Please use this identifier to cite or link to this item: http://localhost:8081/xmlui/handle/123456789/1853
Full metadata record
DC FieldValueLanguage
dc.contributor.authorMishra, Jyoti Prakash-
dc.date.accessioned2014-09-25T16:09:55Z-
dc.date.available2014-09-25T16:09:55Z-
dc.date.issued2008-
dc.identifierPh.Den_US
dc.identifier.urihttp://hdl.handle.net/123456789/1853-
dc.guidePadhy, N. P.-
dc.guideGupta, S.P.-
dc.description.abstractThe major contribution of the present research work is the modelling and simulation of vector control of Induction Motor (IM) drives using different non linear estimation techniques such as Extended Kalman Filter (EKF), Recurrent Neural Network (RNN) and Adaptive Network Fuzzy Inference System (ANFIS). The performance of the proposed direct and indirect vector control schemes are analyzed and improved by considering many issues like parameter identification, elimination of speed sensor and reduction in computational burden with a view to real-time application feasibility. With vector control, the ac drives outperform the dc drives because of higher transient current capability, increased range of speed and lower rotor inertia. The basic idea is to resolve the input current of an induction motor into two quadrature components, one producing flux, and the other, producing torque, and then controlling them independently. Determination of the instantaneous magnitude and position of rotor flux space vector must be known at every instant of time with precision in order to get perfect decoupling between the torque and flux components of current and hence in order to achieve perfect field orientation. In case of direct vector control, the rotor flux can be determined analytically via stator voltage integration. But, due to stator resistance voltage drop, this could affect the performance of the drive at low speed. To overcome this drawback, the rotor flux must then be estimated using a state observer, but this depends on the motor parameters. The slip calculation algorithm also involves the motor parameters in case of indirect vector control. Hence, the performance of vector control may vary considerably over the operational range of the motor due to the variation of machine parameters with temperature, frequency and saturation depending on the operating conditions. As far as the operating conditions in this work is concerned, the stator and rotor inductances; the mutual inductance in the model can be assumed to remain constant. Therefore, variation in motor parameters, particularly the rotor resistance with operating temperature and frequency should be tracked as they occur to get perfect decoupled control. This parameter must be simultaneously estimated with the rotor flux components in order to develop the robust speed, flux, and torque controllers. Further, in vector control, accurate speed information is also necessary to realize high-performance and high-precision speed control. Speed transducers such as shaft-mounted tachogenerators, resolvers, or digital shaft position encoders cannot be mounted in some cases, such as motor drives in hostile environments, high-speed motor drives and drilling applications. Also, they are usually expensive, bulky and reduce the advantages of induction motor drive systems. It would be advantageous to eliminate the need for the speed sensor and the speed is estimated from terminal voltages and currents using rotor flux models. Hence, both speed and slip estimation should be made robust to rotor resistance variation in order to achieve perfect decoupled control. In recent years, different methods based on stochastic estimation theory and Artificial Neural Networks (ANNs) have been proposed. In most of the works, the speed was used as a parameter based on the fourth order model of the machine, by assuming that the electrical and mechanical modes are completely decoupled, which is not true practically. Therefore, it is very difficult to obtain good performance for the entire speed range and transient states as the parameters of induction motor vary with time and operating conditions. Objective of present research work: In view of above discussion, the main objective of the proposed research work is to obtain robust simulation model of speed sensorless vector controlled VSI fed IM drive. This has been realized in MATLAB/SIMULINK environment incorporating following features: 1. Different basic vector controlled VSI fed IM drive schemes based on hysteresis current-controller in stationary reference frame; and synchronous Proportional-Integral (PI) current controller with space vector PWM based voltage controller in stationary frame, have been simulated and analysed under different operating conditions. 2. Robust speed sensorless operation of vector controlled VSI fed induction motor drive with on-line rotor speed, flux and resistance estimations using Extended Kalman Filter (EKF) has been realized. 3. MRAS based robust speed-sensorless vector controlled IM drive using Kalman filter based online training algorithm to train a real time two layered Recurrent Neural Network (RNN) for estimating the speed and rotor resistance simultaneously, has been implemented . 4. A multilayered Recurrent Neural Network (RNN) based direct adaptive control scheme is realized to estimate the rotor speed using on-line gradient descent back-propagation algorithm and the speed-sensorless vector controlled induction motor drive operation is validated. 5. Application of Adaptive Network Fuzzy Inference System (ANFIS) for speed estimation has also been realized in order to implement the speed-sensorless vector controlled induction motor drives operation. Basic vector controlled VSI fed IM drive schemes For the hysteresis current controlled VSI fed drive, the inner loop of the drive is hysteresis current controlled in stator reference frame and the outer speed loop is PI controlled. In case of synchronously current controlled VSI fed drive, the inner current loop is PI controlled in synchronously rotating reference frame with space vector PWM based voltage controller in stationary frame and the outer speed loop is PI controlled. The performance of the proposed schemes have been simulated under various dynamic conditions namely, starting, speed reversal and load perturbation with the assumption that the motor parameters, particularly the rotor resistance does not change during the operation. The simulation of the basic vector controlled induction motor drive is then extended to consider the rotor resistance and/or rotor flux and speed estimations either individually or simultaneously using newly created S-function (system-functions) blocks for studying various issues of high performance drive operations such as parameter insensitivity, elimination of speed sensor and reduction of computational burden. Possible solutions related to above issues have been first attempted using Extended Kalman Filter (EKF) based two nonlinear estimators, but these involve more computational work and also may be difficult for practical implementation. In order to reduce the computational burden along with parameter insensitivity and speed-sensorless operation, different Al-based nonlinear estimators such as Recurrent Neural Networks (RNN) and Adaptive Network Fuzzy Inference Systems (ANFIS) have been proposed in this work to improve the dynamic performances of vector controlled induction motor drive. Finally these AI based estimators are validated in the current controlled based VSI fed speed-sensorless indirect vector drive scheme under different transient and steady-state operating conditions in MATLAB environment. The S-functions provide a powerful mechanism for extending the capabilities of SIMULINK in implementing modelin based on-line identification schemes. It is believed that these Al-based approaches will find increasing application in future mainly due to the fact that the development time of such an estimator is short and the estimator can also be made robust to parameter variations and noise. Robust sensorless vector control with EKF based identification scheme: For the EKF based robust sensorless schemes, two appropriate fifth order discretized mathematical models of induction motor are developed for modelbased on-line estimations. These models consider the coupling of electrical and mechanical modes in order to estimate the rotor flux, speed, and rotor resistance online simultaneously using the EKF algorithm. The excitation signals (control inputs) for the EKF are the outputs of a PWM voltage source inverter used to drive the motor. These control inputs contain components at high frequencies, which are used as the required noise by the Kalman filter. Thus no additional external random noise signals are needed. The performance of the drive has been analyzed using both current-regulated PWM and space vector PWM techniques. The value of rotor resistance in stationary frame d-q induction motor model is varied intentionally, thereby simulating a change in rotor resistance. The EKF algorithms are coded in S-function MATLAB-files and then used in SIMULINK in the form of S-function block for on-line estimations of rotor speed, rotor flux and rotor resistance simultaneously in order to implement robust sensorless direct and indirect vector controlled induction motor drives. Measurement and state covariances are chosen so that both transient and steady-state errors are optimized. The EKF has been shown to reliably converge and track state/parameter changes, even in the presence of harmonics in PWM inverter output. Robust sensorless vector control with RNN based identification scheme: Artificial Neural Network (ANN) is now being used as a powerful tool to obtain nonlinear mathematical model approximations in an induction motor drive system. The performance of an off-line trained ANN approach depends upon the amount and quality of training data used, which in turn depends on system complexity, the range of operating conditions involved and is also sensitive to parameter variations. It is better to use on-line training algorithm for any real time applications. Taking the above ideas into consideration, a robust speed sensorless vector controlled induction motor drive is developed using Kalman IV filter based on-line training algorithm to train a real time partial RNN. The error between the desired state variable of an induction motor and the actual state variable of a neural model is used in the proposed algorithm to adjust the weights of the neural model, so that the actual state variable tracks the desired value. The proposed fast convergence algorithm has the time-varying learning rate and estimates the rotor resistance and speed simultaneously. Unlike the conventional off-line supervised learning, this on-line training of neural network does not require a learning step and the learning process commences as the estimation process starts. The duration of training in every step is definite which is equal to the duration of training algorithm executed once. This makes the estimation process complete in one complete cycle of execution of training algorithm. The estimated speed is then fed back in the speed control loop to realize the speedsensorless vector drive operation. The updated value of rotor resistance is used in the vector controller block to estimate the reference slip-speed and obtain the robust control performance with proposed speed sensorless indirect vector control scheme. It has been observed that the algorithm reliably converges and tracks under various loading conditions, even in the presence of harmonics in PWM inverter output. This realizes that this scheme would be ideal for real time applications. In the above scheme, a two-layered RNN without hidden layer was used and the speed value is not obtained at the output of RNN, but at one of the weights. The weights are adjusted during online training which contain the speed and rotor resistance as parameters. In another scheme, a speed sensorless vector controlled induction motor drive is developed using a multilayered RNN trained with online gradient descent back-propagation algorithm. In this case, with the trial and error procedure, the RNN structure with one hidden layer has been used to obtain the rotor speed value at the output node. The multilayer and the recurrent structure make the speed estimation robust to parameter variations and system noises. The weights of the neurons are continuously modified using online back-propagation training algorithm. The estimated speed is then fed back in the speed control loop to realize the speed-sensorless vector drive operation. The input signals continuously available from the running SIMULINK model of vector drive. Therefore, no off-line training and training patterns are needed. Sensorless vector control with Fuzzy-Neural based identification scheme: It is believed that ANN-based speed estimators will be used in different vector and direct torque controlled drives, in which case the number of hidden layers and hidden nodes has to be obtained by trial and error. However, to overcome the difficulty of finding the appropriate ANN structure (appropriate number of hidden layers and hidden neurons) by trial and error, a fuzzy-neural network can also be used to yield the instantaneous variations of the rotor speed, electromagnetic torque, stator / rotor flux linkages, etc. Since both fuzzy and neural systems are universal function approximators, their combination, the hybrid fuzzy-neural system (or neural-fuzzy system) is also a universal function approximator. The state-of-the-art intelligent control nowadays is at the level of adaptive fuzzy-neural controllers. This approach therefore provides a means of combining the use of imprecise, linguistic information but which has a clear physical significance with formalized mathematical structures and training algorithms. The main advantage of adaptive fuzzy-neural approximators (systems) is that it allows automated design and tuning procedures by using minimum human intervention. Adaptive Neuro-Fuzzy Inference Systems (ANFIS) is a class of adaptive networks that are functionally equivalent to fuzzy inference systems. As the name indicates, a fuzzy inference system is designed systematically using the neural network training method and hence possesses the excellent characteristics of both neural network and fuzzy inference system. As a special neural network, ANFIS can approach all nonlinear-systems with less training data, quicker speed and high precision. In this direction of work, the ANFIS speed estimator is first trained offline for estimating the rotor speed of induction motor drive. The training vectors i.e., Input-output data patterns required for the training of ANFIS are obtained from the mathematical speed estimation model and by running the conventional vector drive MATLAB/SIMULINK model. For this purpose, the mathematical model of induction motor is considered and expression for the rotor speed is also obtained. ANFIS is trained to approximate the equation for the speed. To speed up the convergence, hybrid-learning algorithm, which is a combination of least squares and gradient descent algorithms for training the ANFIS, has been proposed. The performance investigation of ANFIS speed estimator has been carried out in two stages: the first stage involves the process of FIS initialization VI by adopting both grid-partition and subtractive clustering methods separately on the training data set, followed by offline hybrid learning of the ANFIS estimator; the second stage involves the performance validation of offline trained ANFIS when used for online estimation of the speed-feedback signal in a speedsensorless indirect vector control of IM drive scheme. Initialization of the fuzzy inference system (FIS) makes the ANFIS training to start at reduced error measure 'RMSE'. In order to cope up with different dynamic operating conditions, the speed estimator was subjected to learn separately with different sets of input/output training data pairs acquired during: 1) step command starting and speed reversal operation with no-load; 2) step command starting and speed reversal operation with step load perturbation at steady state; and 3) ramp command starting and speed reversal with step load perturbation at steady state. The training process is performed for specified number of epochs. It is observed that the fuzzified network has learnt the required non-linear characteristics (i.e., relation between the speed and the terminal voltage & currents) satisfactorily when simulated under different dynamic conditions such as starting, speed reversal and load perturbations and low-speed operations. Summary: The present work investigates the steady-state and dynamic performance of speed-sensorless vector controlled induction motor drive operation with the application of different nonlinear estimation techniques such as EKF, RNN and ANFIS. In order to reduce the computational burden along with parameter insensitivity and speed-sensorless operation, the Al-based nonlinear estimators such as RNN and ANFIS have been preferred over EKF. The performance of each proposed scheme has been simulated and results are presented with critical evaluation. This research work is likely to make significant contribution in the field of vector control of induction motor drive.en_US
dc.language.isoenen_US
dc.subjectELECTRICAL ENGINEERINGen_US
dc.subjectVECTOR CONTROLLED INDUCTION MOTOR DRIVEen_US
dc.subjectINDUCTION MOTORen_US
dc.subjectRECURRENT NEURAL NETWORKen_US
dc.titleSIMULATION OF AI BASED VECTOR CONTROLLED INDUCTION MOTOR DRIVEen_US
dc.typeDoctoral Thesisen_US
dc.accession.numberG20578en_US
Appears in Collections:DOCTORAL THESES (Electrical Engg)

Files in This Item:
File Description SizeFormat 
SIMULATION OF Al BASED VECTOR CONTROLLED INDUCTION MOTOR DRIVE.pdf10.14 MBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.