Evolutionary Design of Lifting Scheme Wavelet- Packet Adaptive Filters for Elevator Fault Detection Pasquale Arpaia 1 , Ernesto De Matteis 1 , Giuseppe Montenero 1 , Carlo Manna 2 1 - Department of Engineering University of Sannio Benevento, Italy arpaia@unisannio.it 2 - Department of Electrical Engineering University of Naples Federico II Napoli, Italy manna.cnr@gmail.com Abstract—An evolutionary-based procedure for designing adaptive filters based on second-generation wavelet (lifting scheme) packet decomposition for industrial fault detection is presented. The proposed procedure is validated by an experimental case study for induction motor fault diagnosis in an elevator system. Preliminary results on two typologies of faults, broken rotor bars and static air gap eccentricity, are discussed by showing encouraging performance. Keywords: Artificial Intelligence; Fault Diagnosis, Wavelet Transforms. I. INTRODUCTION Elevator systems are typical and conventional transportation tools in modern towns. They provide quick access within buildings, especially for the high-rise one. However, long-time usage of the system increases fault- occurrence probability [1]. This is a critical issue: when a fault on the system occurs , it causes problems both to people living in the building (especially to most ancient) and to workers. For these reasons, companies providing installation and maintenance consider correct inspection and fault detection most critical issues, in order to restore the elevator system rapidly and preserve their business. According to literature survey [1],[3], faults in induction motors used as traction machine for the elevator systems are the most critical faulty conditions. In order to ensure safe operation, timely maintenance and preventive rescue, on-line fault detection on elevator traction machines is a critical task. Fault detection for induction motors are generally classified as electrical faults and mechanical faults. For both of them, literature presents different diagnostic methods, such as temperature measurement, acoustic noise analysis , infrared measurements, and motor current signature analysis (MCSA) [4]. Among these techniques, MCSA [5] turned out to be the most popular for twofold reasons: (i) it is minimally invasive because stator currents are detected from the terminals; and (ii) it can detect both mechanical and electrical faults. The idea behind MCSA is the following: when a fault (mechanical or electrical) in the motor occurs, usually it causes an abnormality in the frequency spectrum of the stator currents. In other words, the fault exhibit itself a feature in the spectrum of the stator current signal, normally easily detected by Fast Fourier Transform (FFT). In lasts years, various faults detection techniques based on MCSA were proposed exploiting Discrete Wavelet Transform (DWT) [6] or Wavelet Packet Decomposition (WPD) [3], combined to FFT in order to obtain an effective features extraction, especially when the stator currents exhibit a complex spectrum due to power electronic drives of the induction motor. For on-line fault detection embedded on low-cost microcontrollers, the aforementioned techniques are not appropriate due to the computational complexity, needing for excessive resources. One of the possible solutions [7] is a WPD based on the lifting scheme [2], also called second-generation wavelet. This method has twofold advantages: (i) the WPD allows a time-frequency localization better than traditional DWT, fostering fault-features identification; and (ii) the second-generation wavelet are less computationally complex than traditional wavelet, because the lifting scheme technique not requires the FFT as design tool for wavelets. The lifting scheme consists in four main steps: split, prediction, compute and update. While the splitting is a trivial task, the last three steps (prediction, compute and update) needs for defining mathematical functions: thus, a detailed description of the signal analyzed is not easy to be achieved. On this basis, in a previous work [8], the authors showed the capability of evolutionary procedures (based on cultural algorithms (CA) [10]) of optimizing effectively the design of adaptive wavelet filters based on lifting scheme. However, from the fault detection point of view, the description of the analyzed signal should as much as possible accurate in order to detect anomalies in frequency when the fault occurred. In this paper. an evolutionary procedure, based on cultural algorithm to design an optimal lifting scheme WPD (LSWPD) of the signal, is proposed. By means of a cultural algorithm- based procedure, for each kind of faults to detect, an optimal LSWPD of a signal capable of highlighting the corresponding features, is designed. II. PROPOSED METHOD In the following, (A) the design problem, and (B) the proposed evolutionary approach are detailed. 978-1-4244-2833-5/10/$25.00 ©2010 IEEE