Modern engineering products are becoming more and more complex, sophisticated and intelligent in physical mechanisms and functional structures, which may result in more failure modes and higher failure probability of the product. Therefore, reliability analysis is critical for modern engineering products, and it often encounters epistemic uncertainty due to the lack of reliability information. Evidence theory is an important model for dealing with reliability analysis with epistemic uncertainty. At present, reliability analysis based on evidence theory has become an important research direction in the field of reliability.

This paper proposes a deep learning method to solve the reliability analysis problem based on evidence theory. The stacked autoencoder is constructed by extracting the spatial location features of the sampled focal elements to achieve high-precision classification of the remaining focal elements, and the confidence interval of reliability are calculated efficiently according to the classification results of the focal elements.  The core innovation of this method is to use the deep learning model to classify focal elements to solve the reliability analysis problem based on evidence theory, and a spatial location features extraction method of focal elements which takes into account the features correlation and integrity is proposed to ensure the classification accuracy and efficiency of the deep learning model. The efficiency and accuracy of the proposed method were demonstrated using numerical examples.