Artificial Intelligence Based Foundry Process Optimization Technology

The arrival of Industry 4.0 has stimulated a significant increase in the digitalization of foundry plants, driven by the improved data-gathering capabilities of machinery. This data is useful to improve the casting process. Casting involves a variety of different factors, such as the metal composition, mold design, and casting conditions, which can influence the final outcome.

Furthermore, casting processes encompass a variety of techniques, including sand casting, investment casting, die casting, and centrifugal casting. Between them, investment casting, also known as lost wax casting, stands out because it has the ability to produce complex and detailed parts. It is widely used in industries such as aerospace, automotive, and medical, where tight tolerances are required. Nevertheless, it can be a time-consuming and expensive process where the state of the parts is unknown until the process is finished. Hence, if there are any defects in the casting, there would be a waste of raw materials and time.

Against this background, we propose a new approach using Artificial Intelligence (AI), where we establish the optimal parametrization of the process to mitigate the occurrence of defects. Our AI workflow is organized into six main modules: I. Data gathering; II. Data preprocessing; III. Model selection; IV. Production dataset building; V. Model prediction; VI. Parametrization selection.

Firstly, data gathering involves collecting machinery parameters from the foundry and associating them with the presence of defects in the casting, the target class.

Secondly, data preprocessing entails segregating the data corresponding to different phases of the production process and normalizing the input. At this stage, the different ranges within which the variables obtained can fluctuate have been determined by applying expert knowledge.

Thirdly, in the model selection process, 24 types of the most common classifiers, such as Naïve Bayes, Random Forest and KNN, are evaluated using 10-fold cross-validation to determine the model that surpasses 70% accuracy with the highest recall. This approach aims to minimize Type 2 errors, which occur when positive classes (defects in casting), are incorrectly classified as negative (no defects in casting).

Then, we build the production dataset from the possible ranges of normalized variables. This dataset includes all possible scenarios of the process, and we get this combination by applying the Cartesian product to each attribute set. The length of the production dataset will depend on the number of attributes and the number of elements that have those attributes.

Later, we perform the model prediction which consists of determining the degree of confidence of the binary target class, in this case, if there are defects or not, and then getting the one with the highest value.

Last, but not the least, we get the parametrization selection by filtering the target class with no defects and ordering by the highest degree of confidence. With this, we get the ranges where there are fewer defects.

Overall, all the models in different phases of the production process got an accuracy over 70%, therefore performing better than humans. In conclusion, with the implementation of our proposed AI workflow, we have demonstrated the capability to enhance the service and productivity of investment casting.