Abstract: Traditional methods of fault detection in batch processes that are data-driven often need to make assumptions about the distribution of the process data. The models in those methods are not accurate in multi-stage partitioning, which have an impact on failure detection rate. A novel one-dimensional convolution autoencoder-Gaussian mixture model (1DC-AE-GMM) method is proposed in this paper to deal with this problem. This method does not require assumptions about the raw data. First, the original data are processed with equal length and scaling and the minimum reconfiguration error is used to train the deep neural network with convolution and multiple middle layers, which carries out phase division and feature extraction automatically and accurately in a non-linear way. Then, the Gaussian mixture model and clustering are set up on the coding layer of the network and the computational quantity of the model is reduced greatly while the feature is extracted. Finally, the global probability detection index is presented with a Markov distance and fault detection is realized. The results show that this method can effectively improve the fault detection rate by simulating experiments on a semiconductor etching process.