Acoustic modelling, data augmentation and feature extraction for in-pipe machine learning applications

Chiantello, Dario Alfredo (2023) Acoustic modelling, data augmentation and feature extraction for in-pipe machine learning applications. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (26MB) | Preview

Abstract

Gathering measurements from infrastructure, private premises, and harsh environments can be difficult and expensive. From this perspective, the development of new machine learning algorithms is strongly affected by the availability of training and test data. We focus on audio archives for in-pipe events. Although several examples of pipe-related applications can be found in the literature, datasets of audio/vibration recordings are much scarcer, and the only references found relate to leakage detection and characterisation. Therefore, this work proposes a methodology to relieve the burden of data collection for acoustic events in deployed pipes. The aim is to maximise the yield of small sets of real recordings and demonstrate how to extract effective features for machine learning. The methodology developed requires the preliminary creation of a soundbank of audio samples gathered with simple weak annotations. For practical reasons, the case study is given by a range of appliances, fittings, and fixtures connected to pipes in domestic environments. The source recordings are low-reverberated audio signals enhanced through a bespoke spectral filter and containing the desired audio fingerprints. The soundbank is then processed to create an arbitrary number of synthetic augmented observations. The data augmentation improves the quality and the quantity of the metadata and automatically creates strong and accurate annotations that are both machine and human-readable. Besides, the implemented processing chain allows precise control of properties such as signal-to-noise ratio, duration of the events, and the number of overlapping events. The inter-class variability is expanded by recombining source audio blocks and adding simulated artificial reverberation obtained through an acoustic model developed for the purpose. Finally, the dataset is synthesised to guarantee separability and balance. A few signal representations are optimised to maximise the classification performance, and the results are reported as a benchmark for future developments. The contribution to the existing knowledge concerns several aspects of the processing chain implemented. A novel quasi-analytic acoustic model is introduced to simulate in-pipe reverberations, adopting a three-layer architecture particularly convenient for batch processing. The first layer includes two algorithms: one for the numerical calculation of the axial wavenumbers and one for the separation of the modes. The latter, in particular, provides a workaround for a problem not explicitly treated in the literature and related to the modal non-orthogonality given by the solid-liquid interface in the analysed domain. A set of results for different waveguides is reported to compare the dispersive behaviour against different mechanical configurations. Two more novel solutions are also included in the second layer of the model and concern the integration of the acoustic sources. Specifically, the amplitudes of the non-orthogonal modal potentials are obtained using either a distance minimisation objective function or by solving an analytical decoupling problem. In both cases, results show that sources sufficiently smooth can be approximated with a limited number of modes keeping the error below 1%. The last layer proposes a bespoke approach for the integration of the acoustic model into the synthesiser as a reverberation simulator. Additional elements of novelty relate to the other blocks of the audio synthesiser. The statistical spectral filter, for instance, is a batch-processing solution for the attenuation of the background noise of the source recordings. The signal-to-noise ratio analysis for both moderate and high noise levels indicates a clear improvement of several decibels against the closest filter example in the literature. The recombination of the audio blocks and the system of fully tracked annotations are also novel extensions of similar approaches recently adopted in other contexts. Moreover, a bespoke synthesis strategy is proposed to guarantee separable and balanced datasets. The last contribution concerns the extraction of convenient sets of audio features. Elements of novelty are introduced for the optimisation of the filter banks of the mel-frequency cepstral coefficients and the scattering wavelet transform. In particular, compared to the respective standard definitions, the average F-score performance of the optimised features is roughly 6% higher in the first case and 2.5% higher for the latter. Finally, the soundbank, the synthetic dataset, and the fundamental blocks of the software library developed are publicly available for further research.