Hybrid intrusion behavior often only occurs in one or more local locations, and there is a certain degree of aggregation in time, which cannot capture its complex features well. Therefore, a real-time detection method for hybrid intrusion behavior in fiber optic sensing networks is proposed. Using the average zero crossing rate and short-term energy as indicators to segment a certain segment of the signal, reducing the accumulated processing delay, and extracting fiber optic sensing signals that may have intrusion behavior. Further extracting signal features through high-order spectral analysis, sample entropy analysis, and singular value analysis, constructing and utilizing multi-layer gradient descent method to train multiple deep neural networks, and inputting the extracted features into the corresponding deep neural networks, and the mixed intrusion behavior detection results are output through the Softmax function. Finally, the improved D-S evidence theory is used to correlate and fuse the detection results output by each deep neural network, implement real-time detection of hybrid intrusion behavior in fiber optic sensing networks. The experimental results show that the proposed method has more accurate intrusion behavior detection results, lower memory and CPU usage.