Gargiulo, Federico (2022) Failure Predictions and Anomaly Detections by Artificial Intelligence with Heterogeneous Measures. [Tesi di dottorato]

Preview

Text Gargiulo_Federico_34.pdf Download (7MB) | Preview

Collection description

The changes in the manufacturing world driven by the revolution of industry 4.0 entail a growing volume of information and the demand for planning in production's tasks and activities. The techniques of measurement, transport, analysis and processing of data are constantly evolving. The technological growth driven by the evolution of industry 4.0 opens up new frontiers of research in the field of predictive maintenance, a theme that is deeply felt in highly competitive production contexts. In this PhD thesis, the topic of predictive maintenance is addressed by means of a research work on measurement techniques, data processing and Machine Learning (ML) algorithms. The research works presented in this thesis have focused on 3 important issues in modern industry. The first work explores the techniques for predictive maintenance on machines for the compression of cryogenic fluids and a solution based on unsupervised machine learning techniques for diagnostics is proposed; the second work concerns techniques for the identification of faults and diagnostics of magic hard disks and a method for the prediction of failures based on decision forests is proposed; the third work focuses on measurement, processing and machine learning techniques for the prevention of failure on three-phase asynchronous electric motors by means of fault detection approach. As for the first research work, a fault detection method exploiting Hidden Markov Models (HMMs) is proposed for fluid machinery without adequate a-priori information about faulty conditions. The method was tested and validated at CERN on screw compressors for cryogenic cooling. As second research work a method to facilitate automated proactive disk replacement is proposed. The method identifies disks with media failures in a production environment and adopts an application of supervised machine learning, based on Regularized Greedy Forest, to predict disk failures. In particular, a proper stage to automatically label (healthy/at-risk) the disks during the training and validation stage is presented along with tuning strategy to optimize the hyperparameters of the associated machine learning classifier. The machine learning model is trained and validated against a large set of 65,000 hard drives in the CERN computer center, and the remarkable achieved results are discussed. Finally a method to identify electrical and mechanical faults in three-phase asynchronous electric is proposed in order to prevent device failures. A measurement strategy and machine learning algorithm, based on Artificial Neural Network, is proposed to properly classify failures. The method is validated on a set of 28 electric motors. The method's performance is compared with common state of art machine learning techniques.

Downloads

Actions (login required)

View Item