Fiber-optic distributed acoustic sensing (DAS) excels in high-quality seismic signal acquisition and detection but is often hindered by noise, significantly reducing signal-to-noise ratio (SNR) and impeding microseismic event detection. Moreover, continuous seismic monitoring campaigns generates huge data volumes. While numerous denoising approaches exist, they often demand substantial computational resources, hindering real-time implementation. We propose an unsupervised neural network to suppress random noise without requiring noisefree ground truth or prior noise characteristics. The network learns to extract features by masking random input traces and reconstructing the target using long-term coherence from neighboring traces. We explore hyperparameter optimization by varying input sample generation, activation functions, scaling methods, and the number of input traces. We evaluate the model by running a detection algorithm on FORGE data and achieve a 43% increase in event detection. We further exploit our algorithm in real-time experiments and achieved within a 90% processing time compared to the data acquisition rate with denoising implemented. Our approaches can be incorporated real-time data acquisition, effectively facilitate the screening and storing the data timeframe with useful information.