The data preprocessing, feature extraction, classifier training and testing play as the key components in a typical fault diagnosis system. This paper proposes a new application of extreme learning machines (ELM) in an integrated manner, where multiple ELM layers play correspondingly different roles in the fault diagnosis framework. The ELM based representational learning framework integrates functions including data preprocessing, feature extraction and dimension reduction. In the novel framework, an ELM based autoencoder is trained to get a hidden layer output weight matrix, which is then used to transform the input data into a new feature representation. Finally, a single layered ELM is applied for fault classification. Compared with existing feature extraction methods, the output weight matrix is treated as the mapping result with weighted distribution of input vector. It avoids wiping off “insignificant” feature information that may convey some undiscovered knowledge. The proposed representational learning framework does not need parameters fine-tuning with iterations. Therefore, the training speed is much faster than the traditional back propagation-based DL or support vector machine method. The experimental tests are carried out on a wind turbine generator simulator, which demonstrates the advantages of this method in both speed and accuracy.