Advanced Fault Diagnosis in Milling Machines Using Acoustic Emission and Transfer Learning
Muhammad Umar, Zahoor Ahmad, Saif Ullah, Faisal Saleem, Muhammad Siddique, Jong-Myon Kim
Abstract
The accurate diagnosis of faults in milling machines is important to ensure manufacturing efficiency and minimize downtime. Acoustic emission (AE) signals, known for their transient and high-frequency nature, provide valuable insights into tool and machine faults. However, their non-stationary characteristics present challenges for traditional analysis methods. This study proposes an innovative framework that combines time-frequency representation, transfer learning, and dimensionality reduction for effective fault diagnosis. AE signals are transformed into scalograms and spectrograms using continuous wavelet transform (CWT) and short-time Fourier transform (STFT), respectively, extracting both localized and global signal characteristics. These visual representations are processed through pre-trained deep learning architectures, EfficientNet-B0 and InceptionV3, to extract high-level features. Dimensionality reduction through uniform manifold approximation and projection (UMAP) further refines these features while preserving useful patterns. Finally, a lightweight k-nearest neighbors (k-NN) classifier is used to distinguish across all classes with high accuracy, achieving an average of 99.60% cross-validation performance. This framework highlights the strength of combining transfer learning with dimensionality reduction for fault diagnosis, providing a computationally efficient and highly accurate solution with significant potential for real-time monitoring and predictive maintenance in advanced manufacturing systems.