Abstract: In order to reveal the nonlinear characteristics of typical defects in carbon steel welds, the welds of six carbon steel test blocks were detected by using ultrasonic phased array detector. Five kinds of typical defect signals were analyzed based on recurrence plot, box dimension, maximum Lyapunov exponent and approximate entropy methods. The results showed that recurrence plot of all the defect signals had complex structures with coexistence of periodicity and randomness, indicating that the propagation of ultrasonic wave in test blocks had nonlinear characteristics, which verified the feasibility of studying defect signals based on nonlinear dynamics theory. The defect signals had relatively stable box dimensions, indicating that the defect signals had fractal characteristics. The maximum Lyapunov exponents of the defect signals were greater than zero, indicating that the defect signals had chaotic characteristics. The difference in approximate entropy of defect signals indicated that the complexity of various defect signals was different. The box dimension values of defect signals were relatively similar which indicated that merely using box dimension cannot accurately classify defect signals. The use of maximum Lyapunov exponent and approximate entropy can effectively identify longitudinal cracks, but the classification effect on the other four types of defect signals was not satisfactory. In order to effectively identify defect signals more accurately, it is necessary to integrate multiple nonlinear characteristic indicators for comprehensive judgment.