Industrial cyber-physical systems are critical infrastructures vulnerable to cyber-attacks. Anomaly and intrusion detection are widely used approaches to enhance the security of these systems. Existing detection methods can be categorized into two families. The first family detects only known attacks. The second family overcomes this limitation but often has a high false positive rate and a long detection time. This paper investigates the second family using side-channel leakages, particularly sound, for high-accuracy detection of intrusions and anomalies in various industrial systems. Despite sound signal's advantages, such as low-cost equipment, minimal computational requirements, and noninvasive measurement. Current sound-based anomaly detection (SAD) methods face challenges such as sensitivity to background noise, unbalanced sound data, computational costs, and detection accuracy. To tackle these issues, we introduce robot-arm sound dataset (RASD) and present a real-time sound-based anomaly detection for industrial systems (SADIS) approach that uses a simple and efficient method to fingerprint expected sound data with reduced dimensions. It employs an autoencoder (AE) for data classification and utilizes the Mahalanobis distance (MD) as an anomaly-scoring function, enhancing detection performance. Our experiments demonstrate that the SADIS approach achieves an average attack detection rate of over 96%, with a detection time of less than 1 second and low computational costs.