Randomly select 25% of the training set data as the verification set, set the model loss value as Mae, and get the loss curve of the training set and the verification set, as shown in Figure 1
It can be seen from the figure that the convergence speed of the training set is faster than that of the verification set. With the increase of the number of iterations, the loss of the model decreases and finally tends to be stable. It is proved that the network structure is designed properly and the model parameters are set reasonably. The prediction results and residual variation trend of training set are shown in Figure 2 and figure 3.
It can be seen from the figure that the prediction effect of the training set is better, and the difference between the actual temperature value and the predicted value is mostly concentrated between – 1 ℃ and 1 ℃. The width of the sliding window is determined to be 150, and the residual error is analyzed to calculate the mean and standard deviation of the residual error in the window. As the window moves, the change trend is shown in Figure 4 and figure 5.
It can be seen from the figure that under the normal state of the fan, the mean value of the residual error predicted by the model is between – 0.408 ℃ and 2.292 ℃, and the maximum absolute value is 2.292 ℃. The standard deviation of the residual is between 0.072 ℃ and 0.966 ℃, and the maximum is 0.990 ℃. According to the formula, the alarm threshold of mean value is ± 4.584 ℃, and the alarm threshold of standard deviation is 1.932 ℃.