Datadriven based fault detection method has become important means for the fault detection of practical industrial processes, however, in practical application it is often limited by the size of process historical data, so that it is difficult to achieve satisfactory fault detection accuracy. In this paper, aiming at this problem a sample space reconstruction strategy is proposed, which constructs the sample pairs of the same or different categories based on random sampling. While the data size is expanded, the strategy transforms complex classification modeling problem into the comparison problem of the similarity among the samples, which effectively reduces the complexity of the task and the amount of the data needed for modeling. Based on the reconstruction strategy, the siamese CNN is introduced and improved, a chemical industrial process fault detection method based on Multiscale Siamese Convolutional Neural Networks (Multiscale Siamese CNN) is proposed. The test results on the TennesseeEastman (TE) process dataset verify the effectiveness of the proposed algorithm. The test results show that the average fault detection accuracy of the proposed algorithm reaches 8966%, which is improved by 8% above compared with that of conventional datadriven fault detection algorithm.