Explanation of the Acoustic Features for Detecting a Cut Interruption in the Laser Cutting Process

The machine learning (ML) algorithm RandOm Convolutional KErnel Transform (ROCKET) is used to recognise a cut interruption during laser cutting to avoid rejects. For this purpose, an audio signal recorded at the laser cutting machine is categorised with help of ROCKET into three classes. These are: Good cut, transition region and cut interruption. However, when using ML algorithms, the user is not given any insight into the algorithm's decisions. It is not possible to develop an understanding of the process, i.e. information about what indicates a cut interruption in the audio signal. For this reason, ML algorithms are often referred to as black box models. In this paper, reROCKET is introduced to make the ROCKET models more transparent. The randomly generated ROCKET kernels are analysed by reverse engineering. The effectiveness of the kernels is determined by calculating the quality of class separation by applying a kernel. The best kernels are then identified and provide information about the features in the audio signal. The performance of a ROCKET model with 500 kernels and a conventional kernel approach is compared. With a single kernel transformation, results close to the prediction of a 500-kernel ROCKET model are obtained. This is particularly advantageous in real-time applications, as the computing time of a convolution is significantly less than that of the 500 kernels.