Mechanical performance prediction of thermoplastic filament wound composite based on virtual sample generation and machine learning

Compared with traditional thermoset carbon fiber composites, thermoplastic composites possess excellent performance, such as high thermal resistance and recyclability. During the laser-assisted in-situ filament winding process of thermoplastic composites, process parameters including laser power, winding speed, mandrel temperature, and winding tension have significant impacts on the mechanical properties of the final parts. To accurately predict the tensile strength of NOL(Navy Ordnance Laboratory) specimens under different in-situ filament winding process conditions while reducing the high experimental costs, this paper proposes a prediction modeling integrating virtual sample generation techniques with machine learning methods. First, a thermoplastic composite winding experimental platform is constructed, and orthogonal experiments are conducted to obtain the original data. Then, the Trend Similarity Assessment (TSA) method is combined with a Backpropagation (BP) neural network to predict the tensile mechanical properties of thermoplastic composite components and verify the model’s accuracy. Finally, the influence of the ratio between original samples and virtual samples on the prediction accuracy of tensile strength is explored. Results shown that the combination of TSA and BP neural networks can improve the prediction accuracy and solve the small sample problem effectively. Moreover, the increasing of original and virtual samples can enhance both the prediction precision and fitting ability of the model.