Beamforming for condition monitoring

Acoustic condition monitoring is an important technology for non-invasive condition monitoring, but using robust array signal processing techniques combined with machine learning is still not a fully explored field. This thesis examines how different microphone arrays and beamforming algorithms affect features commonly used in audio classification. We look at three separate arrays, one uniform rectangular array, and two uniform circular arrays with different diameters and microphone elements. Additionally, we are taking a look at the conventional delay and sum (DAS) beamformer, as well as two adaptive beamformers; a broadband minimum variance distortionless response (MVDR) beamformer using discrete Fourier transform (DFT), and the generalized sidelobe canceller (GSC) using the normalized least means square (NLMS) approach. We show that the size of the array and the number of microphones have a significant impact on its performance regarding directivity and white noise gain. The position of the microphone elements also plays an important role to avoid spatial aliasing. The larger circular array outperforms both of the other arrays, but the smaller arrays have the upside of being easily maneuverable and can be placed in smaller, confined places. The rectangular array has some spatial aliasing for the higher frequencies in the audible range, but as long as it is possible to ensure that the frequency range of our source signal is band-limited to its non-aliasing spectrum, this performs significantly better than the smaller circular array. The three beamformers all perform well when steering is correct, but the adaptive beamformers quickly attenuate the source signal when there is a steering mismatch. The NLMS based GSC beamformer would likely cause problems for a batch based machine learning algorithm due to the continuous updating of its weights. The DFT based MVDR beamformer has issues with phase-shift error if the number of buffered samples is too low, but as long as a sufficient number of samples is used, it performs better than the GSC in every scenario tested. This implies that the MVDR is the preferred beamformer of the two if high directivity is needed. DAS is the safer choice of the beamformers, as there are no parameters to tune for a given environment, and performs well in every tested setting