In Φ-OTDR distributed fiber-optic intrusion detection system, it is essential to obtain sensing results to propose the solution strategy, reduce staff casualties, and property loss in time. Simultaneously, the data-driven deep learning signal classification method has the features of high accuracy and robustness. However, a large number of training samples are required to achieve a better result. To solve this problem, we initially propose a data enhancement method based on noise fusion, which extends the original unbalanced small-sample data into a model-satisfying dataset by designing a spatial correlation noise fusion enhancement (SCNFE). Then, for the signaling network inputs with different time scale, an improved Gramian angular field (GAF) image transformation method is proposed to obtain the samples that satisfy the input scale of the network by introducing the image transformation in the Gram's angle field. Finally, to satisfy the real-time intrusion signal classification of the network on distributed acoustic sensing (DAS) devices, a MobileNetV3-DAS classification network is proposed by combining the efficient channel attention (ECA) attention mechanism. Experiments demonstrate that the spatial correlation noise fusion method proposed in this study can reduce non-equilibrium rate to 1.09. Compared with MobileNetV3, the improved method reduces the model weight and inference time by 29.80% and 10.26%.