Abstract: Aim To diagnose sensor soft and hard failure and recover the failed sensor signal in control system. Methods\ A double level neural network including one main neural network(NN) and several local NNs was proposed. The main NN was responsible for diagnosing sensor failure, with its inputs as sensor signal at time t , and its outputs as sensor signal at time t+1 . And the local NNs were responsible for diagnosing sensor failure and signal recovery for the failed sensor signal, each local NN was corresponding to one sensor signal, with its output as the corresponding sensor signal at time t+1 , and its inputs as the remaining sensor signals at time t . All NNs were trained with on line learning method, that is, working after on line learning. Results and Conclusion\ The proposed method has the abilities of on line learning, low false alarm ratio and identifying multiple sensors failure. The simulation results for the sensor failure diagnosis in a air cushion vechicle show that the method is very useful.