Deep Learning Application in Mechatronics Systems’ Fault Diagnosis, a Case Study of the Demand-Controlled Ventilation and Heating System Ali Behravan, Roman Obermaisser, Mohammad Abboush University of Siegen, Siegen, Germany. Email: ali.behravan@uni-siegen.de Abstract—Mechatronics systems include a vast range of in- terdisciplinary area of electrical and mechanical systems such as heating, ventilation, and air-conditioning systems based on building automation systems are responsible to provide occupants a comfortable and productive environment in buildings. The demand-controlled ventilation system as an advanced control approach in smart buildings is used for the main goal of energy reduction. But, these kinds of systems because of their numerous components such as sensors and actuators are very prone to the faults. Arise of the faults, if they are not detected and diagnosed early, can lead to system’s performance degradation or extra maintenance cost and effort. Nowadays, introducing a suitable generic technique for fault detection and diagnosis is an utmost challenge. The contribution of this paper is to present a novel fault detection and diagnosis framework based on deep learning method using long short-term memory units for a case study of mechatronics systems, demand-controlled ventilation and heating system. This paper presents all the steps including data acquisition, data preprocessing, network model design, model optimization and network model evaluation. Ten types of faults in different classes as well as the healthy data are used to train and evaluate the performance of the designed network model. The results describe a high accuracy (97.4%) with via the designed deep neural network. Also, this study describes the methodology of selecting the optimum parameters of training process by analyzing the effect of each parameter on the training accuracy. Index Terms—Deep Learning, Deep Neural Network, Fault Detection and Diagnosis, HVAC I. I NTRODUCTION In recent years, the efficient use of energy in building sector motivates the researchers to focus on the new technologies such as building automation systems (BAS). Most of these new technologies are based on a mechatronics system platform as they include both mechanical and electrical components. Heating, ventilation, and air conditioning (HVAC) is the major part of BAS from energy consumption perspective. Therefore, this sector absorbed lots of researches during last years. The other parts can be lighting control system, automated home security, appliance control, smart water supply, and irriga- tion. The demand-controlled ventilation system (DCV) besides heating system, plays an important role in energy reduction by the automatic adjustment of ventilation according to the fresh air demand and environment temperature. Brandenmuehl et al. illustrate 15% to 25% of the HVAC system’s energy can be saved by setting the ventilation rates based on the maximum occupancy fresh air requirement [1]. Behravan et al. described thermal dynamic modeling and simulation of a heating system for a multi-zone office building equipped with demand con- trolled ventilation using MATLAB/Simulink [2]. These new technologies are developed based on wireless sensor and actu- ator networks (WSAN) that network components e.g. sensors, actuators, and controller in such network communicate through a wireless network. Authors in the last study, described all the details regarding the configuration of this type of WSAN in reference [3]. Several studies demonstrate arise of faults depending on their type and severity cause waste of energy ranging from 10% to 40%, performance degradation, or excess maintenance effort [4], [5], [6], [7], [8]. The faults may occur in components such as sensors and actuators or in network fabric. Faults can be defined in different aspects of data centric or system centric. Therefore, it is vital to detect and diagnose faults early using an optimized technique to prevent extra maintenance costs and efforts. There are several techniques for fault detection and diagnosis. Some main categories are data- driven methods, knowledge-based methods, and analytical- based methods that each category includes different methods in its sub-levels reported by Katipamula and Brambley [7]. During recent years, various researchers had demonstrated the success of deep learning models in the application of machine health monitoring. Hua et al. proposed an intelligent fault diagnosis network for variable refrigerant flow (VRF) systems using Bayesian belief network (BBN) by determining the suitable network structure and probability distributions of BBN and two typical faults had been taken into consideration for detecting and diagnosing (Refrigerant leakage and Refrig- erant overcharge) [9]. Lee et al. had described a method to detect and diagnose three abnormal states in the air handling unit (AHU) with the popular deep learning model, called Deep Belief Network (DBN), combined with Restricted Boltzmann Machine (RBM). That study had taken three types of faults into account which instances are a fan getting stuck, leakage in the cooling coil valves and low efficiency of heat exchanger. The accuracy of the results of the study’s fault detection and diagnosis was an approximate score of above 95% [10]. Guo et al. developed a fault diagnosis model based on the DBN and investigated its potential to diagnose the faults of the variable refrigerant flow system. The result of the study showed that the fault diagnosis correct rate of the optimized model is 97.7% [11]. In this study, the deep learning approach is developed using