Machine Learning Model for Predicting Production Process Capability in Packaging Process

One of the fundamental elements of qualitative analysis of the production process is the examination of process capability indices Cp and Cpk. Classical methods for evaluating these indices require clear process data, characterized by a normal distribution. However, in practice, data collected from real production processes are often incomplete and do not adhere to a normal distribution. Consequently, applying traditional methods for evaluating Cp and Cpk may lead to false conclusions. The situation becomes even more complicated when data do not meet the assumption of a normal distribution and/or when the process has a specified one-sided tolerance limit. In such cases, an important aspect, especially in processes without an upper tolerance limit, is the lower value of the process capability index Cpl. The article analyzes the impact of the type of silicone spacer on the variance of the heat-sealing process using a contact method, commonly applied in the food industry. To predict the value of Cpl for the strength of foil-paper packaging, classical machine learning models employing the random forest regression algorithm were developed. Due to the lack of practical possibilities to generate a larger number of samples, which is common in manufacturing practice, the range of training data was augmented according to its normal distribution characteristic. Prepared models were tested on real production data. If information about the used silicon spacer was present in training data, the model could be successfully used for example for predicting sealing process parameters.