International Journal of Computer Applications (0975 – 8887) Volume 175– No. 12, August 2020 38 A Machine Learning Approach for Temporal Vibration Analysis Applied to Predictive Maintenance Roberto Alexandre Dias Federal Institute of Santa Catarina – IFSC 950 Av. Mauro Ramos Florianópolis, SC, Brazil Pedro Von Hertwig Federal Institute of Santa Catarina – IFSC 950 Av. Mauro Ramos Florianópolis, SC, Brazil Mário de Noronha Neto Federal Institute of Santa Catarina – IFSC 950 Av. Mauro Ramos Florianópolis, SC, Brazil ABSTRACT The operational condition of a machine affects the quality and efficiency of its work, and letting a problem arrive in a critical state results in negative consequences, which can cause equipment loss and extensive downtime in a factory. The maintenance of machines, therefore, is a concern that came to exist along with the creation of the industry. This work shows the development of a platform that takes advantage of recent advances in sensor technology and machine learning to assist the predictive maintenance process, identifying problems in advance before serious failures can occur. The work proposes a supervisory system which receives high frequency vibration data, stores it, and analyzes the functioning of a machine to classify its behavior as normal or anomalous, generating alerts. The results achieved show that it is appropriate to use machine learning to monitor machines, since well-structured algorithms can detect possible problems before they become apparent to humans. General Terms Predictive maintenance of electrical machines. Vibration monitoring. Machine learning Keywords Vibration analyses. Temporal analysis. Predictive maintenance Machine learning. 1. INTRODUCTION A deterioration of equipment conditions leads to deviations in the production process and a drop in quality. Only proper maintenance can guarantee that the process will not lose quality due to deviations caused by the equipment [1]. In addition to the drop in product quality, the reduction in productivity due to unscheduled downtime is an obvious harm of the lack of proper maintenance. Less obvious is the drop in productivity even if the equipment is not stopped. This condition leads the company to look for the origin of the deficiency in factors such as tooling, materials, and operators; increasing the operational costs in a measurable way, and having effects of non-measurable costs, such as the erosion of the company's image. [1]. Machines are subject to failure in their operation, which incurs high costs due to the replacement of more parts and time lost while a repair is carried out. Preventive maintenance, which takes place at fixed intervals of time according to a manufacturer's recommendation, is one of the maintenance options. It involves changing parts before the time of failure, to reduce the failure rate of the equipment. This practice, however, is not the most efficient because the intervention in healthy equipment generates costs that can be avoided. Predictive maintenance, in contrast, is based on the principle of analyzing the machine's operating conditions to determine whether intervention is required. Thus, maintenance costs are restricted to equipment that in fact presents the possibility of imminent failure. Vibration analysis for diagnosis and condition evaluation has a long history of application to mechanical and energy equipment. Many types of defects increase the vibration level of machines or change their behavior in some way [2]. In this context, “subjective monitoring” is the monitoring that is performed by maintenance personnel using their senses - when a mechanic touches a housing, for example, he can perceive its vibration. Thus, a more experienced mechanic can provide greater precision in diagnosing a machine [3]. Such monitoring techniques, however, have numerous shortcomings: the diagnosis is subjective, depends on experienced personnel and does not provide continuous analysis. In recent years, electrical machine condition monitoring systems have become increasingly efficient and sophisticated. For autonomous monitoring, the vibration must be converted into an electrical signal. Some types of hardware that can be used are piezoelectric accelerometers and MEMS (Micro Electromechanical Systems) [4]. The increasing automation of detection processes has made automated diagnostic and prognostic systems a valuable tool for maintenance personnel and can even replace humans [5]. In this context, machine learning is a tool to aid in the decision-making process that has been gaining popularity due to its ease of adaptation to unfamiliar scenarios and the ability to solve difficult problems solved through simple mathematical modeling. Machine learning approaches have already been used successfully to identify failures in rotating machines [6]. The present work aims to explore the field of using machine learning for application in predictive maintenance. 2. BIBLIOGRAPHIC REVIEW 2.1 Failure prediction methods According to [5] existing methods of failure prediction can be grouped into three main categories:  Traditional reliability method (prediction based on past events - preventive maintenance)  Prognostic method (prediction based on condition monitoring)  Integrated method (prediction based on past events and condition monitoring)