Aiming at the issues of torque value monitoring and variation trend prediction in the process of ship lift synchronous shaft movement， a non-contact measuring method based on strain sensor is designed for achieving synchronous shaft torque real-time and on-line monitoring. Then on the basis of historical monitoring data， the direct-current component of torque signal with variational mode decomposition algorithm is extracted， in order to reduce the amount of torque prediction calculation and make the torque prediction model more suitable for practical working condition. Aiming at the issue of indistinguishable starting point and inconsistent monitoring timing sequence of the synchronous shaft movement process， related fuzzy entropy detection method is proposed. Moreover， by adopting long short-term memory （LSTM） algorithm， the prediction model of synchronous shaft torque is established. Furthermore， the prediction model is verified by comparing with real historical monitoring data. The result shows that using fuzzy entropy to detect the starting point can ameliorate the temporal difference of each sample， and then the accuracy of the prediction model is improved effectively. For all torque monitoring points， the prediction accuracy can be improved by at least 27.5% compared with the foundational threshold judgment prediction method. And at the gearbox connection places， which are the most complex mechanical structure and working conditions of synchronous shaft system， the accuracy of torque prediction is improved by at least 42.9%. So the prediction model can truly predict the variation of synchronous shaft torque. The whole research has a good engineering application value.