In the current complex battlefield environment, low probability of intercept radar signal has been widely used due to their large time-bandwidth product, strong anti-jamming performance, high resolution and low interception. It is difficult to identify the low probability of intercept radar signal by traditional radar reconnaissance methods. Based on the analysis of typical modulation of low probability of intercept radar, a radar signal classification and recognition method based on artificial intelligence was studied. Starting from the time-frequency characteristics of low probability of intercepted radar signals, a multi-window spectrogram analysis method was proposed. In this algorithm, Hermite function was used as the window function of spectrum analysis, and multiple window functions were also used for spectrum analysis. The effective signal with better aggregation is obtained, the noise interference is dispersed, and the time-frequency analysis characteristics of signal modulation characteristics are more obvious through this algorithm. On the basis of multi-window spectrogram, the idea of transfer learning was adopted, and ImageNet-VGG-f neural network was used to complete the task of signal classification and recognition. Experimental results show that the performance of the proposed algorithm is better than the traditional Choi-William distribution and Smooth and Pseudo Wigner-Ville distribution methods at low signal-to-noise ratio.