رویکرد فرایند گرا به کنترل کیفیت خودکار

To guarantee a constant quality of manufactured products, it is necessary to optimise the process parameters immediately when deviations of the workpiece quality have been observed. Established methods of quality management are only able to register quality deviations (like the statistical process control) or to analyse them offline with the help of experts (like failure mode and effects analysis). The presented approach develops a process oriented automated quality control for manufacturing in two steps: First, a local quality controller stabilises the product quality to reference values in the pace of workpiece production. In order to obtain the control law for the process parameters, a learning approach is applied. In the second step, the information exchange between the local controllers is established parallel to the workpiece flow in order to obtain an optimised global process. As an example, the method is applied to quality control in the turning process.

In general, a complete manufacturing process can be modelled as a sequence of local manufacturing subprocesses numbered by i=l, ..., n in Figure 1. Starting with the raw material, the workpieces are processed in the subprocesses in ascending order until the final product is obtained. In each subprocess, the quality of the product is threatened by systematic and random errors. For that reason, the workpiece quality must be measured and a strategy has to be found, to use the measured data in order to compensate at least the systematic errors. In most cases the relation between the process observation and its control is highly nonlinear and classical methods of linear control theory cannot be applied. In most cases, it usually depends on the process knowledge and experience of the operators to adjust the parameters of the machine manually. Quality Control and Quality Management In the recent years, methods of process control have been developed. The statistic process control (SPC) investigates the measured data with statistic methods in order to predict developments within a subprocess for the near future and to give a warning, if a requirement for quality is violated [I]. the failure mode and effects analysis (FMEA) can be used in order to list up possible reasons for failures and to develop suggestions for an improvement [2]. Both methods will support the process operator, but they are not able to define an appropriate control action to optimise the workpiece quality immediately and automatically in a closed loop as shown in Figure 1. Quality Control and Control Theory It is characteristic for a controlled closed loop system, that the determination of the controller output is based on a continuous measurement as well as on reference values [3]. The same situation can be found for the case of quality control: The correction of the controller output is determined continuously in the pace of the process from measurement which quantifies quality characteristics of the product (see Figure 1). It is compared to quality reference values. For the investigated class of production systems, the following properties can be summarised: Since an in-process measurement is impossible or too expensive in most cases, the approach is restricted to a post-process measurement at the end of each subprocess. This can be applied to discrete manufacturing and to batch processes. Hence, the system is discrete in time; the post process measurement limits the time pace. Measurement is related to the product, whereas the controller output is applied to the process. In classical control theory both values are related to the process. It is not sufficient to operate the system in one setpoint. Hence, the relation between measurement and controller output is nonlinear in general In most cases, no mathematical model of this relationship is available. For real plants, multiple inputs and outputs must be considered (MIMO system). For the classical controller design, a mathematical model of the plant is necessary. In the same way the idea of a model based quality control has been developed [4] which uses a model for the interpretation of the measured data. The control unit has to determine its output such that according to the model the measured deviation will be compensated as good as possible. The problem remains that the model has to be developed, which is a time consuming and difficult task in most cases. In section 2 a learning algorithm is presented that is able to obtain a process model from measured data samples. It shows a fast adaptation of the control law within a sequence of a few workpieces. Especially in cases where a sequence of similar workpieces has to be manufactured, the method seems to be advantageous to reach an optimised workpiece quality. Quality Control and Holonic Manufacturing Systems The described quality control of a single subprocess can be improved by an information exchange to the preceding and following subprocess as shown in Figure 2. By this way, the local system receives information about the demands of the complete production chain that can be considered locally. The local subprocess and its control can be regarded as an autonomous unit. It is able to cooperate with other units. These units can be integrated to larger systems, which model production chains.Such a unit is called holon within the concept of Holonic Manufacturing Systems [5], [6]. Therefore, the concept presented in section 3 contributes to a quality related manufacturing holon.

In this paper, a concept of a model based quality control for single processes and for production chains has been presented. It has been shown for the turning process that a control law can be found within a few time steps, even if there is no prior control knowledge. To overcome the difficulty that a process model must be obtained, a simple learning approach using radial basis functions for the approximation of the control law has been developed. Nevertheless, a fast convergence of the adaptation of the control function can only be expected, when the dimension of the input vector is limited. The selection of the dominant quality characteristics and the choice of a sufficient number of radial basis functions needs an understanding of the manufacturing process. For quality control of a process chain, appropriate quality references {d'a}n d the cost functions G") , V'"' must be defined. In order to simplify the application of the concept, a software tool under Matlab has been developed. A software library of standard holons will be implemented for the quality control of more complex manufacturing systems.