Multitask Learning Approach for Bolt Torque Loss Detection Using Variable-Frequency Hammering Cepstral Coefficients

Abstract

Bolt-size-dependent variations in hammering responses tend to affect the classification of acoustic bolt torque loss, thereby making direct transfer from source bolts to a different target bolt size difficult. To address this drawback, this study proposes a variable-frequency hammering cepstral coefficients (VFHCC)-based domain-balanced multitask learning long short-term memory (MTL-LSTM) framework for cross-size bolt torque-loss classification with limited target-bolt calibration. First, hammering cepstral coefficients (HCCs) were introduced as an alternative to conventional Mel-frequency cepstral coefficients (MFCCs). The HCC parameters were optimized using a particle swarm optimization support vector machine (PSO-SVM). Second, variable-frequency hammering was applied to enrich the acoustic representation prior to cepstral feature extraction. Third, a domain-balanced MTL-LSTM model was developed to combine M20/M28 source-bolt data with limited M24 calibration data. The dataset contained 1,650 one-second WAV recordings from M20, M24, and M28 bolts under 11 torque-loss classes. Under the strict source-only protocol, direct transfer from M20/M28 to M24 produced low classification performance, confirming a strong cross-size acoustic domain shift. Under the target-calibrated protocol, VFHCC substantially improved M24 classification compared with single-frequency features. With two M24 calibration samples per class, the proposed method achieved an accuracy of 0.9879 and a macro-F1 score of 0.9875 on the remaining M24 test samples. These results demonstrate that variable-frequency acoustic representation and domain-balanced learning provide an effective strategy for cross-size bolt torque-loss classification when limited target-bolt calibration data are available.