Abstract: To solve the serious signal aliasing problem (due to limited spectrum resources of multi-domain communication, and mutual interference among signals), and deficiencies in existing signal recognition methods (in parameter setting, feature weight allocation, high feature dimension and generalization ability), a signal recognition method based on adaptive neural network algorithm is proposed in this work． This method uses Gammatone time-domain for high-frequency component filtering and gain control, to distinguish useful from useless signals and to suppress aliasing． Choi-Williams Distribution (CWD) and Multi-synchro-squeezing Transform (MSST) are used to obtain the signal time-frequency distribution map． Short-time Fourier transform was combined with window function for synchronous compression, to enhance energy aggregation of time-frequency features． Fisher Score algorithm was used to construct a strong multilayer perceptron (MLP)-based AdaBoost classifier, to simplify the feature space and to extract intrinsic modal features． Misclassification weights are computed, prediction errors are minimized, expansion and corrosion type edge detection operators are combined to locate signal mutation points in operator framing signals, time-frequency distribution graph was subject to adaptive learning to output recognition results． Data show that this method could effectively eliminate redundancy and noise of original seismic hazard signals, time-frequency distribution of the obtained signals is consistent with the actual data． This method is able to significantly distinguish singular spectral eigenvalues of the reconstructed signals under different working conditions, to accurately identify signal mutations at 0.5 s, 2.7 s, and 3.7 s． The present method exhibits high recognition accuracy for different types of seismic waves, shows good filtering and anti-aliasing abilities, provides effective technical support for high-precision signal recognition．