Enhanced CEEMDAN-Based Deep Hybrid Model for Automated Gear Crack Detection

Detecting faults in gearboxes by analyzing vibration signals is a complex task because the vibrations generated by different gears often overlap. This makes it particularly challenging to detect the gear defect when it occurs in one of the intermediate shaft wheels. Hence, the aim of this study is to develop an automated classification and identification system for different levels of gear crack severity.

To achieve this, we employ a multi-step approach. Initially, the acquired signal is pre-processed using an enhanced complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm. Subsequently, we extract a comprehensive set of features from the pre-processed signal, including time domain features, frequency domain features, and entropy-based features. To enhance the feature set, we perform data augmentation. Further, the augmented feature set is subjected to the hybrid classification model to classify the different levels of gear fault. This hybrid model combines bi-directional long-short-term memory (Bi-LSTM) and improved deep belief network (IDBN) classifiers. The scores generated by these classifiers are further processed through a score-level fusion method to produce a unified score.

The results indicate that the combination of Bi-LSTM and IDBN yielded a remarkable accuracy rate of 99.82% while also achieving this high level of accuracy with a minimal computational time of 0.7467. The results indicate that the proposed approach holds promise as a viable and accurate alternative for the precise detection of gear faults.