D.R. Valiullin – Engineer-Programmer, Microwave Electronics LLC (Moscow)
P.N. Zakharov – Ph.D. (Phys.-Math.), Associate Professor, Faculty of Physics M.V.Lomonosov Moscow State University
N.N. Sysoev – Dr.Sc. (Phys.-Math.), Professor, Faculty of Physics M.V.Lomonosov Moscow State University
А.N. Frolov – Deputy Technical Director, JSC «SPE «Radiosvyaz» (Krasnoyarsk)
R.G. Galeev – Dr.Sc. (Eng.), General Director, JSC «SPE «Radiosvyaz» (Krasnoyarsk)
Wireless communication systems operating in multipath environment use equalizers to minimize the BER degradation caused by inter-symbol interference. Existing equalizers such as DFE (decision feedback equalizer) and MLSE (maximum likelihood sequence estimation) prove to be either computationally complex or having large energy loss compared to Shannon limit. The neural network equalizer first introduced in  has computational complexity comparable to that of DFE providing better energy efficiency performance. In  fixed neural network equalizer coefficients were replaced by adaptive coefficients obtained through learning. The coefficients of neurons connections are varied to minimize MSE (mean-square error) with gradient descent. In  the equalizer proposed in  was investigated experimentally with BPSK modulation.
In learning process, regularization allows to avoid retraining. To implement regularization the regularization term which is proportional to the sum of squared equalizer coefficients is added to the expression for MSE. The coefficient of this proportionality is called the regularization parameter. Consequently, the equation for coefficients adaptation is varied. It allows to reduce the coefficients growth, the main reason of retraining.
In this paper, the proposed equalizer with regularization for QPSK modulation was experimentally tested. Test bench consisted of two identical transceiver nodes. Each node included transceiver board connected with digital signal processing board. The proposed equalizer algorithm was implemented in FPGA on DSP board along with automated gain control, matched filtering, carrier recovery, clock recovery algorithms. The results of experiments have shown that regularization improves energy efficiency of the equalizer’s second iteration for Eb/N0 < 2 dB range.
- Myburgh H.S., Olivier J.C. Near-optimal low complexity MLSE equalization // IEEE. Wireless Communications and Networking Conference. 2008. P. 226–230.
- Valiullin D.R., Zaharov P.N. Ekvalajzer na osnove nejronnyh setej s obucheniem v mnogoluchevom kanale // Uspehi sovre-mennoj radioelektroniki. 2016. № 11. S. 200–202.
- Valiullin D.R., Zaharov P.N. Eksperimental'nye issledovaniya ekvalajzera na osnove nejronnyh setej s obucheniem v mnogoluchevom radiokanale // Zhurnal radioelektroniki. 2017. № 12.
- Dzh. Prokis Cifrovaya svyaz': Per. s angl. M.: Radio i svyaz'. 2000.