DEVELOPMENTS IN ANN BASED CONTROLLERS FOR ON-LINE PROCESS CONTROL APPLICATIONS

Abstract

Process control systems use controllers with adjustable settings which control and coordinate the process operations over a wide range of operating conditions. The classical feedback control theory is the basis for the development of simple automatic control systems using P, PI, and PID controllers. Their easily comprehensible principles and relatively simple implementations have been the main reason for their wider applications in process industry. To compensate for parameter variations in the plant as well as to adopt to the changes in the environment, controller settings are tuned after the control system has been installed using trial and error method. For the significantly changed conditions of the process, the controllers must be retuned to obtain satisfactory control. In conventional control, linear approximation of the plant properties itself has some disadvantages, like linear approximation becomes computationally impractical, if the plant is very complex and highly dynamic and also there are difficulties in adapting itself to changing plant parameters. Thus the classical controllers have drawbacks of approximating the properties of the system and are very sensitive to the variations in the process plant parameters. In order to tackle these difficulties, in recent years, there has been extensive interest in adding adaptive features in conventional controllers. The adaptive control system automatically adjusts the controller settings to compensate for unanticipated changes in the process environment. A good number of adaptive control techniques like direct adaptive control, indirect adaptive control and model reference adaptive control has been developed and are in use in real time process plant applications. The ability to adapt automatically to variations in the plant dynamics and environment has made such adaptive controllers increasingly important for various applications. Adaptive controllers require mathematical modeling of the plant which is sometimes difficult and impracticable in some cases The inaccuracy in the modeling of the plant leads to degraded performance of the controller. These controllers also contain some parameters those make it difficult to handle by the field engineers/process operators for parameter settings and control actions. Sometimes, the adaptive control techniques also find difficulty in dealing with unknown structures and nonlinearities of IV the plants. To overcome the above mentioned limitations of the classical and adaptive controllers, the work has been started towards the development of ANN based intelligent controllers in the recent times. An intelligent control system possesses in built adaptation or learning and decision making capability to meet the desired performances over a wide range of changes in the process parameters. ANN based controllers have the following important features: (i) They learns the system characteristics through nonlinear mapping, (ii) Control strategies are designed in model free environment, (iii) Lack of proper knowledge of the plant dynamics does not pose any hindrance in their development and implementation, (iv) The adaptation and learning in these controllers are obtained through off-line and on-line weight adaptation to keep the network updated to the current environment, and (v) The neural network processes multilple number of inputs and outputs and thus is readily applicable to multivariable systems. Keeping above in mind, it has been planned in this work to investigate the various aspects of neural network concepts and to modify and implement them for real time process control applications. Direct adaptive control, model reference adaptive control, indirect adaptive control and cascade control strategies have been implemented using different types ofANN structures. These control techniques have been tested on simulated system models as well as on a continuously stirred tank (CST) system. The CST process represents a continuously stirred tank reactor (CSTR) which is commonly used in the process industry and is nonlinear in nature. It is one of the most suitable process for testing the performance ofANN based control configurations. The experimental set-up of the CST process has been designed, fabricated and installed in the laboratory. It is a time-varying process and is most suitable for implementing the neural network based control strategies. The process setup consists of associated electronic hardware such as signal-conditioning circuit for conditioning the temperature signals from the process, analog-to-digital converter and digital-to-analog converters to interface the experimental process with a personnel computer, voltage-to- current converter to convert output of digital to analog converter into a current signal and current-to-pressure converter to convert the current signal into a corresponding pneumatic (pressure) signal and thus to interface a linear pneumatic control valve with the computer. The dynamics of the spring actuator of the control valve is faster than that of the CST process, and hence, the dynamics of the actuator is not considered in the design of controller. The mathematical model of the CST process is a second order system witha time delay, but as one of the time constants is smaller in value, the model is assumedto be a first order system with time delay. The first controller i.e. hybrid direct neural adaptive controller (HDNC) has been developed by using a linear feedback compensator in parallel with a direct neural network adaptive controller (DNC). In this work, the effects of number of neurons in the hidden layer and learning rate parameter on the system response have been studied. It is observed that the system performance does not always improve with the increase in the number of neurons in the hidden layer or by taking the higher value of learning rate parameter, instead, it increases the overshoot in the system response and the response becomes oscillatory. For lesser number of hidden layer neurons or for lower value of learning rate, the speed of convergence is slowed down. An iterative procedure has been developed for optimal selection of the learning rate and number of neuron in the hidden layer for minimum overshoot in the process output with fast convergence. The controller has been tested on the simulated process models and also on a real-time CST process. The performance of the hybrid direct adaptive neural controller is compared with the performance of a direct neural adaptive controller. The hybrid direct neural adaptive controller performs better compared to the direct neural adaptive controller under the performance indices of set point tracking, disturbance in the CST process, recovery speed, minimum overshoot and integral square error. The hybrid direct neural adaptive controller compensates for the limitations of the linear controller by adjusting the appropriate control signal and improves the overall performance during total operating region. These controllers are well suited for slow processes with known dynamics. In these controllers, the error between the plant and desired plant outputs is VI used for the network training. There is necessity to know the Jacobian of the plant dynamics for adjusting the network parameters. Selected learning rate has got different effect at different operating points and also there is need to select the parameters of the linear feedback compensator. In order to overcome some of the limitations of HDNC, a second modified indirect neural adaptive controller (MIDNAC) has been developed by using an indirect adaptive control method using Gaussian neural network for on-line estimation of nonlinear function. The results of MIDNAC are compared with the indirect neural adaptive control (IDNAC). A Gaussian neural network reduces the computational efforts of modifying the connection weights between hidden layer and input layer. A method for adaptive learning rate has been developed to overcome the difficulties associated with the optimal selection of learning rate parameter. Also a linear control law is obtained to overcome the limitations of on-line identification and control in neural network based adaptive control system, in addition to the control signal obtained from the estimated parameters. The unknown functions of the system are continuously estimated and are used for adaptive tuning of the nonlinear control law parameters. When the identified parameters approaches to their true values, the control parameters also approach their desired values, therefore, the control mechanism comes out to be robust in nature. This control strategy has the advantage as it does not require the learning of inverse dynamics of the controlled plant. In this techniques parallel identification of the plant is carried out using neural network to estimate the unknown functions of the model. The proposed controller produces the control signal of containing the spikes. The magnitude of the spikes in the control signal depends on the value of parameter used in linear control law. The performance of the controller has been evaluated on first and second order simulated models, on a model of non-linear process and also on an experimental set-up of the CST process. It has been observed that the modified indirect neural adaptive control configuration gives improved performance compared to indirect neural adaptive control configuration in terms of recovery speed and integral square error. vii The HDNC and MIDNAC are suitable when the desired output of the system is specified in terms of set points. But if the desired performance of a system is given in terms of a reference model, then these controllers are not suitable. Therefore, the second phase of the work has been carried out to develop a modified direct neural model reference adaptive controller (MDNMRAC) which forces the plant to behave like the reference model. The controller has been developed by modifying a direct neural.model reference adaptive controller (DNMRAC) by using a feedforward neural network for forward modeling of the plant. The network is trained using error, which is the difference between the outputs of the plant and the neural network model. The control law has been constructed using trained network parameters and tracking error. It has been found that the performance greatly improves with modifications both in terms of transient as well as steady state behaviors. This performance has been evaluated on the simulated process models and also on an experimental setup of a CST process. To evaluate the performances on an experimental setup, the temperature control of a CST process has been carried out. The robustness of the system has been confirmed for the set-point tracking. The performance of controllers has also been evaluated under the influence of disturbances. The performance of the MDNMRAC as compared to DNMRAC has been found better in terms of integral square error. The HDNC requires plant dynamics and choice of constants of LFBC on the other hand MIDNAC does not require these parameters, but both are suitable for setpoint tacking type of processes. MDNMRAC requires the plant model and are suitable for model output tracking types of processes only. In all these controllers, the disturbances arising from manipulated variable are taken care-off only after they have already influenced the process output. Thus there is a necessity to develop such a controller which can take care of the disturbances in an effective manner before their influences on the plant output. Therefore, an ANN based cascade controller has been developed wherein the disturbance arising within the secondary loop is taken care-off before it influences the process output. In cascade control of a CSTR process, there is one manipulated variable (cooling jacket flow rate) and more than one measurements (reactor and cooling jacket temperatures). The controller has been developed by using a vm feedforward neural network with variable network parameters in cascade control of systems (NCC). Addition of a linear feedback compensator has resulted in the improvement of both the transient as well as steady state behaviors. The controller uses an adaptive learning algorithm. Also a control technique that uses a neural controller for implementing the primary and secondary controllers and a linear feedback compensator (NCC with LFBC) has been successfully tested. On the basis of the results on the simulated model as well as on an actual experimental set up, it can be stated that an efficientneural control system is realized by using variable learning parameters. Also by incorporating a linear feedback compensator inthe ANN control system, in parallel with the primary and secondary controllers, the performance is enhanced to a greater extent. The test results show that an ANN based cascade controller having a linear feedback • compensator in parallel with the primary and secondary controllers gives better performance. On the basis of the comparison of the performance of the developed controllers with the conventional cascade controller (CCC), it can be stated that the results with this controller are better and indicate high promise of its industrial applications. The performance of the developed controller is comparable in respect of the set point tracking and better in respect of the load disturbance to conventional cascade controller. In the present work, four adaptive ANN based controllers using different adaptive control strategies have been successfully developed and implemented. The performances of all the controllers have been studied both on the simulated models and also an experimental setup of a CST process. The experimental setup was designed, fabricated and installed in the laboratory as a part of the present work. The controllers have been tested in different trials with variable set points and disturbances of different nature. The comparison of various neural control configurations has been carried out in terms of the ISE. It has been observed that the performances of HDNC, MIDNAC and MDNMRAC are better then DNC, IDNAC, and DNMRAC respectively. It is also observed that MIDNAC gives better performance compared to DNC, HDNC, IDNAC, IX DNMRAC, MDNMRAC, CCC, NCC, NCC with LFBC for set point tracking. For the disturbance rejection the performance of NCC is better in comparison to CCC. The NCC with LFBC gives the best performance compared to all for disturbance rejection. Besides providing solutions to practical implementation of ANN based controllers, the work done in this thesis also opens new directions to carryout further investigations on the performances of different type of controllers.