Deep Learning-Driven Data Curation and Model Interpretation for Smart Manufacturing

Chinese Journal of Mechanical Engineering

Table 1 Representative techniques for data curation

Data denoising		Ref.	Outlier detection		Ref.	Data imputation		Ref.
Projection-based	Local geometric projection	[36]	Data-level	Autoencoder	[48, 55]	Time series	Recurrent neural network	[62, 65, 66]
Frequency-based	Empirical model decomposition; Wavelet transform	[37‒40]	Data-level	Autoencoder	[48, 55]	Time series	Recurrent neural network	[62, 65, 66]
Noise-assisted	Stochastic resonance	[41, 42]	Model-level	Probabilistic neural network; Temperature scaling; Input perturbation	[56‒61]	Image	Convolutional neural network; Hybrid approach	[67, 68]
Data-driven/hybrid	Generative prior; Unrolled optimization	[46, 47]	Model-level		[56‒61]	Image	Convolutional neural network; Hybrid approach	[67, 68]

Data balancing		Ref.	Data annotation		Ref.
Data interpolation	Synthetic minority over-sampling technique	[75]	Image annotation	Fully convolutional network; U-Net; Mask region-based CNN	[30] [79, 80]
Generative model	Variational autoencoder; Generative adversarial network	[76] [29]	Natural language processing	Word embedding; Transformer; BERT	[83‒88]