Anomaly Detection Using Acoustic Measurement Data: Test Case of Diffuser Clogging on a Single-Stage Centrifugal Compressor

This paper presents an experimental study conducted on a single-stage centrifugal compressor to investigate the possibilities of using noninvasive, easily implementable free-field microphones as data basis for a predictive maintenance approach. The approach combines statistical and time-resolved spectral evaluation methods to detect anomalies at an early stage without influencing the operation of the machine. Doing this, efficiency of service planning is increased and downtimes as well as costs are reduced. The experimental study covers the anomaly of diffuser clogging, which occurs at compressors that are operating in industry branches with poor environmental conditions like polluted air. In real-life operation the phenomenon of diffuser clogging is a creeping process over a period of time which leads to a decrease in efficiency and unexpected downtimes when the clogging is progressed considerably. To investigate this phenomenon, a part of the radial diffuser passage of a singlestage centrifugal compressor is modified using 3D-printed inserts. Several types of inserts with varying thickness are designed to simulate the different stages of clogging. The acoustic measurement data is then used to elaborate identification features like spectral patterns that are capable of detecting and locating the anomaly. It is shown that spectral energy content changes in fixed frequency bands, making it possible to distinguish different test cases. Furthermore, the gradually increasing blockage of the diffuser passage provides insight about the emerging effects on the flow path and influence on other machine parts. Ultimately, the results are compared with a previous test case that covers the misalignment of inlet guide vane blades validating that unique fault-specific spectral fingerprints evolve for different anomalies. This reinforces the potential of employing frequency analysis on noninvasive, easily implementable free-field microphones as data input for predictive maintenance applications.