Anomaly occurrences in mechanical equipment within industry 4.0 may lead to massive systems shut down, jeopardizing the safety of the machinery and its surrounding human operator(s) and environment, as well as the severe economic implications succeeding the faults and their associated damage. Various mechanical tools are mostly placed in harsh and ruthless environments, where the machines are consistently vulnerable to many fault types connected to their functionality nature. Hence, not only the machines and their components are prone to anomalies, but also the sensors attached to them necessary to collect viable signals to monitor and report the overall machine health and behavioural changes. Those sensors may likewise fail and carry out various anomalies.
This thesis elucidates a full research and analytical implementation of component and sensor faults detection and diagnosis, utilizing numerous machine and deep learning approaches in application of a hydraulic system extracted from a hydraulic test rig. It is unfortunate that hydraulic systems are rarely approached for anomaly detection subject comparing to other mechanical machines in the past decade. Specifically, comprehensive systems that cover all aspects of anomaly detection in hydraulic systems, which includes both sensor and component faults, essential feature engineering methods, and innovative detection algorithms based on the latest technologies such as, the application of deep learning.
In this work, three main contributions to anomaly detection in hydraulic systems extracted from a hydraulic test rig are thoroughly achieved. Firstly, we provided a combination of LSTM autoencoders and supervised machine and deep learning methodologies to perform two separate stages of fault detection and diagnosis. The two phases are condensed by: (1) the detection phase using the LSTM autoencoder. Followed by (2) the fault diagnosis phase represented by the classification schema. The previously mentioned framework is applied to component and sensor faults in hydraulic systems, deployed in the form of two in-depth applicational experiments. In the detection phase declared by the classification process, diversified machine and deep learning supervised methods are compared and analysed for their component and sensor fault detection performance in hydraulic systems. In addition, we provided comparisons of plentiful feature engineering techniques in the time-domain, to showcase the influence of each feature engineering method on its corresponding supervised classifiers in the detection phase. Secondly, we provided an unprecedented feature selection method called Recursive k-means Silhouette Elimination (RkSE), and it is deployed to perform feature selection for component fault classification in multi-variate hydraulic test rig dataset. Moreover, RkSE is utilized as a window compression method when deployed to achieve sensor fault identification in univariate sliding window-structured datasets. Finally, an innovative application of Random Forests (RF) in a hybrid architecture between data-driven and model-based diagnosis approaches is introduced and applied to hydraulic systems for dynamic diagnostic rules generation.