The invention relates to a fault-tolerant method and system for tactile communication based on hybrid prediction. When sending data, by predicting and compressing the data to be sent, only sending data frames whose relative deviation between the predicted value and the real value is greater than the dead zone parameter, so as to reduce The amount of data that needs to be sent; when receiving data, if missing data is encountered, the count is increased by 1, and the current count value n is compared with the previously received prediction threshold L. If n>L, zero-order hold is used The prediction algorithm predicts the missing data, otherwise the LSTM model is used to predict the missing data; if the data is received, the count value is cleared, and the predicted value at this time is the real value, and the newly received prediction threshold L is obtained at the same time. The present invention enables the receiving end to independently select a relatively conservative prediction algorithm by setting the continuous prediction threshold, reduces the impact of accumulated errors caused by continuous prediction, and further improves the stability of the tactile communication system.