Simulation on neural networks for DUET-based delay estimation of abrasive debris signal separation

Aliasing signals generate when two or more abrasive particles pass through an inductive debris detection sensor simultaneously, which will lead to an accumulative error for further diagnosis and prognosis of machinery equipment. The degenerate unmixing estimation technique (DUET) is an effective method for dividing aliasing signals into original sources and getting a more accurate number of the superimposed wear debris. By using the two-dimensional weighted histogram, two key parameters are estimated, which directly influences the following accuracy of source separation. To promote the precision of the parameter delay, neural networks methods including feedforward, cascade-forward, auto encoder (AE), sparse auto encoder (SAE), convolutional neural networks (CNN) are attempted and compared by using the simulative data. Different data structures are used for the testing and the result shows that the delays give the lowest mean square error (MSE) with the two-layer CNN.