Experimental validation on the Harbin Institute of Technology aero-engine inter-shaft bearing dataset shows that the proposed model achieves 97% diagnostic accuracy under extreme noise conditions (SNR = -5 dB).
Abstract
To address the challenges of difficult fault feature extraction and feature aliasing in aero-engine inter-shaft bearings under strong noise conditions, this paper proposes a fusion diagnostic method that integrates a dual-scale one-dimensional convolutional neural network, a multi-head self-attention Transformer, and a bidirectional gated recurrent unit. The method employs a three-stage progressive network architecture for end-to-end fault diagnosis. The dual-scale 1DCNN extracts local temporal features from vibration signals, and batch normalization and dropout are applied to stabilize training and reduce potential overfitting. The Transformer encoder models dependencies among the extracted feature representations, supporting the representation of fault-sensitive features. The BiGRU captures bidirectional temporal dependencies in the fault evolution process. Experimental validation on the Harbin Institute of Technology aero-engine inter-shaft bearing dataset shows that the proposed model achieves 97% diagnostic accuracy under extreme noise conditions (SNR = -5 dB). Compared with existing methods, these results indicate that the proposed network effectively maintains diagnostic performance under controlled noise conditions.
To improve the fault diagnosis accuracy of marine propulsion shaft bearings under variable working conditions and with limited samples, a Multi-scale Deep Multi-source Subdomain Adaptation Network (MDMSAN) is proposed. First, DCGAN is employed to augment source domain samples. A shared feature extraction network integrating multi-scale convolutions, LSTM, and ECA mechanism is constructed to extract domain-invariant features. Private feature extractors are designed for each source-target domain pair, utilizing LMMD for fine-grained subdomain alignment. A transferability-aware weighted fusion strategy integrates multi-source classification results. Experiments on CWRU and PT500 datasets demonstrate that MDMSAN achieves average accuracies of 99.92% and 99.73% under variable working conditions, and maintains 98.85% accuracy with only 1/16 of the target samples, significantly outperforming existing methods.
Zhipeng Wang, Wei Yuan· Advances in Engineering Tech...· 0 citations
A multi-scale linear attention multi-source subdomain adaptation network (MLAMSAN) that integrates the multi-scale linear attention (MLA) that can achieve fault diagnosis under cross operating conditions through subdomain feature alignments that exhibits the superior diagnostic performance and the strong generalization ability.
Zheng Han, Yuqi Fan, Yaping Wang et al.· Engineering Research Express· 0 citations
This smart fault diagnosis method based on the Time Convolution Network - Bidirectional Gated Recurrent Unit - Attention Model (TCN-BiGRU-Attention) can achieve high-precision and stable fault diagnosis for rolling bearings, providing an effective intelligent diagnosis solution for engineering applications.
Mingli Li, Zhu Yuan· Frontiers of Mechanical Engi...· 0 citations
The validation results on multiple typical bearing fault datasets show that the proposed MorletConv CNN model is characterised by enhanced physical interpretability and generalisation ability while maintaining high diagnostic accuracy, providing new ideas and method support for achieving highly reliable rolling bearing fault diagnosis.
Taoyang Zhan, Kang Han, Yuhan Huang et al.· Insight - Non-Destructive Te...· 0 citations
Multi-sensor data fusion is widely recognized as a key enabler for reliable aero-engine condition monitoring under complex operating conditions. However, this paradigm inherently increases system complexity. This paper conducts a systematic empirical study to evaluate the practical efficacy and boundaries of deep feature mining using only a single channel. We propose a Local-Global State Space Model Network (LG-SSMNet), which uses multi-scale 1D CNNs to fit local transient pulse responses and employs a Selective State Space Model (Mamba) to construct a long-term periodic integrator, achieving feature decoupling within a single channel. To address the discretization divergence issue of continuous vibration signals, an empirical stabilization configuration involving dual normalization and stratified learning rates is introduced. Deep quantitative analysis on the HIT aero-engine bearing dataset shows that the four acceleration channels achieve stable accuracies over 99.6% (approaching 100%), while displacement channels show significant divergence. The underperforming channel (Sensor 1) reached only 91.31% due to a critical missed detection rate of outer race faults (4.3% misclassified as healthy). Through joint empirical evidence using ROC-AUC, F1-scores, and qualitative feature visualization, we demonstrate that with reasonably selected acceleration sensors, single-channel deep mining can serve as a practical alternative to multi-channel fusion.
Lei Zhang, Z. Ren· 2026 IEEE International Conf...· 0 citations
Abstract. Timely fault diagnosis in turbomachinery plays an important role in avoiding disastrous failures, reducing the number of unexpected downtimes, and optimizing the maintenance cycle in industrial applications that use a lot of energy. The paper suggests a new hybrid deep learning architecture, combining one-dimensional convolutional neural networks (1D-CNN) with bidirectional long short-term memory (BiLSTM) networks to the problem of automatic detection and early forecasting of mechanical faults based on raw vibration signals. The combination of a multi-domain feature extraction strategy, which covers time-domain statistical indicators, spectral descriptors based on the fast fourier transform (FFT), and continuous wavelet transform (CWT) scalogram representations into a single 128-dimensional feature vector is used as the input of the model. The suggested CNN-BiLSTM design effectively extracts both local spectrotemporal features and long-range sequential dynamics of the rotating machinery vibration data. Experiments on the CWRU Bearing Dataset and a custom turbomachine testbed dataset show that the proposed model can achieve a mean classification accuracy of 97.6% with clean signal conditions and 95.1% at a signal-to-noise ratio of 10 dB, and outperforms standalone CNN, BiLSTM, SVM, ANN, and Random Forest baselines by up to 10.2 The findings determine the appropriateness of the proposed framework on edge-deployable and real-time condition monitoring systems.
N. Bharani· Materials Research Proceedin...· 0 citations