In the actual operation of machinery, the normal data are abundant and the fault data are rare. The recognition rate of the minority class is low when the convolutional neural network is used to process these imbalanced data. To solve this problem, an imbalanced fault diagnosis method for machinery based on the cost sensitive convolutional neural network is proposed. Firstly, the intrinsic performance state knowledge is achieved in raw data of machinery through multilevel convolution and pooling operations. Then, the intrinsic performance state knowledge is mapped to mechanical health by fully connected layer. Finally, the cost sensitive loss function is used to set a large cost to the misclassification of the minority class. The effective classification of mechanical imbalanced data is realized. The proposed method is evaluated by tool monitoring data and bearing monitoring data with different imbalanced ratio. Compared with the traditional convolutional neural networks, experimental results show that the recognition rate of minority samples in imbalanced datasets has been improved by more than 22%.