The application of machine learning method based on ultrasonic signals in tensile and shear performance evaluation of titanium steel explosive composite bars is studied. This paper proposes a probabilistic neural network (PNN) evaluation and classification method based on the eigenvalues of ultrasound signals. Firstly, 120 samples of workpiece are taken as the object to obtain the full sequence A-scan signals of water immersion ultrasonic testing. The signals are analyzed in time domain and improved covariance power spectral density estimation. Six characteristic values are used as PNN input: depth of the composite layer, reflection frequency of the upper composite layer, spectral energy, reflection frequency of the lower composite layer, spectral energy, and attenuation of the secondary reflected wave on the lower surface. Then, a tensile test is performed on the workpiece sample to obtain the tensile and shear strength values as the PNN output. Finally, a classification training model is established based on 96 sample characteristic signals and tensile and shear strength values. The remaining 24 samples are used as the test set, and the tensile and shear strength values of these samples are classified and predicted. Experimental results show that the accuracy rate of 24 consecutive predictions is 94.35%. This article finds new ideas for the fast and full coverage evaluation of the tensile and shear properties of titanium steel explosive composite bars.