The invention discloses a hybrid machine learning signal classification method based on PCA dimensionality reduction, comprising the following steps: Step 1, for linear frequency modulation signal LFM, binary phase shift keying BPSK signal, binary frequency shift keying 2FSK and four-phase For phase-shift keying QPSK signals, according to the instantaneous autocorrelation classification method, set the zero-crossing threshold and standard deviation threshold to separate the LFM signal, QPSK signal from other signals; step 2, the second-level classification, for the remaining signal BPSK The signal and the 2FSK signal adopt three characteristics of normalized amplitude duty cycle, normalized central instantaneous phase absolute value variance and normalized central instantaneous frequency absolute value variance, and use principal component analysis PCA algorithm to realize feature dimensionality reduction; steps 3. The objective function of the optimal classification is obtained by using the SVM classifier, and the distinction between BPSK and 2FSK signals is realized. The invention adopts the machine learning technology to realize classification, has high degree of automation and good classification effect.