Adaptive receivers for removing linear and nonlinear intersymbol interference by mean of time delay neural nets (AR-TDNN)

Abstract

We review in this paper several classical equalizers for data transmission and show that they are especial cases of a general time delay neural net, TDNN. The main point of this paper is t o show that the solution to the deconvolution problem for many transmission channel types, including linear and non-linear distortions, can be solved by mean of TDNN. Alternativelly TDNN can be seen as a generalization o f currently used equalizers and as tools for discovering new ones. _-INTRODUCTIONIntersymbol interference, ISI. i n h i g h speed data communication is one of the main factors which limitate the data throughput. the other one is the noise Usually, the receivers deal with the first impairment by mean of adaptive filters which cancel IS1 trying at the same time to maintain the noise at enough low level) These filters can operate at symbol rate, the T equalizers, in both ronfigurations teed forward or feed-back, or f r a c t i o d l y spaced equalize*\ operating at a fraction of the symbol 'ate Alternativelly. the impulse response of the channel model is known, it is alw possible to use the Viterbi's receiver which do n o t invert the channel neither cancel intersymbol interferences This algorithm basically search, in an optimum way, the sequence of symbols which generated the received signal closest to the received one 191.All of these receivers are based on the important fact that the channel behaviour is linear. When non-linear effects are present and consequently non-linear intersymbol interference, different systems have t o be used, for instance those based on Volterra seriesJ11, [21 In this paper we are going t o study neural nets [31. [41, based equalizers Inputs t o these nets are signals and i t s delayed samples, we call to this nets, following [lo], time delay neural nets, TDNN Theses equalizers WIII be used for both linear and non-linear channels Let a(n) be the sequence of transmited symbols, a(n) in the unidimensional case are the set o f the 2M symbols:{ f 1, f 2, f 3, ? M) In this paper we are going to consider only the case of M = 1 From the receiver end the channel is characterized, aftef sampling at baud rate, as a combination of linear and non-linear filters A n example of channel model t o be considered is that one characterized by two linear filters with a non-linear filter without memory in betweek Special cases will include only linea filters or a linear filter and the non-linear one. See Figure 1 This work was supported by the PLANICYT under the Project "Tratamiento Avanzado de la Informaci6n". General channel model I l a(") = f ? 1, ' 2 , . ' M } a(n) AR TDNN Figure 1. Problem definition In the proposed equalizer the signal x(n) and its time delayed samples are processed by a neural net. Then, the input t o this net i s t h e set of samples x(n-L). x ( n -L + l ) , . . . x(n),x(n + 1).x(n + L l when the symbol n has to be estimated L I S selected in accordance wmth the channel memory Dete,.+ed symbols can be used as addnllonal inputs tr the net We are going to consider several class~rdl ...

Igualación adaptiva de canales de comunicación de variación rápida con redes neuronales

“…Muchos aspectos prácticos están relacionados con la velocidad de convergencia, inmunidad al ruido, tolerancia a la fluctuación de la portadora y complejidad, se deben analizar para evaluar la facilidad de usar TDNN como receptores. Se han obtenido resultados que muestran que los Receptores Adaptivos con TDNN (AR-TDNN) han podido resolver el problema de la deconvolución [36].…”