The detection of weak fault signals is realized by applying non-resonant parametric excitation to Lu chaotic system. Periodic signals far larger than the characteristic frequency of the system are taken as input, the sharp change of the system attractor is observed, and the existence of weak signals is detected. The conditions that the parameters of the detection system should meet are deduced by using the averaging method and Lyapunov method. Numerical research shows that the detection system can quickly reach a stable state. When the amplitude of the detected weak signals changes slightly, the attractor state changes abruptly, and the system is sensitive to the input. Through the "butterfly effect" of the system, feature signals that are difficult to extract by traditional methods can be detected. And the system has strong anti-interference performance.