Unsupervised Anomaly Detection in Multivariate Time Series Using Variational Autoencoders

Sammanfattning: In this master’s thesis, a novel unsupervised anomaly detection tool was developed in collaboration with Sandvik Rock Processing to assist engineers and experts in analyzing large amounts of sensor data from cone crushers used in the stone crushing industry. The tool focuses on analyzing power, pressure, and CSS sensor data. A crucial preprocessing step was implemented to algorithmically identify operation segments of sufficient length, differentiating between off, idle, continuous, and discontinuous states based on power usage. The Variational Autoencoder (VAE) employed a unique architecture with two 1D convolutions in the encoder and 1D transposed convolution in the decoder, utilizing parallel kernel sizes of 2 and 15 to capture both short-term and long-term patterns in the data. The decoder also incorporated a polynomial trend block to enhance the reconstruction. The VAE was trained on well-behaved operation segments to identify anomalous behaviour through the reconstruction error metric, Mean Absolute Percentage Error (MAPE). The anomaly detection tool achieved an F1 score of 0.89, 0.75, and 0.92 for the different sensors when tested with labelled anomalies provided by Sandvik. Despite the challenge of limited labelled data, the tool successfully identifies the worst operation segments and can be utilized for deriving useful operation metrics. The main benefit of implementing this tool in the context of Sandvik Rock Processing’s operations is the significant acceleration of sensor data analysis and the ability to highlight areas of concern for engineers and experts. Potential future improvements include using a larger dataset for training, more rigorous testing of hyperparameters, and better data collection to account for factors such as machine models and expected operating pressure and power.