It is difficult to obtain labeled fault data while there are a multitude of unlabeled data available in industrial process

so how to utilize data information effectively is an important focal point in the field of fault diagnosis. A new semi-supervised learning method

reconstruction-based semi-supervised extreme learning machine (RSELM)

was proposed for more sufficient data mining and information usage. Compared to traditional semi-supervised ELM

RSELM replaced random input weight in hidden layer with output weight

which was obtained by ELM auto-encoder (ELM-AE)

such that data feature was extracterd more effectively. Since data could be reconstructed linearly by its neighbors

a self-adaptive reconstruction graph of neighboring data in combination with connection weight of optimal labeled data better reflected data structure information. A novel objective function preserving local structure information was further built to train classifier effectively. Simulation experiment on standard datasets and TE process demonstrated effectiveness of the proposed algorithm.