Variable-Length Genetic Algorithm-Based Blind Deconvolution for Railway Bearing Fault Diagnosis

Blind deconvolution (BD) methods have shown significant promise in machinery fault diagnosis, yet their performance remains constrained by fixed filter length and susceptibility to local optima. To tackle this problem, this article presents a variable-length genetic algorithm (VGA) to optimize the blind filter for railway bearing fault diagnosis under complex interference. Unlike existing works, an attempt is made to determine the length and coefficients of the optimal blind filter in a general framework by introducing a cutting-splicing operator to replace the crossover to achieve chromosome length variation during the optimization iteration. Based on several popular BD approaches, including minimum entropy deconvolution (MED), maximum correlated kurtosis deconvolution (MCKD), maximum Gini index deconvolution (MGID), and maximum envelope spectral sparsity deconvolution (MESSD), their variable-length optimized version, i.e., VGA-MED, VGA-MCKD, VGA-MGID, VGA-MESSD, are proposed and rigorously compared with the original version and the fix-length optimized version, i.e., GA-MED, GA-MCKD, GA-MGID, and GA-MESSD by both simulated and experimental railway bearing faulty signals. The results demonstrated the effectiveness and adaptability of the novel methods in recovering fault-induced repetitive transients. Furthermore, the adaptability and generalizability of the given framework make it a robust solution for BD-based fault diagnosis in industrial applications.