Anomaly Detection in Iron Ore Conveyor Belts

The presence of foreign objects on iron ore conveyor belts poses significant challenges to operational efficiency and safety in the mining industry. These contaminants, such as debris, scrap, or harder minerals, can damage equipment, disrupt operations, and increase costs. Traditional supervised learning methods for anomaly detection, such as YOLO and U-Net, are limited in their ability to address this issue due to the variability, rarity, and unpredictable nature of these anomalies, as well as the extensive labeled datasets required for training. This paper investigates the application of Variational Autoencoders (VAEs) for unsupervised anomaly detection in iron ore conveyor belt systems. By training the VAE exclusively on data representing normal operating conditions, the model captures a latent representation of non-anomalous operation, enabling it to identify deviations indicative of potential anomalies. The study evaluates the impact of varying latent space dimensions (8, 16, 64, and 128) on model performance and incorporates Gaussian Mixture Models (GMM) and Linear Discriminant Analysis (LDA) for additional clustering and analysis of latent features. Results indicate that the 64-dimensional latent space configuration, paired with GMM, achieves the best performance, with precision and recall exceeding 93 %. This method minimizes the reliance on labeled anomalous data and presents a robust solution to enhance safety and efficiency in mining operations.