Advanced Quality Control in Foundry Manufacturing Process

摘要

One of the most complex manufacturing processes is the foundry process which consists of several sub-processes linked together to a complex process chain. The foundry process is defined by the individual dependencies between the process parameters and variables as well as by several elements of uncertainty which affect the stability of any sub-process applied and, consequently, product quality. The final quality of a cast part is made in a process situation where anything and anyone has direct control and correcting possibilities over the process. This fact induces strictly the demand on an extremely accurate and in advance control of all sub-processes. To this end, it is necessary to know and finally to control all causes and factors that contribute to the final quality of the cast part. The integrated control of this process, therefore, is a complex and challenging task. The main requirement of foundries is tools which are able to perform data analysis under consideration of any influences of sub-processes on manufacturing quality. This implicates highest flexibility regarding the availability and applicability of different data analyzing tools. During the last 20 years, several different methods have been developed for data analyzing in manufacturing processes. These are mainly Neural Networks (NN), Bayesian Networks (BN), further statistical data analyses and expert systems. Most of these tools have been designed for a specific application and do not have the ability to be applied on foundry problems. Due to the specific algorithms of analyses each of them provides separate results which are strongly related mostly to one sub-process. Correlation and interaction between several processes, parameters and variables are difficult to establish. An Intelligent process control by self-adjusting systems which automatically trace back specific product quality attributes to input variables and fitting these to specified output results is hardly available. In many applications of these methodologies, it has also been substantiated that it is not possible to meet the demands of an intelligent process control using only a single tool, or simple combination of different tools. This is the basic problem using standard control tools in complex manufacturing processes like foundries. Against this background, we present here a new prediction methodology for foundry process that, on the basis of several machine learning methods, is able to perform a comprehensive analysis as well as preparation of data before it is provided as input to prediction tools. This methodology is implemented as software EIDOminer. The weaker tools, that is, which cannot predict good enough quality characteristics of the existing production process, are switched off automatically by so-called Supervisor in IAM of EIDOminer software. The Supervisor determines, for a given process, which analysis method or combination of methods achieve best prediction results and provide a single optimal output result. After learning from the quality data of a product, the software can give prediction in a run time. The software is tested in two German foundry enterprises and the implemented approach shows significant results.