Hybrid Deep Learning Based on Domain Expertise and Multimodality for Chatter Detection

Overview & Synopsis

Background Machine learning has significantly contributed to vibration-based chatter detection. However, supervised approaches generally require large amounts of labeled data and substantial expert involvement, which can be time-consuming and difficult to implement in industrial environments. This study investigates the use of unsupervised learning combined with domain expertise and multiple representations of vibration signals to facilitate the identification and interpretation of machining states. Methods Real machining vibration signals acquired under industrial conditions were transformed into temporal representations, frequency spectra obtained using the Fast Fourier Transform (FFT), spectrograms with different time-frequency resolutions, and structured combinations of these representations. The analysis involved K-Means, K-Medoids and agglomerative hierarchical clustering, dimensionality reduction using PCA and UMAP, as well as Autoencoder, Variational Autoencoder, and pre-trained VGG and ResNet models. The resulting clusters were assessed using clustering metrics and expert evaluation to examine both their geometrical quality and their interpretability from a machining perspective. ResultsThe results show that good geometrical cluster separation does not necessarily correspond to easy expert interpretation. Spectrogram-based modalities provided the most satisfactory results, particularly when combined with K-Means clustering. The expert evaluation showed that spectrograms were the most suitable representations for efficient labeling, with the best spectrogram configuration achieving a global score of 8. Spectro-temporal representations also provided satisfactory results, with scores of 5 and 6 for the best configurations, whereas the temporal representation alone obtained a score of 0. K-Means generally outperformed K-Medoids and agglomerative clustering across the evaluated signal modalities. Conclusions The findings demonstrate the importance of combining quantitative clustering assessment with domain expertise when analyzing machining vibration signals. In particular, spectrogram-based representations provided the most suitable modalities for expert interpretation, while the results also showed that increasing model or representation complexity does not necessarily improve clustering interpretability. The proposed framework therefore highlights the complementary role of unsupervised artificial intelligence methods and expert knowledge in identifying meaningful structures in machining vibration data.