Aiming at nonlinearity, multi phases and 3D data matrixes in batch processes, a fault detection method using a batch image-based convolutional autoencoder is proposed. Process data of each batch is considered as a grayscale image and is input to the convolutional autoencoder(CAE) directly for representation learning. Data variation in each batch can be regarded as the texture change of the image. Information loss caused by 3D data unfolding to 2D is avoided. Meanwhile, variable correlation is effectively extracted using global modeling with no need to phase division. Convolution operation extracts local conjunction features, and using a autoencoder is an efficient way for unsupervised learning. Then the one-class support vector methold(OCSVM) is used to constructe monitoring statistic and calculate control limit for fault detection. By applying the proposed method on the Pensim simulation and recombinant human granulocyte colony-stimulating factor (rhG-CSF) fermentation process, the effectiveness is demonstrated.