Data Efficient Deep Learning for Rolling Element Bearing Fault Diagnosis

Abstract

Roller element bearings are responsible for 50 to 60 percent of rotating machinery failures, making reliable fault diagnosis critical for industrial machinery health monitoring. This thesis demonstrates that state of the art performance in bearing fault diagnosis does not require massive datasets, but rather well-designed data collection, preprocessing, and loading methods. Focusing on the entire data lifecycle from acquisition to model input, this work challenges the notion that extensive data is required to train machine learning (ML) models for condition monitoring by showing that carefully structured and efficiently loaded data can deliver superior cross domain generalization even with limited samples. Benchmark publicly available bearing datasets are analyzed, revealing weak cross domain coverage that restrict the transferability of ML models to industrial scenarios. To address this gap, a new bearing dataset design is proposed under controlled conditions with naturally developed faults, and a novel selective embedding data loading method is introduced, which alternates sensor inputs in a structured and alternating fashion to maximize training diversity and prevent overfitting without increasing data volume. Theoretical foundations connecting selective embedding to gradient variance reduction, effective sample size, and generalization bounds are established to explain why the method works. Cross domain experiments are conducted across heavy machinery, manufacturing, and railway datasets to confirm that selective embedding can achieve over 90% classification accuracy under challenging domain splits, outperforming both single-channel and parallel loading methods that require significantly more data or computational effort. These findings establish that data efficiency, not data size, is the key to scalable, transferable, and computationally affordable deep learning for industrial bearing fault diagnosis.