Continuous online flaws detection with photodiode signal and melt pool temperature based on deep learning in laser powder bed fusion.

• Setting the acquisition rate of the photodiode sensor can avoid photodiode signals aliasing. • Robust correlation can be established between photodiode signal and average melt pool temperature through the neural network. • Time series model is suitable for flaw detection based on the melt pool photodiode signal. The continuous online flaws detection of the melt pool plays a significant role in assessing the quality of the printed sample in metal additive manufacturing (AM). However, the continuous detection of melt pool flaws faces great challenges because the suitable continuous acquisition system and real-time processing algorithm for melt pool feature extraction are difficult to establish. In this work, a continuous online flaws detection method combining the photodiode signal and melt pool temperature based on deep learning algorithms in the Laser Powder Bed Fusion (LPBF) is proposed. A multi-signal fusion system is designed to synchronously obtain the photodiode signal and temperature signal of the melt pool. The characteristics of the melt pool photodiode signal under the action of pulsed laser and the characteristics of the melt pool temperature are explored. The aliasing of photodiode signals is found due to the different settings of pulse laser frequency and photodiode sensor acquisition rate in AM. According to the two principles that the melt pool temperature is related to the flaw formation, and accurate flaws identification can be carried out based on the melt pool temperature field, the Back Propagation Neural Network (BPNN), Stacked Sparse AutoEncoder (SSAE), and Long Short-Term Memory (LSTM) are used to build the correlation model between the photodiode signal and the average melt pool temperature, and the flaw detection is carried out through the average melt pool temperature error. Therefore, once the correlation model is established, continuous online flaws detection can be realized only by using photodiode signals. The results demonstrated that a robust correlation can be established between the photodiode signal and the average melt pool temperature through the neural network, and the correlation error can be as low as 2.2 %. The detection of flaws is simplified into a binary classification problem through a reasonable threshold setting, and the LSTM with 74.39 % detection accuracy is more suitable for flaws detection based on the photodiode signal of the melt pool. [ABSTRACT FROM AUTHOR]