Oövervakad anomalidetektering med hjälp av dimensionalitetsreduktion och syntetiska data i livsmedelsproduktion

Sammanfattning

The introduction of artificial intelligence into industries has significantly changed how industrial processesare monitored, analysed, and optimised, enabling early detection of process deviations. Traditional modelsoften struggle to perform effectively on high-dimensional, mixed-type industrial datasets that lack labels.This research investigates unsupervised anomaly detection for batch-level monitoring in an industrialcheese production process using real-world data from Tetra Pak. The study evaluates how differentdimensionality reduction techniques - Gower distance, Principal Component Analysis (PCA), FactorAnalysis of Mixed Data (FAMD) and Uniform Manifold Approximation and Projection (UMAP) affect theperformance of DBSCAN in mixed-type industrial data. This study also examines the impact of syntheticinlier data augmentation on hybrid models that combine DBSCAN or Isolation Forest with Autoencodersfor anomaly detection when the dataset is small. The methods are evaluated using clustering metrics,reconstruction-error plot, classification metrics, and expert-validated annotations. The results show thatdata representation plays a critical role in anomaly detection performance, with FAMD providing the moststable and informative feature space for DBSCAN. Synthetic data augmentation improves the robustnessof Autoencoder-based models by reducing false positives and producing clearer separation between normaland anomalous batches. With the cheese technologist’s annotation, it is identified that the hybrid modelsperform better than standalone models. The findings show that combining appropriate dimensionalityreduction techniques with hybrid unsupervised models enables more reliable early detection of subtledeviations in industrial cheese production.