No abstract

Aim: This work proposes a workflow monitoring sensor observations over time to identify and predict relevant changes or anomalies in the cure cycle (CC) industrial process. CC is a procedure developed in an autoclave consisting of applying high temperatures to provide composite materials. Knowing anomalies in advance could improve efficiency and avoid product discard due to poor quality, benefiting sustainability and the environment. Methods: The proposed workflow exploits machine learning techniques for monitoring and early validating the CC process according to the time-temperature constraints in a real industrial case study. It uses CC's data produced by the thermocouples in the autoclave along the cycle to train an LSTM model. Fast Low-cost Online Semantic Segmentation algorithm is used for better characterizing the time series of temperature. The final objective is predicting future temperatures minute by minute to forecast if the cure will satisfy the constraints of quality control or raise the alerts for eventually recovering the process. Results: Experimentation, conducted on 142 time series (of 550 measurements, on average), shows that the framework identifies invalid CCs with significant precision and recall values after the first 2 hours of the process. Conclusion: By acting as an early-alerting system for the quality control office, the proposal aims to reduce defect rates and resource usage, bringing positive environmental impacts. Moreover, the framework could be adapted to other manufacturing targets by adopting specific datasets and tuning thresholds.

No abstract