When high-speed acquisition systems handle data streams reaching tens of GSa/s, the limitations of real-time processing speed prevent the system from detecting occasional abnormal signals in real time, leading to signal omissions. Traditional anomaly detection relies on the a priori characteristics of the signal. However, these methods have low capture efficiency for episodic abnormal signals with unclear characteristics and irregular morphology. Thus, this article proposes a real-time anomaly detection method using empirical mode decomposition (EMD) to improve the system′s ability to capture anomalous signals. Firstly, the feature point extracted based on edge features is used as the start point of EMD, reducing the complexity of anomaly detection. Secondly, the non-noise intrinsic mode functions obtained from the EMD are used to reconstruct the normal signal template, and anomaly detection is carried out based on the degree of match between the test signal and the normal signal template. Finally, a parallel EMD is implemented in hardware to improve anomaly detection efficiency. By detecting anomalies in the modulated signals, the real-time anomaly capture rate of the proposed method is 95%, which represents a significant improvement over the traditional method.