Self-supervised time–frequency feature alignment for process monitoring of cyber–physical CNC machines

Abstract

Self-supervised learning excels at uncovering latent features from incomplete data, thereby providing robust support for downstream applications. Capitalizing on this strength, a growing number of fault diagnosis models have been developed to monitor CNC machine tools, which are essential to modern manufacturing. These machines operate under demanding conditions – characterized by high speeds and heavy loads – and consequently generate mechanical signals with pronounced nonlinearity. Such inherent nonlinearity poses significant challenges for conventional feature extraction methods, necessitating advanced self-supervised techniques to effectively capture and interpret the underlying fault-related features for reliable condition monitoring. In this research, we introduce a self-supervised time–frequency feature alignment (STFA) algorithm for monitoring the manufacturing processes of industrial CNC machine tools. The STFA algorithm initially employs two domain-specific modules to extract time–frequency features from surveillance signals. A modern CNN is utilized to extract spatiotemporal information from the time domain, while a multi-scale CNN captures multi-granular features from the frequency domain. Subsequently, a dedicated time–frequency feature alignment module (TFAM) maps these features into a unified space, thereby exploiting their complementarity and enabling a more comprehensive representation. The STFA algorithm is trained through a dual-stage process—first, a pre-training phase to establish robust feature representations from unlabeled data, followed by a fine-tuning stage using a limited number of labeled samples to adapt the model for precise fault diagnosis. The effectiveness of the proposed STFA algorithm is validated using two manufacturing datasets collected from industrial CNC machine tools.