Monitoring of Product Conformity Based on Machine Learning Prediction Evaluation

In Predictive Maintenance (PM) of manufacturing industries, the main objective is to improve production time and increase the quantity of produced parts. However, product quality management is less addressed. Indeed, each non-conform part represents a direct loss, underlining the importance of product conformity assessment. This research provides a new methodology for assessing the impact of systemic parameters, such as the characteristics of raw materials and the state of machines, on the quality of the final product. The methodology begins with the collection and preparation of industrial data which, due to their specific nature, require detailed processing to meet the required quality standards. We then analyze and compare three distinct neural architectures: Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and an LSTM-GRU. These models were deployed to compare their capacity to predict the global quality of the part directly, compared to an indirect prediction which requires the determination of the conformity of each dimension of the part. The implementation of this methodology was carried out on a real dataset, provided by the French company SCODER, which reveals promising prospects for the PM sector. The results highlight the effectiveness of using the LSTM model to predict 3D measurements and improve quality management in manufacturing production processes.