Hybrid attribute conditional adversarial denoising autoencoder for zero-shot classification of mechanical intelligent fault diagnosis

Abstract

Data-based intelligent fault diagnosis method is a research hotspot in modern mechanical systems. However, due to practical limitations, fault samples under all working conditions cannot be obtained, which would cause the data-based model lack of particular training data, resulting in unsatisfied testing performance. Therefore, zero-shot classification of mechanical intelligent fault diagnosis is a very practical work. Inspired by the zero-shot learning method, hybrid attribute conditional adversarial denoising autoencoder (CADAE), which uses hybrid attribute as condition, is proposed to solve the zero-shot classification problem. CADAE consists of three network modules: an encoder, a generator and a discriminator. The discriminator is applied to control the data distribution of hidden layer encoded by the encoder, and we add hybrid attribute condition into hidden layer to control the reconstruction process of generator. Finally, the generator module of the trained CADAE would be used to generate samples to train a classifier for missing classes. The proposed method is verified with three datasets under different data missing conditions. The results show that the proposed method could effectively solve the zero-shot classification problem with high classification accuracy exceeds 95%.