A Smart CEEMDAN, Bessel Transform and CNN-Based Scheme for Compound Gear-Bearing Fault Diagnosis

In rotating machinery, compound faults occur when multiple faults happen simultaneously, making them challenging to detect. This study targets the diagnosis of compound gear-bearing faults, focusing on overcoming the limitations of current time–frequency transforms in detecting compound faults, the absence of sufficient denoising techniques before time–frequency domain conversion, and the complexity in selecting precise noise-free components.

The proposed methodology integrates complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), Bessel transform, and convolutional neural network (CNN). Initially, CEEMDAN decomposes the raw vibration signals, followed by the selection of noise-free intrinsic mode functions (IMFs) based on their cross-correlation with the raw signal. To prevent information loss during denoising, an optimized cross-correlation threshold is determined. Subsequently, the Bessel transform facilitates time–frequency transformation. Lastly, the CNN serves as the fault classification algorithm.

The proposed scheme is validated through a case study, resulting in an impressive average classification accuracy of 97.5%. This outcome underscores the method’s effectiveness, showcasing superior diagnostic precision compared to conventional decomposition and time–frequency transform-based techniques.

Incorporating CEEMDAN, Bessel transform, and CNN, the method introduced emerges as a robust and efficient solution for diagnosing compound gear-bearing faults. The achieved high classification accuracy in the case study affirms the effectiveness and superiority of this approach over conventional methods.