Research Article  Open Access
Sensorless Speed Control of Permanent Magnet Synchronous Motors by Neural Network Algorithm
Abstract
The sliding mode control has the merits with respect to the variation of the disturbance and robustness. In this paper, the sensorless slidingmode observer with least mean squared error approach for permanent magnet synchronous motor (PMSM) to detect the rotor position by counter electromotive force and then compute motor speed is designed and implemented. In addition, the neural network control is also used to compensate the PI gain tuning to increase the speed accuracy without regarding the errors of the current measurement and motor noise. In this paper, a digital signal processor TMS320F2812 utilizes its highspeed ADC module to get current feedback information and thus to estimate the rotor position and takes advantage of the builtin modules to achieve SVPWM current control so that the senseless speed control will be accomplished. The correctness and effectiveness of the proposed control system will be verified from the experimental results.
1. Introduction
Nowadays, the permanent magnet synchronous motor (PMSM) is more widely applied than before in the servo control systems. For servo control systems, their excellent performance comes from using position sensors for feedback control. However, the disadvantages of shaft sensors limit the applications, such as system cost increasing, motor size, and reliability decreasing in applications such as air conditioner compressors, where the environment is highly humid and hot. As a result, extensive research has been conducted on overcoming these difficulties by eliminating the position sensors in servomotor systems [1].
Two kinds of sensorless control methods are popularly used in servo control systems, fundamental model based method and saliency based method. Fundamental model based method uses observers, such as back electromotive force (EMF) observer and flux linkage observer, to estimate the rotor position. Meanwhile, the saliency based method consists of continuous signal injection and transient voltage vector injection [2]. Some authors proposed to estimate speed and position of PMSM by backEMF or flux linkage but it is hard at low speed operation and standstill because back EMF amplitude is approaching zero. Other references compared several methods and recommended continuous signal injection method for low operating speed because of its simple hardware configuration. However, the torque ripple and acoustic noise by continuous signal injection are larger than other methods.
A flux linkage estimation method [3] is developed by first measuring the stator linetoline voltages and stator phase current to obtain the back EMF space vector. This vector in turn is used to yield the angle of the flux linkage vector. However, it is affected by integrator drift in low speed and its accuracy is highly sensitive to parameter variation. Another observer based method [4] is presented that removes the dependency on mechanical parameters; however there is still the need for an electrical model of the machine. The tracking control problem is addressed for a sensorless PMSM with unknown constant load torque. Assuming that only stator currents and voltages are available for feedback, a novel sixth order nonlinear adaptive control algorithm is designed for the PMSM [5]. However, the proposed control algorithm is tested just by simulation. A highspeed slidingmode observer [6] is proposed for PMSM, which estimates the rotor position and the angular velocity from the back EMF. Carriersignalinjectionbased control methods [7] are proposed for sensorless control. Fundamental pulse width modulation excitation is used to improve sensorless control of a permanent magnet machine [8]. No additional signal injection or separate test vectors are required. However, modification is needed when narrow PWM voltage vectors occur.
Recently, artificial neural networks (ANNs) have attracted much attention to their possible use in a wide range of engineering applications like power electronics or motor drives [9–15]. The use of ANNs is motivated by their useful features, like having a parallel distributed architecture, being able to identify nonlinear system dynamics, and having the ability to learn, generalize, and adapt. All of these features justify the use of ANNs for motor drive applications, including sensorless control.
This paper presents a neural network based approach to the sensorless control of the PMSM. The basic premise of the method is that an ANN provides a very efficient mapping structure for the nonlinear PMSM. By measuring the phase currents and applying them as inputs to the estimator by least mean squares, the rotor angle is estimated, thereby facilitating the elimination of the rotor position sensor. An NN with back propagation algorithm is considered to compensate the PI gains for speed control.
The paper is organized as follows for further discussion. Section 2 introduces system description including PMSM modeling, neural network with LMS compensation for rotor angle estimation, and speed control by ANN. Section 3 consists of simulation by Matlab/Simulink and experiments. Finally, a conclusion is given in Section 4.
2. Servo System Design
There are three parts for the proposed servo control system design, PMSM modeling, slidingmode observer (SMO) with LMS approach, and PI speed control compensated by neural networks.
2.1. PMSM Modeling
By Park and Clarke transformations, the voltage equations of PMSM from the stationary abc frame to the  frame and the rotating  frame will be given as [9–15] where are  and axis voltages; are  and axis currents; is flux linkage due to the permanent magnets; is the electric speed; and and are the resistance and inductance. The  and axis back EMFs are given as
The use of the Park transformation implies using the rotor position. However, the actual rotor position is not known in a sensorless application and needs to be estimated. Figure 1 shows the block diagram of sensorless PMSM control system, which will be simulated by Matlab/Simulink and experimented to verify the design.
2.2. SlidingMode Observer with LMS Approach
Control systems with slidingmode control (SMC) may have better performance of smaller settling time, less or no overshoot, and faster tracking ability. In addition, slidingmode observer (SMO) can provide information of rotor position and speed estimation. The dynamic equations of SMO are given as follows: where and are the estimated variables of and and is observer gain. The sliding vector for the system is and the defined Lyapunov function is The observer gain will be designed to satisfy Lyapunov’s stability theorem, , as the system trajectory approaches to the sliding hyperplane, . As a result, we have and the estimated rotor angle is
In order to compensate the rotor position estimation error due to the process of motor speed tracking and noise, an adaptive linear element neural network (NN) structure shown in Figure 2 with least mean square approach is adopted. Artificial NNs are similar to biological NNs in the sense that they are based on the same principle of operation based on highly parallel structure and acquiring knowledge through a learning process [16]. The building blocks of an ANN are simple computational nodes, called neurons. They are connected into a network structure by weighted connections, much like the synaptic connections in the human brain. Although the components of an NN are simple and few, the structure and type of these networks vary greatly, determined by the ever increasing interest in this research field.
In Figure 2, are input variables, are weights, is the bias, is averaged error, and is the reference angle from the speed command, respectively. If the estimated error at time instant is (), by LMS approach, we have the follow equations [9–16]: where is the learning rate.
2.3. PI Speed Control Compensated by Neural Networks
The proportionalintegral (PI) control is first considered in the system, where and are the proportional gain and integral gain, respectively. PI controller is widely used in the industrial applications. However, it cannot cope with the load variation or parameter variation well. In the paper, the ANN with back propagation algorithm is utilized to compensate the PI control quantity as shown in Figure 1.
The output of the th neuron at the th layer is calculated as
The error function is defined as
The conjugated gradient method is considered to find the minimum value of the error function (11).
3. Simulation and Experimental Results
The system simulation is programmed by Matlab/Simulink. The parameters of PMSM 8CB75 are listed in Table 1. After training, the chosen topology that gives good performance with minimal resources is a 2hiddenneuron structure for the speed estimating neural network. The activation functions in the hidden layers of the networks are hyperbolic tangent sigmoid functions and those of the output layer are purely linear transfer functions. The values of these weights are given as input layer to hidden layer weights: , hidden layer to output layer weights: , hidden neurons’ thresholds: , and output neuron’s threshold: 0.6560. The final mean squared error of is reached after 20 training epochs for the neural networks. The gain constants of PI control are and , respectively. These parameters are applied to the simulation and experimentation.

Figures 3 and 4 display the rotor angle estimation with SMO and with SMO plus LMS at the motor speed of 100 rpm, respectively. It is easy to find the accuracy of little difference by simulation. However, for experimentation, the great difference between the estimated values can be shown in Figures 5 and 6. One of the reasons is lack of estimation of initial rotor position, which is our next research topic. As a result, there are no complete triangle position waveforms from zero degree to 360 degrees in the figures. In addition, the step speed responses of 200 rpm without and with NN compensation are depicted in Figures 7 and 8. The latter has better performance of less overshoot and settling time. For experimentation, the noload speed responses of 200 rpm without (a) and with NN compensation (b) and the speed errors (c) (blue for no compensation) are displayed in Figure 9. Simultaneously, the step speed responses of 200 rpm with 1 kg disc load are shown in Figure 10. The speed responses without compensation display both the computing errors and the noise during motor rotation. The one by NN control has almost no overshoot, very short settling time, and zero steadystate error. The speed step responses of 100150100 rpm without (a) and with NN compensation (b) under loading of 1 kg disc are shown in Figure 11. The effectiveness of the proposed algorithm shows more clearly.
(a)
(b)
(c)
(a)
(b)
(c)
(a)
(b)
4. Conclusions
This paper proposes a sensorless control system for PMSM by presenting an approach based on neural networks to compensate both the estimated position error and PI control gains. The suggested method is useful for applications to reduce cost. The simulation and experimental results show that the controlled system is capable of estimating rotor angle and motor speed within acceptable limits for many applications.
Conflict of Interests
The authors declare that there is no conflict of interests regarding the publication of this paper.
Acknowledgments
The authors would like to express their appreciation to Ministry of Science and Technology, Taiwan, under Contract no. NSC 1002632E218001MY3 and MOST 1032221E218033 for their financial support.
References
 B. NahidMobarakeh, F. MeibodyTabar, and F.M. Sargos, “Back EMF estimationbased sensorless control of PMSM: robustness with respect to measurement errors and inverter irregularities,” IEEE Transactions on Industry Applications, vol. 43, no. 2, pp. 485–494, 2007. View at: Publisher Site  Google Scholar
 J. M. Liu and Z. Q. Zhu, “Novel sensorless control strategy with injection of high frequency pulsating carrier signal into stationary reference frame,” IEEE Transactions on Industry Applications, vol. 50, no. 4, pp. 2574–2583, 2013. View at: Google Scholar
 R. Wu and G. R. Slemon, “A permanent magnet motor drive without a shaft sensor,” IEEE Transactions on Industry Applications, vol. 27, no. 5, pp. 1005–1011, 1991. View at: Publisher Site  Google Scholar
 T. D. Batzel and K. Y. Lee, “Starting method for sensorless operation of slotless permanent magnet synchronous machines,” in Proceedings of the IEEE Power Engineering Society Summer Meeting, pp. 1243–1247, Alberta, Canada, 1999. View at: Google Scholar
 P. Tomei and C. M. Verrelli, “Observerbased speed tracking control for sensorless permanent magnet synchronous motors with unknown load torque,” IEEE Transactions on Automatic Control, vol. 56, no. 6, pp. 1484–1488, 2011. View at: Publisher Site  Google Scholar  MathSciNet
 H. Kim, J. Son, and J. Lee, “A highspeed slidingmode observer for the sensorless speed control of a PMSM,” IEEE Transactions on Industrial Electronics, vol. 58, no. 9, pp. 4069–4077, 2011. View at: Publisher Site  Google Scholar
 Z. Q. Zhu and L. M. Gong, “Investigation of effectiveness of sensorless operation in carriersignalinjectionbased sensorlesscontrol methods,” IEEE Transactions on Industrial Electronics, vol. 58, no. 8, pp. 3431–3439, 2011. View at: Publisher Site  Google Scholar
 Y. Hua, “Improved sensorless control of a permanent magnet machine using fundamental pulse width modulation excitation,” IET Electric Power Applications, vol. 5, no. 4, pp. 359–370, 2011. View at: Publisher Site  Google Scholar
 T. D. Batzel and K. Y. Lee, “An approach to sensorless operation of the permanentmagnet synchronous motor using diagonally recurrent neural networks,” IEEE Transactions on Energy Conversion, vol. 18, no. 1, pp. 100–106, 2003. View at: Publisher Site  Google Scholar
 F. F. M. ElSousy, “Hybrid ${H}_{\infty}$based waveletneuralnetwork tracking control for permanentmagnet synchronous motor servo drives,” IEEE Transactions on Industrial Electronics, vol. 57, no. 9, pp. 3157–3166, 2010. View at: Publisher Site  Google Scholar
 H.J. Guo, S. Sagawa, T. Watanabe, and O. Ichinokura, “Sensorless driving method of permanentmagnet synchronous motors based on neural networks,” IEEE Transactions on Magnetics, vol. 39, no. 5, pp. 3247–3249, 2003. View at: Publisher Site  Google Scholar
 F.J. Lin, L.T. Teng, and H. Chu, “Modified Elman neural network controller with improved particle swarm optimisation for linear synchronous motor drive,” IET Electric Power Applications, vol. 2, no. 3, pp. 201–214, 2008. View at: Publisher Site  Google Scholar
 J. Yang, Y. Wu, Y. Yu, and W. Zhao, “Permanent magnet synchronous motor control based on retina neural network,” in Proceedings of the International Conference on EProduct EService and EEntertainment (ICEEE '10), Henan, China, November 2010. View at: Publisher Site  Google Scholar
 H. Li, J. Wang, S. S. Gu, and T. Yang, “A neuralnetworkbased adaptive estimator of rotor position and speed for permanent magnet synchronous motor,” in Proceedings of the 5th International Electrical Machines and Systems (ICEMS '01), vol. 2, pp. 735–738, Shenyang, China, August 2001. View at: Google Scholar
 Y. A.R. I. Mohamed, “Design and implementation of a robust currentcontrol scheme for a PMSM vector drive with a simple adaptive disturbance observer,” IEEE Transactions on Industrial Electronics, vol. 54, no. 4, pp. 1981–1988, 2007. View at: Publisher Site  Google Scholar
 L. H. Tsoukalas and R. E. Uhrig, Fuzzy and Neural Approaches in Engineering, John Wiley & Sons, New York, NY, USA, 1997.
Copyright
Copyright © 2014 MingShyan Wang et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.