A stochastic configuration network (SCN) introduces inequality constraints to limit the assignment of input weights and biases. The network can approximate arbitrary mathematical function and data model as the number of hidden nodes gradually increases. In the process of SCN construction, the properties of the network itself and the ill-posed and ill-conditioned problems of the sample data may cause over-fitting of the network model. This study proposes an improved SCN model based on the Dropout technology, called Dropout-SCN, to improve the recognition accuracy of the network model by adaptively constraining the output weight distribution. We then perform a verification using optical fiber data. Compared with the traditional SCN and L2 norm regularized SCN models, the Dropout-SCN model has a lower test error, which effectively prevents the network over-fitting problem and improves the recognition accuracy of the intrusion signals in the optical fiber pre-warning system.