0278-0046 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2016.2582729, IEEE Transactions on Industrial Electronics IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 1 Abstract—Early detection of the motor faults is essential and Artificial Neural Networks (ANNs) are widely used for this purpose. The typical systems usually encapsulate two distinct blocks: feature extraction and classification. Such fixed and hand-crafted features may be a sub-optimal choice and require a significant computational cost that will prevent their usage for real- time applications. In this paper, we propose a fast and accurate motor condition monitoring and early fault detection system using 1D Convolutional Neural Networks (CNNs) that has an inherent adaptive design to fuse the feature extraction and classification phases of the motor fault detection into a single learning body. The proposed approach is directly applicable to the raw data (signal) and thus eliminates the need for a separate feature extraction algorithm resulting in more efficient systems in terms of both speed and hardware. Experimental results obtained using real motor data demonstrate the effectiveness of the proposed method for real-time motor condition monitoring. Index Terms—Convolutional Neural Networks; Motor Current Signature Analysis I. INTRODUCTION OTOR fault detection and diagnosis methods can be divided into three major categories: model-based, signal-based, and knowledge-based. Model-based methods use mathematical models describing the normal operating conditions of the induction motors [1]. In model-based methods, fault diagnosis algorithms are developed to monitor the consistency between the measured outputs of the practical systems and the model-predicted outputs [2]. The main advantage of a model-based method is that the fault diagnosis is very straightforward if the model parameter has a one-to- one mapping with the physical coefficients [3]. The signal- based methods usually employ one of four main classes of Manuscript received December 10, 2015; revised April 21, 2016; accepted May 8, 2016. T. Ince, L. Eren and M. Askar are with the Electrical & Electronics Engineering Department, Izmir University of Economics, Izmir, Turkey (e-mail: turker.ince@ieu.edu.tr, levent.eren@ieu.edu.tr, murat.askar@ieu.edu.tr). S. Kiranyaz is with the Electrical Engineering Department, Qatar University, Doha, Qatar (e-mail: mkiranyaz@qu.edu.qa). M. Gabbouj is with the Department of Signal Processing, Tampere University of Technology, Tampere, Finland (e-mail: moncef.gabbouj@tut.fi). signal processing techniques [4]: time-domain analysis [5],[7], frequency-domain analysis [8],[9], enhanced frequency analysis [10],[11], and time–frequency analysis techniques [12],[14]. The signal-based systems do not require an explicit or complete model of the system but their performance may degrade when working in an unknown or unbalanced condition. It is a well-known fact that as the complexity of advanced signal processing tools used increases, fault detection capability is increased together with the computational cost [6]. The knowledge-based systems may be divided into two groups: qualitative methods on the basis of symbolic intelligence and quantitative methods on the basis of machine learning intelligence [3]. The qualitative methods include fault trees, diagraphs, and expert systems whereas quantitative methods have both unsupervised learning systems such as K-means, C-means, nearest neighbor, principal component analysis (PCA), and self- organizing maps (SOM), and supervised learning systems such as artificial neural networks (ANN), fuzzy logic (FL), support vector machines (SVM), partial least squares (PLS), and hybrid systems. The hybrid systems may be more suitable for complex fault detection problems where the features are extracted from statistical projection methods such as PCA and PLS, or signal processing methods such as fast Fourier transform (FFT) and wavelet transform (WT). The performance of knowledge- based methods relies on training data and quality of selected features heavily. In several studies [15]-[27] different features are proposed. The selected features are presented to classifiers as inputs. Diagnosis of electric stator faults in induction machines using an ANN based approach is proposed in [17]. Machine fault conditions were predicted with less than 2.4% error using only 13 training data patterns and 9 validation data patterns. In [18], Li et al presented a neural-network based motor bearing fault diagnosis system using time and frequency based features, which achieved average detection rates between 88.75% and 96.25% for different number of hidden neurons. In [21], a neural-network-based fault prediction scheme without using any machine parameter or speed information is presented. Speed is estimated from measured terminal voltage and current. With minimal tuning of the neural network, induction machines of different power ratings can be accommodated, and 93% or more detection performance is achieved. In [22], two types of neural detectors, feedforward multi-layer perceptron (MLP) and self-organized Kohonen’s Real-Time Motor Fault Detection by 1D Convolutional Neural Networks Turker Ince, Member, IEEE, Serkan Kiranyaz, Senior Member, IEEE, Levent Eren, Member, IEEE, Murat Askar and Moncef Gabbouj, Fellow Member, IEEE M