Anomaly Detection and Condition Monitoring of UAV Motors and Propellers Farhad Pourpanah, Bin Zhang, Rui MA and Qi Hao * School of Computer Science and Engineering Southern University of Science and Technology, Shenzhen, China, 518055 Emails: farhad.086@gmail.com, 11649159@mail.sustc.edu.cn, mar@sustc.edu.cn and haoq@sustc.edu.cn Abstract—An early detection of fault components is crucial for unmanned aerial vehicles (UAVs). The goal of this paper is to develop a monitoring system to early detect possible faults of UAV motors and propellers. Motor current signature analysis (MCSA) approach is used to analyze the stator current signals under different conditions. Then, fuzzy adaptive resonance (Fuzzy ART) neural network (NN), which is an unsupervised learning scheme, is employed to judge whether motors are operating in normal or faulty condition. In addition, the vibration signature analysis (VSA) technique is employed to monitor the UAV propellers. A Q-learning-based Fuzzy ARTMAP NN is used to learn extracted statistical features, and the Genetic algorithm (GA) is used to select an optimal subset of features through an off-line manner in order to reduce computational time. The experimental results validated the effectiveness of the proposed model in detecting faults of UAV motors and propellers as compared with CART, KNN, NB and SVM. I. I NTRODUCTION Unmanned aerial vehicles (UAVs) are being increasingly used in many applications owing to their versatility, high ma- neuverability, simple structure, vertical take-off and landing, low cost and easy maintenance [1] [2]. However, failure in any part of UAV can cause catastrophic damages and safety concerns [3]. Therefore, it is crucial to develop a system that continuously monitors important components of the UAV and detect possible faults early. The most common UAV faults include insufficient battery capacity, loss of communication, motors and propellers prob- lems [4]. Among them, the probability of occurring faults in motors and propellers are more than other parts when UAV is flying. However, there are few researchs on UAV motors and propellers fault detection. For example, a motor anomaly detection system for UAVs is proposed in [4]. The proposed system employed reinforcement learning to ludge the condition of motors using temperature of the motors. In [5], support vector machine (SVM) is employed to classify the acceleration signals which are measured by on-board inertial unit measurement (IMU) sensors. Conventional motor fault detection techniques are based on electrical and/or mechanical monitoring methods [6]. Mechan- * Corresponding author This work is partially supported by The Science and Technology Innovation Committee of Shenzhen Municipality (No. CKFW2016041415372174, and No. GJHZ20170314114424152); and The National Natural Science Founda- tion of China (No. 61773197). Current sensor Accelerometer Harmonics Statistical features Classifier Model Feature selection GA Fault Detection Clustering Healthy or broken Normal or abnormal Fig. 1: Fault detection system diagram ical monitoring methods may cause load critical problems [7], and are not effective methods for UAV motor fault detection. Nevertheless, current monitoring methods, in particular motor current signature analysis (MCSA) approach, can be imple- mented with small-size and low-cost sensors [8]. In addition, accelerometers have been extensively used in UAVs for flight stabilization [9] and rotating machine to detect faults [10]. This paper presents a novel framework for condition mon- itoring and fault detection of UAV motors and propellers (Fig. 1). Two methods, i.e., MCSA and VSA, are employed to obtain data and extract features from motors current and propellers vibration signals, res[ectively. Then, those extracted features are used to train unsupervised and supervised learning algorithms, separately. Finally, the trained models, i.e., Fuzzy ART [11] and QFAM-GA [12] [13], are used to continuously monitor the UAV motors and propellers. II. THE PROPOSED CONDITION MONITORING SYSTEM This section describes the proposed UAV motors and pro- pellers condition monitoring system. As shown in Fig. 1, it consists of two parts, i.e., motors and propellers. Firstly, three current sensors (e.g., one for each phase of a three- phase motor) and a 3-axis accelerometer are used to obtain current and vibration signals related to each UAV motor and propeller, respectively. Those sensors are connected to Arduino as data collection unit, and the acquired signals are stored in the computer via network connection. The observed current signals are transferred into their respective power spectral density (PSD) using fast Fourier transform (FFT) to extract their 1st, 3rd, 5th, 7th and 9th harmonics. Simultaneously, nine time domain statistical features, as reported in [14], are extracted from those vibration signals. Then, the fuzzy ART NN [11] (Fig. 2) is used to categorize extracted harmonics into a number of clusters, as follows: Firstly, fuzzy ART determines the similarity level of the complemented-coded of 978-1-5386-4707-3/18/$31.00 ©2018 IEEE