It is found that the mechanical properties of welded joints is mainly related to the welding heating input and complicated mutual effects in multi composition welding material. According to this principle, a soft sensor model based on the BP neural network (BPNN) is designed. The soft sensing BPNN has been constructed by use of the BP network toolkit of the Matlab software. The dates of five kinds of Low-Ally steels mechanical properties under different welding thermal cycle which used to train and test the BPNN have been obtained by welding thermal simulator. In this way, the soft sensing model for welding joint mechanical properties testing has been established. The BPNN has been tested by use of testing samples from the obtained data. By comparing the predicting date and the actual date, it shows that the soft sensing model has acceptable precision. The soft sensing method provides a new way for predication of mechanical properties of welding joints.