Mechanism-Data Fusion for Centrifugal Compressor Aerodynamic Failure Diagnostics Based on Bayesian Extreme Learning Machine

Centrifugal compressors (CCs) are critical equipment for compressing and transporting various gases in chemical production. As a type of vane compressor, their stable and reliable operation can be compromised by issues such as surge, rotating stall, and other malfunctions. Traditional diagnostic approaches for CCs typically depend on a singular signal, such as vibration analysis, which can be one-sided and overlook substantial amounts of information. Additionally, multi-source-signal fusion diagnostic methods for CCs are relatively scarce. With this in mind, a mechanism-data fusion diagnostic approach for CC aerodynamic failure based upon Bayesian extreme learning machine (BELM) is proposed in this article. Firstly, the vibratory data is processed using wavelet packet decomposition (WPD) and various statistical features are extracted from different scales. The Relief-F algorithm is then employed to select the most significant feature set, which are subsequently leveraged to train the BELM data-driven model. Next, mechanism-knowledge-based indices of process parameter signals (e.g., pressure, flow) are designed and calculated, and a mechanism-based diagnostic model can be derived via utilizing the fuzzy inference system (FIS). In the implementation stage, the sparse representation of FIS results is conducted at first, followed by the diagnostic information fusion attained through the optimization of BELM parameters. Finally, experimental validations on CC datasets as well as comparison analyses with other information fusion methods sufficiently certify the effectiveness and superiority of the developed approach.