A Deep Learning-based GAN-LSTM Framework for Friction Failure Detection and Acoustic Emission Signal Enhancement in Machine Engines

PeerJ Computer Science  ·  2025

Guojun Bai; Yang Yu; Yue Zhang;

HIGHLIGHT

State-of-the-art convolutional neural network (CNN) and long short-term memory (LSTM)-based models do not accurately improve the quality of signals, resulting in misclassifications and decreased diagnostic accuracy. To overcome the above limitations, this article introduces a new generative adversarial network (GAN)-LSTM-based framework combining Wasserstein generative adversarial networks with gradient penalty (WGAN-GP) for AE signal improvement and LSTM networks for sequential fault classification.

RELATED

Papers      Patents     Grants     Venues     Experts    Cite    Share    Save

ABSTRACT

Acoustic emission (AE) signal-based fault detection has become a crucial predictive maintenance and industrial monitoring technique, which allows for the early diagnosis of failures in machines. Current fault classification methods are plagued by high noise disturbance, weak generalization, and poor temporal dependency modeling of AE signals. State-of-the-art convolutional neural network (CNN) and long short-term memory (LSTM)-based models do not accurately improve the quality of signals, resulting in misclassifications and decreased diagnostic accuracy. To overcome the above limitations, this article introduces a new generative adversarial network (GAN)-LSTM-based framework combining Wasserstein generative adversarial networks with gradient penalty (WGAN-GP) for AE signal improvement and LSTM networks for sequential fault classification. The suggested methodology employs a three-stage process: (1) AE signal denoising using WGAN-GP-based AE for eliminating unnecessary noise while retaining fundamental fault characteristics, (2) short-time Fourier transform (STFT) feature extraction for an improved frequency-time domain representation, and (3) fault classification using an LSTM-based model for precise health condition prediction of machines. Three benchmark AE datasets (EAE-I, REB-II, and TAD-III) were used to perform experimental validation, proving the model to outperform baseline strategies. The model proposed has 97.0% accuracy on REB-II, 94.4% on EAE-I, and 96.2% on TAD-III, outperforming the best baseline models by a wide margin with a maximum of 93.0% accuracy. Additionally, the proposed method has a precision of 0.98, a recall of 0.97, and an F1-score of 0.975, reflecting its strength in identifying intricate fault patterns. The ablation study also confirms that GAN-based signal enhancement is responsible for a 7–9% gain in classification performance, demonstrating its efficiency in eliminating noise distortions without losing important AE signal features.

 Paraphrase     Questionfy     Wikify     Simplify     Summarize