Abstract: A convolutional neural network is constructed to solve the problem of detecting a sinusoidal signal in a non-uniform time series with periodic gaps, also containing Gaussian white noise. It is shown that this model approximates the posterior probability of signal presence, and the neural network model runs significantly faster than classical Bayesian numerical modeling. Some pitfalls of this method are discussed, in particular the problem of specifying initial approximations for the neural network weights, as well as possible development prospects.