2194 IEEE TRANSACTIONS ON MAGNETICS, VOL. 33, NO. 5, SEPTEMBER 1997 Network for Detection at Strong Nonlinear Intersymbol Interference F. Obernosterer, W.F. Oehme, A. Sutor Department of Electrical Engineering, University of Erlangen, 91058 Erlangen, Germany Abstract-As recording density rises read signals are increas- ingly distorted by nonlinear intersymbol interference (ISI). Against this background an artificial neural network with a new decision making scheme has been set up and trained to work as a detector. Tests have been performed with experimentally captured read signals from a modified disk drive with magneto- resistive (MR) read heads. In comparison with multi-level decision feedback equalization (MDFE) the detection results show superior performance at extremely high linear recording densities. An error rate of 41W6 has been achieved at a user density D, = 3.5. We describe the architecture and the training procedure of the neural network and present detection results. I. INTRODUCTION At very high linear recording densities the read signal of the magnetic thin-film disk recording system is massively distorted by nonlinear intersymbol interference (ISI) [I]. In addition the total noise is known to be neither stationary nor Gaussian [2],[3]. For these reasons the established detection schemes for linear channels with stationary noise exhibit degraded performance [4]. There are several proposed solutions for the detection problem in the presence of nonlinear ISI. Biglieri et al. [5] applied a receiver structure that Compensates for nonlinearities based on a Volterra channel model. Lee et al. 161 investigated receivers which are based on a partial erasure model. A common difficulty of most of these schemes is the need for an explicit description of the nonlinear channel including nonstationary noise, where approximations may cause detector performance loss. In order to overcome this problem artificial neural networks have been used for nonlinear filtering in digital magnetic recording. They Rave the capability to "learn" a complicated and nonlinear relation between input and output variables at the level of signal samples. Nair and Moon applied a multilayer perceptron as nonlinear equalizer to the digital magnetic recording channel [7]. In contrast to their approach the neural network presented in this paper does not new decision making scheme is proposed which is based on the comparison of three output variables. The detector is implemented in software and has been tested with samples of real world read signals. In the following we describe the detector with its training procedure and show its detection performance in comparison with an MDFE detector. equalioe the readback pulse to a certain target response A Manuscript received January 31, 1997. F. Ohernosterer, +49 9131 857610, fax +49 9131 302951, obe@lte.e-technik.uni-erlangen.de; W.F. Oehme, +49 9131 857192. fax +49 9 13 I 30295 1 ~ oehmeC3lte.e-technik.uni-erlangen.de. A. Data Acquisition The signal sequences for network training and evaluation of the detection performance have been generated by a commercial 3.5" 1 GB-drive with MR read heads. The channel electronics were modified in order to allow recording of arbitrary binary data patterns at externally determined symbol frequencies. Only the innermost track has been used for recording to avoid bandwidth limitations caused by the pre-amplifier circuit. The width of the isolated pulse measured at 50 % amplitude (PW50) is 50 ns. Rate 2/3- RLE(1,7) encoding has been applied throughout the investi- gation. The achieved relative user data density of D, = 3.5 corresponds to a channel symbol rate of 105 Mbit/s. The pre- amplified read signal was low pass filtered at 45 MHz and digitized at 1 GS/s. The symbol clock was recovered from the samples by software. No gain control was applied. 11. THE NEURAL NETWORK DETECTOR An artificial neural network is a massively parallel distributed processor which consists of interconnected simple computing cells (units) allocated in layers. A single unit typically consists of an adder for summing the input signals and a subsequent nonlinear activation function cp (Fig. 1). The strengths of the connections between units are called weights and determine the mapping behaviour of a neural network. Mathematically a set of weights at the input of a layer j can be represented by the matrix Wj. The weights are adjusted during the training process in order to attain the desired system behaviour. Therefore, a set of training patterns (stimuli) has to be applied, which consist of an input vector and the corresponding desired response. Thus the neural network learns from examples. A. The Structure ofthe Neural Network Detector The basic structure of our Neural Network Detector (NND, Fig. 2) is a double layer feedforward network (perceptron) z=cp( c WV% ) V z 0 U" Fig. 1. Structure of a processing unit. 00 0 1997 IEEE 00 18-9464/97$10