The gear degeneration evaluation technology plays an important role in maintaining safety of various equipment operation. The traditional gear degeneration evaluation methods are susceptible to feature extraction and data pre-processing tricks. The methods based on the generative model use raw observations to perform evaluation. And the human factors can be effectively reduced. However, traditional generative models, such as variational autoencoder (VAE), are limited by poor performance in marginal probability density evaluation. In this study, multivariate invertible deep probabilistic learning (MIDPL) is proposed, which can establish the connection between a given distribution and an unknown observation distribution by stacking learnable invertible transformation. The marginal probability density evaluation of the multi observation sequence can be realized through the given distribution. The proposed MIDPL model is evaluated by gear degeneration experiments. Compared with VAE, the evaluation errors of MIDPL for gear pitting dataset and gear breaking dataset are reduced by 30. 92% and 69. 25% , respectively. The proposed MIDPL can achieve more accurate and stable degeneration evaluation.