以類神經網路為基礎的通訊等化器之研究

Abstract

在實際的通訊系統中需要以資料等化器來回復一個失真信號的原始波形，近來許多以類神經網路為基礎的等化器設計被應用在嚴重失真的信號回復。在本文，我們提出了一個新的類神經網路模式，使用一個多變數冪級數函數做為人工神經元的集成函數，應用於多層認知器結構倒傳遞類神經網路，由於對應的訓練演算法是以最陡坡降法推導，故收斂解存在。與使用一階多變數多項式當集成函數的傳統方法相比較，這個新方法的樣本空間分割邊界，將由傳統的片段線性分割變成片段非線性分割，傳統的多層認知器結構倒傳遞類神經網路可視為這個新方法的一個線性特別解。因此，可說這個新的模式是一般化的多層認知器結構倒傳遞類神經網路，與其他的片段線性分割的方法相比較，這個新方法因為具有片段非線性分割樣本空間的能力，所以在應用上有更大的彈性。 在有線通訊系統中，我們將以多層認知器結構倒傳遞類神經網路為基礎的通道等化方法應用於不同的地方，在資料速率大於通道頻寬十倍左右的有線頻寬受限通道上，與傳統上使用最小均方誤差演算法為基礎的決策回授等化器相比較，以多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器可提供比較好的效能、容忍比較多的取樣時脈歪斜和允許比較大的通道嚮應變異。然而對於具有非線性失真的嚴重碼際干擾通道來說，以多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器還有改善的空間，使用一般化的多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器可以提供更好的效能。在多條平行的有線頻寬受限通道上，我們採用多輸入多輸出的以多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器和多輸入多輸出的以一般化多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器，同時抑制碼際干擾、串音干擾和背景雜訊。同樣的，多輸入多輸出的以一般化多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器優於多輸入多輸出的以多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器，而且多輸入多輸出的以多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器又優於使用最小均方誤差演算法為基礎的決策回授等化器。 對於無線通訊系統，我們提出一個以多層認知器結構倒傳遞類神經網路為基礎的改進方法。在多路徑平坦衰減通道中，我們應用軟性輸出和軟性決策回授的結構於一個以多層認知器結構倒傳遞類神經網路為基礎的通道等化器，並於其後串接一個軟性決策通道解碼器以改進整體的效能。此外，利用輸出層神經元的轉移函數尺度因子的最佳化，以及在訓練型樣加入少量的隨機擾動，可以進一步的改善以多層認知器結構倒傳遞類神經網路為基礎的軟性決策回授等化器的效能。由模擬結果，在多路徑平坦衰減通道中，使用包含位元交錯的軟性決策通道解碼器時，以多層認知器結構倒傳遞類神經網路為基礎的軟性決策回授等化器的效能優於以多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器和軟性輸出的以多層認知器結構倒傳遞類神經網路為基礎的決策回授等化器。

In practical communication systems, it is necessary to apply data equalizers to recover the original waveform from the distorted one. Recently, various equalizer designs based on artificial neural networks have been studied to the severely distorting signal recoveries. In this study, we propose a new neural network model that applies a multivariate power series as the summation function of the MLP/BP neural networks. The corresponding training algorithm is deduced by the gradient steepest descent method; consequently, the convergence solutions exist. Compared to the conventional approach using a first order multivariate polynomial, the boundaries separating the pattern space change from piecewise linear into piecewise nonlinear. The traditional method is a special case of the proposed model. Therefore, this new model is a generalized MLP/BP neural network that is more flexible than other piecewise linear approaches because of the nonlinear separating pattern space. For wireline communications, we apply the MLP/BP-based channel equalization schemes to different applications. In wireline band-limited channels that the data rate is about ten times as much as the channel bandwidth, the MLP/BP-based DFEs provide better performance, tolerate more sampling clock skew, and permit larger channel response variance than LMS DFEs. However, the BER performance of the traditional MLP/BP-based DFEs is not good enough for the severe ISI channels with nonlinear distortions. In such channels, the generalized MLP/BP-based DFEs can outperform the traditional MLP/BP-based DFEs that do better than the LMS DFEs. In wireline band-limited parallel channels, the MIMO MLP/BP-based DFEs and the MIMO GMLP/BP-based DFEs can suppress ISI, CCI and AWGN, simultaneously. By the simulation results, the MIMO GMLP/BP-based DFE can yield a substantial improvement over the MIMO MLP/BP-based DFE that perform better than the LMS DFEs in such channels. For wireless communications, a modified approach, which is also based on the MLP/BP neural network, is presented. We apply the soft output and the soft decision feedback structure to the MLP/BP-based channel equalization scheme that concatenates with the soft decision channel decoder to improve whole performance on multi-path fading channels. Moreover, the performance of the MLP/BP-based soft DFE is also increased with the optimal scaling factor searching of the transfer function in the output layer of the MLP/BP neural networks and extra small random disturbances added to the training data. By the simulations, the MLP/BP-based soft DFEs with bit-interleaved TCM outperform the MLP/BP-based DFEs with bit-interleaved TCM and the soft output MLP/BP-based DFEs with bit-interleaved TCM in multi-path fading channels.