Multivariate anomaly detection in quality control of paper manufacturing

Abstract

The purpose of this thesis was to develop an autoencoder-based anomaly detection method for multivariate quality control system data and evaluate different architectures in detecting anomalies from selected properties. Vanilla, denoising, deep, variational and convolution autoencoders were compared using synthetic scan data containing both pointwise anomalies exceeding quality limits as well as subtle structural anomalies within limits. The study also tested models’ generalization capability to unseen product data using transfer leaning and robustness against added measurement noise.

The results showed that with vanilla and denoising autoencoders it was already possible to detect over 95% of all anomalies. From the advanced model structures, variational autoencoder showed good but varying performance due to its probabilistic nature, while deep and convolutional autoencoder exhibited comparable performance with simple models. Autoencoders outperformed statistical process control in detection performance and multivariate analysis proved to be effective compared to univariate analysis. Simulations applying generalized fine-tuned models with transfer learning resulted in comparable or better performance compared to nominal models and noise robustness test showed that model complexity can increase noise sensitivity. Structured anomaly profiles led to lower reconstruction errors than normal profiles, requiring custom detection strategies based on absolute deviation. Overall, the findings demonstrated the potential of using autoencoders for unsupervised anomaly detection tasks and highlighted the importance of model design and evaluation. The developed framework thus offers a basis for designing a future anomaly detection application for real product environments.