A Novel Incipient Fault Diagnosis Method for Analogue Circuits Based on an MLDLCN

Incipient faults in analogue circuits used in complex electrical systems are hard to diagnose due to weak fault features. To improve the reliability and maintainability of analogue circuits in complex electrical systems, a novel incipient soft fault diagnosis method for analogue circuits based on a multilayer dictionary learning and coding network is proposed, including feature preprocessing, linear dictionary feature encoding, and classification modules. In the first module, time–frequency analyses are performed using continuous wavelet transforms to demonstrate the spectrum maps of the fault signals, while scale-invariant feature transforms are used to enhance local features and obtain the keypoint descriptors of the time–frequency spectrum. In the second module, fault features are obtained by locally constrained linear coding (LLC) method using complete dictionaries from the keypoint descriptors acquired in the previous module, which are captured by linear combination of several adjacent atoms in the dictionary learning. To address the limitations of single-layer dictionary learning methods in complete extraction of fault features, a multilayer learning method is used to get richer fault information and improve the diagnosis accuracy. Finally, the linear output features are captured through pooling and fully connected layers. In the third module, the linear features acquired in the second module are quickly classified with simple linear classifiers. The experimental results demonstrate that the proposed method outperforms existing fault diagnosis methods. In order to verify the effectiveness of the proposed method in analogue circuit fault diagnosis, the Sallen–Key filter circuit and four-op-amp biquadratic filter circuit, which are widely used in the field, are selected as experimental circuits in this paper. Specifically, when the component fault values are offset by 20% of their nominal value, the proposed method achieves accuracies of 99.20% for the Sallen–Key bandpass filter circuit and 98.48% for the four-op-amp biquadratic filter circuit.