Aluminum alloy (Al) is a commonly lightweight material, which has wide applications in the development of modern railways and automobiles. Friction stir lap welding (FSLW) is a commonly lightweight material connection technology, which avoids the defects produced by traditional fusion welding methods. The pin tool of FSLW is not replaced in time after wear will influence welding quality and waste resources. To reduce the unnecessary losses, this paper proposes a FSLW pin tool wear division model based on the convolutional neural network (CNN), bidirectional long-short term memory (BiLSTM) and Attention module. Firstly, the wear of the pin tool is divided into four levels based on the macro morphologies of the pin tool, welds surface appearance, internal defects and tensile strength. Then, the time–frequency characteristics of vibration signals are extracted by power spectral density (PSD) and wavelet packet transformation (WPT). The wear characteristics datasets of the pin tool are built by these features. More detailed information can be provided for the wear prediction of the pin tool. Finally, the CNN-BiLSTM-Attention network is established to classify the pin tool wear levels. CNN and BiLSTM are used to extract effective information of wear characteristics in spatial and temporal directions, respectively. Attention mechanism is used to distribute different weights to choose critical features and then enhance the role of important features in the calculation process. The wear features datasets are normalized and input into the classification model. The classification result shows that the average accuracy of the CNN-BiLSTM-Attention model reaches 95.8 %, which is 5.7 % higher than that of the CNN-BiLSTM model. Compared with the mature classification model, it has higher accuracy and anti-noise ability. Therefore, a FSLW pin tool wear division model is established based on the CNN-BiLSTM-Attention neural network, which the pin tool wear information during FSLW could be predicted more detailed, and the pin tool is managed more efficiently.
Read full abstract