前記歪み識別手段は、識別された歪みの確率分布の統計情報を学習することによって前記歪みパラメータを学習するように構成された、請求項1に記載のノイズ除去システム。
前記歪み識別手段は、前記入力信号の少なくとも1つの送信シンボルに従って、前記入力信号をクラスタ化及び分離するように構成された、請求項1に記載のノイズ除去システム。
前記関数は、前記光送信機に入力される少なくとも1つの信号の歪みを補償し、又は、前記光受信機から出力された少なくとも1つの信号の歪みを補償する、請求項6に記載の信号処理システム。
前記信号修正手段は、前記補償手段からの、学習済みの関数の精度を示す少なくとも1つのフィードバックを用いて、前記信号修正手段の出力を適応的に調整する、請求項6又は7に記載の信号処理システム。
光受信機102は、光送信機101から光送信信号を受信すると、受信手段102aが、コヒーレント復調(coherent recovery)等の光復調手法を用いて、光信号を復調する。そして、復調された光信号は、アナログデジタル変換器102bに供給される。ADC102bは、アナログ光信号をデジタル信号に変換する。DSP(Digital Signal Processing)102cは、デジタル信号を処理する。DSP102cは、同期やフィルタリング等の様々な信号処理を含む。
光送信機101の処理は、理想的ではないため、望ましくない歪みを引き起こす。これらの歪みは、伝送性能に影響を与えるため、適切に補償される必要がある。入力信号を事前に補償する構成要素を追加するために、事前補償手段103が用いられる。
上述した方法は、光受信機102の後に補償関数が実装される事後補償のシナリオとして実装することもできる。このシナリオでは、補償手段は、光受信機102の後に実装され、光送信機101の前には実装されない。
第1の課題は、収集された信号データからの学習する歪み補償関数が不正確であることである。第1の課題が生じる理由は、歪み成分の一部が本質的に不定であるため、学習不可能なものであることである。換言すると、歪みの一部は、ノイズ、又はそれを不定にする分布に従う他の要因である可能性があるため、学習することができない。
第2の問題は、学習済の歪み補償関数が、信号データが収集された動作環境にのみ適していることである。第2の課題が生じる原因は、機能学習が不正確に学習して、動作条件が変わると変化する可能性のある動作条件のノイズプロファイルを補償することである。動作条件のノイズプロファイルとは、システムの特性、例えば、SNRや他のノイズ特性等を意味する。入力電力や他のシステム条件が変わると、このノイズプロファイルも変わるため、補償関数も変更する必要がある。
第3の問題は、歪み補償関数を正確に学習させるために、大量のデータが必要になることである。第3の課題が生じる理由は、非決定性の影響を取り除き、また、決定性の影響のために正確な歪み補償関数を学習させるために、大量のデータが必要になるからである。
第1の効果は、歪み補償関数の正確な学習が可能になることである。第1の効果の理由は、ノイズ除去手段が、信号データにおける学習不能による影響又は非決定性の影響を取り除き、又は低減できるからである。
第2の効果は、歪み特性が、ブラインド方法又は教師なしの方法で正確に学習され、補償プロファイルの学習に使用されることである。第2の効果の理由は、歪み識別手段が、提供されたラベルなしの信号データから歪み特性を正確に識別できるからである。
第3の効果は、歪み特性の学習データの要求が低減されることである。この効果の理由は、ノイズ除去手段が、データ内の学習不能な要素又は不定の要素を取り除くため、正確な歪み補償関数の推測に必要なデータが少なくなるからである。
歪み識別手段201は、信号の統計的な特性を特徴付けることを可能にするパラメータを学習する教師なしクラスの学習アルゴリズムを含む。学習アルゴリズムの教師なしクラスは、ラベルなしで信号から対象(objective)を学習することを目的とする。この場合のラベルは、実際の送信信号(すなわち、デジタル送信信号(グランドトゥルース))である。対象は、シンボルの平均又は分散等の信号の特性を、ブラインドで、すなわち、ラベルを用いず、教師なしで学習することである。具体的には、K次の信号コンスタレーションの場合、典型的な学習アルゴリズムは、k個の固有の送信シンボルのそれぞれに対応する平均(すなわち、中心)や分散等のパラメータを学習することを目的とする。さらに多くのパラメータをこの歪み識別手段201に学習させて、信号を特徴付けることもできる。様々な教師なし学習アルゴリズムを採用することができる。例えば、学習アルゴリズムは、信号の特徴の特性が、ガウス分布クラスに属するとの仮定の下で信号の特徴を学習することを目的とする混合ガウスモデル(GMM)とし得る。パラメータを学習すると想定される分布は、システムの動作条件に従って調整することができる。換言すると、ノイズが属する可能性のある分布に応じて、教師なし学習アルゴリズムとして適切な混合モデルを採用することができる。
信号修正手段202の処理の選択は、2つのモード、すなわち、デジタル送信信号を使用する第1のモードと、デジタル送信信号を使用しない第2のモードが有り得る。デジタル送信信号を使用する第1のモードでは、信号修正手段202は、実際の信号のデジタル送信信号を用いて、信号の修正/トリミングを設定する。第2のモードは、実際の信号のデジタル送信信号を得ることが困難又は利用できない状況で使用することができる。デジタル送信信号を使用しない第2のモードでは、信号修正手段202は、歪み識別手段201からの各シンボルについての尤度推定を用いて、「擬似的なデジタル送信信号」を形成する。そして、信号修正手段202は、「擬似的なデジタル送信信号」を用いて、信号の修正/トリミングを設定する。
これらの歪み識別手段201及び信号修正手段202は相互に、光受信機102からのデータのノイズを除去するように動作し、特に、事前補償手段203へのデータを改善することにより、ノイズ除去システム200の性能の向上を実現する。
さらに、ステップS103では、ブラインド教師なしクラスタリングアルゴリズム(Blind Unsupervised Clustering Algorithm)のトレーニングが行われ、すなわち、教師なし学習アルゴリズムとして上述した教師なしクラスタリングアルゴリズムを実装する歪み識別手段201が、ステップS102で収集された受信信号出力データに対して実行される。歪み識別手段201は、受信信号を元のデジタル送信信号のシンボルとして分類することを目的として、受信したコンスタレーションの特性を学習する。一般に、歪み識別手段201で説明される教師なし学習アルゴリズムは、複数の信号を分類(すなわち、クラスタ化)することを目的とする。我々は、このアルゴリズムを用いて、複数の信号を分類するために用いられるパラメータを抽出する。抽出されたパラメータは、信号修正手段202に提供される。
ステップS105では、事前補償手段203のトレーニングが行われる。このステップでは、この修正された信号が、事前補償手段203に提供され、トレーニングが、デジタル送信信号を用いて行われる。事前補償手段203は、光送信機101から利用可能なデジタル送信信号を使用することができる。
トレーニングが収束した後、ステップS106では、デジタル送信信号からの損失の評価が、損失関数によって行われる。デジタル送信信号からの損失の評価は、ニューラルネットワークのトレーニングにおけるような、自動化されたプロセスである。デジタル送信信号からの損失の評価は、事前補償手段203の適合/精度を評価することである。評価基準は、「平均二乗誤差損失」のような損失関数とし得る。この損失は、学習された補償出力と、利用可能なデジタル送信信号から生成された期待される理想値とを比較することによって評価される。この損失評価方法は、事前補償手段203の構造の実装、例えば、フィルタ、メモリ多項式又は人工ニューラルネットワーク等に応じて、異なり得る。損失関数からの出力は、収束及び最小損失を保証するために、バックプロパゲーション等の反復アルゴリズムを用いた重み関数を修正するために使用される。
ステップS107では、バックツーバック(b2b)接続の評価は、損失関数を用いた簡単な評価又はBER(Bit Error Ratio)の計算によって行われる。このプロセスは、手動でもよく、定義された関数によって自動化してもよい。このステップでは、事前補償手段203は、最初にバックツーバックセットアップに実装される。ステップS107では、完全なシステムの性能が評価される。すなわち、光送信機101が、バックツーバックセットアップにおいて光受信機102のセットアップに接続され、事前歪み関数を使用して、又は使用せずに、評価が行われる。このステップは、事前補償手段203による歪みの軽減によって実現された性能の改善を評価するために行われる。
光送信機101と光受信機102の間に接続された光チャネルが存在する場合、ステップS108において、チャネルの配置が行われる。換言すると、光ファイバが、光送信機101と光受信機102の間に接続される。この配置は、手動又は自動で行うことができる。光送信機101によって生じた歪みが、歪み最小化戦略(ストラテジー)の学習に使用される信号データにのみ存在するために、バックツーバックセットアップは、初期のステップの間に用いられる。しかしながら、オンラインの通信システムは、少なくとも1つのチャネル、例えば、光送信機101と光受信機102の間に配置された光ファイバケーブル等を含み得る。したがって、配置の段階では、学習済みの補償関数が、光ファイバを含む光通信システムに実装される。
上述したように、歪み識別手段201は、光受信機102からの入力信号の歪みを、教師なしの方法で識別し、識別された歪みを示す歪みパラメータを出力する。信号修正手段202は、デジタル送信信号を使用し、又はデジタル送信信号を使用せずに、光送信機101に入力される信号を、出力された歪みパラメータを用いて修正し、当該信号のノイズを除去する。修正された信号は、光通信システムで使用される信号の歪みを補償する歪み関数が、デジタル送信信号を用いて学習するために使用する。したがって、光送信機101によって生じる歪みが軽減される。
上述したように、事前補償手段203は、デジタル送信信号と、修正された信号を用いて、事前補償手段203に入力された少なくとも1つの信号の歪みを補償する関数を学習させる。したがって、光送信機101によって生じる歪みが軽減される。
コンスタレーション001は、信号から分布統計情報を抽出した後のコンスタレーションを示す。コンスタレーション002は、信号統計情報を修正することによって信号の特性を修正した後のコンスタレーションを示す。
第2の実施形態は、補償ブロックが光受信機102の後に配置される事後的な補償の状況において、提案されたノイズ除去システム300が、どのように実装されるかを示すことを目的とする。
提案されたこれらの手段は、光送信機101及び光受信機102を含む既存の従来技術の構成100に含まれる。我々は、事前補償手段203の代わりに、事後補償手段303を用いた事後的な補償のスキームについて着目する。事後補償手段303は、少なくとも1つの入力信号の歪みを補償する補償手段に相当する。
事後補償手段303は、事前補償手段103と同様であり、ニューラルネットワーク、フィルタ、メモリ多項式、及び利用可能なデータから所望の関数を推論するための他の技術を用いて、動作することができる。
歪み識別手段301は、信号の統計的な特性の特徴付けを可能にするパラメータを学習する、教師なしクラスの学習アルゴリズムを含む。学習アルゴリズムの教師なしクラスは、ラベルを用いずに、信号から対象を学習することを目的とする。この場合のラベルは、実際の送信信号(すなわち、デジタル送信信号)である。K次の信号コンスタレーションに関して、典型的な学習アルゴリズムは、k個の固有の送信シンボルのそれぞれに対応する平均(すなわち、中心)や分散等のパラメータを学習することを目的とする。この歪み識別手段301が、さらに多くのパラメータを学習して信号を特徴付けることもできる。学習アルゴリズムは、混合ガウスモデル(GMM)とすることができ、これは、信号の特徴の特性がガウス分布クラスに属するという仮定の下で、信号の特性を学習することを目的とする。システムの動作条件に従って、パラメータを学習すると仮定される分布を調整することができる。
これらの歪み識別手段301及び信号修正手段302は相互に、光受信機102からのデータのノイズを除去するように動作し、特に、事後補償手段303へのデータを改善することにより、ノイズ除去システム300の性能の改善を実現する。
さらに、ステップS203では、ブラインド教師なしクラスタリングアルゴリズムのトレーニングが行われ、すなわち、教師なし学習アルゴリズムとして上述した教師なしクラスタリングアルゴリズムを実装する歪み識別手段301が、ステップS202で収集された出力データに対して実行される。
ステップS205では、事後補償手段303のトレーニングが行われる。このステップでは、信号修正手段302から出力された、この修正された信号が、事後補償手段303に提供され、ここで、修正された信号から元の信号を抽出することを目的として、デジタル送信信号を用いて、トレーニングが行われる。この目的は、第1の例示的な実施形態における事前補償手段203の入力に対する事前補償を学習する目的とは対照的である。このトレーニング/適合(フィッティング)には、信号修正手段302からの入力シンボルパターン及び修正された信号パターンを含むデータセットを用いて事後補償を学習する既知の任意の方法を採用することができる。
トレーニングが収束した後、ステップS206では、デジタル送信信号からの損失の評価が行われる。ステップS207では、バックツーバック(b2b)接続の評価が行われる。このステップでは、事後補償手段303が、バックツーバックセットアップにおいて最初に実装される。
歪み識別手段401は、第1の例示的な実施形態の歪み識別手段201と同一であり、信号修正手段402は、信頼区間Tを設定するための追加の入力を使用する信号修正手段202と同様である。歪み識別手段401及び信号修正手段402は、ノイズ除去システム400に相当する。
事前補償手段403は、入力によって実現される適合のレベル又は精度を示す追加の出力を伴う事前補償手段203に相当する。フィードバックは、損失関数とし得る事前補償手段403からの損失関数の出力に相当する。損失関数の出力は、当該出力と、デジタル送信信号に基づいて期待される出力を比較することにより、学習済みの関数の精度を示す。このフィードバックは、信号修正手段402によって使用されるパラメータ、例えば、信頼区間「T」等を調整するために使用される。この信頼区間「T」の調整は、次の反復処理において損失関数値をフィードバックすることを目的として行われ、すなわち、事前補償手段403が、信号修正手段402からの新しい出力を用いてトレーニングされる。これにより、信号修正手段402からの出力もまた、事前補償手段403の以前の状態に依存することが保証される。
さらに、ステップS303では、ブラインド教師なしクラスタリングアルゴリズムのトレーニングが行われ、すなわち、教師なし学習アルゴリズムとして上述した教師なしクラスタリングアルゴリズムを実装する歪み識別手段401が、ステップS302で収集された出力データに対して実行される。
ステップS305では、事前補償手段403のトレーニングが行われる。このステップでは、この修正された信号が、事前補償手段403に提供され、トレーニングが、デジタル送信信号を用いて行われる。トレーニングが収束した後、ステップS306でデジタル送信信号からの損失の評価が行われる。
ステップS306からの損失に基づいて、次のステップが決定される。ステップS307では、損失が、期待される範囲内であるか否か判定される。ステップS307は、ノイズ除去システム400の損失判定手段によって実行され得る。この損失判定手段は、少なくとも1つのプログラムを有する少なくとも1つのコンピュータによって実現できる。この期待される範囲は、損失の許容範囲に対応する。損失関数/精度基準は、損失を評価するために使用される。選択された損失関数が「平均二乗誤差」の場合、評価対象の損失は、既定の損失閾値と比較される。評価対象の損失が、この既定値より小さい場合、損失が、期待される範囲内であると判断される。
損失が、期待される範囲内でない場合(NO)、ステップS308において、信号修正手段402で使用される信頼区間Tが、事前補償手段403からのフィードバックに基づいて更新される。ステップS308は、ノイズ除去システム400の更新手段によって実行され得る。更新手段は、少なくとも1つのプログラムを有する少なくとも1つのコンピュータによって実現できる。このプロセスでは、補償手段からのフィードバックは、損失関数の出力等の値であり、これは、デジタル送信信号に対する学習済みの関数の適合を示す。この推定は、信頼区間で行われたデータの信号の修正が正確であるか否かを示す。フィードバックとして使用される損失関数の値が高い場合、信号修正手段からの新たな出力によって、損失関数の値が減少し、又は適合が改善するように、使用される信頼区間が適切に調整される。そして、この処理は、ステップS304に戻る。
損失が、期待される範囲内の場合(YES)、ステップS309に処理が進む。ステップS309では、バックツーバック(b2b)接続の評価が行われる。このステップでは、事前補償手段403が、バックツーバックセットアップにおいて最初に実装される。
本実施形態に示す歪み識別手段501は、歪み識別手段301と同一であり、信号修正手段502は、信頼区間「T」の設定に使用される事後補償手段503からの追加の入力(フィードバック)を除き、信号修正手段302と同一である。歪み識別手段501及び信号修正手段502は、ノイズ除去システム500に相当する。
事後補償手段503は、実現された適合の精度の程度を示す追加の出力を伴う事後補償手段303と同様の手段である。フィードバックは、損失関数とし得る事後補償手段503からの損失関数の出力に相当し得る。このフィードバックは、信頼区間「T」のような信号修正手段502のパラメータを調整するために使用される。
さらに、ステップS403では、ブラインド教師なしクラスタリングアルゴリズムのトレーニングが行われ、すなわち、教師なし学習アルゴリズムとして上述した教師なしクラスタリングアルゴリズムを実装する歪み識別手段501が、ステップS402で収集された出力データに対して実行される。
ステップS405では、事後補償手段303のトレーニングが行われる。このステップでは、この修正された信号が、事後補償手段503に提供され、デジタル送信信号を用いて、トレーニングが行われる。
ステップS406からの損失に基づいて、次のステップが決定される。ステップS407では、損失が、期待される範囲内か否か判定される。ステップS407は、ノイズ除去システム500の損失判定手段によって行うことができる。損失判定手段は、少なくとも1つのプログラムを有する少なくとも1つのコンピュータによって実現することができる。この期待される範囲は、損失の許容範囲に対応する。損失が、期待される範囲内でない場合(NO)、ステップS408で、信号修正手段502で使用される信頼区間Tが、事後補償手段503からのフィードバックに基づいて更新される。ステップS408は、ノイズ除去システム500の更新手段によって実行してもよい。更新手段は、少なくとも1つのプログラムを有する少なくとも1つのコンピュータによって実現することができる。そして、この処理は、ステップS404に戻る。損失が、期待される範囲内である場合(YES)、処理は、ステップS409に進み、バックツーバックセットアップにおいて、事後補償手段503が最初に実装される。
まず、光送信機101及び光受信機102は、バックツーバックのデータ接続で接続される。この例では、16-QAMコンスタレーションに属する入力シンボルが送信される。光受信機102からの出力は、ADC102bを通過した、コヒーレント受信機からの信号であり、これはさらに、フィルタリング、サンプリングを含むアルゴリズムを有するDSP102cによって処理される。
次いで、光受信機102からの出力は、歪み識別手段401に提供される。歪みがガウス分布に従うとの仮定に基づき、歪み識別手段401が実行される。したがって、歪み識別手段401で使用される教師なしクラスタリングアルゴリズム(UCA:Unsupervised Clustering Algorithm)は、混合ガウスモデル(GMM)である。この場合のGMMは、コンスタレーションの順序と等しい16個のクラスタを有するように設定される。GMMは、16個のクラスタ中心の座標と、2次元座標における中心を基準とした分散を含む解に収束する信号を使用する。この情報は、クラスタの1つ(すなわち、16個の送信シンボルのうち1個)に属する信号の可能性を推定するために使用できる。
この修正の後、信号修正手段402の出力は、事前補償手段403に提供され、ここで、この信号データは、入力シンボルパターン(デジタル送信信号)と共に、適切な補償関数を適合させるために使用される。この適合は、FIRフィルタ係数、メモリ若しくはボルテラ多項式係数、又はニューラルネットワークの重みを学習するために使用できる。
この適合の後、適合の損失又は不正確さを評価してもよい。損失が許容可能な閾値を超える場合、信頼閾値Tの値が修正され、信号修正手段402は、損失が許容範囲内になるまで、信号を修正する。
本発明は、光送信機の効果に関する事前歪みシステムに適用可能である。具体的には、信号からノイズ及び他の望ましくない特性を除去することを目的とする信号修正プログラムである。本発明は、光送信機の効果を補償する事後補償システムにも適用可能である。本発明は、事前モード又は事後モードのいずれかの光通信の効果を補償することを目的とする光通信にも適用可能である。
他の種々の変形は、当業者には明らかであり、ここでは、これ以上詳細に説明しない。提案された特許は、最初のバックツーバックデータ接続及びトレーニングの代わりに、チャネルを備えた完全な光通信システムに直接実装できることに留意されたい。換言すると、歪み補償関数は、完全な光通信システムからのデータを用いて学習することができる。
プログラムは、コンピュータに読み込まれた場合に、実施形態で説明された1又はそれ以上の機能をコンピュータに行わせるための命令群(又はソフトウェアコード)を含む。プログラムは、非一時的なコンピュータ可読媒体又は実体のある記憶媒体に格納されてもよい。限定ではなく例として、コンピュータ可読媒体又は実体のある記憶媒体は、random-access memory(RAM)、read-only memory(ROM)、フラッシュメモリ、solid-state drive(SSD)又はその他のメモリ技術、CD-ROM、digital versatile disk(DVD)、Blu-ray(登録商標)ディスク又はその他の光ディスクストレージ、磁気カセット、磁気テープ、磁気ディスクストレージ又はその他の磁気ストレージデバイスを含む。プログラムは、一時的なコンピュータ可読媒体又は通信媒体上で送信されてもよい。限定ではなく例として、一時的なコンピュータ可読媒体又は通信媒体は、電気的、光学的、音響的、又はその他の形式の伝搬信号を含む。
[Title of the Invention] A DE-NOISING SYSTEM, A SIGNAL PROCESSING SYSTEM, A DE-NOISING METHOD, AND A DE-NOISING PROGRAM
[TECHNICAL FIELD]
[0001]
The present invention relates to a signal de-noising system, a signal processing system, a de-noising method, and a de-noising program for compensation of distortion, in particular a compensation of distortion in optical communication.
[BACKGROUND ART]
[0002]
An example of a conventional distortion compensation system is disclosed in Patent Literature 1. As illustrated in FIG.1, an optical communication system denoted by 100 includes an optical transmitter 101, an optical receiver 102 and a pre-compensation means 103. The optical transmitter 101 includes a DAC (Digital to Analog Convertor) 101a and transmit means 101b. The transmit signal inputted to the optical transmitter 101 is through various sub devices such as DAC 101a which converts the input signal in digital form into analog form for transmission. The transmit means 101b may include a driving amplifier, Mach Zehnder modulator which are indicated in 101b.
[0003]
When the optical receiver 102 receives an optical transmit signal from the optical transmitter 101, the receive means 102a recovers the optical signal by using an optical recovery method such as coherent recovery. The recovered optical signal is then provided to Analog to Digital Convertor 102b. The ADC 102b converts the analog optical signal to a digital signal. The DSP (Digital Signal Processing) 102c processes digital signal. The DSP 102c includes various signal processing operations such as synchronization and filtering.
[0004]
The process in the optical transmitter 101 is not ideal and thus induces distortion which are not desirable. These distortions impact the transmission performance, and therefore need to be compensated adequately. The pre-compensation means 103 is utilized to add pre-compensated component to the input.
[0005]
The conventional system having such a structure operates as follows. Initially the input signal which consist of the encoded message to be transmitted is given to the DAC 101a of the optical transmitter 101. This analog input is then converted to an optical signal and sent to the optical receiver 102. The optical signal at the receiver 102 is recovered by the means 102a. The input to the optical transmitter 101 and output from the receiver 102 are utilized to learn a pre-compensation function in the pre-compensation means 103. This may be achieved by utilizing various function learning methods on the input and output data such as filter coefficient learning, memory polynomial and neural networks. This may be as depicted in Non-Patent Literature 1 where neural network based function learning is described. As described in FIG. 1, the pre-compensation means 103 comprising the compensation function learnt by using the input and output data may be implemented before the optical transmitter 101 in the implementation phase.
[0006]
The above-described method can also be implemented as a post compensation scenario where the compensation function is implemented after the optical receiver 102. In this scenario a compensation means is implemented after the optical receiver 102 and not before the optical transmitter 101.
[Citation List]
[Patent Literature]
[0007]
[Patent Literature 1] US7756421B2, Electrical domain compensation of non-linear effects in an optical communications system
[Non-Patent Literature]
[0008]
[Non-Patent Literature 1] G. Paryanti, H. Faig, L. Rokach and D. Sadot, "A Direct Learning Approach for Neural Network Based Pre-Distortion for Coherent Nonlinear Optical Transmitter," in Journal of Lightwave Technology, vol. 38, no. 15, pp. 3883-3896, 1 Aug.1, 2020, doi: 10.1109/JLT.2020.2983229
[SUMMARY]
[Problem to be solved by the Invention]
[0009]
A first problem is distortion compensation function inaccurately learning from collected signal data. The reason for the occurrence of the first problem is that parts of the distortion components of the distortion are non-deterministic in nature and hence not learnable. In other words, some parts of the distortion may be noise or other elements which follow a distribution making it non-deterministic and therefore they are not learnable.
[0010]
A second problem is that the learnt distortion compensation function is suitable only for the operating environment in which signal data is collected. The reason for the occurrence of the second problem is the function learning might have inaccurately learnt to compensate for operation condition noise profile which is likely to change when the operating condition is changed. Operating condition noise profile refers to the properties of system such as SNR and other noise characteristics. As input power or any other system condition is varied, this noise profile is bound to change and therefore the compensation function may also need to be changed.
[0011]
A third problem is that large amount of data is required to accurately learn the distortion compensation function. The reason for the occurrence of the third problem is that large amount of data is needed to eliminate non-deterministic effects and learn an accurate distortion compensation function for the deterministic effects.
[Objective of the Invention]
[0012]
An objective of this disclosure is to provide a de-noising system, a signal processing system, a de-noising method and de-noising program capable of solving at least one of the above-described problems.
[Means for Solving the Problem]
[0013]
A de-noising system for an optical communication system including an optical transmitter and an optical receiver is proposed. The de-noising system includes:
a distortion identify means that identifies a distortion of input signals from the optical receiver in an unsupervised manner to output distortion parameters indicating the distortion identified; and
a signal trim means that utilizes the outputted distortion parameters to modify signal inputted to the optical transmitter with or without the assistance of the ground truth.
[0014]
A signal processing system for an optical communication system including an optical transmitter and an optical receiver is proposed. The de-noising system includes:
a distortion identify means that identifies a distortion of input signals from the optical receiver in an unsupervised manner to output distortion parameters indicating the distortion identified;
a signal trim means that utilizes the outputted distortion parameters to modify signal inputted to the optical transmitter; and
a compensation means that utilizes the modified signal along with the ground truth to learn a function to compensate a distortion of at least one signal inputted to the compensation means.
[0015]
A de-noising method performed by a computer for an optical communication system including an optical transmitter and an optical receiver is proposed. The method includes:
identifying a distortion of input signals from the optical receiver in an unsupervised manner to output distortion parameters indicating the distortion identified; and
utilizing the outputted distortion parameters to modify signal inputted to the optical transmitter with or without the assistance of the ground truth.
[0016]
A de-noising program for an optical communication system including an optical transmitter and an optical receiver is proposed. The program causes a computer to execute:
identifying a distortion of input signals from the optical receiver in an unsupervised manner to output distortion parameters indicating the distortion identified; and
utilizing the outputted distortion parameters to modify signal inputted to the optical transmitter with or without the assistance of the ground truth.
[Effect of the Invention]
[0017]
A first effect is to ensure that accurate learning of the distortion compensation function is possible. The reason for the first effect is that the de-noising means is able to eliminate or reduce the un-learnable or non-deterministic effect in the signal data.
[0018]
A second effect is to ensure that the distortion characteristics are accurately learnt in a blind or unsupervised manner and used to learn the compensation profile. The reason for the second effect is that the distortion identify means is able to accurately identify the distortion characteristics from the provided unlabeled signal data.
[0019]
A third effect is to ensure that the distortion characteristics learning data requirement is reduced. The reason for the effect is that the de-noising means eliminates un-learnable or non-deterministic components in the data so less data is required to infer the accurate distortion compensation function.
[BRIEF DESCRIPTION OF THE DRAWINGS]
[0020]
[FIG. 1] A block diagram illustrating the structure of the prior art: an optical communication system with a pre-compensation.
[FIG. 2] A diagram illustrating a first example embodiment with pre-compensation implemented along with our proposed method.
[FIG. 3] A flow diagram illustrating the flow of the operation of the first example embodiment.
[FIG. 4] Main components of a de-noising system.
[FIG. 5] Main components of a signal processing system.
[FIG. 6] A diagram illustrating a second example embodiment with post compensation implemented along with our proposed method.
[FIG. 7] A flow diagram illustrating the flow of the operation of the second example embodiment.
[FIG. 8] A diagram illustrating a third example embodiment with pre-distortion implemented along with our proposed method.
[FIG. 9] A flow diagram illustrating the flow of the operation of the third example embodiment.
[FIG. 10] A diagram illustrating a fourth example embodiment with post compensation implemented along with our proposed method.
[FIG. 11] A flow diagram illustrating the flow of the operation of the fourth example embodiment.
[FIG. 12] A diagram indicating the functioning of our proposed means for a QPSK constellation.
[FIG. 13] A diagram indicating the distortion characteristic learning for a 16QAM optical b2b setup.
[FIG. 14] An example of pseudo code summarizing the third example embodiment.
[EXAMPLE EMBODIMENT]
[First Example Embodiment]
[Explanation of Structure]
[0021]
First, a first example embodiment of the invention is elaborated below referring to the accompanying drawings.
[0022]
Referring to FIG. 2, an optical communication system 20 according to the first example embodiment is described. The optical communication system 20 includes a distortion identify means 201, a signal trim means 202 and a pre-compensation means 203. The distortion identify means 201 and the signal trim means 202 correspond to a de-noising system 200. These means 201 to 203 are embodied by at least a computer with at least a program. The computer may include various IC such as CPU (Central Processing Unit), processor, data processing device, FPGA (Field Programmable Gate Array), and ASIC (Application Specific Integrated Circuit).
[0023]
These proposed means 201,202 may be included into the existing prior art setup of the system 100 including the optical transmitter 101, the optical receiver 102 and the pre-compensation means 103. The means 203 may be identical to the pre-compensation means 103 as in FIG. 1.
[0024]
Those means generally operate as follows.
[0025]
The means 201 includes a learning algorithm of the unsupervised class that learns parameters that enable it to characterize the signal statistical properties. The unsupervised class of learning algorithms aim to learn the objective from the signal without any label. The label in this case could be the actual transmit signal (i.e. ground truth). The objective is to learn blindly the signal characteristics such as means and variance for the symbols, i.e. without assistance of label/unsupervised. Specifically, for a k-th order signal constellation, a typical learning algorithm aims to learn parameters such as mean (i.e. center) and variance corresponding to each of the k unique transmit symbols. More parameters may also be learnt to characterize the signal by this means 201. A variety of unsupervised learning algorithms may be adopted. For example, the learning algorithm could be a Gaussian mixture model (GMM) that aim to learn the signal characteristics under the assumption that signal characteristics properties belong to the Gaussian distribution class. According to the system operating condition, the distribution being assumed to learn the parameters can be adjusted. In other words, an appropriate mixture model as the unsupervised learning algorithm may be adopted depending on a distribution the noise may belong to.
[0026]
The means 202 utilizes the parameters learnt from the means 201, which are then used in order to modify the signal. This modification may be undertaken by using properties associated with the distribution along with the learned parameters identified from the previous means 201. The identification of the distribution to which the distorted signal most appropriately corresponds to is a critical step achieved by the means 201 and facilitates the further modification of the signal in the means 202. One of such properties is the “confidence interval” of the distribution that indicates the likelihood that a signal point corresponds to a particular symbol from among the k input symbols. This modification may be done using the ground truth about the actual signal using a confidence interval threshold “T” as follows.
Equation 1
Where f( ) is the signal modifying function which is dependent on the input from the means 201 (i.e. y), the confidence interval T, the parameters learned from the means 201 as “w” and the ground truth as “x”.
[0027]
vThe means 202 may also modify the signal without the ground truth as follows.
Equation 2
The de-noise is implemented by the means 201 and 202 at the equation 1 and 2.
[0028]
The selection of operation of the means 202 can be 2 modes: a first mode with ground truth and a second mode without ground truth. In the first mode with ground truth, the means 202 uses the actual signal ground truth to set the signal modification/trimming. The second mode may be used in scenarios where the actual signal ground truth is difficult to obtain or unavailable. In the second mode without ground truth, the means 202 uses the likelihood estimation for each symbol from the means 201 to form a “pseudo ground truth”. And the means 202 uses the “pseudo ground truth” to set the signal modification/trimming.
[0029]
The output from the means 202 is provided to the pre-compensation means 203. The pre-compensation means 203 corresponds to a compensation means to compensate a distortion of at least one input signal.
[0030]
Those means 201 and 202 mutually operate in such a way that de-noise the data from the optical receiver 102 to achieve the performance improvement for the system 200 especially by improving the data to the means 203.
[Description of Operation]
[0031]
Next, referring to flowcharts in FIG. 3, the general operation of the first example embodiment is elaborated.
[0032]
First, in step S101 in FIG. 3, back to back (b2b) connect is established, that is optical setup is run without any compensation in a back to back setup with the optical transmitter 101 and the optical receiver 102 connected. In this setup, the input transmit symbols are passed through the optical transmitter 101 and the optical receiver 102 and then received symbol which is the output from the optical receiver 102 is collected by the means 201. The input transmit pattern consists of the message to be transmitted encoded with appropriate constellation symbols. This input signal gets converted to electrical form then to optical form and then finally back to electrical form in the optical receiver 102 which is converted back to the digital form at the output of the optical receiver 102. Then, in step S102 both the input signal to the optical transmitter 101 and output signal from the optical receiver 102 are collected. The output signal to the optical receiver 102 is utilized by the means 201. The input signal to the optical transmitter 101 may be utilized by the means 202 and the means 203. The means 202 may or may not utilize the input signal depending on whether signal trimming is supervised or unsupervised.
[0033]
Further, in step S103, Blind Unsupervised Clustering Algorithm training is done, that is the means 201 which may implement an Unsupervised Clustering Algorithm as described before as unsupervised learning algorithm is run on the collected received signal output data in step S102. The means 201 learns the characteristics of the received constellation with the objective of classifying the received signals as the original ground truth symbol. Generally, an Unsupervised learning algorithm as described in the means 201 aims to classify (i.e. cluster) the signals. We repurpose this algorithm to just extract the parameters that are used to classify the signals. The extracted parameters are passed to the means 202.
[0034]
In step S104, modification of dataset (i.e. signals) for confidence region T
i is done. In this step, the output and learned parameters from the means 201 are is passed on to the means 202 in order to de-noise and modify the signal. This step involves utilizing the parameters from the means 201 as described in the Explanation of structure section in order to generate the output for the means 202. This output is modified signals.
[0035]
In step S105, training of the means 203 is done. In this step, this modified signal is passed onto the means 203 where training is done along with the ground truth. The means 203 may use the ground truth available from the optical transmitter 101.
[0036]
After convergence of training evaluation of loss from the ground truth is done in step S106 by a loss function. The evaluation of loss from the ground truth is an automated process such as in neural network training. The evaluation of loss from the ground truth is to evaluate the fit/accuracy of the pre-compensation means 203. The evaluation criteria may be a loss function such as “mean square error loss”. This loss is evaluated by comparing the learnt compensation output with the expected ideal value generated from the available ground truth. This loss evaluation methodology may change depending on the implementation of the structure of the means 203 such as filter, memory polynomial or Artificial Neural Network. Output from the loss function is used to modify the weights function using iterative algorithms such as back propagation to ensure convergence and minimal loss.
[0037]
In step S107, evaluation of back to back (b2b) connect is done by a simple evaluation using loss function or computing BER (Bit Error Ratio). This process may be manual or automated with well-defined functions. In this step, the means 203 is implemented first in the b2b setup. In step S107, the performance of the complete system is evaluated. That is, the optical transmitter 101 is connected to the optical receiver 102 setup in the back to back arrangement and evaluation is done with and without the presence of the pre-distortion function. This step is done to validate the performance improvement achieved due to the mitigation of distortions by the means 203.
[0038]
In step S108, in presence of the optical channel which is connected in between the optical transmitter 101 and the optical receiver 102 deployment with the channel is done. In other words, the optical fiber is connected between the optical transmitter 101 and the optical receiver 102. This deployment can be done manually or automatically. The b2b setup is utilized during the initial steps so that distortions induced by the optical transmitter 101 are only present in the signal data used for learning the distortion mitigation strategy. However, an online communication system may include at least a channel such as an optical fiber cable placed between the optical transmitter 101 and the optical receiver 102. Therefore, in the deployment stage, learnt compensation function is implemented in an optical communication system with optical fiber.
[0039]
FIG. 4 describes main components of the de-noising system 200 for an optical communication system including the optical transmitter 101 and the optical receiver 102. The de-noising system 200 includes the distortion identify means 201 and the signal trim means 202.
[0040]
As described above, the distortion identify means 201 identifies a distortion of input signals from the optical receiver 102 in an unsupervised manner to output distortion parameters indicating the distortion identified. The signal trim means 202 utilizes the outputted distortion parameters to modify signal inputted to the optical transmitter 101 with or without the assistance of the ground truth with the objective of denoising the signal. The modified signal is used for a distortion function compensating distortions of signals used in the optical communication system to learn along with the ground truth. Therefore, distortions induced the optical transmitter 101 are mitigated.
[0041]
Further, the distortion identify means 201 is configured to learn the distortion parameters by learning statistical information for probability distributions of the distortion identified.
[0042]
Further, the distortion identify means 201 is configured to learn the statistical information with learning algorithms.
[0043]
Further, the distortion identify means 201 is configured to cluster and separate the input signal according to at least one transmit symbol of the input signal.
[0044]
Further, the signal trim means 202 is configured to modify the signal based on statistical properties of the signal, the statistical properties include likelihood of the signal.
[0045]
FIG. 5 describes main components of a signal processing system for an optical communication system including the optical transmitter 101 and the optical receiver 102. The signal processing system includes the distortion identify means 201 and the signal trim means 202, and the compensation means 203.
[0046]
As described above, the compensation means 203 utilizes the modified signal along with the ground truth to learn a function to compensate distortions of at least one signal inputted to the compensation means 203. Therefore, distortions induced the optical transmitter 101 are mitigated.
[0047]
Further, the function of the compensation means 203 compensates a distortion of at least one signal inputted to the optical transmitter.
[Description of Effect]
[0048]
Next, the effect of the first example embodiment is described with aid from FIG. 12. FIG. 12 illustrates an ideal QPSK constellation of de-noising system according to this disclosure. In Fig.12, the character d refers to the variance associated with the distribution for the labeled symbol before the signal trimming means. The character d’ refers to the variance associated with the distribution for the symbols after application of the signal trim means.
[0049]
The constellation 001 indicates a constellation after extraction of the distribution statistical information from signal. The constellation 002 indicates a constellation after modification of the signal characteristics by modifying the signal statistics.
[0050]
As the first example embodiment is configured in such a manner that distortion characteristics is learned by the algorithm in the means 201 as indicated in 001 in FIG. 12, it is possible to learn the distortion characteristics as learnable parameters.
[0051]
In addition, the example embodiment is configured in such a manner that the component of the distortion can be trimmed according to the distribution identified by the means 202 and some fixed confidence threshold related to the likelihood function as in 002 in FIG. 12, which enables the pre-compensation means 203 to be more accurate in modelling and require less data to converge to an accurate pre-distortion profile. Note that this de-noising occurs for all the symbols in Fig 12, that is all symbols in Fig12 b will be demolished.
[Second Example Embodiment]
[Explanation of Structure]
[0052]
Next, a second example embodiment of the invention is elaborated below referring to the accompanying drawings.
[0053]
Referring to FIG. 6, an optical communication system 30 according to the second example embodiment is described. The optical communication system 30 includes a distortion identify means 301, a signal trim means 302 and a post compensation means 303. The distortion identify means 301 and the signal trim means 302 correspond to a de-noising system 300. These means 301 to 303 are embodied by at least a computer with at least a program. The computer may include various IC such as CPU, processor, data processing device, FPGA, and ASIC.
[0054]
The second embodiment aims to showcase how the proposed de-noising system 300 is implemented in a post compensation scenario where a compensation block is placed after the optical receiver 102.
[0055]
The proposed means are included into the existing prior art setup 100 that includes the optical transmitter 101, the optical receiver 102. Instead of a pre-compensation means 203, we look at a post compensation scheme with a post compensation means 303. The post compensation means 303 corresponds to a compensation means to compensate a distortion of at least one input signal.
[0056]
Note that the post compensation means may indicate a variation from the prior art in FIG. 1 but this does not represent a novelty. Post compensation is a fairly well-known technique for optical system effect compensation. The means 301 and 302 implemented before the means 303 indicate the implementation of the proposal in this embodiment.
[0057]
The means 303 is similar to the means 103 and may be operated using Neural networks, filters, memory polynomial and other techniques which aim to infer the desired function from the available data.
[0058]
Those means generally operate as follows.
[0059]
The means 301 consists a learning algorithm of the unsupervised class that learns parameters that enable it to characterize the signal statistical properties. The unsupervised class of learning algorithms aim to learn the objective from the signal without any label. The label in this case could be the actual transmit signal (i.e. ground truth). For a K-th order signal constellation, a typical learning algorithm aims to learn parameters such as mean (i.e. center) and variance corresponding to each of the k unique transmit symbols. More parameters may also be learnt to characterize the signal by this means 301. The learning algorithm could be a Gaussian mixture model (GMM) that aim to learn the signal characteristics under the assumption that signal characteristics properties belong to the Gaussian distribution class. According to the system operating condition, the distribution being assumed to learn the parameters can be adjusted.
[0060]
The means 302 utilizes the parameters learnt from the means 301 which are then used in order to modify the signal. This modification may be undertaken by using properties associated with the distribution along with the learned parameters. One of such properties is the “confidence interval” of the distribution that indicates the likelihood that a signal point corresponds to a particular symbol from among the k input symbols. This modification may be done using the ground truth about the actual signal using a confidence interval threshold “T” as follows.
Equation 3
Where f() is the signal modifying function which is dependent on the input from the means 301 (i.e. y), the confidence interval T, the parameters learned from the means 301 as “w” and the ground Truth as “x”.
[0061]
The means 302 may also be utilized without the ground truth as follows.
Equation 4
[0062]
The output from the means 302 is provided to the means 303.
[0063]
In an alternative implementation, the means 301, 302, 303 may be utilized jointly to learn the most optimal settings to compensate the optical distortions in a post compensation manner. For a neural network based means 303, this joint learning may be implemented by a multi-loss objective function which involves the output and parameters from the means 301 and 302. Multi-loss objective function corresponds to scenario where weighted combination of more than one loss function is used in training a neural network. A possible combination of two loss functions could be mean square error and the likelihood function for the ground truth symbol. The likelihood function would derive parameters from the means 301 and 302 in this scenario. This could be implemented as follows.
Equation 5
Where g( ) is a function to make the likelihood function suitable as a loss function to ensure convergence of training. l
1, l
2 are scalar weights used for the weighted combination of the loss functions.
[0064]
Those means 301 and 302 mutually operate in such a way that de-noise the data from the optical receiver 102 to achieve the performance improvement for system 300 especially by improving the data to the means 303.
[Description of Operation]
[0065]
Next, referring to flowcharts in FIG. 7, the general operation of the second example embodiment is elaborated.
[0066]
First, in step S201 in FIG. 7, back to back (b2b) connect is established, that is optical setup is run without any compensation in a back to back setup. In this step, the input transmit symbols are passed through the optical transmitter 101 and the optical receiver 102 and received symbol is the output from the optical receiver 102. Then, in step S202 the output data is collected by the means 301.
[0067]
Further, in step S203, Blind Unsupervised Clustering Algorithm training is done, that is the means 301 which may implement an Unsupervised Clustering Algorithm as described before as unsupervised learning algorithm is run on the collected output data in step S202.
[0068]
In step S204, modification of dataset (i.e. signals) for confidence region T
i is done. In this step, learned parameters from the means 301 is utilized by means 302 in order to de-noise and modify the received signal output.
[0069]
In step S205, training of the means 303 is done. In this step, this modified signal output from the means 302 is passed onto the post compensation means 303 where training is done along with the ground truth with an objective to retrieve the original signal from the modified signal. This objective is in contrast to the objective of learning the pre-compensation on the input of the means 203 in the first example embodiment. This training/fitting could utilize any of the known methods for learning post compensation using the dataset which contains the input symbol pattern and the modified signal pattern from the means 302.
[0070]
After convergence of training evaluation of loss from the ground truth is done in step S206. In step S207, evaluation of back to back (b2b) connect is done. In this step, the means 303 is implemented first in the b2b setup.
[0071]
In step S208, in presence of the optical channel which is connected in between the optical transmitter 101 and the optical receiver 102 deployment with the channel is done.
[Description of Effect]
[0072]
Next, the effect of the second example embodiment is described with aid from FIG. 12.
[0073]
As the second example embodiment is configured in such a manner that distortion characteristics is learned by the algorithm in the means 301 as indicated in 001 in FIG. 12, it is possible to learn the distortion characteristics as learnable parameters which include statistical information of the signal constellation.
[0074]
In addition, the second example embodiment is configured in such a manner that the component of the distortion can be trimmed as in 002 in FIG. 12 according to the distribution identified by the means 302 using the learned information from the means 301 along with a fixed confidence parameter related to the likelihood, which enables the post compensation means 303 to be more accurate in modelling and require less data to converge.
[0075]
Further, the function of the post compensation means 303 compensates a distortion of at least one signal outputted from the optical receiver.
[Third Example Embodiment]
[Explanation of Structure]
[0076]
Next, a third example embodiment is elaborated referring to the accompanying drawings. Referring to FIG. 8, an optical communication system 40 according to the third example embodiment is described. The third example embodiment is identical to the first example embodiment in every respect except for a feedback from a pre-compensation means 403 to the means 402.
[0077]
A distortion identify means 401 is identical to the means 201 of the first example embodiment while the signal trim means 402 is similar to means 202 with an additional input used to set the confidence interval “T”. The distortion identify means 401 and the signal trim means 402 correspond to a de-noising system 400.
[0078]
The means 403 is equivalent to the means 203 with an additional output to indicate the level or accuracy of fitting achieved from the input. The feedback may correspond to the loss function output from the means 403 which could be a loss function. The loss function output indicates the accuracy of the learnt function by comparing the output to the expected output based on the ground truth. This feedback is used to adjust parameters used by the means 402 such as confidence interval “T”. This adjustment of the confidence interval “T” is done with an objective to feedback loss function value in the next iteration, that is the means 403 is trained with new output from the means 402. This ensures that output from the means 402 is also dependent on the previous state of the means 403.
[Description of Operation]
[0079]
At least a program according to a third example embodiment is loaded onto at least a computer to control the operation of the computer. The computer may include various IC such as CPU (Central Processing Unit), processor, data processing device, FPGA (Field Programmable Gate Array), and ASIC (Application Specific Integrated Circuit). Under the control of the program, the computer performs the following processes, which are identical to those processes which are performed by the computers of the first example embodiment.
[0080]
Next, referring to flowcharts in FIG. 9, the general operation of the third example embodiment is elaborated.
[0081]
First, in step S301 in FIG. 9, back to back (b2b) connect is established, that is optical setup is run without any compensation in a back to back setup. In this step, the input transmit symbols are passed through the optical transmitter 101 and the optical receiver 102 and received symbol is the output from the optical receiver 102. Then, in step S302 the output data is collected by the means 401.
[0082]
Further, in step S303, Blind Unsupervised Clustering Algorithm training is done, that is the means 401 which may implement an Unsupervised Clustering Algorithm as described before as unsupervised learning algorithm is run on the collected output data in step S302.
[0083]
In step S304, modification of dataset (i.e. signals) for confidence region T
i is done. In this step, learned parameters from the means 401 is utilized by the means 402 along with the received signal and the confidence parameter “T” in order to de-noise the signal.
[0084]
In step S305, training of the means 403 is done. In this step, this modified signal is passed onto the means 403 where training is done along with the ground truth. After convergence of training the evaluation of loss from the ground truth is done in step S306.
[0085]
Based on the loss from step S306, the next step is decided. In step S307, it is determined whether the loss is within the expected range. The step S307 may be done by a loss determining means in the de-noising system 400. The loss determining means can be embodied by at least a computer with at least a program. This expected range correspond to acceptable range of the loss. A loss function/accuracy metric is utilized to evaluate the loss. In case the selected loss function is “mean square error”, loss evaluated would be compared to a preset loss threshold. If the evaluated loss is less than this preset value, the loss is determined to be within the expected range.
[0086]
If the loss is not within the expected range (NO), then the confidence interval T used in the means 402 is updated based on feedback from the means 403 in step S308. The step S308 may be done by an update means in the de-noising system 400. The update means can be embodied by at least a computer with at least a program. In this process, the feedback from the compensation mechanism is a value such as the loss function output which gives an indication of the fit of the learned function to the ground truth. This estimate gives an indication of whether the signal trimming on the data done with a confidence interval is accurate. If the loss function value used as feedback is high, the confidence interval used is adjusted appropriately so that the new output from the signal trim means leads to a reduced loss function value or better fit. And this process returns to step S304.
[0087]
In case the loss is within the expected range (YES), the process proceeds to step S309. In step S309, evaluation of back to back (b2b) connect is done. In this step, the means 403 is implemented first in the b2b setup.
[0088]
In step S310, in presence of the optical channel which is connected in between the optical transmitter 101 and the optical receiver 102 deployment with the channel is done.
[Description of Effect]
[0089]
Next, the effect of the third example embodiment is described with aid from FIG. 12 illustrating the effects on a QPSK signal constellation.
[0090]
As the third example embodiment is configured in such a manner that distortion characteristics is learned by the algorithm in the means 401 as indicated in 001 in FIG. 12, it is possible to learn the distortion characteristics as learnable parameters which include the statistical information such as means and variance.
[0091]
In addition, the third example embodiment is configured in such a manner that the component of the distortion can be trimmed according to the distribution identified by the means 402 as in 002 in FIG. 12 with feedback from the means 403. In other words, the signal trim means 402 utilize at least one feedback indicating an accuracy of the learnt function from the compensation means 403 to adaptively adjust output of the signal trim means. This enables the pre-compensation means 403 to be more accurate in modelling and require less data to converge with enhanced accuracy.
[Fourth Example Embodiment]
[Explanation of Structure]
[0092]
Next, a fourth example embodiment is elaborated referring to the accompanying drawings. Referring to FIG. 10, an optical communication system 50 according to the fourth example embodiment is described. The fourth example embodiment illustrating an implementation which is identical to the second example embodiment in every respect except for a post compensation means 503.
[0093]
A distortion identify means 501 presented in this embodiment is identical to the means 301 while a signal trim means 502 in identical to the means 302 except for an additional input (feedback) from the means 503 which is used to set the confidence interval “T”. The distortion identify means 501 and the signal trim means 502 correspond to a de-noising system 500.
[0094]
The means 503 is a similar means as the means 303 with an additional output indicating the degree of accuracy of fit achieved. The feedback may correspond to the loss function output from the means 503 which could be a loss function. This feedback is used to adjust the means 502 parameters such as confidence interval “T”.
[Description of Operation]
[0095]
At least a program according to the fourth example embodiment is loaded onto at least a computer to control the operation of the computer. The computer may include various IC such as CPU, processor, data processing device, FPGA, and ASIC. Under the control of the program, the computer performs the following processes, which are identical to those processes which are performed by the computers of the second example embodiment.
[0096]
Next, referring to a flowchart in FIG. 11, the general operation of the fourth example embodiment is elaborated.
[0097]
First, in step S401 in FIG. 11, back to back (b2b) connect is established, that is optical setup is run without any compensation in a back to back setup. In this step, the input transmit symbols are passed through the optical transmitter 101 and the optical receiver 102 and received symbol is the output from the optical receiver 102. Then, in step S402 the output data is collected.
[0098]
Further, in step S403, Blind Unsupervised Clustering Algorithm training is done, that is the means 501 which may implement an Unsupervised Clustering Algorithm as described before as unsupervised learning algorithm is run on the collected output data in step S402.
[0099]
In step S404, modification of dataset (i.e. signals) for confidence region T
i is done. In this step, learned parameters from the means 501 is utilized by the means 502 along with the received signal output in order to de-noise the signal.
[0100]
In step S405, training of the means 303 is done. In this step, this modified signal is passed onto the post compensation means 503 where training is done along with the ground truth.
[0101]
After convergence of training the evaluation of loss from the ground truth is done in step S406.
[0102]
Based on the loss from step S406, the next step is decided. In step S407, it is determined whether the loss is within the expected range. The step S407 may be done by a loss determining means in the de-noising system 500. The loss determining means can be embodied by at least a computer with at least a program. This expected range correspond to acceptable range of the loss. If the loss is not within the expected range (NO), then the confidence interval “T” used in the means 502 is updated based on feedback from the means 503 in step S408. The step S408 may be done by an update means in the de-noising system 500. The update means can be embodied by at least a computer with at least a program. And this process returns to step S404. In case the loss is within the expected range (YES), the process proceeds to step S409 where the means 503 is implemented first in the b2b setup.
[0103]
In step S410, in presence of the optical channel which is connected in between the optical transmitter 101 and the optical receiver 102 deployment with the channel is done.
[Description of Effect]
[0104]
Next, the effect of the fourth example embodiment is described with aid from the illustration in FIG. 12 indicating the proposed processing for QPSK signals.
[0105]
As the fourth example embodiment is configured in such a manner that distortion characteristics is learned by the algorithm in the means 501 as indicated in 001 in FIG. 12, it is possible to learn the distortion characteristics as learnable parameters corresponding to the statistical information which may include mean and divergence from mean.
[0106]
In addition, the fourth example embodiment is configured in such a manner that the component of the distortion can be trimmed according to the distribution identified by the means 502 as in 002 in FIG. 12 so that the statistical information of the signal is changed. This enables the post compensation means 503 to be more accurate in modelling and require less data to converge. Along with the feedback from the means 503, the correct statistical information is utilized to adapt the signal so that accurate fitting of the post compensation means 503 is achieved.
[Example]
[0107]
Next, the operation of a mode for carrying out the present invention is described by way of a concrete example which elaborates the third example embodiment.
[0108]
As illustrated in FIG. 8, a pre-compensation means 403 is connected to the optical transmitter 101 with the objective of mitigating distortions preliminary.
[0109]
First, the optical transmitter 101 and the optical receiver 102 are connected in a back to back data connection. In this example an input symbols belonging to the 16-QAM constellation is transmitted. The output from the optical receiver 102 is the signal from the coherent receiver passed through an ADC 102b which is further processed by DSP 102c having algorithms including filtering, sampling.
[0110]
The output from the optical receiver 102 is then passed to the means 401. The means 401 is implemented with the assumption that the distortions follow the Gaussian distribution. Therefore, the Unsupervised Clustering Algorithm (UCA) utilized in the means 401 is the Gaussian Mixture model (GMM). The GMM in this case is set to have 16 clusters equal to the constellation order. The GMM utilizes the signal to converge at a solution that includes the coordinates of the 16 cluster centers along with variance around the center in 2-D coordinates. This information can be utilized to estimate the likelihood of a signal belonging to one of the clusters (i.e. 1 of the 16 transmit symbols).
[0111]
In FIG. 13, the clustering of 16- QAM Received constellation from a 32 Gbaud b2b optical setup with significant non-linear distortions is described. The GMM is able to cluster the symbols and also identify the 16 cluster centers accurately. Let the cluster centers be demoted as C
i and the Variance as V
i for the i-th cluster. The signal along with the learned parameters from the GMM are passed on to the means 402. A simple signal trimming with the aid of the GMM parameters is proposed in this example as follows.
Equation 6
Where w corresponds to a scalar weight, the function L( ) indicates confidence region coordinates for a given variance V
i and confidence threshold T. The function L( ) may be a simple likelihood function for a Gaussian distribution with the input variance and confidence boundary. C
i corresponding to a signal x may be identified by computing the closest center to x or using additional information such as the ground truth (actual transmit symbol for x).
[0112]
After this modification, the output of the means 402 is to be passed on to the means 403 where this signal data along with input symbol pattern (ground truth) is used to fit an appropriate compensation function. This fitting may be used to learn a FIR filter coefficients, memory or Volterra polynomial coefficients or even a neural network weights.
[0113]
After this fitting, the loss or inaccuracy of the fit may be evaluated. If the loss is more than an acceptable threshold, the value of the confidence threshold T is modified and the means 402 modify the signal until the loss is within acceptable range.
[0114]
In FIG. 14, a pseudo code representing the above description is shown with the Digital Pre-distortion (DPD) corresponding to the pre-compensation means. Note that DPD may be replaced with the post compensation means. After the loss is within acceptable range, this fitted function is adopted and implemented in the flow before the Tx means 101b so that input pattern is pre-distortion to mitigate the distortion effects accurately. Since the exact function to be learnt is extracted accurately from the available data by de-noising (elimination of noise un-learnable and probabilistic components) using the means 401 and 402, the signal data volume required for function fitting will be significantly less.
[0115]
All other embodiments similar to the above mention example for the third example embodiment may be implemented.
[Industrial Applicability]
[0116]
The present invention is applicable to a pre-distortion system for optical transmitter effects. Specifically, a signal modifying program with the objective of removing the noise and other undesired characteristics from the signal. The present invention is also applicable to a post compensation system that compensates for the optical transmitter effects. The present invention may also be applicable to optical communication with the objective of compensating the optical communication effects in either pre or post mode.
[0117]
Various other modifications will be apparent to those skilled in the art and will not be described in further detail here. Note that the proposed patent may also be implemented directly on the complete optical communication system with channel instead of initial back to back data connection and training. In other words, the distortion compensation function may use data from the complete optical communication system to learn.
[0118]
The program includes instructions (or software codes) that, when loaded into a computer, cause the computer to perform one or more of the functions described in the embodiments. The program may be stored in a non-transitory computer readable medium or a tangible storage medium. By way of example, and not a limitation, non-transitory computer readable media or tangible storage media can include a random-access memory (RAM), a read-only memory (ROM), a flash memory, a solid-state drive (SSD) or other types of memory technologies, a CD-ROM, a digital versatile disc (DVD), a Blu-ray disc or other types of optical disc storage, and magnetic cassettes, magnetic tape, magnetic disk storage or other types of magnetic storage devices. The program may be transmitted on a transitory computer readable medium or a communication medium. By way of example, and not a limitation, transitory computer readable media or communication media can include electrical, optical, acoustical, or other forms of propagated signals.
[Reference Signs List]
[0119]
10,20,30,40,50 Optical communication system
101 Optical transmitter
101a Digital to Analog Converter
101b Transmit means
102 Optical receiver
102a Receive means
102b Analog to Digital Convertor
102c DSP
103,203,403 Pre-compensation means
200,300,400,500 A de-noising system
303,503 Post compensation means
201,301,401,501 Distortion Identify means
202,302,402,502 Signal Trimming means