Anomaly detection of multivariate time series for industrial machinery

In industrial contexts, anomaly detection is crucial for identifying deviations from normal operating conditions, ensuring proactive maintenance, minimising downtime, and optimising the reliability and efficiency of industrial processes. Advanced machinery, constantly monitored by diverse sensors, generates multiple temporal sequences of data that can be analysed for evaluating performance. Within this setting, this thesis delves into a real-world scenario, focusing on the analysis of multivariate time series data produced by a set of filling machines for dairy products. Initially, the study conducts an in-depth analysis of the signals generated by the machinery and dedicates itself to collecting and preprocessing a dataset for following analyses. Subsequently, the Temporal Fusion Transformer (TFT), a cutting-edge deep learning model, is employed to effectively capture the complex temporal patterns inherent in industrial process signals and detect anomalies within the dataset. By addressing the challenge of dealing with intricate real-world data, this research aims to unravel their latent complexities and enhance anomaly detection precision through the utilisation of advanced deep learning models.