Gearbox Fault Diagnosis Using Stacked Multisensor Image Fusion and Self-Normalizing Convolutional Neural Networks

The accurate diagnosis of gearbox faults in industrial machinery is essential for ensuring operational efficiency and preventing costly downtime. This paper introduces a novel diagnostic framework that employs advanced signal processing and deep learning techniques to elevate fault detection accuracy and reliability in gearbox systems. By leveraging stacked Synchrosqueezing Wavelet Transforms (SSWT) for multi-sensor image fusion, our approach begins with transforming raw vibration signals from multiple sensors into high-resolution Time-Frequency (TF) representations, essential for detecting subtle mechanical anomalies. These images are then processed using Self-Normalizing Convolutional Neural Networks (SCNNs), which exploit their inherent normalization capabilities to maintain consistency across network layers and mitigate the vanishing gradient problem. The spatial features extracted are further analyzed through Gated Recurrent Units (GRUs), adept at modeling long-term dependencies and capturing fault progression dynamics. Enhancement of this process with a Multihead Attention layer allows selective focusing on salient features, significantly improving diagnostic precision. Validated on two public gearbox datasets, our method achieved 100% accuracy on both, advancing gearbox fault diagnostics and setting a new benchmark for industrial fault detection methods.