Acoustic logging data plays an important role in horizon calibration and reservoir inversion. However, due to the limitations of equipment and geological environment, the actual acoustic logging curve is often distorted. To provide reliable data support for oil and gas exploration and improve the accuracy of reservoir prediction, a well-to-seismic joint inversion method based on parameter optimization residual network is proposed to reconstruct the distorted acoustic logging curve. Considering that the traditional artificial neural network cannot express the strong nonlinear relationship between well and seismic, this method uses the residual network (ResNet) in deep learning to build an intelligent inversion model. Through network design, parameter selection and model training, a better mapping expression between well and seismic can be found. Considering the characteristics of logging curve and the deficiency of MSE loss, a cost-sensitive loss function Fusion is designed to further improve the overall inversion accuracy of the model. Experiments are carried out on real seismic data and logging data, compared with the inversion results of Fully Connected Neural Network (FCNN) and Multiple Linear Regression (MLR), it shows that the accuracy of the acoustic logging curves inverted by the proposed method is higher, the correlation coefficient reaches 0.912, and the root mean square error is reduced to 13.399. Using the proposed Fusion loss to invert the acoustic logging curve, the correlation coefficient increases by 2.5%, and the root mean square error decreases by 17.4%.