Gas compressor stations can maintain the natural gas pressure in long distance pipelines. Gas leakages are classified as category 1 hazards and pose a significant risk in compressor stations. However, the existing leak detection technologies are unsuitable because of the slow response and high cost. Therefore, this paper presents a gas leak detection system based on acoustic waves and deep learning. Specifically, an explosion-proof microphone array with 30 channels is designed and installed in a compressor station. Accordingly, a multi-channel frequency Transformer (MCFT) is proposed to extract useful information from acoustic waves and classify leak conditions. Experiments are performed using a dataset (10 categories) collected in the compressor station. The results reveal that the accuracy and leak detection rate reach 99.09% and 99.98%, respectively, while the false alarm rate declines to 0.2%. Compared with the existing state-of-the-art deep learning methods, the proposed MCFT exhibits significant advantages when applied to a real-world dataset. The robustness and efficacy of the proposed system are demonstrated via sensitivity studies using a number of microphones and hyperparameters of MCFT. A real-time detection scheme further validates that the proposed system can provide fast gas leak detection and ensure the process safety of pipeline transportation.