A Novel Intelligent Fault Diagnosis Method for Self-Priming Centrifugal Pumps

Abstract

The real-time diagnostic monitoring of self-priming centrifugal pumps is essential to ensure their safe operation. Nevertheless, owing to the intricate structure and complex operational conditions inherent in such pumps, existing fault diagnosis methods encounter challenges in effectively extracting crucial fault feature information and accurately identifying fault types. Consequently, this paper introduces an intelligent fault diagnosis method tailored for self-priming centrifugal pumps. The approach amalgamates refined time-shift multiscale fluctuation dispersion entropy, cosine pairwise-constrained supervised manifold mapping, and adaptive chaotic Aquila optimization support vector machine techniques. To begin with, refined time-shift multiscale fluctuation dispersion entropy is employed to extract fault-related features, adeptly mitigating concerns related to entropy domain deviations and instability. Subsequently, the application of cosine pairwise-constrained supervised manifold mapping serves to reduce the dimensionality of the extracted fault features, thereby bolstering the efficiency and precision of the ensuing identification process. Ultimately, the utilization of an adaptive chaotic Aquila optimization support vector machine facilitates intelligent fault classification, leading to enhanced accuracy in fault identification. The experimental findings unequivocally affirm the efficacy of the proposed method in accurately discerning among various fault types in self-priming centrifugal pumps, achieving an exceptional recognition rate of 100%. Moreover, it is noteworthy that the average correct recognition rate achieved by the proposed method surpasses that of five existing intelligent fault diagnosis techniques by a significant margin, registering a notable increase of 15.97%.

Gaussian-kernel weighted neighborhood preserving embedding algorithm and its application in fault detection

“…Table 2 details 21 different fault types, with faults 3, 9, and 15 consistently exhibiting low FDRs across different The these failures lies in minimal data changes during their occurrence, posing a significant obstacle to most data-driven monitoring approaches [36]. In more recent years, deep learning has gained widespread use in fault detection and diagnosis [37][38][39][40]. Therefore, this experiment not only compares multivariate statistical methods, including manifold learning, but also introduces deep learning for verification.…”