Latent Dimensions of Auto-Encoder as Robust Features for Inter-Conditional Bearing Fault Diagnosis

Condition-based maintenance (CBM) is becoming a necessity in modern manufacturing units. Particular focus is given to predicting bearing conditions as they are known to be the major reason for machine down time. With the open-source availability of different datasets from various sources and certain data-driven models, the research community has achieved good results for diagnosing faults in bearing fault datasets. However, existing data-driven fault diagnosis methods do not focus on the changing conditions of a machine or assume all conditional data are available all the time. In reality, conditions vary over time. This variability can be based on the measurement noise and operating conditions of the monitored machines such as radial load, axial load, rotation speed, etc. Moreover, the availability of the data measured in varying operating conditions is scarce, as it is not always feasible to collect in-process data in every possible condition or setting. Considering such a scenario, it is necessary to develop methodologies that are robust to conditional variability, i.e., methodologies to transfer the learning from one condition to another without prior knowledge of the variability. This paper proposes the usage of latent values of an auto-encoder as robust features for inter-conditional fault classification. The proposed robust classification method MLCAE-KNN is implemented in three steps. First, the time series data are transformed using Fast Fourier Transform. Using the transformed data of any one condition, a Multi-Layer Convolutional Auto-Encoder (MLCAE) is trained. Next, a K-Nearest Neighbors (KNN) classifier is trained based on the latent features of MLCAE. The so-trained MLCAE-KNN is then used to predict the fault class of any new observation from a new condition. The results of using the latent features of the Auto-Encoder show superior inter-conditional classification robustness and superior accuracies compared to the state-of-the-art.