Abstract: Aiming at the problems of the weakness of echo, signal aliasing and the influence of noise, which make it difficult to detect the defects of the flip chip accurately, a high-frequency ultrasonic defect detection method based on liner frequency modulated (LFM) excitation and autoregressive spectrum extrapolation (ARSE) is proposed. A simulation model for high-frequency ultrasonic defect detection is established. Based on the response of ultrasonic probe, the LFM signal is designed as the probe excitation signal, and the pre-modulated LFM signal is used as the reference signal to compress the echo signal, suppress the noise, and increase the signal-to-noise ratio by 12 dB on average. Using the modified covariance method, the autoregressive coefficients of effective frequency band can be calculated. The effective frequency band of pulse compressed ultrasonic signal is extrapolated to separate the aliasing echo signal. The B-scan image of the flip chip is acquired through parametric scanning, followed by pulse compression to reduce noise in the B-scan matrix. Autoregressive coefficients are then computed and used for spectrum extrapolation. Additionally, a moving average filter is applied to further suppress noise bright spots. Compared to the simulation model of intact flip chips, as well as those with defects such as crack and bubble, the proposed method significantly enhances the time-domain resolution of B-scan images while maintaining good performance even under −4 dB noise interference. By considering the sound velocity of the material, the peak echo time of the B-scan matrix can be extracted to calculate the embedded depth of defects. The calculation error for determining the embedded depth of crack and bubble models of flip chip is less than 5%, enabling the identification and localization of solder bump defects.