Elimination of end effects in LMD by Bi-LSTM regression network and applications for rolling element bearings characteristic extraction under different loading conditions

End effects of Local Mean Decomposition (LMD) are regarded as a typical problem leading to a distorted decomposed waveform and interfere with the extraction of characteristics. This paper proposes a novel self-adaptive point extension approach based on a Bidirectional Long Short-Term Memory (Bi-LSTM) regression network to eliminate this problem. This approach divides the existing samples into two parts and conducts two training processes, in which the first-training obtains the optimal network initialization parameters and the second-training gets the final extension network to identify the correct extremum. A simulated signal is used to demonstrate the advantages of the proposed approach over BPNN, LSTM, and characteristic segment approaches. The standard LMD method is combined with the proposed extension to form an improved LMD algorithm (ILMD). Finally, ILMD is applied to three experimental vibration signals which are collected from different loading conditions. The results demonstrate that ILMD can accurately extract failure and rotational characteristic frequencies of rolling element bearings with higher amplitude, and accordingly, the error caused by end effects does not influence the extracted information. (C) 2020 Elsevier Inc. All rights reserved.