Comparative Analysis between BP and LVQ Neural Networks for the Classification of Fly Height Failure Patterns in HDD Manufacturing Process Thanapong Thanasarn International College King Mongkut’s Institute of Technology Ladkrabang (KMITL), Bangkok, Thailand thanapong.thanasarn@seagate.com Chanon Warisarn College of Data Storage Innovation King Mongkut’s Institute of Technology Ladkrabang (KMITL), Bangkok, Thailand kwchanon@kmitl.ac.th Abstract— In this paper, Learning Vector Quantization neural network (LVQ) and Back Propagation neural network (BP) algorithms were used for fly height failure pattern classification in the manufacturing process of Hard Disk Drive (HDD). Six types of failure patterns were chosen from fly height measurement, to be recognized, including three causes of failure: poor mechanics on Head Gimbal Assembly (HGA), contamination on slider Air Bearing Surface (ABS), and head instability. In the experiment, data from fly height failure was used to learn the pattern recognition for both neural network algorithms. The simulation results show that the classification performance of fly height failure patterns based on LVQ neural network is better than BP neural networks; LVQ can achieve the best of overall accuracy at 90.7% while BP reaches only 81.8%. Keywords - Learning Vector Quantization Neural Network; Back Propagation Neural Network; Hard Disk Drive; Fly Height. I. INTRODUCTION In the manufacturing process of Hard Disk Drive (HDD), after assembly process, HDD requires functional testing to calculate all of the drive operations, such as the calculation of areal density with bit per inch (BPI) and track per inch (TPI), the measurement of bit error rate (BER) performance, the MR bias calculation, and the fly height measurement, etc. Fly height testing is the measurement of the spacing distance between a recording head and a media by applying voltage values into a heater element of the recording head until the recording head touchdown on the media and also provides the heater voltage values in each data zone that are appropriate to perform read and write operations [1], [2]. In general, the failure analysis on failed heads from fly height measurement must be analyzed using the fly height profile to classify the causes of failure. It was observed that each fly height profile can indicate the cause of each problem. For example, poor mechanics on Head Gimbal Assembly (HGA), found contamination on the slider Air Bearing Surface (ABS), and a head instability problem. Currently, however, fly height profile failure is investigated by failure analysis engineers over a long period of time. Therefore, ways to reduce the engineer’s work with failure analysis are needed for the HDD manufacturing process to be more efficient. Recently, many researchers have applied neural network models for pattern recognition in several fields, such as: face image recognition, the speaker identification, the classification of faults in electrical systems, and the recognition of electrocardiogram (ECG) patterns [3]-[12]. However, neural network models have never been applied with fly height failure pattern classification in the manufacturing process. Therefore, this paper proposes two classification methods by using Learning Vector Quantization neural network (LVQ) and Back Propagation neural network (BP) to support failure analysis. The paper is organized as follows: After presentation of the system structure for the classification of fly height failure patterns of LVQ and BP model in Section-II, Simulation results are given in Section-III. Finally, Section-IV is a conclusion of the results. II. METHODOLOGY In this work, the spacing distance between recording head and media are considered in terms of the voltage values that are applied into the heater element of the recording head. It affects the writer and reader element protrude until the contact area touchdown on the media, which is detected by the servo detector system. Data is received from the fly height measure for this study, including heater voltage values in the Digital to Analog Converter (DAC) unit for both read and write operations in each location of the data zone on the media. Nevertheless, to reduce the complexity of input datasets, only data from write operation for pattern classification were chosen. The six types of failure pattern that are found on failed heads are shown in Figure 1. The x-axis is the heater voltage (in DAC unit) and the y-axis is the data zone of the media in order of the outer disk towards the inner disk, which are composed of 10 data points for each head. In this study, data was collected from failure analysis results, which have been distinguished by type of pattern for 1,000 samples per type. Then, there are 6,000 failure samples to supervise networks and test classification accuracy between LVQ and BP neural networks.