Multimodal 1D CNN for delamination prediction in CFRP drilling process with industrial robots

There is a growing demand for carbon fiber-reinforced plastics (CFRPs) in the aerospace and automotive industries. Consequently, the assembly and repair of CFRP components has garnered considerable attention. However, at industrial sites, there are potential difficulties in the CFRP drilling process. These challenges arise from the anisotropic properties of CFRP and the limitations of using an optical microscope to assess the quality of millions of drilled holes. Therefore, this study introduced an advanced indirect prediction method for the CFRP hole quality based on multisensor data. During the drilling process, data including force, torque, acceleration, voltage, current, sound, and images of the hole exits were acquired via a robotic machining system. The delamination factors, Fd and Fa, which quantify the quality of the hole, can be calculated from the hole images. Preprocessing was employed to segment the sensor data into discrete drilling trials and extract spectral features from the data. This paper proposed a multimodal one-dimensional convolutional neural network (1D CNN) that predicts delamination factors from time series multi-sensor data. A case study, using a test set of 100 trials, validated the proposed model. The performance of the proposed model was evaluated by the mean squared error (MSE) and inference time, yielding 8.42 ± 0.167 × 10−2 and 3.48 ± 0.192 s for Fd, and 6.1 ± 0.729 × 10−2 and 3.13 ± 0.098 s for Fa, respectively, across the entire test set with 100 trials. This result exceeds those of previous studies in terms of both the MSE value and inference time for multimodality, emphasizing the predictive accuracy and real-time operational capabilities of the proposed model in industrial settings. Furthermore, this approach provides practicality and flexibility, facilitating the easy integration or removal of sensors through data-driven preprocessing and a multimodal learning scheme.