Compressor Fleet Diagnostics by Employing MLP-Generated Heatmaps

Reciprocating compressors are renowned for their versatility and widespread application across various installations. Ensuring their reliable operation is paramount, as failures can lead to costly unscheduled downtimes, safety hazards, and significant production losses. Consequently, the identification of both healthy and faulty operating conditions is essential. This study addresses this challenge by means of the integration of multi-layer perceptron (MLP) models for cross-regression analysis with heatmap visualization. In this study, we propose a comprehensive framework, of which the first step is MLP training with seemingly healthy data, targeting feature regression of acceleration, velocity, and fast Fourier transform (FFT) datasets. This is followed by the visualization of discrepancies between model predictions and field data in the form of heatmaps, thus facilitating the straightforward identification of both magnitude and location of such discrepancies. The methodology is systematically applied across various datasets, representing three different reciprocating compressors of the same product family, with different architectures and time frame lengths, to evaluate its effectiveness as a tool for fleet diagnostics. Additionally, one of the three compressors is also located on a different site. The data covers more than two years in total, encompassing variables that range from crank angle degrees, i.e., acceleration and velocity, to frequency domain representations, i.e., FFT data. The results demonstrate the capability of the proposed framework for detecting faulty operation across both the frequency and crank angle degree domains and across different machines, making the proposed framework an attractive solution for automated, cloud-based fleet management.