Because the different batch of batch processes has different raw materials and changeable control conditions, measurement data may lead to drift and it’s difficult to obtain complete sampling data at any time during the reaction process. So a multi-way kernel T-PLS (MKT-PLS) algorithm was proposed to improve the fault diagnosis accuracy of batch processes. This algorithm firstly unfolds three dimensional process data matrix by sampling time sequence, then fills process variable data according to certain rules to form complete sample for data missing problem, so obtained appropriate data matrix is used to fault detection by MKT-PLS algorithm. Simulation results of Pensim V 2.0 simulation platform show that the fault detection rate of the proposed algorithm is higher than the other algorithm for detecting faults affect the quality of final products. This algorithm is more suitable to monitor the real-time batch process.