Clustering and Anomaly Detection in Industrial IoT Data

The Industrial Internet o fThings(IoT) has revolutionized industrial operations by enabling real-time data collection and analysis through smart sensors and connected devices. This digital transformation is particularly crucial in the manufacturing industries, where predictive maintenance and efficiency optimization are essential. However, the massive volume of high frequency data collected by different sensors generated in IoT environments presents challenges related to data quality, dimensionality and computational efficiency. Addressing these challenges requires advanced data driven approaches, particularly in clustering and anomaly detection contexts. This thesis is inspired by the internship conducted at a leading company in the production of pumps, where clustering and deep learning techniques were applied to IoT data coming from a particular type of pumps. These smart pumps capture many operational parameters and thus generate a large amount of time-series and multivariate data, in particular, the dataset consists of millions of records and theretofore presented significant challenges due to missing or inconsistent values, sensor drift and data imbalance. The need to segment pump behaviors into distinct categories to extract meaningful insights suggest me to exploit different clustering algorithms. The second part of the thesis involves analyzing anomaly detection by exploring AT-DCAEP (Attention-based Dual-channel Autoencoder with External Prediction) model. This model is designed for multivariate time-series anomaly detection by combining spatial and temporalfeature learning. The results help to develop effective clustering and anomaly detection methodologies for IoT applications, which improve predictive maintenance strategies and operational efficiency in industrial settings.