Leakage Detection in Pneumatic Systems with Machine Learning and Upstream Single-Point Signals

Pneumatic systems are widely utilized in industrial manufacturing plants. Leakage is the most common fault and the dominating way of energy waste in pneumatic systems. Due to the low investment cost and high reliability of pneumatic systems, it is significant to detect leakage faults with a minimal number of cheap sensors. In this study, the feasibility of locating the leakages in 11 positions of a typical pneumatic system is verified with machine learning methods and the measured signal at a single point upstream. Both external and internal leakages of different pneumatic components are considered. Feature extraction is conducted using a one-dimensional convolutional neural network (1D CNN). Various machine learning classifiers, including GPC (Gaussian Process Classification), SVM (Support Vector Machine), KNN (k-Nearest Neighbour), CART (Classification and Regression Tree), MLP (Multi-Layer Perceptron), and RF (Random Forest) are used for fault classification and comparison. Class Activation Mapping (CAM) is calculated for the visualization of decision-making processes. The results show that it is feasible and convenient to detect and locate the multipoint leakages in a pneumatic system by analysing signal measured at a sole upstream point with the help of machine learning methods. The methodology can be extrapolated to and applied in more complex pneumatic systems.