Abstract: At present no measure is available to predict the trend of crack propagation online during ingot forming. To solve the problem, the soft sensing technique was applied to the semi-continuous electromagnetic casting process of Al-alloys, where the auxiliary parameters available to direct online sensing were chosen to develop a soft sensing model of cracks on the basis of multilayer feed forward neural network. The outputs from the model are the quantized values of crack, which means that the quantized values of ingot cracks are indirectly censored online and the trend of crack propagation is predicted. Experiment/simulation results revealed that the predicted values of crack from the model conform well the values actually measured, thus reflecting well the actual trend of cracks so as to make online prediction/control of cracks available.