Faults explanation based on a machine learning model for predictive maintenance purposes

In this paper, an explainable data-driven based approach is proposed in order to perform fault detection and explanation of its potential causes. This approach is based on two main layers. The first layer is based on the use of an autoencoder with a view to detect the occurrence of a fault. The second layer explains the causes that generated the fault. This explanation is provided as the deviations in the input features or variables leading to generate the fault. These deviations can be used then by the maintenance manager in order to define the maintenance actions to be performed. The proposed approach is illustrated and evaluated using the well-known Tennessee Eastman process benchmark.