Memory-guided reconstruction for generalized zero-shot industrial fault diagnosis

Abstract

Recently, zero-shot learning (ZSL) has emerged as a promising method in the industrial fault diagnosis. However, restricted by the strong bias problem, unseen class faults tend to be classified as seen class faults in the generalized zero-shot learning (GZSL) task. To address this issue, a novel method called memory-guided reconstruction (MGR) is proposed for generalized zero-shot industrial fault diagnosis. In MGR, memory prototypes of seen classes are first learned by a self-organizing map (SOM) and stored in a memory module. During the training, the encoding of a sample is obtained from the encoder as a query. Instead of directly reconstructing from this query, a support memory aggregated from relevant memory prototypes of the query is delivered to the decoder for reconstruction. A specific memory alignment matrix is designed to measure the consistency between the query and support memory. At the test stage, unseen classes tend to produce higher reconstruction errors than seen classes because the support memory is acquired from seen class memory prototypes. A new “identify-classify” learning paradigm is adopted: identify the domain (i.e. seen or unseen) of the test sample based on the strengthened reconstruction error, and further classifythe sample within the identified domain. Extensive experiments on the benchmark dataset demonstrate the significant superiority of MGR. Moreover, MGR achieves competitive results compared to supervised learning methods. The code of MGR is available at https://github.com/htz-ecust/memory-guided-autoencoder.