1 Fault Diagnosis in a Distributed Motor Network using Artificial Neural Network Saud Altaf, Adnan Al-Anbuky, Hamid GholamHosseini Sensor Network and Smart Environment Research Centre (SeNSe), School of Engineering, Auckland University of Technology, Auckland - New Zealand {saltaf, adnan.anbuky, hgholamh}@aut.ac.nz Abstract—Signature analysis methods have been proven to deliver good results in the laboratory environment and successfully applied to isolated motors. The influence of fault signal on a non-faulty motor may be interpreted as faulty condition of the healthy motor. Therefore, it is difficult to identify a motor fault within a network and precisely identify the type of fault. This paper presents a supervised distributed Artificial Neural Network (ANN) that is able to identify multiple fault types such as broken rotor bar (BRB) or air gap eccentricity faults as well as the location of fault event within an industrial motor networks. Features are extracted from the current signal, based on different frequency components and associated amplitude values with each fault type. A set of significant fault features such as synchronized speed, rotor slip, the amplitude value of each fault frequency components, the Root Mean Square (RMS) and Crest Factor (CF) value are used to train the ANN using Back Propagation (BP) algorithm. The simulation results show that the proposed technique is able to identify the type and location of fault events within a distributed motor network. The proposed architecture works well with the selection of a significant feature sets and accurate fault detection result has been achieved. Classification performance was satisfactory for healthy and faulty conditions including fault type identification. Keywords—Distributed Motor Network; Artificial Neural Network; Fault Identification and Localization; Feature Extraction. I. Introduction During the last two decades, the condition monitoring and fault diagnosis of induction motors have gained good interest. This has helped in improving the overall industrial system reliability. In an industrial environment, the power system network acts as a conduit for a motor fault signals. The influence of faulty motor’s signal from a non-faulty motor may indicate a faulty condition of the healthy motor. In this uncertain situation, it is difficult to identify the motor fault within a network and precisely identify the type of fault. But, if the motor’s behavior is dynamically captured through predefined features, the combined effect of motors on a power network system’s performance can be more clearly visualized. This may assist in the measurement of the behavioral performance. Most motor faults are associated with sidebands at characteristic frequencies on motor current spectra [1]. A fault severity level can be helpful in locating a fault source. Rotor related faults are the easiest to detect in induction motors. This can be done through current based signal analysis on the associated double slip frequency sidebands in the current spectrum of the fundamental supply frequency [2]. This represents the severity of rotor fault, related to the higher value of magnitude of these sidebands. When a rotor fault takes place, changes in shape, magnitude and frequency of the waveform are expected. These changes can be associated with the physical phenomena causing the event. Throughout the event the disturbed waveform experiences several non-stationary and stationary stages [3]. Both normal operation of network components and short-circuits can be the cause of voltage disturbances in distributed network of motors. Many studies have been presented in research area of fault diagnosis using isolated induction motors [4-7]. Artificial Neural Network (ANN) is reported as being a knowledge-based technique for single motor fault diagnosis. These studies perform the diagnoses by mapping different fault symptoms in an isolated motor to produce the diagnosis decision. Eldinet et al [8] presented a diagnosis system based on ANN, which applies the RMS measurements of current, voltage and speed to train the ANN in diagnosis of motor rotor faults. Voltage faults are only identified in a steady state condition, not in a dynamic load condition. Another study was presented by Arabaci [9], based on the influence of the rotor fault on current in the frequency domain, using ANN in a steady motor operating condition. This study demonstrated the possible symptoms of significant frequency components on the frequency spectrum related to the broken rotor bar fault. These symptoms are used as an input matrix using the supervised ANN architecture. The proposed technique concluded that the process of rotor fault diagnosis and discrimination between each fault with reasonable accuracy. Drira et al [10] present a rotor fault model using Fast Fourier Transform (FFT) and the supervised ANN learning method. Significant features (RMS, crest factor, highest magnitude, etc.) are extracted from the current spectrum and all possible magnitude highest sidebands