Deep Learning vs. Classical Modeling of Processes for Fault Detection in Industrial Heating-Cooling Systems

Faults in heating-cooling systems can often be observed by changes in temperature. Such faults can be detected and identified by modeling thermodynamic behavior. In classical models, physical equations with fixed or trainable parameters are used to model this behavior. They are limited in non-linear complexity and the number of parameters to be estimated. They also usually require the involvement of expert knowledge. In this paper, a deep learning approach is presented for modeling thermodynamic behavior without explicitly modeling the physical properties. The modeled artificial neural network (ANN) can predict the temperature based on other influencing variables. A comparison with a mathematical-physical model (MM) shows that the ANN can reproduce temperature changes similarly good when sufficiently data is available. With increasing prediction windows, the ANN even outperformed the MM model for most states. Both models can detect certain heating faults by comparing the measured and predicted temperatures. Finally, we demonstrate the diagnostic capabilities of our methods by injecting a fault into the system.