Compressed sensing can effectively relieve the burden of data storage and transmission for mechanical condition monitor? ing. However， this method exists some problems such as low compression efficiency and slow signal reconstruction process in the application of fault diagnosis. In this paper， using the corresponding relationship between autoencoder and compressed sensing， a novel fault feature extraction method of the rolling bearing in the compressed domain based on the deep convolutional measurement network is proposed. For the problem that noise-free fault signal samples are difficult to obtain， a dataset construction method based on the fault mechanism is proposed. The model trained on this dataset is suitable for bearing signals under different working conditions A deep convolutional denoising autoencoder （DCDAE） is constructed， in which the number of layers is determined by the required signal compression rate and the frequency of the hidden layer corresponds to that of the original signal. The fully trained encoding sub-network of DCDAE， named deep convolutional measurement network （DCMN）， is used to compress the rolling bearing vibration signal instead of the traditional measurement matrix， and then the fault features are directly extracted in the compressed domain. The effectiveness of the proposed dataset construction method and the compressed domain feature extraction method are analyzed through the simulations. The rolling bearing experimental signals further verify that the deep convolutional measurement network trained by the proposed method has good generalization and can effectively extract fault features for fault diagnosis in the compressed domain with a compression ratio far lower than that of the traditional compressed sensing method.