Audio feature augmentation for bolt looseness classification in data-deficient scenarios using machine learning

Existing machine learning (ML) models have limitations in classifying bolt looseness in data-deficient scenarios. In this article, a novel audio-feature augmentation was established to detect percussion-based bolt looseness in data-deficient scenarios using ML approaches. An n-time augmentation was performed on the important features extracted from the percussion audio signals using correlation and eigenvalue analyses. The accuracy and generalisation ability of bolt looseness prediction of (i) k-nearest neighbour, (ii) multi-layer perceptron, (iii) multinomial logistic regression, (iv) random forest, and (v) hybrid (ML) methods were demonstrated using statistical metrics. One of the key observations from the present study is that the application of eight-fold cross-validation on the hybrid model trained with 20-time feature augmentation demonstrated the best overall scenario in enhancing bolt looseness prediction. It exhibited a maximum increase in accuracy of 18.0% in identifying bolt looseness levels in data-deficient scenarios compared to without feature augmentation. Further, the eight-fold cross-validation analysis indicated an accuracy fluctuation drop from 33.4% to 5.6%, indicating consistent prediction when trained with 20-time feature augmentation as compared to without feature augmentation. Overall, the present investigations manifested the verifiable potential of the audio feature augmentation approach to enhance bolt looseness prediction in data-deficient scenarios.