Wavelet Transformer Hybrid Model for Real Time Industrial Fault Detection

Abstract

The fault detection of reliable industrial acoustic signal processing is a challenging problem because non-stationary noise, complicated vibration patterns, and the flaws of traditional signal analysis methods make the detection of harmful noise difficult. The conventional machine learning and classical signal processing methods can be very weak at capturing both the local time-frequency features and the global temporal patterns in acoustic fault signals. Additionally, the current deep learning models are either spectral feature-based or time-dependent models, which leads to poor diagnostic quality and low resilience to changing industrial settings. To fill this research gap, this paper suggests Adaptive Wavelet -Transformer Hybrid Model (AWTHM) in detecting faults in industrial machines on real time. The framework proposed is a combination of discrete wavelet transform (DWT) used to extract multi resolution time-frequency features and a lightweight transformer architecture to learn long-range associations of acoustic signals. Adaptive attention mechanism is also included to improve discriminative feature learning and noise artifact rejection. The model is tested on a benchmark industrial acoustic fault dataset of various machine operating conditions as well as fault categories. The experimental validation is done in a Python based deep learning environment with cross-validation and comparative benchmarking with conventional CNN, LSTM and SVM classifiers. The proposed AWTHM is classified with accuracy of 97.8%, precision of 97.2%, recall of 96.9% and the ability of the method to detect faults 22% better than the existing techniques. The key value of this work is that it succeeds in integrating effectively wavelet-based signal decomposition and transformer-based temporal learning to achieve robust and scalable acoustic fault diagnosis.