Project: Data Driven Engineering

Up to 10 % energy cost savings at launch

Project: Data Driven Engineering - Up to 10 % energy cost savings at launch | MDESIGN Consulting Services

Problem definition

In the field of casting production, the final quality of the manufactured products is subject to a multitude of process and machine parameters. Contrary to what some people think, massive components are particularly sensitive to different influencing variables during the manufacturing process. In these cases, it is not only the attempt to define the decisive process variables in order to achieve a high-quality end product that is problematic, but also the mostly missing test possibilities to carry out a final 100% test for quality control.

In the manufacturing chains, not only measurable variables such as temperatures, cooling times, vibrations, ambient conditions or machine parameters must be taken into account. A decisive factor is the experience of the machine operators, who actively react to deviating ambient conditions or visible faults within the manufacturing process.
The digitalization of these procedures offers a future-proof perspective, whereby the two main areas of dynamic measured variables and the subjective assessment of the employees must not be considered strictly separately.



The acquired experience in the area of casting production can neither be formulated by the machine operators in the classical sense, nor can it be concretely programmed by defined rules. This complicates the digitization goal for automated quality optimization and inspection. By using machine learning methods, existing data sources are to be used to reformulate company knowledge into applicable structures.
By using these methods, an enormous amount of data on process parameters can be evaluated simultaneously and used as a support with predictions on possible sources of error. In this way, the experts with many years of experience in handling the machines are not forced to monitor hundreds of parameters, but are able to focus on the essential parameters in order to intervene in the processes in a professional, targeted manner.

This measure is intended to reduce production errors caused, for example, by cooling processes and times that are partially not adapted to the structure, belt speeds or the chemical composition of the base material. These influences are to be detected as early as possible during production in order to minimize follow-up costs due to components that do not meet quality standards.



The implementation and integration of the assisting MachineLearning algorithms took place promptly after completion of the technical fundamentals and learned methods. The dynamics of the algorithms through the further learning process under the ongoing production line brings future security and more accurate prediction prospects in the subsequent years of further product lines of cast components.
Already at the time of launch using comparatively low training data, good predictions regarding the expected manufacturing quality can be achieved. By means of these analyses, especially in the area of series part production, individual parts with a high probability of defects can already be specifically examined or completely removed from the overall production chain. With the saving of subsequent production steps on potentially unusable parts, a large part of the expenses for the realization of the MachineLearning project could be covered in the short term.

While this level of knowledge and development significantly facilitates the supervision and monitoring of the process, a more advanced objective will be achieved as the production series progress and the training content of the ML method increases. Here, using series-capable measures, e.g. sound examinations or photo recordings, a non-destructive 100% inspection of the end products can be carried out with the help of the MachineLearning algorithms.
The decisive factor for success is the use of the competence of the experts to assess the end products, who are now in a position to intensively examine selected components instead of making a statement about the series product by random sample analysis.