A Denoising Based Autoassociative Model for Robust Sensor Monitoring in Nuclear Power Plants

Table 2

Summary of DAASM hyperparameters.

Hyperparameter type	Tested hyperparameter values	Successful hyperparameter against optimum model
Pretrained DAE units	3	3
Network architecture	: input layer neurons; : bottleneck layer neurons; : output layer neurons; : encoder cascade; : decoder cascade; : number of layers; : neurons per layer
Learning rate for unsupervised pretraining		[,
Learning rate for supervised training	Scheduled learning rate based on training error monitoring:
Mean pretraining error for each hidden layer	Corresponding to minima observed during cross validation
Weight decay,
Momentum,
Input corruption level,	Corrupted input fraction: Gaussian corruption (% of sensor’s nominal value):	Input fraction: 25–35]%; Gaussian noise level: [0.10–0.25]
Dropout fraction in DAE-3		0.1