In order to improve the survivability of optical networks, this paper proposes a fault service recovery algorithm based on service delay sensitivity and alarm feedback measurement. The algorithm first applies the link multi-fault location model based on hybrid evolutionary neural network to accurately locate the fault location from a large number of alarm information. Then, the time delay tolerance of the faulty business is evaluated, and the recovery sequence of the faulty business is arranged based on this. In addition, the algorithm also evaluates the reliability of each link according to the type and quantity of alarm information on each link, and gives priority to the link with higher reliability when choosing the recovery path for the fault service. The simulation results show that the fault location accuracy of this algorithm is as high as 91%, the relative recovery delay of fault service decreased by 48.83% and the reliability of recovery path increased by 6.05% on average.