| Summary: | The increasing availability of data, gathered by sensors and intelligent machines, is changing the way decisions are made in the manufacturing sector. In particular, based on predictive approach and facilitated by the nowadays growing capabilities of hardware, cloud-based solutions, and new learning approaches, maintenance can be scheduled—over cell engagement and resource monitoring—when required, for minimizing (or managing) unexpected equipment failures, improving uptime through less aggressive maintenance schedules, shortening unplanned downtime, reducing excess (direct and indirect) cost, reducing long-term damage to machines and processes, and improve safety plans. With access to increased levels of data (and over learning mechanisms), companies have the capability to conduct statistical tests using machine learning algorithms, in order to uncover root causes of problems previously unknown. This study analyses the maturity level and contributions of machine learning methods for predictive maintenance. An upward trend in publications for predictive maintenance using machine learning techniques was identified with the USA and China leading. A mapping study—steady set until early 2019 data—was employed as a formal and well-structured method to synthesize material and to report on pervasive areas of research. Type of equipment, sensors, and data are mapped to properly assist new researchers in positioning new research activities in the domain of smart maintenance. Hence, in this paper, we focus on data-driven methods for predictive maintenance (PdM) with a comprehensive survey on applications and methods until, for the sake of commenting on stable proposal, 2019 (early included). An equal repartition between evaluation and validation studies was identified, this being a symptom of an immature but growing research area. In addition, the type of contribution is mainly in the form of models and methodologies. Vibrational signal was marked as the most used data set for diagnosis in manufacturing machinery monitoring; furthermore, supervised learning is reported as the most used predictive approach (ensemble learning is growing fast). Neural networks, followed by random forests and support vector machines, were identified as the most applied methods encompassing 40% of publications, of which 67% related to deep neural network with long short-term memory predominance. Notwithstanding, there is no robust approach (no one reported optimal performance over different case tests) that works best for every problem. We finally conclude the research in this area is moving fast to gather a separate focused analysis over the last two years (whenever stable implementations will appear).
|
|---|
