Table 4 Comparison of the proposed work with other parameter-based transfer learning methods used for domain generalisation for efficient machine fault diagnosis.

He et al.³⁹

Improved deep transfer autoencoder

Raw time-series vibration data

Gearbox fault dataset

Same source and target domain datasets but with varying operating condition

90% accuracy when target domain has significant changes in operating conditions

Lu et al.⁴⁰

AlexNet

Spectrograms

Bearing dataset (CWRU)

Source domain: non-manufacturing data (ImageNet), target domain: manufacturing (bearing data)

99.7% accuracy when target domain has significant changes in operating conditions

Wang et al.⁴¹

VGG19

Scalograms

Bearing dataset (CWRU)

Source domain: non-manufacturing data (ImageNet), target domain: manufacturing (bearing data)

93% accuracy when target domain comprises different fault types and severities

Li et al.⁴²

CNN

Raw time-series vibration data

Bearing dataset (CWRU), gearbox fault dataset

Source and target domain datasets with variour operating conditions and distinct fault components

\(>90\)% accuracy when all CNN layers are fine-tuned with target domain data (lower accuracy when only last layer is fine-tuned)

Chen et al.⁴³

CNN

Raw time-series vibration data

Bearing dataset (CWRU), gearbox fault dataset, and lab generated bearing data

Source and target domain datasets across different machines and varying operating conditions

99% accuracy with all training samples (accuracy number drops with reduced number of training samples)

This work

Light-weight CNN

Scalograms

Bearing dataset (CWRU), run-to-failure bearing data (IMS), lab generated data with low precision sensor (LPS)

Source and target domain datasets across different machines and sensors of different precision

98% accuracy for high precision, unlabelled sensor data in target domain and 96.6% accuracy with labelled, low precision sensor data in target domain, \(>95\)% accuracy even with halving number of samples used for retraining