Automated classification of tribological faults of alternative systems with the use of unsupervised artificial neural networks

Abstract

Preventing, anticipating, avoiding failures in electromechanical systems are demands that have challenged researchers and engineering professionals for decades. Electromechanical systems present tribological processes that result in fatigue of materials and consequent loss of efficiency or even usefulness of machines and equipment. Several techniques are used in an attempt to minimize the inherent losses of these systems through the analysis of signals from the equipment studied and the consequences of these wastes at unexpected moments, such as an aircraft in flight or a drilling rig in an oil well. Among them we can mention vibration analysis, acoustic pressure measurement, temperature monitoring, particle analysis of lubricating oil etc. However, electromechanical systems are complex and may exhibit unexpected behavior. Reliability-centric maintenance requires ever faster, more efficient and robust technological resources to ensure its efficiency and effectiveness. Artificial neural networks (ANN) are computational tools that find applicability in several segments of the research and signal analysis, where it is necessary to handle large amounts of data, associating statistics and computation in the optimization of dynamic processes and a high degree of reliability. They are artificial intelligence systems that have the ability to learn, are robust to failures, and can deliver real-time results. This work aims at the use of artificial neural networks to treat signals from the monitoring of tribological parameters using a test bench to simulate contact failures in an air compressor in order to create an automated fault detection and classification system, unsupervised, with the use of self-organized maps, or SOM, applied to the preventive and predictive maintenance of electromechanical processes.