1998年，Torkkola提出了反饋分離結(jié)構(gòu)的信息最大化算法，該算法對(duì)應(yīng)于穩(wěn)定的IIR濾波器，只能用于最小相位混合。而真實(shí)的環(huán)境往往包括非最小相位傳輸系統(tǒng)。本文在分析反饋分離算法的基礎(chǔ)上，提出了一種在頻域上用前饋結(jié)構(gòu)的信息最大化算法，該算法對(duì)應(yīng)于非因果FIR濾波器，為非最小相位系統(tǒng)的盲分離問題提供了一個(gè)解決方案。最后，我們將演示用這種方法分離實(shí)際環(huán)境下的兩個(gè)自然信號(hào)。
關(guān)鍵詞:盲源分離(BSS);信息最大化;卷積混合;自然梯度

Abstract: In 1998 , Torkkola gave information maximization rules of a feedback architecture which can only separate minimum phase mixtures,though working well on simulated data,they fail in real environment which usually involve non-minimum phase transfer systems. This paper analyses the rules of the feedback architecture and presents an informax maxization algorithm of a feedforward architecture in the frequency domain which corresponds to a non-causal FIR.The new algorithm can solve the non-minimum phase mixing problem well.At last, simulation experiments will prove that two natural signals in a real environment can be separated successfully.
Keywords: blind source separation(BSS); information maximization; convolved mixtures;
natural gradient