Neural Network Integrated With Classical Signal Processing Techniques for Generating Optimized Square Envelope Spectrum and Its Application in Machine Condition Monitoring

Optimized square envelope spectrum (OSES) is a newly proposed potential tool for machine condition monitoring. However, OSES lacks an adaptive noise cancellation module, which makes it unable to handle strong noise datasets. The original version of OSES suggests performing bandpass filtering on the dataset in advance to eliminate noise interference. However, the selection of parameters (cutoff frequencies) for bandpass filters heavily relies on expert experience, which greatly limits the flexibility of OSES. To solve this problem, we attempted to integrate a finite impulse response (FIR) filtering module into the OSES algorithm and construct a new OSES generation method with adaptive filtering capability. In addition, we have designed and implemented a neural network version for OSES generation by the networking of classic signal processing techniques (such as filtering, Hilbert envelope transform, and Fourier transform). During the training process, the new network can automatically learn the preprocessing filter and generate the OSES, eliminate the need for manual determination of bandpass filter, and expand the use scenario of OSES. In addition, 3D-OSES is constructed using an online training strategy for long-term health monitoring. The neural network can be trained without fault samples, and the diagnosis results can be confirmed by identifying the fault characteristic frequency (FCF) and its harmonics in the OSES. Results demonstrate that the new OSES generation method can accurately locate the differential spectral structure between the health sample and the fault sample, stably generate the OSES, and improve the ability of the OSES to handle complex practical engineering problems.