Early Defect Detection in Thermo-Stamping Process Using Set Encoding for Nonlinear Dynamics Identification
Topic(s) : Manufacturing
Co-authors :
Hamza TARIN (FRANCE), Sebastien COMAS-CARDONA (FRANCE), Domenico BORZACCHIELLO , Philippe LE BOT (FRANCE), Yves LE GUENNEC (FRANCE)Abstract :
Early defect detection plays a crucial role in the thermo-stamping process, for economical (decrease of non-destructive tests costs for instance) but also for technical (optimization of process set-up) reasons.
Deep-learning methods have witnessed remarkable scientific interest. These techniques, referred to as "black boxes" are applied to many issues in industry 4.0, such as enabling prognostics for anticipating failures and optimizing production efficiency. Despite their effectiveness, these methods have limitations in terms of providing qualitative interpretability and ensuring accurate predictions beyond the available data, alongside with a significant dependency on the quality and quantity of data.
In this work, we present an online adaptative model based on physics informed machine learning techniques [1] and sparse identification of nonlinear dynamics such as SINDy [2]. These type of models integrate empirical data, symmetry constraints, and physical or intuitively plausible constraints during the training process. Our model relies also on set encoding [3, 4] architecture for sequence data (time series for instance), enabling accurate and online adaptive model identification for dynamic systems with variable input data sizes.
By incorporating diverse data types and leveraging known physics we first train the model offline. It will be then utilized for real-time prediction of material properties, and extrapolation of state variables throughout the various process phases, enabling effective and early defect detection.
Deep-learning methods have witnessed remarkable scientific interest. These techniques, referred to as "black boxes" are applied to many issues in industry 4.0, such as enabling prognostics for anticipating failures and optimizing production efficiency. Despite their effectiveness, these methods have limitations in terms of providing qualitative interpretability and ensuring accurate predictions beyond the available data, alongside with a significant dependency on the quality and quantity of data.
In this work, we present an online adaptative model based on physics informed machine learning techniques [1] and sparse identification of nonlinear dynamics such as SINDy [2]. These type of models integrate empirical data, symmetry constraints, and physical or intuitively plausible constraints during the training process. Our model relies also on set encoding [3, 4] architecture for sequence data (time series for instance), enabling accurate and online adaptive model identification for dynamic systems with variable input data sizes.
By incorporating diverse data types and leveraging known physics we first train the model offline. It will be then utilized for real-time prediction of material properties, and extrapolation of state variables throughout the various process phases, enabling effective and early defect detection.