Failure diagnosis of a compressor subjected to surge events: A data-driven framework

Due to higher reliability and safety requirements, the importance of condition monitoring and failure diagnosis has progressively cleared up. In this context, being able to properly deal with noise and data reduction is fundamental for improving failure diagnosis and assuring safe operations. These tasks are particularly difficult in presence of many non-stationary and non-linear signals. Accordingly, this paper proposes a failure diagnosis methodology that integrates Empirical Mode Decomposition (EMD) and Neighborhood Component Analysis (NCA) for noise removal and data reduction. While noise detection and reduction techniques are established to reduce the uncertainties integrated with data acquisition, traditional approaches that cannot capture the non-stationary and non-linear nature of data might result in higher uncertainty. As a validated denoising method, EMD is applied to cope with the previous limitations. The NCA overcomes typical limitations such as imposing class distributions. After data pre-processing, the diagnosis is performed through a Random Forest. The methodology is tested on real data of a compressor subjected to surge, showing an accuracy higher than 97%. Moreover, the surge accuracy is close to 95%, while the regime accuracy is higher than 97%. The developed framework could assist practitioners in evaluating the condition of assets and, accordingly, planning maintenance.