Pneumatic systems are widely used in industrial manufacturing sectors. With the increasing penetration of intelligent and sustainable manufacturing, energy efficiency and fault diagnosis are being more and more significant. There is no exception even more urgent for pneumatic systems. In this study, a low-cost fault diagnosis concept for pneumatic systems is proposed by introducing exergy and machine learning. This concept is preliminarily verified in a typical simple pneumatic system with two parallel-installed cylinders. Stacked Auto-encoder (SAE) and one-dimensional convolutional neural network (1D CNN) are used for feature extraction of pressure, flow, and exergy data. Various machine learning algorithms, including Gaussian Process Classification (GPC), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Classification and Regression Trees (CART), Multi-Layer Perceptron (MLP), and Random Forest (RF) are used for learning and classifying different levels of external, internal, and compound leakage faults. Besides, the gradient-weighted class activation maps (Grad-CAM) are generated for interpreting the excellent performance of 1D CNN. The results show that it is feasible to accurately diagnose faults of multiple downstream cylinders with only one upstream sensor, thereby achieving low-cost fault diagnosis. Compared with pressure and flow data, the exergy data performs higher stability in accuracy and is less sensitive to different algorithms. Thus, we extrapolate that it is highly possible to implement fault diagnosis of complex pneumatic systems by deeply analyzing the energy data collected in energy management systems.