Predictive Maintenance and IoT
Conventional preventive maintenance (PvM) is schedule-based maintenance with the same time interval for reducing the likelihood of equipment failure. The PvM assumes the likelihood of failure increases with the usage and age of the equipment. Hence, Malik (1979) proposed the “Proportional Age Reduction” model that, for a piece of equipment with age t, its post-maintenance age can be reduced from t to t/𝛽, where 𝛽=(1,∞). Age may not be a good indicator, and the PvM approach may result in the unnecessary replacement of some equipment components after a predetermined period. On the other hand, if the equipment runs with abnormal energy consumption, speed, or temperature before replacement, PvM cannot provide the replacement decision.
To address the above shortcomings of PvM, Predictive maintenance (PdM) enabled by the industrial Internet of Things (IoT) sensors is proposed by which we can identify maintenance issues in real-time. The PdM approach is prediction-based maintenance. The equipment is continuously monitored by various sensors (IoT sensors), which generate data in real-time to predict the next error. PdM can then be conducted before the failure occurs. In particular, PdM is invaluable if the equipment and the corresponding processes are too expensive to have any critical damages.
Predictive Maintenance and ML
Since PdM comes with high potential in reducing maintenance costs, it has recently attracted a certain number of researchers applying machine learning methods to PdM, such as Samatas et al. (2021), Ayvaz and Alpay (2021), Florian et al. (2021), Chen et al. (2020), Çınar et al. (2020), Sahal et al. (2020), Ruiz-Sarmiento et al. (2020), Wan et al. (2017), Susto et al. (2014), and Susto et al. (2012).
According to the review paper of Carvalho et al. (2019), the popular machine learning methods for PdM are Random Forests (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and k-means. They also concluded that “there is no preference for an equipment to perform PdM strategies, and vibration signals are the most common data used to design PdM models, there is a preference for real data to build PdM models.”
Ayvaz and Alpay (2021) developed a predictive maintenance model for personal care goods production lines, such as baby care, feminine hygiene, and home care products, in Turkey. They evaluated six algorithms, four ensembles (Random Forest, XGBoost, Gradient Boosting, and AdaBoost), and two constituent machine learning algorithms (Neural Network and Support Vector Regression), developing a model consisting of one dependent output and 47 independent inputs. They also explored the performance of Multiple Regression, Lasso Regression, and Ridge Regression. Due to the non-linear nature of the problem, these three regression-based methods could not capture any variance in the data.