely to be used to reduce vehicle weight [1-2]. Therefore, analyzing the characteristics of leaf springs with high precision is a necessity [3]. The desire for the high ride comfort of passengers has been increasing over the years. For this reason, automotive manufacturers are investing more in research and development projects for new suspension types [4-6]. Different intelligent control methods like neural network, fuzzy logic have been used for that purpose [7-8]. In this paper, an Adaptive Neuro-Fuzzy Inference System is used to improve the accuracy of the MBS model of leaf spring test bench which is a complex servo hydraulic system. This test bench has a complicated coupling relationship and nonlinear feature. It is used to find the forces acting on the leaf springs when certain inputs are applied and to investigate the effects of these forces on the durability of the leaf springs. Simpack as a powerful multi-body simulation software is used for modeling the test bench. Experiments are conducted on the actual leaf spring test bench and a large amount of dataset is gathered by using data acquisition system. However, the simulation results and actual system results differ for the same inputs. The aim of this study is to solve this correlation problem. There are different approaches that have been used in literature. Hao et al. [9] designed a leaf spring test bench to conduct different types of tests like permanent deformation, stiffness and fatigue. They used PID (Proportion Integration Differentiation) controller to increase the precision of these experiments. Tagawa et al. [10] correlated their test results analytically by formulating the forcedeflection characteristics of leaf springs. Tank et al. [11] correlated CAE (Computer Aided Engineering) and Rig test results for stress and stiffness by using FEA (Finite Element Analysis) approach. Ghuku et al. [12] used image processing technique to obtain more precise results. In this study, leaf springs are tested under different load conditions on the servo-hydraulic test bench. Virtual prototype of the leaf spring test bench created in the Simpack environment is improved with the data acquired from the experiments. Nevertheless, the results of the multi-body simulations did not fully correlate with the experiment results. This phenomenon has been investigated by the authors and an offline FLC is used to improve the results in the previous studies [3]. That solution was unique to the current situation and it required feedback signal. In this study, in addition to fuzzy logic, artificial neural networks have also become a part of the solution process to obtain a more generic result which is independent of both feedback and case. The main contribution of this paper is, by using an intelligent control method, ANFIS, to find a generic solution for the correlation problem of this highly nonlinear test system. For this purpose, the power of Simpack in modelling and the power of Matlab in controls are combined. Co-simulations are performed to obtain more accurate results by using their superior sides. The remainder of the paper is organized as follows: ANFIS architecture is described in Section 2. Design and implementation of ANFIS is explained in Section 3. Related leaf spring test bench is identified in Section 4. Simulation results and discussions are presented in Section 5. Conclusions derived from this study are provided in Section 6. 2. Methodology 2.1. ANFIS Architecture ANFIS is a widely applied AI that combines the advantages of both Neural Networks (NN) and Fuzzy Logic (FL). It is generally used for complex and nonlinear systems in various fields. Denai et al. [13] used ANFIS approach to control neuromuscular system with large uncertainties and highly nonlinearity. Garcia et al. [14] designed an ANFIS based energy management system which consists of battery, renewable energy sources and hydrogen. Kurnaz et al. [15] controlled the autonomous behaviour of unmanned aerial vehicle by using ANFIS methodology. ANFIS architecture was first proposed by Jang [16] in 1993 as shown in Figure 1 where circle represents a fixed node and square represents an adaptive node. It uses the NN learning algorithm to generate a Tagaki-Sugeno type Fuzzy Inference System that approaches a nonlinear system with a variety of linear systems. Fuzzy rules and
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
171 3
membership functions (MFs) are obtained by training the system using experimental data sets [17]. In order to determine the parameters of the adaptive system, back propagation or hybrid learning methods are used in the learning process.
Fig. 1. ANFIS Architecture
ANFIS has Multi Input Single Output (MISO) structure and it consists of five layers. In Figure 1, there are two inputs and one output where A1 and A2 are MFs of input x and B1 and B2 are MFs of input y. First layer represents the MFs of each input. In the second layer, AND operation is performed with the outputs of the Layer 1 and rules are formed as follows; • •
Rule 1: If x is 𝐴𝐴𝐴𝐴1 and y is 𝐵𝐵𝐵𝐵1 , then f1 = p1x + q1y + r1,
Rule 2: if x is 𝐴𝐴𝐴𝐴2 and y is 𝐵𝐵𝐵𝐵2 , then f2 = p2x + q2y + r2
where p1, q1, r1 and p2, q2, r2 are the parameters of the output functions. Outputs of each node are called as firing strength of a rule. In the third layer, firing strengths are normalized. The output of the fourth layer is the first order Sugeno function where all the nodes are adaptive. These nodes find the consequent parameters. The last layer is the defuzzification layer which calculates the crisp output of the ANFIS. 2.2. MANFIS Architecture There are different ways to obtain multi-output by using ANFIS structure. One of them is using Multi-output Adaptive Neuro-Fuzzy Inference System (MANFIS) which combines as many ANFIS as desired. This problem requires three outputs. Thus, three ANFIS are integrated into the simulation model to generate the MANFIS architecture [18], as shown in Figure 2. ANFIS
F1
ANFIS
F2
ANFIS
F3
X
Y
Fig. 2. MANFIS Structure
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
172 4
3. Leaf Spring Test Bench SAE (Society of Automotive Engineers) is a globally active professional association and standards organization. This international organization offers the recommended applications and technical standards for the construction, design, and testing of vehicle components. Leaf springs can be tested under SAE standards as well as national testing standards. Permanent deformation, fatigue life estimation, durability and strength tests are some of the tests conducted on leaf springs. The most efficient results are obtained from full vehicle tests because multiaxial forces (vertical, lateral, longitudinal) act on the components when they interact with each other. However, component-based tests are highly preferred because they are faster, easier and more cost-effective than these tests [19]. On the other hand, virtual tests allow designers to test products without physical prototypes. This method not only shortens the testing cycle but also considerably reduces the cost. Fatigue lifetime is defined as the number of deflection cycles that will withstand a leaf spring without failure [20]. According to design rules, maximum-minimum test load, the desired number of cycles and the frequency are specified. Tests are conducted until failure occurs. If the tests reach or exceed the specified cycles, it means that the leaf spring has passed the required fatigue test. In this study, tests are carried out on a component-based test system which is designed to estimate the fatigue lifetime of leaf springs. First, signals are applied to the test bench to observe the movements of the individual axes. In addition to that, the signals that represents different road conditions were applied to the system.These challenging roads simulates the desired fatigue life for the durability of the vehicle. It is important to select the amplitude and frequency ranges to better demonstrate the rough road durability during the tests [21]. Data sets are gathered for different road scenarios by data acquisition system. The MBS model of the test bench is trained by using these experimental data sets during the correlation process. 4. Design and Implementation of ANFIS For this study, various tests have been performed by applying different road data at a frequency of 0.5 Hz on the actual test bench. Leaf springs are exposed to different loads during the experiments and large amounts of data are collected. 10% of these data is used to train the system and 90% is used to evaluate the results. A hybrid learning algorithm is used for training the ANFIS. In this hybrid algorithm, every epoch is a combination of a forward and backward pass. Hybrid algorithm uses least squares fitting method in the forward pass and gradient descent method in the backward pass. During the trial and error procedure, different numbers and types of MFs are integrated into the system for each piston of the test bench. It is decided that five to eight MFs are enough for this system and Gaussmf and Gauss2mf type MFs give the minimum error. The co-simulation model is designed as shown in Figure 3. As mentioned in the previous sections, the test system has five servo-hydraulic pistons. Analysis from the simulation model showed that the three pistons were not functioning properly. Therefore, intelligent control is designed and performed for these three pistons. Piston1 Disp
ANFIS
Piston1 Force
Piston2 Disp
ANFIS
Piston2 Force
Piston3 Disp
ANFIS
Piston3 Force
SIMAT
Piston1 Disp
To Workspace
Piston2 Disp
To Workspace
Piston3 Disp
To Workspace
Fig. 3. Block Diagram of the Co-simulation Model
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
173 5
Virtual prototype of test bench is extracted into the SIMAT block in order to connect Matlab and Simpack. During the simulation, Matlab and Simpack are in charge of the control loop and the dynamics of the mechanical system, respectively. The connection between two programs are secured via Ethernet Network using a TCP/IP communication. SIMAT is an interface that can only be actuated when torque and force are given as an input. [22] For that reason, inputs of the SIMAT are pistons forces. It is trained to find the corresponding forces of the displacement inputs. As the hybrid learning algorithm repeats inside the ANFIS, it improves the control parameters to get the required control performance. 5. Simulation Results and Discussions In this study, each ANFIS model is trained for the relevant piston. All ANFIS models are combined as MANFIS structure to control all the pistons simultaneously. The results of experiment on the actual test system, Simpack simulation and co-simulation for Piston 1, 2, 3 are shown in Figure 4, 5 and 6 respectively. All the graphs are normalized. The goodness of fit method is used to find how well co-simulation results fit the experiment results. Normalized Root Mean Square Error (NRMSE) method is used to evaluate the improvement. Root Mean Square Error (RMSE) measures the error between predicted values of the model and the actual values observed for n different predictions and times t by using Eq. 1 [23].
RMSE =
∑
n t =1
($ y t − yt ) 2 n
(1)
Normalization of RMSE facilitates the comparison of data sets. It is expressed in Eq. 2.
NRMSE =
RMSE ymax − ymin
(2)
It is calculated that the results are improved by 88,54% for Piston 1, 74,38% for Piston 2 and 73,78% for Piston 3.
Fig. 4. Comparison of ANFIS (Co-simulation)& Experiment& Simpack Simulation Results for Piston 1
174 6
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
Fig. 5. Comparison of ANFIS(Co-simulation)& Experiment& Simpack Simulation Results for Piston 2
Fig. 6. Comparison of ANFIS(Co-simulation)& Experiment& Simpack Simulation Results for Piston 3
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
175 7
6. Conclusion This paper is focused on solving the correlation problem between experiments and simulations results of leaf spring test bench by integrating an intelligent controller, ANFIS. Virtual prototype of the 5-DOF test bench is created in Simpack environment. After the simulations performed, different displacements were observed when the same experimental force inputs were applied by the pistons. It is realized that three pistons of the test bench were not working properly. Although several improvements were made on the MBS model, a controller was still needed. The Simpack does not include a control design tool, however MATLAB / Simulink with its comprehensive and powerful control library, has a co-simulation interface, the SIMAT, that can simultaneously exchanges data with Simpack. Therefore, SIMAT made it possible to take advantage of the superior features of both software for this problem. Then the test data collected from the actual system is used to train adaptive neuro fuzzy system for each piston within MATLAB. The MANFIS structure is integrated to the Simpack simulations where a multi-output system is required. The results of the co-simulations demonstrate that ANFIS considerably improved the simulation results and minimized the error. Piston performances are improved 88,54% for Piston 1, 74,38% for Piston 2 and 73,78% for Piston 3. 7. Acknowledgment We would like to express our appreciations to Y. Emre Erginsoy and İ. Oğuz Er for their valuable and constructive suggestions about this research work. References [1]
P. Kothari, and A. Patel (2014) “A Review Paper on Design & Analysis of Leaf Spring”, International Journal of Engineering Research
[2]
P. Solanki, and A. K. Kaviti (2015) “Study and Review on the Analysis of Leaf Spring”, International Journal for Scientific Research
& Technology 3(3). & Development 3(9). [3]
E. Üstünışık, A. Kırlı, İ. O. Er, Y. E. Erginsoy (2019) “Improvement of the Simulation Model for High Accuracy Leaf Spring Test
[4]
A. Kırlı (2015) “Design and Application of an Active Suspension System on a 6 DOF Half Vehicle Model” IEEE XXV International
[5]
B. Ekici (2005) “Multi Response Optimisation in a Three Link Leaf Spring Model”, International Journal of Vehicle Design 38(4):
[6]
D. Liu, H. Chen, R. Jiang, W. Liu (2010) “Study of Ride Comfort of Active Suspension Based on Model Reference Neural Network
Bench by Implementing Fuzzy Logic” Intelligent and Fuzzy Systems Conference. Conference on Automation Technologies (ICAT), Sarajevo, 2015, pp. 1-6. 326-346. Control System”, Sixth International Conference on Natural Computation 4:1860-1864. [7]
N. Al-Holou, T. Lahdhiri, D. Joo, J. Weaver, and F. Al-Abbas (2010) “Sliding Mode Neural Network Inference Fuzzy Logic Control
[8]
N. M. Ghazaly, A. O. Moaaz (2014) “The Future Development and Analysis of Vehicle Active Suspension System”, IOSR Journal of
for Active Suspension Systems”, IEEE Trans. Fuzzy Syst. 10 (2): 234–246. Mechanical and Civil Engineering 11(5):19-25. [9]
K. Hao, W. Xu, B. Li, X. Fen (2014) “Design of Test bench for Automotive Leaf Spring”, Advanced Materials Research: 1718-1722.
[10] Y.Tagawa, S. Komatsu, and M. Nagai (1995) `Suspension Control System by Model Matching Method', JSAE Review. [11] R. Tank, S. Kurna (2015) “Investigation of Stresses and Deflection in Multi Stage Leaf Spring of Heavy Duty Vehicle by FEM and Its Experimental Verification”, SAE Technical Paper. [12] S. Ghuku, K. N. Saha (2016) “Design Development and Performance Analysis of Leaf Spring Testing Set up in Elastic Domain”, Journal of the Association of Engineers 86. [13] M. A. Denai, F. Palis, A. Zeghbib (2004) “Anfis Based Modelling and Control of Nonlinear Systems:A Tutorial”, IEEE International Conference of Systems, Man and Cybernetics. [14] P. García, C. A. García, L. M. Fernández, F. Llorens, F. Jurado (2014) “ANFIS Based Control of a Grid-Connected Hybrid System Integrating Renewable Energies, Hydrogen and Batteries”, IEEE Transactions on Industrial Informatics 10(2): 1107 – 1117. [15] S. Kurnaz, O. Çetin, O. Kaynak (2010) “Adaptive Neuro-fuzzy Inference System Based Autonomous Flight Control of Unmanned Air Vehicles”, Expert Systems with Applications 37: 1229-1234.
8
Elif Üstünışık et al. / Procedia Computer Science 158 (2019) 169–176 Elif Üstünışık, Ahmet Kırlı / Procedia Computer Science 00 (2019) 000–000
176
[16] T.O.S. Hanafy (2014) “A New Algorithm to Model Highly Nonlinear System based CANFIS”, International Journal of Computer Applications 94(17). [17] J.-S. R. Jang (1993) “ANFIS: Adaptive Network Based Fuzzy Inference System,” IEEE Trans. Syst. Man Cybern 23(3): 665–685. [18] A. Hajian, H. Zomorrodian, P. Styles (2011) “Depth Estimation of Cavities from Microgravity Data Through Multi Adaptive Neuro Fuzzy Interference System”, Near Surface 2011-The 17th European Meeting of Environmental and Engineering Geophysics. [19] M. Duruş, L. Kırkayak, A. Ceyhan, K. Kozan (2015) “Fatigue Life Prediction of Z Type Leaf Spring and New Approach to Verification Method”, 3rd International Conference on Material and Component Performance under Variable Amplitude Loading: 143-150. [20] S. Mouleeswaran, S. Vijayarangan (2006) “Static Analysis and Fatigue Life Prediction of Steel and Composite Leaf Spring for Light Passenger Vehicles”, Journal of Scientific & Industrial Research 66: 128-134. [21] P. Jayakumar, J. Alanoly and R. Johnson (2005) “Three Link Leaf Spring Model for Road Loads”, Society of Automotive Engineers Warrendale, PA, SAE Paper. [22] Simpack Homepage, http://www.simpack.com [23] S. Oba, M. Sato, I. Takemasa, M. Monden, K. Matsubara, S. Ishii (2003) “A Bayesian Missing Value Estimation Method for Gene Expression Profile Data”, Bioinformatics 19(6): 2088-2096.