Deep Multiscale Soft-Threshold Support Vector Data Description for Enhanced Heavy-Duty Gas Turbine Generator Sets’ Anomaly Detection

Table 1

Configuration of the main encoder backbone network parameters.

Layer	Algorithm structure	Parameters
Branch 1	Convolution, batch normalization, ReLU activation	Cov1d (133, 128, 3, 2, 1)
	Convolution, batch normalization, ReLU activation	Cov1d (128, 64, 3, 2, 1)
	Convolution, batch normalization, ReLU activation	Cov1d (64, 32, 3, 2, 1)
Branch 2	Convolution, batch normalization, ReLU activation	Cov1d (133, 128, 5, 2, 2)
	Convolution, batch normalization, ReLU activation	Cov1d (128, 64, 5, 2, 2)
	Convolution, batch normalization, ReLU activation	Cov1d (64, 32, 5, 2, 2)
Branch 3	Convolution, batch normalization, ReLU activation	Cov1d (133, 128, 7, 2, 3)
	Convolution, batch normalization, ReLU activation	Cov1d (128, 64, 7, 2, 3)
	Convolution, batch normalization, ReLU activation	Cov1d (64, 32, 7, 2, 3)
GAP layer	Global average pooling	C:32
Feature fusion layer	Feature concatenation	C:96
Soft thresholding activation layer	Soft thresholding activation module	C:96