Multidimensional fast nonlinear blind deconvolution network for bearing compound features extraction

The uneven stress distribution and abnormal load caused by a single bearing fault often lead to another new fault. The weak features of the new fault are either aliased with the existing fault features and ignored, or directly covered by irrelevant interference components. To achieve separation and extraction of compound faults, multidimensional fast nonlinear blind deconvolution network (MFNBD-net) is proposed. Firstly, fast nonlinear blind deconvolution (FNBD) is extended to MFNBD based on the principle of multi-dimensional blind deconvolution to obtain the potential of decoupling composite features. Then, uniform multidimensional initialization for indicating the convergence direction is introduced to enhance the performance of multi-feature extraction. Next, based on the uniformity of harmonic distribution, trimmed envelope spectrum kurtosis for guiding the elimination of irrelevant and repetitive components in multi-dimensional output is proposed. Finally, adaptive nonlinear transformation and filter waveform penalty are incorporated into the deconvolution process and MFNBD-net is proposed. Simulation and experiments show that MFNBD-net has advantages in multi-dimensional feature decoupling and robustness, and it is a promising composite feature extraction tool.