Control system architectures that feature a closed feedback loop involving the process, sensors, controller, and actuators are a step beyond the traditional approach. Control algorithms, such as the PID (proportional, integral, derivative) algorithm, reside in the controller, which is a special type of digital computer. The challenge is to control what is basically a dynamic analog process (e.g., machining) with discrete digital logic. The control architecture must be designed to ensure that the process can always operate optimally under the presence of various uncertainties. Thus there may be multiple layers of feedback control loops (e.g., servo-control loops around the machinery itself, control loops around the tool and the workpiece for fine adjustment of operating condition, etc.).
The design of feedback control algorithms is affected by a number of factors. The key issue arises from the dynamics involved in the process. The controlled variable does not respond instantaneously to the controlling input, which results in a characteristic response curve with dynamic delay. Process models implemented in the controller must manage the dynamics properly. A fixed-parameter controller has difficulty in keeping up with the nonlinear, time-varying behavior of a process. Good control performance at one operating condition can give way to poor performance at another operating condition. These models may be further constrained by the amount of bandwidth for the feedback loop (i.e., the closed-loop response speed) and the product specifications, such as error tolerance.
Control algorithms currently used in manufacturing are commonly simple PID control algorithms that use a low-order transfer function model. This technology is adequate for traditional machining operations in which the machining speed is low. Also, the performance limitation of these PID controllers provides for only a low level of closed-loop performance in unit manufacturing processes, which is reflected in the final product quality level.
Sophisticated control architectures are required for modem unit processes that inherently possess time-varying, nonlinear process dynamics and are high performance in terms of speed and control accuracy (e.g., high-speed machining). For instance, a manufacturing manager recently observed that “a high-speed spindle is worthless unless the machine can feed fast enough to exploit it and the cnc [computer numerical control] is fast enough to keep everything under control” (Coleman, 1992). Advances in control theory, as well as those in microprocessor and digital signal processing technology achieved over the last several decades, can be and should be utilized aggressively to face these new challenges in modern manufacturing.
There are several advanced control methodologies applicable to manufacturing process control. One type of controller adjusts set points as a result of data received from sensor arrays (Hardt, 1993; Ulsoy and Koren, 1993). For example, a model-based adaptive controller1 could employ an algorithm to compensate for dimensional errors induced by thermal distortion of workpieces.
Adaptive and robust control theory has been an active research topic for the past two decades. The research addresses the problem of how to attain optimum system performance when a process model is not known precisely in advance, the operating conditions are variable, the process parameters vary nonlinearly during operation, etc. The philosophy behind adaptive control theory is that the controller must adapt its control gains so that the overall system remains at or near the optimal condition in spite of varying process dynamics. Adaptive control is a key element to provide flexibility to unit manufacturing processes, which must be responsive to rapidly changing needs of products. On the other hand, the philosophy behind robust control theory is that a fixed-gain controller should be selected, so that the performance of the overall system remains acceptable under variations of process dynamics.
There is significant potential benefit to applying adaptive and robust process controllers. Disturbance observer theory, a robust control methodology, has been shown to be ideally combined with feed-forward control algorithms to provide high accuracy performance for servo-systems, which are essential in high-speed machining. Successful experimental results have been reported for adaptive force control in machining and adaptive weld-pool control in welding.
Important forms of adaptive control are the self-tuning controllers that were developed to overcome the limitations of fixed-point controllers in responding to time-varying process dynamics, variable operating conditions, nonlinear process dynamics, and lack of operator expertise during control-loop commissioning. Self-tuning controllers use process identification algorithms to estimate or track the time variation of key process parameters in real time. Based on these results, control parameters are computed in real time to ensure optimal system performance.
Two entirely different types of self-tuning controllers have been developed—expert systems and process models. An expert system consists of a set of rules that are derived from the knowledge of experienced process engineers and operators. A fuzzy logic controller is a viable candidate to translate human
knowledge to control strategies and algorithms suitable for computer implementation. Advantages of the expert system approach are that it is robust and thus additional rules can be readily added and that a process model is not required. But there are some disadvantages. The expert system usually is developed using a particular controller structure and thus cannot be readily ported to another type of controller. Also, the rule base itself can not be readily analyzed.
The model-based controller uses rigorously defined performance criteria, and hence mathematical analysis of these criteria is possible. It can be adapted for implementation in different controller structures, and it may be used for process diagnostics, such as to locate a failed sensor or actuator. The disadvantages include the chance that the model structure may not match the physical process; for example, an actuator dead zone or backlash could cause underestimation of process gain. Also, rapidly occurring process changes can cause problems if the model execution time is too slow.
The controller of the future will most likely incorporate both the expert system and the model-based technologies. A critical issue is the customization of these modem control algorithms to specific manufacturing applications such that stable performance results.
“Internal state” has been a key concept in modem control, and control theory has been advanced together with estimation theory. Estimation theory provides methodologies to estimate “state,” which may not be directly measured. The estimated state can be utilized for state feedback control as well as for monitoring and failure detection.
Learning control can be used to learn the optimum control input through repeated trials (Dagli, 1994). When unit processes repeat the same task, this control methodology fine tunes the controller’s performance. For example, in injection molding, the piston speed must be controlled so that the flow of molten plastic reaches all parts of the mold and no voids are created. Learning-control algorithms can be combined with simulation models and operational data to evaluate the performance of each trial injection. The time profile of the piston speed is adjusted after every trial until a quality product is produced. The number of trials required depends on the complexity of the process. This type of scheme also has been tested in machining to compensate for low-velocity friction forces. It has been demonstrated that a dozen or so trials are sufficient to construct a compensation signal to remove undesirable glitches, which are visible in the part geometry as irregularities in machining that are caused mainly by static friction.
Intelligent control has received increasing attention over the past few years. Intelligent control systems have the ability, to varying degrees, to find strategies autonomously in an uncertain environment. Intelligent controllers rely on a knowledge base, which may contain experts’ knowledge about operations of unit
processes. The knowledge base may come from process study and modeling and may be updated by a learning mechanism during operation. Intelligent controllers may provide a signal to switch operational modes for a process responding to sensor outputs. The development of strategies for intelligent controllers includes expert systems.
Introduction of a new controller usually requires some modification to other machine functions. For example, in some cases the controlling input (i.e., manipulated variable) for adaptive force control in machining is the tool feedrate. The adjustment of feedrate requires coordination of the machine tool’s servo-controllers as well as its computer numerical control functions. Traditional computer numerical controls generate reference signals for servo-loops after linear and circular interpolation in accordance with the tool feedrate supplied by part programmers. However, this approach will not be feasible if the feedrate is varied in real time.
