Sliding window analysis of residual error prediction for fan gearbox model based on LSTM network

The LSTM model is trained by selecting the normal operation data of the unit. When the gearbox bearing is in normal working condition, the operation data of the unit is relatively stable, and the prediction error of LSTM model is relatively small. When the bearing has potential faults, with the increase of running time, the fault degree will be aggravated, and the bearing temperature characteristics will deviate from the normal working range, which will lead to the prediction error of LSTM model becoming larger. In order to avoid false alarms and continuously analyze the characteristics of prediction residuals, this paper uses sliding window analysis method to process the prediction residuals, which can effectively eliminate the influence of random factors on the variation of residuals.

Take a sliding window with width n and calculate the mean and standard deviation of the residuals contained in the window

Where: EI is the ith residual in the window.

The sliding window is used to analyze and process the residuals of the training set, and the maximum absolute value of the mean and the maximum standard deviation are obtained by calculation and statistics, which are recorded as xmax and Smax

In the formula, K1 and K2 are usually taken as 2. When the mean value and standard deviation exceed the alarm threshold, the fault alarm signal will be sent out.

Scroll to Top