Abstract: After welding the large components of prefabricated building steel structures, ultrasonic testing of welds is susceptible to mixed noise interference, and its layer by layer attenuation characteristics lead to unstable signal quality, making it difficult to effectively extract detection features, inaccurate classification of multiple types of defects, and errors in weld defect detection. Therefore, a non-destructive testing method for welding seam defects of prefabricated steel structure components is proposed. Construct an ultrasonic transmission model, analyze the reflection, scattering, and other behaviors of ultrasonic waves during propagation in steel structures, consider key factors such as scattering attenuation, and obtain the ultrasonic attenuation coefficient of steel structural components. In response to the interference of instrument scattering noise and ultrasonic scattering noise in on-site signal collection, the wavelet decomposition method is introduced, and a soft threshold strategy is adopted to dynamically adjust the threshold by integrating the statistical characteristics of noise and the trend of scale changes, effectively improving signal quality and achieving defect feature recognition under noise interference. Using support vector machine to search for the optimal classification hyperplane in high-dimensional feature space, constructing a convex quadratic optimization model and introducing kernel function, and adopting a one-to-one strategy to achieve accurate classification of multiple types of defects. Based on the classification results, the relative position of the defect is determined by calculating the time difference between the emitted wave and the reflected wave of the defect, combined with the constant speed of ultrasonic waves, to clearly determine whether there are defects in the weld and achieve non-destructive testing of defects. The experimental results show that the proposed method has a signal-to-noise ratio of 109dB for weld seam detection signals, and the detected crack length is highly close to the actual value, indicating a significant advantage in detection performance.