Being non signal processing student I have a limited understanding of concepts.
I have a continuous periodic bearing faulty signal of 10 seconds duration sampled at 12 kHz. So I have 120,000 discrete time samples in 10 sec signal. I have utilized some machine learning techniques (Convolution Neural Network) to classify faulty signals to the non faulty signals.Based on RPM, one shaft cycle need 400 time samples approximately. Some research papers have considered the signal by segmenting by 1024 window size (non overlapping) to create 117 segments from 10 sec signal which are utilized as input to the network. So here in each of the 117 segments have approximately 2.553 shaft cycle (1024/400) information. I am confused because but no one mentioned why they are segmenting for such duration
My question: Is the information from one shaft cycle enough to provide all details ? Because number of discrete time samples in one shaft cycle are always below the sampling frequency. If I segment it for 10 shaft cycles (for example), is there any explanation for what difference will it make?
Normally to analyse what is inside the signal are there any minimum duration's standards defined ?