Deep learning-based tensile strength prediction in fused deposition modeling

As one of the transformative technologies, Additive Manufacturing (AM) has been facing the challenge of product property inconsistency, which prevents its adoption in various critical applications. Various studies have been dedicated to building the predictive models for improved AM quality control. However, the distinct layer-by-layer printing process and the related cyclic layer thermal history present significant difficulty in model accuracy and reliability. This paper presents a data-driven predictive model by taking into account the printing process in a deep learning network structure, using Fused Deposition Modeling (FDM) as a representative case study. Temperature and vibration data are measured to reveal the layer-wise thermal and mechanical activities as well as the process variation, and the inter-relationship among different printing layers is characterized by a Long Short-term Memory (LSTM) network. The sensing data, combined with process parameters and material property, are then fused for predicting the tensile strength of manufactured parts. To quantify the influence of each process parameters on the prediction result, a modified Layer-wise Relevance Propagation (LRP) is investigated. Experimental evaluation has shown that the LSTM-based predictive model outperforms several machine learning techniques, such as Support Vector Regression and Random Forest. The relevance distribution revealed by LRP is found useful for interpreting the prediction result.