TADUFMA: Transformer-based Adaptive Denoising and Unified Feature Modeling for Multi-condition Anomaly detection in Computerized Flat Knitting Machines

Accurate anomaly monitoring is vital for the stable operation of computerized flat knitting machines. However, practical challenges remain: complex multi-factor noise disrupts data stability, and rapid changes in operation conditions hinder the ability of model to extract critical features, thereby limiting detection accuracy and robustness. This paper proposes transformer-based adaptive denoising and unified feature modeling for multi-condition anomaly detection in computerized flat knitting machines, a novel anomaly detection framework comprising two-stage models. In the first stage, transformer-diffusion synergistic adaptive denoising integrates the strengths of the Transformer architecture and diffusion models, employing adversarial training to dynamically eliminate complex noise from multifactor interference. In the second stage, variational autoencoder (VAE)-based fusion attention multi-condition anomaly detection combines VAE, convolutional neural networks, and multi-domain unified attention modules to extract diversified process features. It constructs a causality-guided VAE to map multi-condition data into a unified latent space, performing anomaly detection through reconstruction errors. The experimental validation of real-world computerized flat knitting machine production data demonstrates that our method achieves a 17.7% improvement in anomaly detection over state-of-the-art approaches. The proposed framework provides an innovative solution for multi-condition feature modeling and industrial anomaly detection.