Skip to content

Adaptive Wavelet Denoising Combined with Improved Transformer for Industrial Equipment Vibration Fault Diagnosis

Adaptive Wavelet Denoising Combined with Improved Transformer for Industrial Equipment Vibration Fault Diagnosis

Author: Javier Manuel Ortiza, Bram Jan van der Meerb

Corresponding Author : Javier Manuel Ortiz
E‑mail: javier.ortiz@upm.es

a Department of Signal Theory and Communications, Universidad Politécnica de Madrid, Ciudad Universitaria s/n, Madrid, 28040, Spain
b Department of Mechanical Engineering, Delft University of Technology, Mekelweg 2, Delft, 2628 CD, Netherlands

Journal Metadata: American International Journal of Scientific & Technical Research, Vol 1, Issue 2, 2026, Pages 1‑8

ISSN 3144‑2153

Abstract

Vibration signal analysis serves as one of the most mainstream technical routes for condition monitoring and fault diagnosis of rotating machinery such as rolling bearings in modern industrial systems. In practical industrial scenarios, vibration signals collected by sensors are inevitably mixed with strong background noise generated by mechanical coupling, electromagnetic interference and complex working condition fluctuations. The weak fault impulse features characterizing early bearing damage are often submerged in noise, which brings great challenges to the accurate identification of fault categories. Traditional fault diagnosis methods rely heavily on manually designed feature extraction rules, and their generalization ability under variable working conditions is limited. In recent years, Transformer models based on self‑attention mechanism have achieved remarkable performance in time series analysis tasks, but the vanilla Transformer still suffers from excessive computational overhead and interference from redundant background features when directly applied to vibration signal diagnosis. To solve the above limitations, this paper constructs an industrial vibration fault diagnosis framework integrating adaptive wavelet denoising and improved Transformer. The adaptive wavelet threshold denoising algorithm is firstly adopted to suppress high‑frequency noise components in raw vibration signals and retain the transient impact characteristics related to bearing faults. Afterwards, an improved Transformer model with gated multi‑head attention is proposed to realize adaptive feature screening and long‑range dependency mining of denoised vibration sequences. The gated unit embedded before multi‑head attention can automatically evaluate the importance of different signal segments, suppress the response of redundant noise features, and guide attention heads to focus on the time segments containing fault impulses. Comparative experiments are carried out on the public bearing fault dataset, and the diagnostic accuracy of the proposed method is compared with classic machine learning models, convolutional neural networks and original Transformer. The experimental results demonstrate that the proposed integrated framework achieves superior classification accuracy and stronger anti‑noise robustness, especially for early weak bearing fault samples. This work provides a feasible technical reference for intelligent fault diagnosis of rotating machinery under noisy industrial environments.

Keywords: wavelet denoising; adaptive threshold; Transformer; gated multi‑head attention; vibration signal; fault diagnosis

Full research content, tables, references and complete data analysis are contained in the attached full-text PDF file.

© 2026 Javier Manuel Ortiz. This article is distributed under the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium,
provided appropriate credit to the original author and source is given.