Edwin Lughofer, Kurt Pichler,
"Data-driven prediction of possible quality deterioration in injection molding processes"
, in Applied Soft Computing, Vol. 150, Elvesier, 1-2024, ISSN: 1568-4946
Original Titel:
Data-driven prediction of possible quality deterioration in injection molding processes
Sprache des Titels:
Englisch
Original Kurzfassung:
We propose an approach for the automated prediction of possible quality deteriorations at injection molding machines using data-driven models. This approach relies on data solely recorded during the regular production phase without the need to (i.) collect data from anomalous phases, which may be cost-intensive due to production waste, (ii.) perform time-intensive (manual) annotation cycles for assigning class labels, and (iii.) know specific fault modes in advance. Our approach embeds two main concepts: the first is the establishment of causal relations between process variables, which serve as online residual generators. The multi-variate residual signals are analyzed using an advanced independent component analysis, whose reconstruction characteristics serve as a control signal indicating a potential anomaly of any type/mode, upon violation of the internal dependence structure. No parameters are needed for the online analysis, which in turn avoids the use of (pre-collected) data from anomalous phases, which is typically required to tune the parameters properly. The second concept performs a direct prediction of quality criteria, which are permanently by-measured through the usage of an optical inspection system of the produced/molded parts. Time series-based trends of the process data are used to establish quality prediction models based on a specific time series transformation technique. It preserves the local structure of the data to ensure it can be effectively combined with predictive fuzzy systems training. To account for system dynamics, the prior trained prediction models can be automatically updated on the fly with new measurements. To further increase flexibility in the case of higher dynamics, as was the case in our application, a dynamic lazy learning (dLL) approach was developed in combination with partial least squares. This technique is able to permanently update the reference data base and establish a local model per new query sample with reduced input dimensionality. Our approach was successfully evaluated during the real injection molding processes of bottle caps: several real-occurring anomalies and significant changes in material and production parameters were successfully found over a horizon of 7162 production cycles. Furthermore, three essential quality criteria were predicted with a correlation of 0.7, 0.89 and 0.98 between predicted and observed trend lines, with mean absolute error rates below 5%. This makes the detection of quality deteriorations possible at an early stage.