Machinery diagnostic techniques for maintenance optimization

One of the future challenges of machinery diagnostics and prognosticsis to prepare for the Internet of Things (IoT), where it is possible to change and improve existing approaches drastically. An intensifying application of the IoT will increase the use of embedded sensors and, therefore, create a demand for diagnostic tools where manual workis minimized and is mainly handled by smart algorithms. The auto-mated anomaly detection of large assets and their components with a system of smart algorithms needs proper optimization. Foremost, it is critical to avoid machinery failures, since they can interrupt production, cause unbearable production losses for the business and, even more, can put the lives of personnel in danger if a catastrophic failure occurs. On the other hand, if all the components are repeatedly creating false alarms, the verification of these incidents may be overwhelming. This research studied how a one-class SVM can be optimized by tuning the algorithm to function properly by taking the criticality of the system into consideration. Another topic dealt with was how a one-class SVM can be used for identifying the location of faults by carefully selecting proper input features. Furthermore, a method was tested where a variational Bayesian for Gaussian mixture algorithm was used for pre-processing and separating the condition monitoring data into operation mode classes. Later these classes can be used for improving the time for acquiring the condition monitoring data or to give more information as to how prognostic algorithms should be selected. In addition, a method was tested which involved the use of a Random Forest for feature selection and for the creation of indifference to load or other similar external factors by comparing separate classes with each other. Overall, the idea is that all of these tech-niques can be combined and merged in order to improve machinery diagnostic tools and prepare for the coming era of digitalizatio