About the project

This project investigated the possibility of determining the wear of a grinding belt and the presence of a fault in a belt grinder via acoustic measurements using AI models and mathematical methods. The aim was to enable a quicker intervention in the event of a fault and optimize the replacement intervals of the grinding belts.

Initially, various grinding processes with different parameters such as cutting speed, different grit sizes, processing materials and belt wear were repeated several times. The grinding processes were recorded using a microphone on the grinding machine. The grinding recordings were processed to make them suitable for use in a machine learning project. The results show that Mel Frequency Cepstral Coefficients (MFCC) would be more suitable for this problem than LFCC, IMFCC, Mel spectrograms and spectrograms.

The research found that autoencoders, a special form of artificial neural network, can learn and reconstruct the grinding noise. In the event of an error or increased belt wear, the reconstruction error of the networks increases, which can be used as an indicator of anomalies and to determine wear levels.

Artificial neural networks have also been tested as classification models. In addition to the wear level, these can detect other attributes from the grinding process, such as grit size or cutting speed.

Alongside artificial neural networks, mathematical methods such as principal component analysis (PCA) also make it possible to determine belt wear. Visualizing the PCA results also provides a better understanding of the problem to be solved and the grinding process itself.