To improve the road state recognition efficacy and the ride comfort control performance of the active suspension system, a model predictive control (MPC) based active suspension control method is proposed based on road state recognition by the residual convolutional neural networks with split-attention (ResNeSt). First, the road state recognition algorithm scheme with respect to the active suspension control is established by the ResNeSt network considering the multi-path split attention mechanism. The proposed network is trained and tested via utilizing the cross-entropy objective loss function and the AdamW gradient descent algorithm. Then, the MPC-based active suspension control algorithm is developed based on road state recognition. Specifically, the prediction model is derived from the discrete state space equation and the objective function is constructed by the performance indexes of the prediction outputs and the control inputs. Moreover, the weighting matrix values are determined by the recognized road surface results. The objective function of the MPC is transformed into the quadratic programming (QP) to obtain the global optimum while satisfying the constraints. Comparative studies of the passive suspension and the LQR control are performed to demonstrate the effectiveness of the proposed architecture. The results show that the ResNeSt network is capable of identifying different kinds of road states with guaranteed precision and computational performance. The proposed active suspension control algorithm provides satisfactory real-time and transient active control of the suspension based on different road conditions. Compared with the LQG algorithm, quantitative results for the mean RMS values of spring mass acceleration, suspension dynamic deflection and tire dynamic load are reduced by 36.56%、32.99% and 36.28%.