A CEEMD-Based Improved Wavelet Threshold Denoising Method for Bearing and Gear Fault Diagnosis of Train Bogie Transmission Systems

The bogie transmission system is the core power chain of a railway vehicle, and the health of its traction-motor bearings, gearbox bearings and gears directly determines running safety. Early-stage fault impulses, however, are weak and are usually buried in strong background vibration from gear meshing, electromagnetic interference and wheel-rail excitation, which makes fault feature extraction extremely difficult. This paper proposes a cascade denoising method that combines complementary ensemble empirical mode decomposition (CEEMD) with an improved wavelet threshold function for the fault diagnosis of train bogie transmission systems. The vibration signal is decomposed by CEEMD, sensitive intrinsic mode functions are selected by the kurtosis criterion, the reconstructed signal is denoised by a piecewise improved wavelet threshold function whose transition band is controlled by a parameter alpha, and Hilbert envelope demodulation extracts the fault characteristic frequencies. The method is validated on simulated inner-race fault signals at four input noise levels (−10 to +5 dB) and on the public BJTU-RAO bogie dataset covering five health states and three operating speeds. At −10 dB input SNR the output SNR is improved from −10.0 dB to −2.2 dB, close to the best soft-threshold result, and at 0 dB the proposed method outperforms both the hard and soft thresholds. On the real bogie data, the method improves the fault-frequency identification SNR of the tooth-root crack by 3.6 dB and of the rolling-element fault by 1.6 dB at 20 Hz supply (with smaller gains at 40/60 Hz) and clearly preserves the 293.2 Hz outer-race fault frequency. The whole pipeline processes a 2 s, 64 kHz record within about 3 s, which is potentially applicable to quasi-real-time monitoring with a processing time of about 3 s per 2 s record. It should be noted that the improvement is fault-dependent: on already-visible strong faults the universal threshold may over-smooth the feature, and on pure Gaussian white noise direct wavelet thresholding without CEEMD can outperform the full cascade. The method is therefore recommended for weak incipient fault detection rather than as a universal denoiser.