JP4306299B2 - Control method and control apparatus for compensating for periodic disturbance in servo system - Google Patents

Description

【００３５】
このＲＲＯは位置の外乱として現れるが、これはラプラス変換Ｆ（ｓ）で図２（ｂ）に示すように力の外乱ｆ（ｔ）と等価である。
【００３６】
【数２】

【００３７】
ここで、Ｒ（ｓ）はＲＲＯ ｒ（ｔ）のラプラス変換である。したがって、
【００３８】
【数３】

【００３９】
ここで、Ｆ_ｉ＝−ｉ^２ω^２Ｒ_ｉ／ｋ_ｖｋ_ｙである。式（１）で指数がゼロの第１項は、定数であるため２重の微分後に消える。式（３）は、さらに、下記式（４）のように記述することができる。
【００４０】
【数４】

【００４１】
ここで、
【００４２】
【数５】

【００４３】
であり、且つこれらは、以下で定められる。
【００４４】
Ｘ（ｔ）＝［sinωt sin2ωt ・・・ sinnωt cosωt cos2ωt ・・・ cosnωt］^ＴＷ＝［F_１cosj1 F_２cosj2 ・・・ F_ｎcosjn F_１sinj1 F_２sinj2 ・・・ F_ｎsinjn］^Ｔ
三角関数の特性によって、次の行列運動方程式が得られる。
【００４５】
【数６】

【００４６】
ここで、
【００４７】
【数７】

【００４８】
且つ、
【００４９】
【数８】

【００５０】
である。
【００５１】
式（４）及び式（５）は、周期的外乱が三角多項式で表現され得ることを示しており、この三角多項式は、図３のニューラルネットワークで表される。図３のニューラルネットワークは２層になっている。すなわち、（２ｎ＋１）個のニューロンを有する隠れ層と、１つのニューロンからなる出力層の２層である。ニューロン３０３を除き、Ｘ_１，．．．，Ｘ_ｎ，Ｘ_ｎ＋１，．．．，Ｘ_２ｎで示されている隠れ層ニューロンの出力は、式（４）のＸのそれぞれの成分に等しい。
【００５２】
図３で、ニューロン３０３以外の各隠れ層ニューロンは、入力を、それぞれの値ａ_１，．．．，ａ_ｎ，ａ_ｎ＋１，．．．，ａ_２ｎで乗算する（この乗算は、素子３００として示されている）という簡単な機能を果たし、その結果を積分器３０１に入力して、隠れ層ニューロンの出力を生成する。ニューロン３０３は、特別なニューロンであり、一定の出力として１を出力し、出力層のバイアスとして作用する。ニューロン３０３は、他の隠れ層ニューロンと等しく取り扱うことが可能だが、その値ａ_０＝０の場合は、
【００５３】
【数９】

【００５４】
となる。
【００５５】
隠れ層からの各出力Ｘ_０，．．．，Ｘ_ｎ，Ｘ_ｎ＋１，．．．，Ｘ_２ｎは、１つのニューロンからなる出力層に入力される。この１つのニューロンは、入力された各出力Ｘ_０，．．．，Ｘ_ｎ，Ｘ_ｎ＋１，．．．，Ｘ_２ｎのそれぞれを、対応する重み付け値Ｗ_０，．．．，Ｗ_ｎ，Ｗ_ｎ＋１，．．．，Ｗ_２ｎで乗算して（３０４で示す乗算器で）、それぞれ結果３０６を得る。この結果３０６を加算器３０７を使用して合計し、ｆ_Ｎ（ｔ，Ｗ（ｔ））である出力３０８を得る。
【００５６】
調整可能な重みＷ_０，．．．，Ｗ_ｎ，Ｗ_ｎ＋１，．．．，Ｗ_２ｎの値は、後述する学習機構により変化させることが可能であり、動作３０５として図３に示されている。
【００５７】
このニューラルネットワークの構造は、所謂「放射状ニューラルネットワーク（Radial Neural Network）」に極めて類似している。その理由としては、このニューラルネットワークが、各フィードフォワードパスに、１つだけしか調整可能な重みを有しておらず、出力前に出口関数がないことがあげられる。すなわち、出力ニューロンの出力は、重み付けされた入力の合計にすぎず、その関数ではない。このニューラルネットワークと放射状ニューラルネットワークとの相違は、図３のニューラルネットワークが、それ自身の動作によって出力層への入力を三角関数の形式で得ることである。このため、図３のニューラルネットワークをハイブリッド三角関数ニューラルネットワーク（ＨＴＮＮ）と称する。
【００５８】
ＨＴＮＮについて、以下２点の注釈を加える。
【００５９】
１．ＨＴＮＮは、任意の周期的信号に近似することができる。これは、次のように数学的に表すことができる。周期Ｔを有する任意の連続周期的関数ｆ（ｔ＋Ｔ）＝ｆ（ｔ）、及び任意のδ＞０において、有限の値ｎ及び重み付けベクトルＷ^＊が存在し、下記式（７）のようになる。
【００６０】
【数１０】

【００６１】
２．周期Ｔ、及びｎ未満の次元を有する全三角多項式の中で、ＨＴＮＮは可能な限り最良の平均平方近似値をｆに与える。
【００６２】
以下に、図３のＨＴＮＮの学習機構及びシステム安定性について説明する。
【００６３】
図２（ｂ）から、ｙ＝ｙ_ｍならば、ＨＴＮＮを有しないＨＤＤサーボの原動力は、下記式（８）のように記述することができる。
【００６４】
【数１１】

【００６５】
公称サーボ制御器２００が、線形比例微分制御器である場合、下記式（９）が得られる。
【００６６】
【数１２】

【００６７】
ここで、ｅ＝ｙ^＊−ｙは、位置誤差（position error）である。
【００６８】
この制御器のゲインは下記式（１０）のとおりである。
【００６９】
【数１３】

【００７０】
ここで、ｋ_ｃ＞０，β_０＞０及びβ_１＞０は、下記式（１１）で定義される合成誤差（composite error）の係数である。
【００７１】
【数１４】

【００７２】
したがって、上記式（８）乃至式（１０）を考慮して、式（１１）を微分することによって、下記式（１２）が得られる。
【００７３】
【数１５】

【００７４】
ここで、ｆ_Ｒ（ｔ）＝−ｋ_ｖｋ_ｙｆ（ｔ）である。
【００７５】
ＨＴＮＮ２２２の出力ｆ_Ｎ（ｔ，Ｗ（ｔ））を考慮すると、下記式（１３）が得られる。
【００７６】
【数１６】

【００７７】
安定性の設計の場合は、下記式（１４）のように、リアプノフ関数が構築される。
【００７８】
【数１７】

【００７９】
ここで、Ｅ_Ｗ（ｔ）＝Ｗ^＊−Ｗ（ｔ）は重み誤差（weight error）である。
【００８０】
ＨＴＮＮ２２２が、連続時間で（例えば、アナログ回路によって）動作する場合は、ＨＴＮＮ２２２の学習ルールは、ここでは下記式（１５）となるように選ばれる。
【００８１】
【数１８】

【００８２】
ここで、
【００８３】
【数１９】

【００８４】
である。
【００８５】
したがって、ＨＴＮＮ２２２の動作、及び学習アルゴリズムは図４に概略的に示されているようになる。ここで、ボックス４００はＨＴＮＮ２２２を表し、ボックス４０１は学習機構２２１の動作を示す。ＨＴＮＮ２２２は、ＸをＡで乗算する連続ループを有し、Ｘは学習機構４０１にフィードバックされる。したがって、Ｖ（ｔ）の時間微分は、下記式（１６）のとおりである。
【００８６】
【数２０】

【００８７】
よって、
【００８８】
【数２１】

【００８９】
上記式（１７ａ）が満たされる時、式（１７ｂ）に示す関係が成り立つ。
【００９０】
したがって、システム合成誤差は、β_１δ／ｋ_ｃに収束するが、これは換言すると、誤差ｅ（ｔ）は次の線形微分方程式（１８）に従って収束することを意味する。
【００９１】
【数２２】

【００９２】
式（１８）の解は、下記式（１９）のとおりである。
【００９３】
【数２３】

【００９４】
式（１９）は、システムの誤差ｅ（ｔ）が、その初期値ｅ（０）から、β_１δ／（β_０ｋ_ｃ）のバイアスで、正の定数β_１δ／（β_０ｋ_ｃ）に向かって、β_０／β_１の指数定数で、指数関数的に収束することを示している。
【００９５】
以下に、ＨＴＮＮ及び学習機構のパラメータをどのようにして選択するかという点について説明する。
【００９６】
ＨＴＮＮの隠れ層ニューロンの初期状態、すなわちＸ（０）は、状態Ｘ（ｔ）が三角関数的であることを確実にするように割り当てられるべきであり、つまり下記式（２０）のようになる。
【００９７】
【数２４】

【００９８】
ＨＴＮＮ内の重みＷ（０）の初期値は、ＨＴＮＮがゼロ初期出力を有するように、ゼロに設定されてもよい。
【００９９】
【数２５】

【０１００】
ＨＴＮＮの次元ｎは、近似精度δに影響し、さらに、全体のシステム補償性能に影響する。一般に、ｎが大きいと、結果的にδが小さくなるので、良好な補償である。しかしながら、ＨＴＮＮの次元ｎは、実際には制限がある。例えば、ＨＴＮＮがアナログＶＬＳＩで実現される場合は、集積度で制限され、また、ＨＴＮＮがソフトウェアで実現される場合は、ＤＳＰの計算能力で制限されることがある。ＨＴＮＮの次元ｎを選択する現実的な方法は、誤差信号に対する高調波解析によるものである。誤差信号の高調波成分の内容から、ＨＴＮＮが全ての主な高調波成分を打ち消すことが確実となるように、適切なｎを選ぶことができる。ｎを設定する他の方法としては、ハードウェア又はソフトウェアの制限を考慮した上で、充分な補償性能が達成されるまでｎを初期値から増加させるというものがある。この方法は、ＨＴＮＮが動作設定前に、又はオンラインアダプテーションによって行われる。
【０１０１】
次に、パラメータβ_０、β_１及びｋ_ｃの設定方法について説明する。
【０１０２】
上記式（１０）から、下記式（２２）が得られる。
【０１０３】
【数２６】

【０１０４】
リニアシステムが臨界状態で動作するようにリニア制御器が設計されていると仮定した場合、すなわち、
【０１０５】
【数２７】

【０１０６】
であると仮定すると、下記式（２４）が得られる。
【０１０７】
【数２８】

【０１０８】
β_１＝ｋ_１と設定すると、β_０＝２ｋ_０となる。よって、式（９）及び式（１１）から、下記式（２５）が得られる。
【０１０９】
【数２９】

【０１１０】
学習機構において式（２５）を使用すると、通常現実的には問題となるｅ（ｔ）の微分の測定が避けられる。
【０１１１】
ここで、ハードウェア又はソフトウェアのいずれか一方を用いた実施の形態について説明する。
【０１１２】
先ず、ＤＳＰを使用してのソフトウェアによる実施の形態、及びアナログＶＬＳＩでのハードウェアによる実施の形態を以下に詳細に説明する。この場合、上述の連続方程式はもはや適切ではないので、代わりに、式（５）を下記式（２６）のように変換する。
【０１１３】
【数３０】

【０１１４】
ここで、Ｔはサンプリング時間で、
【０１１５】
【数３１】

【０１１６】
である。Ｘ_ｄ（ｋ）の初期値は
Ｘ_ｄ（０）＝［０_１×ｎ Ｉ_１×ｎ］^Ｔ
である。
【０１１７】
学習ルールを表す式（１５）は、下記式（２７）で置換される。
【０１１８】
【数３２】

【０１１９】
ＤＳＰソフトウェアにより学習機構が実現されるＨＴＮＮの概略ブロック図を図５に示す。ここで、５００はニューラルネットワーク本体であり、５０１は学習機構である。この場合、全体の構造は図４のものに似ているが、ブロック５００において、行列Ａが行列Ｄに置き換えられ、積分器はｚ^−１で示されている遅延ユニットに置き換えられている。同様に、学習機構に対応するブロック５０１において、積分器は遅延ユニットｚ^−１に置き換えられ、ブロック５０１の出力はこの遅延ユニットの合計、すなわち、式（２７）の結果で、前のサイクルにおけるＷである。
【０１２０】
次に、アナログＶＬＳＩによる実施の形態について説明する。この場合、上記式（５）及び式（１５）は有効であり、上記式（２６）及び式（２７）に置き換えられることはない。一実施例が図６に示されており、これは参考文献［６］（S. P. DeWeerth et al., "A neuron-based pulse servo for motion control", Proc. IEEE Conf. Robotics and Automation (Cincinnati, OH), 1990, pp1698-1703）、［７］(C. A. Mead, Analog VLSI and neural systems, Reading MA, Addison-Wesley, 1989)で発表されている設計手法に従っている。
【０１２１】
図６の左上には相互コンダクタンス増幅器（上記文献［７］参照）と呼ばれる部品６００を示している。部品６００は出力電流ｉ_０を有しており、この電流は、複数ある入力端子のうちの２つ入力端子間の入力電圧ｖ_ｉ、他の端子へのバイアス電流ｉ_ｂ、及び利得Ｋの関数であり、下記式（２８）で求められる。
【０１２２】
【数３３】

【０１２３】
図６には、また、４象限乗算器と呼ばれる部品６０１が示されている。部品６０１は出力電流ｉ_０を有し、この電流は２つの入力電圧ｖ_ｉ１、ｖ_ｉ２並びにバイアス電流ｉ_ｂの関数であり、下記式（２９）で求められる。
【０１２４】
【数３４】

【０１２５】
図６の上部はニューラルネットワーク本体であり、下部は学習機構である。
【０１２６】
ニューラルネットワーク本体は、６１０で示される２ｎ個のニューロンを隠れ層に含み、１つのニューロン６２０を出力層に含んでいる。ニューロン６１０からの出力電圧は、ニューロン６２０の入力の一部をなす。ニューロン６１０は、隠れ層ニューロンを構成する素子６１１、６１２、６１３及び６１４を有する。ニューロン６２０は、隠れ層ニューロンの出力を対応する重み付け値で乗算し、その合計をして電圧出力ｖ_０を求める素子６２１及び６２２を有する。
【０１２７】
６４０として示されている定電圧ｖ_ｘ０は、学習素子６３０に対応する部分によって生成される重み付け値ｗ_０を介して、ニューロン６２０内のいずれかの乗算器（例えば図６の素子６２１）のバイアス電流によってニューロン６２０にバイアスを与える。学習素子６３０は、乗算器６２１で使用するための重み付け値を生成する。
【０１２８】
誤差信号ｅ（ｔ）は、増幅器６５０として示される部品を使用して生成され、増幅器６５０は図示の抵抗配列で、２つの入力ｙ及びｙ^＊に接続されている。
【０１２９】
この動作を詳細に説明する。抵抗値Ｒを有する抵抗器６１３を考慮しつつ、相互コンダクタンス増幅器６１２の定義を利用すると、下記式（３０）が得られる。
【０１３０】
【数３５】

【０１３１】
ここで、フィードバック行列［α_ｉ］は、ニューロン６１０内の各ノードを接続、又はその接続を切ることによって決まる。相互コンダクタンス増幅器６１２及びキャパシタＣからなる追従積分器（follower integrator）６１４を使用すると、Ｋ及びＲの正しい選択について、下記式（３１）が得られる。
【０１３２】
【数３６】

【０１３３】
乗算器６２１は、下記式（３２）に従って、隠れ層ニューロンの出力を学習した重みｗ_ｉで乗算する。
【０１３４】
【数３７】

【０１３５】
ここで、重みは学習素子６３０により求められ、ｉ_ｂ１＝ｗ_０である。
【０１３６】
ＨＴＮＮの全出力は、２ｎ全てからの電流の合計を取り、且つ増幅器６２２を使用して、電流ｉ_０を電圧信号Ｖ_０に変換することによって求められる。
【０１３７】
【数３８】

【０１３８】
各学習素子６３０は、下記式（３４）に従って、対応する重み値を連続的に更新する。
【０１３９】
【数３９】

【０１４０】
こうして、式（４）中のＨＴＮＮは、アナログＶＬＳＩで実現される。
【０１４１】
上述のように、本発明の実施例によれば、位置決めサーボシステムのシステム性能について、周期的外乱が及ぼす影響を補償することが可能な方法及び装置を提供することができる。
【０１４２】
【発明の効果】
本発明に係る制御装置及び制御方法は、システマティックに実現することができ、周期的外乱の必要な成分の全てが容易に補償可能である。
【０１４３】
また本発明に係る制御装置及び制御方法は、環境変化を学習し、それらに適応する学習能力を有している。
【０１４４】
さらに、本発明に係る制御装置及び制御方法は、ＤＳＰのソフトウェアによってのみならず、ハードウェアによっても実行することができ、サーボＤＳＰの制限は、本発明の補償効果に影響しない。
【０１４５】
また、アナログＶＬＳＩで実現できるので、全ＨＴＮＮ及びその学習機構を、１つのチップに集積することができる。
【０１４６】
本明細書においては、本発明の実施の形態を詳細に説明したが、当業者にとって明らかであるように、本発明は上述した実施の形態に限定される物ではなく、本発明の範囲内で多様な変形が可能であることは言うまでもない。
【０１４７】
（参考文献）
下記参考文献の開示内容を全て本願に引用して援用する：
［１］A. Sacks, M. Bodson and W. Messer, "Advanced method for repeatable runout compensation", IEEE Trans. on Magnetics, vol. 31, no. 2, pp. 1031-1036.
［２］Y. Onuki, H. Ishioka and H. Yada, "Repeatable runout compensation for disk drives using multi-loop adaptive feedforward compensation", SICE ’98, Chiba Japan, pp. 1093-1098.
［３］M. Bodson, A. Sacks and P. Khosla, "Harmonic generation in adaptive feedforward cancellation schemes", IEEE Trans. on Automatic Control, vol. 39, no. 9, pp. 1939-1994
［４］Y. Lou, "Observer-based compensation of repeatable run-outs in hard disk drives", Proceedings of the 9^ｔｈ Sony Research Forum (1999) p453-458.
［５］Y. Onuki and H. Ishioka, "Compensation for repeatable tracking errors in hard drives using discrete-time repetitive controllers", Proceedings of the 6^ｔｈ International Workshop on advanced motion control, March 2000, Nagoya, Japan, pp 490-495.
［６］S. P. DeWeerth et al., "A neuron-based pulse servo for motion control," Proc. IEEE Int Conf. Robotics and Automation (Cincinnati, OH), 1990, pp 1698-1703.
［７］C. A. Mead, Analog VLSI and neural systems, Reading MA, Addison-Wesley, 1989.
【図面の簡単な説明】
【図１】ハードディスクドライブのサーボシステムにおける周期的外乱の影響を示すグラフである。
【図２】（ａ）はハードディスクドライブのサーボシステムを制御する本発明の実施の形態の構造を示す概略図であり、（ｂ）は同様の図において周期的外乱を力信号として表したものである。
【図３】図２のニューラルネットワークの構造を示す概略図である。
【図４】図３のニューラルネットワーク、及び学習機構の概略を示し、ニューラルネットワークは、連続システムで表されている。
【図５】ＤＳＰソフトウェアで実行される図３のニューラルネットワーク、及びその学習機構の概略ブロック図である。
【図６】アナログＶＬＳＩによる図３のニューラルネットワークの実施例を示す図である。
【符号の説明】
２００ 公称サーボ制御器、２１０ システム、２１１ 積分器、２１２ 積分器、２２０ ユニット、２２１ 学習機構、２２２ ＨＴＮＮ、３００ 素子、３０１ 積分器、３０３ ニューロン、３０４ 乗算器、３０６ 結果、３０７加算素子、３０８ 出力、４００ ＨＴＮＮ、４０１ 学習機構、５００ ニューラルネットワーク本体、５０１ 学習機構、６００ 相互コンダクタンス増幅器、６０１ ４象限乗算器、６１０ ニューロン、６２０ ニューロン、６３０ 学習部品、６４０ 定電圧、６５０ 増幅器[0001]
BACKGROUND OF THE INVENTION
  The present invention is, for example,Such as disk drive read / write head controlOperation of controlled system with positioning drive mechanismThe present invention relates to a control method for controlling. The present invention further includes a control device for executing the above-described method, and the control device.PositioningRelated to servo system.
[0002]
[Prior art]
Periodic disturbances are found in many servo systems and degrade system performance. A prime example is a read / write head servo system in a disk drive for data recording, where the main function of this servo is to accurately place the head on a selected track of a rotating disk. Is to maintain. Positioning inaccuracy due to the effects of periodic disturbances is called Repeatable Run-Out (RRO), and is one of the major obstacles in improving system positioning accuracy. ing.
[0003]
FIG. 1 shows a 3.5 "hard disk drive position error after a so-called track following process where the read / write head moves to the desired track and tries to maintain the position on the track. Indicates a signal (position error signal).
[0004]
When the spindle disk is decentered, the influence causes a periodic disturbance at the head position. As a result, the conventional servo reliably suppresses the residual error to within ± 25% of the track width. I can't. However, the required position accuracy is a residual of less than ± 10% of the track width.
[0005]
In a disk drive in which a disk is removable, the RRO is larger because the disk is more eccentric. This kind of problem with periodic disturbances also occurs in machine tools and rotating machines including robot manipulators that perform repetitive tasks. In addition, some non-linearities in actuators can be modeled as periodic disturbances.
[0006]
A common feature of these disturbances is that they are periodic (ie, repetitive). This period is well known or can be analyzed or measured.
[0007]
[Problems to be solved by the invention]
It can be seen that the residual e (t) shown in FIG. 1 is periodic and the fundamental frequency of the residual is 75 Hz, which perfectly matches the rotational speed of the disk.
[0008]
Conventionally, as a method of canceling RRO, there is a method using a notch filter as a part of a feedback loop. Since the notch filter is inserted in series in the feedback loop, it affects all signals having a notch frequency in the loop. Therefore, when the notch filter is inserted in the closed loop, the system characteristics are changed, and the configuration of the servo controller becomes more complicated. Furthermore, the main problem when using a notch filter is that the method is model-based and the notch filter is designed with fixed parameters selected based on information about the disturbance. Since the disturbance can vary depending on the operating state, compensation by the notch filter is not always effective.
[0009]
Another solution for RRO is Adaptive Feedforward Cancellation (AFC) (reference [1] A. Sacks, M. Bodson And W. Messer, "Advanced method for repeatable runout compensation". ", IEEE Trans. On Magnetics, vol. 31, No. 2, pp. 1031-1036", [2] Y. Onuki, H. Ishioka and H. Yada, "Repeatable runout compensation for disk drives using multi-loop adaptive feedforward compensation ", SICE '98, Chiba Japan, pp. 1093-1098, [3] M. Bodson, A. Sacks and P. Khosla," Harmonic Generation in adaptive feedforward cancellation schemes ", IEEE Trans. On Automatic Control, vol 39, No. 9, pp. 1939-1994) Unlike the notch filter method, AFC operates in the feedforward path and does not change the system characteristics.
[0010]
However, it has been found that the AFC algorithm designed to remove a predetermined number of frequency components may generate higher-order harmonics (see the above document [3]) and may deteriorate the compensation effect. ing.
[0011]
Furthermore, the so-called “corrected AFC” requires a phase shift, but if the phase shift cannot be accurately specified, it causes a residual error (see the above-mentioned documents [1] and [2]).
[0012]
In order to realize AFC, sine and cosine functions are required, and these are usually look-up tables of a digital signal processor (DSP). ) Is executed. Lookup tables take up a lot of memory space and make programming complicated.
[0013]
Reference [4] (Y. Lou, "Observer-based compensation of repeatable run-outs in hard disk drives", Proceedings of the 9^th In Sony Research Forum (1999) p 453-458), the present inventor has proposed an observer-based feedforward compensation (hereinafter referred to as OFC) method. The RRO is treated as an additional system state that is observed by the state observer. The observed signal is then used to create a feed forward compensation loop to remove the effect of RRO. The problem with OFC is that it requires two extra state variables for each harmonic component of the RRO, which makes it difficult to fully compensate for the harmonic components.
[0014]
Another mechanism for removing periodic disturbances is a repetitive controller, which is described in Reference [5] (Y. Onuki and H. Ishioka, “Compensation for repeatable tracking errors in hard drives using discrete-time repetitive controllers”. ", Proceedings of the 6^th International Workshop on advanced motion control, March 2000, Nagoya, Japan, pp 490-495). As described in the above document, the iterative controller can compensate for higher harmonics than AFC, and the iterative controller's computational load is only a quarter of AFC. However, for each harmonic, the compensation effect by the iterative controller is smaller than that by AFC. The repetitive controller used in reference [5] above has an attenuation factor of 45 dB for the fundamental harmonic and lower attenuation for the other higher harmonics, while AFC is each compensated harmonic. Has an attenuation factor greater than 60 dB.
[0015]
In addition, all the methods described above are performed by software on a DSP or microprocessor. Thus, DSP or microprocessor limitations (ie limited computational speed and / or word length) may limit the ability to compensate for disturbances.
[0016]
The present invention provides a new and useful method and apparatus for controlling a servo system. The present invention also provides a method and apparatus that can effectively, reliably, simply and economically compensate for periodic disturbances in a servo system.
[0017]
[Means for Solving the Problems]
  The present invention proposes to control a servo system under the influence of periodic noise using a neural network comprising two layers (hidden layer and output layer).
That is, the present invention provides a control method for controlling the operation of a controlled system having a positioning drive mechanism, wherein a measurement signal indicating the operating state of the controlled system output from the controlled system is used as a reference signal by a comparison unit. In comparison, the system includes a plurality of hidden layer neuron units that generate a position error signal indicating a position error in the operation of the controlled system and output a predetermined function with respect to the input. A hidden layer that is arranged in pairs and configured so that the output of one neuron unit of each pair is input to the other neuron unit, and the sum of weighted outputs of the plurality of hidden layer neuron units, respectively. An output layer including an output layer neuron unit for outputting, in the output layer neuron unit By adjusting the weighting value when adding the outputs of multiple hidden layer neuron units, it is subject to the influence of periodic disturbances that are input to a neural network that outputs an arbitrary approximate periodic signal to the input. Based on the position error signal generated by the comparison means from the measurement signal indicating the operating state of the controlled system in the above, the neural network learning mechanism learns the influence of the periodic disturbance, and the output layer neuron unit A periodic signal that compensates for the influence of periodic disturbance in the controlled system is output from the output layer neuron unit of the neural network by adjusting a weighting value when adding the outputs of the plurality of hidden layer neuron units. The neural network to control the ratio A position error signal generated by the means is supplied to a servo controller to generate a correction signal for correcting a position error in the operation of the controlled system indicated by the position error signal, and generated by the servo controller. The correction signal is added and combined with the periodic signal output from the output layer neuron unit by the adding means, and the control signal obtained by adding and combining the correction signal and the periodic signal is supplied to the controlled system by the adding means. The positioning drive mechanism of the controlled system is driven in a state in which the influence of the periodic disturbance is compensated.
Further, the present invention provides a control device for controlling the operation of a controlled system having a positioning drive mechanism by comparing a measurement signal output from the controlled system and indicating the operating state of the controlled system with a reference signal. Comparing means for generating a position error signal indicating a position error in the operation of the controlled system; and driving the positioning drive mechanism of the controlled system and indicating the controlled state indicated by the position error signal generated by the comparing means. A servo controller for generating a correction signal for correcting a position error in the operation of the system, and a plurality of hidden layer neuron units each receiving a position error signal generated by the comparison means and outputting a predetermined function for the input. The plurality of hidden layer neuron units are arranged in pairs, and one neuron unit of each pair Has a hidden layer configured to be input to the other neuron unit, and an output layer including an output layer neuron unit that outputs a sum of weighted outputs of the plurality of hidden layer neuron units. Adjusting a weighting value when adding the outputs of the plurality of hidden layer neuron units in the output layer neuron unit, and outputting a periodic signal arbitrarily approximated to the input, and a period Based on the position error signal generated by the comparison means from the measurement signal indicating the operating state of the controlled system under the influence of the external disturbance and input to the neural network, the influence of the periodic disturbance is learned, The outputs of the plurality of hidden layer neuron units in the output layer neuron unit A neural network that controls the neural network so as to output a periodic signal from the output layer neuron unit of the neural network to compensate for the influence of periodic disturbances in the controlled system by adjusting a weighting value when calculating A network learning mechanism; and an adding means for adding and synthesizing the correction signal generated by the servo controller and the periodic signal output from the output layer neuron unit, and the correction means and the periodic signal are added by the adding means. A control signal obtained by addition and synthesis is supplied to the controlled system, and the positioning drive mechanism of the controlled system is driven in a state where the influence of the periodic disturbance is compensated.
Furthermore, the present invention is a positioning servo system comprising a controlled system having a positioning drive mechanism and a control device that controls the operation of the controlled system, wherein the control device is output from the controlled system. Comparing a measurement signal indicating an operation state of the controlled system with a reference signal to generate a position error signal indicating a position error in the operation of the controlled system, and a positioning drive mechanism of the controlled system. A servo controller for driving and generating a correction signal for correcting a position error in the operation of the controlled system indicated by the position error signal generated by the comparison means, and a position error signal generated by the comparison means are input. A plurality of hidden layer neuron units each outputting a predetermined function with respect to the input. Layer neuron units are arranged in pairs, and the output of one neuron unit of each pair is input to the other neuron unit, and the outputs of the plurality of hidden layer neuron units are weighted respectively. An output layer including an output layer neuron unit that outputs the added sum, and by adjusting a weight value when adding outputs of the plurality of hidden layer neuron units in the output layer neuron unit, On the other hand, it is generated by the comparison means from a neural network that outputs an arbitrarily approximated periodic signal and a measurement signal indicating the operating state of the controlled system under the influence of periodic disturbance, and is input to the neural network. Based on the position error signal, the output layer neuron unit A periodic signal that compensates for the influence of periodic disturbances in the controlled system is output from the output layer neuron unit of the neural network by adjusting the weighting value when adding the outputs of several hidden layer neuron units The neural network learning mechanism for learning the influence of the periodic disturbance and controlling the neural network, the correction signal generated by the servo controller, and the periodic signal output from the output layer neuron unit Adding means for adding and synthesizing, supplying a control signal obtained by adding and synthesizing the correction signal and the periodic signal by the adding means to the controlled system, and compensating the influence of the periodic disturbance in the state The positioning drive mechanism of the controlled system is driven.
[0019]
The present invention is based on the characteristic property of periodic disturbances that can be expanded in a trigonometric series. In effect, each pair of hidden layer neurons can compensate for one of the components of the trigonometric series. Therefore, if there are a sufficient number of hidden layer units, periodic disturbances of any dimension can be accurately compensated.
[0020]
The neural network can be automatically tuned. That is, it has a learning ability to learn and adapt to environmental changes such as changes in periodic noise.
[0021]
As will be described later, since the control method and control apparatus of the present invention can be executed not only by DSP software but also by hardware, the obtained compensation is not limited to the limitation of the servo DSP.
[0022]
Specifically, the control method and control apparatus of the present invention can be realized by an analog circuit, preferably an integrated circuit such as a VLSI. In particular, the neural network and its learning system can be integrated in the chip, for example an additional part integrated in an existing driver chip.
[0023]
An advantage of the present invention over AFC is that the AFC sine / cosine look-up table is replaced with hidden layer self-feedback. This makes the present invention simpler and more reliable.
[0024]
Furthermore, the learning mechanism of the present invention allows the neural network to cope with changing operating conditions. This makes the preferred embodiment of the present invention more resistant.
[0025]
Furthermore, since the physical meaning of the structure is clear, the present invention can compensate for harmonics of any dimension of periodic disturbances by simply adding more neurons. This makes the preferred embodiment of the present invention more flexible.
[0026]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, a control method and a control apparatus according to the present invention will be described in detail with reference to the drawings. Note that the present invention is not limited to the following embodiments.
[0027]
First, FIG. 2A shows an embodiment of the present invention. This embodiment is, for example, a hard disk drive servo system.
[0028]
  This embodimentServo systemHas a system 210 that is controlled in response to the read / write head. This system 210 is affected by the disturbance r. Note that all signals in this specification are a function of time. The measurement signal ym is output from the system 210 and indicates the state of the system 210. In order to determine the position error signal e which is the input of the nominal servo controller 200, the measurement signal y_mIsBy the comparator 100Reference signal y^*Compared with The nominal servo controller 200 outputs an instruction signal u that drives the actuator. Nominal servo controller 200 is a servo controller designed for a nominal system, i.e., system 210 without disturbance r.
[0029]
  This servo systemFurther comprises a unit 220 having a neural network 222. In this specification, a neural network (which will be described in detail later) is referred to as a hybrid trigonometric neural network (hereinafter referred to as HTNN), each having (2n + 1) outputs each represented by a vector X. It has a hidden layer consisting of neurons and an output layer that generates a single output fN. The HTNN 222 is controlled based on a group of (2n + 1) weight values described as a vector W. The vector W is set by a learning mechanism (LM) 221. The learning mechanism 221 is supplied with the output X and the instantaneous value e of the hidden layer neuron.
[0030]
  The instruction signal u which is the output of the nominal servo controller 200 isAdder 110Is added to the output fN of HTNN 222 to form a control signal that is input to the controlled system 210. Here, the controlled system 210 is modeled as two integrators 211 and 212.
[0031]
When the system 210 is a hard disk drive (HDD), the system 210 includes a movable actuator, an arm with a suspension, a drive mechanism that enables movement for positioning the actuator, and a voice coil motor. None of these components are shown, but since all have known structures, it will be apparent to those skilled in the art.
[0032]
A periodic disturbance, referred to herein as repetitive rotational runout (RRO), is denoted r. In the case of an HDD, it appears as a position disturbance caused by disk eccentricity, rotation axis deformation, or incomplete servo writing process.
[0033]
Since this RRO is periodic, it can be approximated by a triangular polynomial as shown in the following formula (1).
[0034]
[Expression 1]

[0035]
This RRO appears as a position disturbance, which is equivalent to a force disturbance f (t) as shown in FIG. 2B by Laplace transform F (s).
[0036]
[Expression 2]

[0037]
Here, R (s) is a Laplace transform of RRO r (t). Therefore,
[0038]
[Equation 3]

[0039]
Where F_i= -I²ω²R_i/ K_vk_yIt is. The first term with an index of zero in equation (1) is a constant and therefore disappears after double differentiation. Formula (3) can be further described as the following formula (4).
[0040]
[Expression 4]

[0041]
here,
[0042]
[Equation 5]

[0043]
And these are defined below.
[0044]
X (t) = [sinωt sin2ωt ... sinnωt cosωt cos2ωt ... cosnωt]^TW = [F₁cosj1 F₂cosj2 ・ ・ ・ F_ncosjn F₁sinj1 F₂sinj2 ・ ・ ・ F_nsinjn]^T
The following matrix equation of motion is obtained by the characteristics of the trigonometric function.
[0045]
[Formula 6]

[0046]
here,
[0047]
[Expression 7]

[0048]
and,
[0049]
[Equation 8]

[0050]
It is.
[0051]
Equations (4) and (5) show that the periodic disturbance can be represented by a triangular polynomial, and this triangular polynomial is represented by the neural network of FIG. The neural network of FIG. 3 has two layers. That is, there are two layers: a hidden layer having (2n + 1) neurons and an output layer consisting of one neuron. Except for neuron 303, X₁,. . . , X_n, X_{n + 1},. . . , X_2nThe output of the hidden layer neuron shown by is equal to each component of X in equation (4).
[0052]
In FIG. 3, each hidden layer neuron other than the neuron 303 receives the input value a.₁,. . . , A_n, A_{n + 1},. . . , A_2nWhich performs the simple function of multiplying by (this multiplication is shown as element 300) and inputs the result to integrator 301 to generate the output of the hidden layer neuron. The neuron 303 is a special neuron that outputs 1 as a constant output and acts as a bias of the output layer. The neuron 303 can be treated equally with other hidden layer neurons, but its value a₀If = 0,
[0053]
[Equation 9]

[0054]
It becomes.
[0055]
Each output X from the hidden layer₀,. . . , X_n, X_{n + 1},. . . , X_2nAre input to an output layer consisting of one neuron. This one neuron has each input output X₀,. . . , X_n, X_{n + 1},. . . , X_2nFor each of the corresponding weight values W₀,. . . , W_n, W_{n + 1},. . . , W_2nRespectively (with a multiplier indicated by 304) to obtain a result 306 respectively. The results 306 are summed using an adder 307 and f_NAn output 308 is obtained that is (t, W (t)).
[0056]
Adjustable weight W₀,. . . , W_n, W_{n + 1},. . . , W_2nThe value of can be changed by a learning mechanism, which will be described later, and is shown in FIG.
[0057]
The structure of this neural network is very similar to the so-called “Radial Neural Network”. The reason is that this neural network has only one adjustable weight in each feedforward path and there is no exit function before output. That is, the output of the output neuron is only the sum of the weighted inputs, not its function. The difference between this neural network and the radial neural network is that the neural network of FIG. 3 obtains the input to the output layer in the form of a trigonometric function by its own operation. For this reason, the neural network of FIG. 3 is referred to as a hybrid trigonometric function neural network (HTNN).
[0058]
For HTNN, add the following two notes.
[0059]
1. HTNN can approximate any periodic signal. This can be expressed mathematically as follows: For any continuous periodic function f (t + T) = f (t) with period T, and for any δ> 0, a finite value n and a weighting vector W^*And is represented by the following formula (7).
[0060]
[Expression 10]

[0061]
2. Among all trigonometric polynomials with period T and dimensions less than n, HTNN gives f the best mean square approximation possible.
[0062]
Hereinafter, the learning mechanism and system stability of the HTNN of FIG. 3 will be described.
[0063]
From FIG. 2 (b), y = y_mThen, the driving force of the HDD servo that does not have HTNN can be described as the following equation (8).
[0064]
## EQU11 ##

[0065]
When the nominal servo controller 200 is a linear proportional derivative controller, the following equation (9) is obtained.
[0066]
[Expression 12]

[0067]
Where e = y^*-Y is a position error.
[0068]
The gain of this controller is as shown in the following formula (10).
[0069]
[Formula 13]

[0070]
Where k_c> 0, β₀> 0 and β₁> 0 is a coefficient of a composite error defined by the following equation (11).
[0071]
[Expression 14]

[0072]
Therefore, the following formula (12) is obtained by differentiating the formula (11) in consideration of the above formulas (8) to (10).
[0073]
[Expression 15]

[0074]
Where f_R(T) = − k_vk_yf (t).
[0075]
Output f of HTNN 222_NConsidering (t, W (t)), the following formula (13) is obtained.
[0076]
[Expression 16]

[0077]
In the case of stability design, a Lyapunov function is constructed as shown in the following equation (14).
[0078]
[Expression 17]

[0079]
Where E_W(T) = W^*-W (t) is a weight error.
[0080]
When the HTNN 222 operates in a continuous time (for example, by an analog circuit), the learning rule of the HTNN 222 is selected so that the following equation (15) is satisfied.
[0081]
[Formula 18]

[0082]
here,
[0083]
[Equation 19]

[0084]
It is.
[0085]
Therefore, the operation of HTNN 222 and the learning algorithm are as shown schematically in FIG. Here, box 400 represents HTNN 222, and box 401 represents the operation of learning mechanism 221. The HTNN 222 has a continuous loop that multiplies X by A, and X is fed back to the learning mechanism 401. Therefore, the time derivative of V (t) is as shown in the following formula (16).
[0086]
[Expression 20]

[0087]
Therefore,
[0088]
[Expression 21]

[0089]
When the above formula (17a) is satisfied, the relationship shown in the formula (17b) is established.
[0090]
Therefore, the system synthesis error is β₁δ / k_cThis means that the error e (t) converges according to the following linear differential equation (18).
[0091]
[Expression 22]

[0092]
The solution of equation (18) is as shown in equation (19) below.
[0093]
[Expression 23]

[0094]
Equation (19) shows that the error e (t) of the system is calculated from the initial value e (0) by β₁δ / (β₀k_c) Bias, positive constant β₁δ / (β₀k_c), Β₀/ Β₁The exponential constant indicates that it converges exponentially.
[0095]
The following describes how to select HTNN and learning mechanism parameters.
[0096]
The initial state of the HTNN hidden layer neuron, ie, X (0), should be assigned to ensure that the state X (t) is trigonometric, ie: .
[0097]
[Expression 24]

[0098]
The initial value of the weight W (0) in the HTNN may be set to zero so that the HTNN has a zero initial output.
[0099]
[Expression 25]

[0100]
The dimension n of HTNN affects the approximation accuracy δ, and further affects the overall system compensation performance. In general, when n is large, δ is small as a result. However, the dimension n of HTNN is actually limited. For example, when HTNN is realized by an analog VLSI, it is limited by the degree of integration, and when HTNN is realized by software, it may be limited by the calculation capability of the DSP. A realistic way to select the dimension n of HTNN is by harmonic analysis on the error signal. From the contents of the harmonic components of the error signal, an appropriate n can be selected to ensure that HTNN cancels all the main harmonic components. Another method of setting n is to increase n from the initial value until sufficient compensation performance is achieved, taking into account hardware or software limitations. This method is performed before the HTNN sets the operation or by online adaptation.
[0101]
Next, the parameter β₀, Β₁And k_cThe setting method of will be described.
[0102]
From the above formula (10), the following formula (22) is obtained.
[0103]
[Equation 26]

[0104]
Assuming that the linear controller is designed so that the linear system operates in a critical state, i.e.
[0105]
[Expression 27]

[0106]
Assuming that, the following equation (24) is obtained.
[0107]
[Expression 28]

[0108]
β₁= K₁To set β₀= 2k₀It becomes. Therefore, the following equation (25) is obtained from the equations (9) and (11).
[0109]
[Expression 29]

[0110]
Using equation (25) in the learning mechanism avoids the measurement of e (t) derivative, which is usually a problem in practice.
[0111]
Here, an embodiment using either hardware or software will be described.
[0112]
First, an embodiment using software using a DSP and an embodiment using hardware in an analog VLSI will be described in detail below. In this case, the above continuous equation is no longer appropriate, and instead, equation (5) is transformed into equation (26) below.
[0113]
[30]

[0114]
Where T is the sampling time,
[0115]
[31]

[0116]
It is. X_dThe initial value of (k) is
X_d(0) = [0_1xn  I_1xn]^T
It is.
[0117]
Expression (15) representing the learning rule is replaced with the following expression (27).
[0118]
[Expression 32]

[0119]
FIG. 5 shows a schematic block diagram of HTNN in which a learning mechanism is realized by DSP software. Here, 500 is a neural network body, and 501 is a learning mechanism. In this case, the overall structure is similar to that of FIG. 4, but in block 500 matrix A is replaced by matrix D and the integrator is z^-1Has been replaced with the delay unit shown in. Similarly, in block 501 corresponding to the learning mechanism, the integrator is a delay unit z.^-1And the output of block 501 is the sum of this delay unit, ie, W in the previous cycle with the result of equation (27).
[0120]
Next, an embodiment using analog VLSI will be described. In this case, the above formulas (5) and (15) are effective and are not replaced by the above formulas (26) and (27). An example is shown in FIG. 6, which is described in reference [6] (SP DeWeerth et al., “A neuron-based pulse servo for motion control”, Proc. IEEE Conf. Robotics and Automation (Cincinnati, OH , 1990, pp 1698-1703), [7] (CA Mead, Analog VLSI and neural systems, Reading MA, Addison-Wesley, 1989).
[0121]
6 shows a component 600 called a transconductance amplifier (see the above document [7]). Component 600 is output current i₀This current is an input voltage v between two input terminals of the plurality of input terminals._i, Bias current i to other terminals_b, And a function of the gain K, which are obtained by the following equation (28).
[0122]
[Expression 33]

[0123]
FIG. 6 also shows a component 601 called a four quadrant multiplier. The component 601 has an output current i₀This current has two input voltages v_i1, V_i2And bias current i_bWhich is determined by the following equation (29).
[0124]
[Expression 34]

[0125]
The upper part of FIG. 6 is a neural network body, and the lower part is a learning mechanism.
[0126]
The neural network body includes 2n neurons indicated by 610 in the hidden layer and one neuron 620 in the output layer. The output voltage from neuron 610 forms part of the input of neuron 620. The neuron 610 includes elements 611, 612, 613, and 614 that constitute hidden layer neurons. The neuron 620 multiplies the output of the hidden layer neuron by the corresponding weight value, and sums the result to obtain the voltage output v.₀Elements 621 and 622 for obtaining
[0127]
Constant voltage v shown as 640_x0Is the weighting value w generated by the part corresponding to the learning element 630₀, The neuron 620 is biased by the bias current of any multiplier in neuron 620 (eg, element 621 in FIG. 6). Learning element 630 generates a weighting value for use in multiplier 621.
[0128]
The error signal e (t) is generated using the components shown as amplifier 650, which is the resistor array shown and has two inputs y and y.^*It is connected to the.
[0129]
This operation will be described in detail. When the definition of the transconductance amplifier 612 is used while considering the resistor 613 having the resistance value R, the following equation (30) is obtained.
[0130]
[Expression 35]

[0131]
Where the feedback matrix [α_i] Is determined by connecting or disconnecting each node in the neuron 610. Using a follower integrator 614 consisting of a transconductance amplifier 612 and a capacitor C, the following equation (31) is obtained for the correct selection of K and R:
[0132]
[Expression 36]

[0133]
The multiplier 621 uses the weight w obtained by learning the output of the hidden layer neuron according to the following equation (32)._iMultiply by
[0134]
[Expression 37]

[0135]
Here, the weight is determined by the learning element 630 and i_b1= W₀It is.
[0136]
The total output of HTNN is the sum of the currents from all 2n and using amplifier 622, the current i₀Voltage signal V₀Is obtained by converting to
[0137]
[Formula 38]

[0138]
Each learning element 630 continuously updates the corresponding weight value according to the following equation (34).
[0139]
[39]

[0140]
Thus, HTNN in equation (4) is realized by analog VLSI.
[0141]
As described above, according to the embodiments of the present invention, it is possible to provide a method and an apparatus capable of compensating the influence of periodic disturbance on the system performance of the positioning servo system.
[0142]
【The invention's effect】
The control device and the control method according to the present invention can be systematically realized, and all components that require periodic disturbance can be easily compensated.
[0143]
In addition, the control device and the control method according to the present invention have a learning ability to learn and adapt to environmental changes.
[0144]
Furthermore, the control device and the control method according to the present invention can be executed not only by the DSP software but also by hardware, and the limitation of the servo DSP does not affect the compensation effect of the present invention.
[0145]
Further, since it can be realized by analog VLSI, all HTNNs and their learning mechanisms can be integrated on one chip.
[0146]
In the present specification, embodiments of the present invention have been described in detail. However, as will be apparent to those skilled in the art, the present invention is not limited to the above-described embodiments, and is within the scope of the present invention. It goes without saying that various modifications are possible.
[0147]
(References)
The entire disclosure of the following references is incorporated herein by reference:
[1] A. Sacks, M. Bodson and W. Messer, "Advanced method for repeatable runout compensation", IEEE Trans. On Magnetics, vol. 31, no. 2, pp. 1031-1036.
[2] Y. Onuki, H. Ishioka and H. Yada, "Repeatable runout compensation for disk drives using multi-loop adaptive feedforward compensation", SICE '98, Chiba Japan, pp. 1093-1098.
[3] M. Bodson, A. Sacks and P. Khosla, "Harmonic generation in adaptive feedforward cancellation schemes", IEEE Trans. On Automatic Control, vol. 39, no. 9, pp. 1939-1994
[4] Y. Lou, "Observer-based compensation of repeatable run-outs in hard disk drives", Proceedings of the 9^th Sony Research Forum (1999) p453-458.
[5] Y. Onuki and H. Ishioka, "Compensation for repeatable tracking errors in hard drives using discrete-time repetitive controllers", Proceedings of the 6^th International Workshop on advanced motion control, March 2000, Nagoya, Japan, pp 490-495.
[6] S. P. DeWeerth et al., "A neuron-based pulse servo for motion control," Proc. IEEE Int Conf. Robotics and Automation (Cincinnati, OH), 1990, pp 1698-1703.
[7] C. A. Mead, Analog VLSI and neural systems, Reading MA, Addison-Wesley, 1989.
[Brief description of the drawings]
FIG. 1 is a graph showing the influence of periodic disturbances in a servo system of a hard disk drive.
FIG. 2A is a schematic diagram showing the structure of an embodiment of the present invention for controlling a servo system of a hard disk drive, and FIG. 2B shows periodic disturbances as force signals in the same diagram. is there.
FIG. 3 is a schematic diagram showing the structure of the neural network of FIG. 2;
4 shows an outline of the neural network of FIG. 3 and the learning mechanism, where the neural network is represented by a continuous system.
FIG. 5 is a schematic block diagram of the neural network of FIG. 3 and its learning mechanism executed by DSP software.
6 is a diagram showing an example of the neural network of FIG. 3 using analog VLSI.
[Explanation of symbols]
200 nominal servo controller, 210 system, 211 integrator, 212 integrator, 220 units, 221 learning mechanism, 222 HTNN, 300 elements, 301 integrator, 303 neurons, 304 multiplier, 306 result, 307 adder element, 308 outputs , 400 HTNN, 401 learning mechanism, 500 neural network body, 501 learning mechanism, 600 transconductance amplifier, 601 4-quadrant multiplier, 610 neuron, 620 neuron, 630 learning component, 640 constant voltage, 650 amplifier

Claims (10)

In a control method for controlling the operation of a controlled system having a positioning drive mechanism ,
A measurement signal indicating the operating state of the controlled system output from the controlled system is compared with a reference signal by comparison means to generate a position error signal indicating a position error in the operation of the controlled system ,
It includes a plurality of hidden layer neuron units that each output a predetermined function for input, the plurality of hidden layer neuron units are arranged in pairs, and the output of one neuron unit of each pair is input to the other neuron unit has a hidden layer that is so that arrangement, and an output layer comprising an output neuron unit which outputs a sum obtained by adding to each weighted outputs of several hidden layer neuron units plurality, said in the output layer neuron unit By adjusting the weighting value when adding the outputs of multiple hidden layer neuron units, it is under the influence of periodic disturbances that are input to a neural network that outputs an arbitrary approximate periodic signal to the input. Generated by the comparison means from a measurement signal indicating the operating state of the controlled system Based on the position error signal, the neural network learning mechanism learns the influence of the periodic disturbance and adjusts the weight value when adding the outputs of the plurality of hidden layer neuron units in the output layer neuron unit. Thereby controlling the neural network to output a periodic signal from the output layer neuron unit of the neural network that compensates for the effects of periodic disturbances in the controlled system ,
Supplying a position error signal generated by the comparison means to a servo controller to generate a correction signal for correcting a position error in the operation of the controlled system indicated by the position error signal ;
Adding synthesized by a periodic signal and adding means output a correction signal generated by the servo controller from the output layer neuron unit,
A control signal obtained by adding and synthesizing the correction signal and the periodic signal by the adding means is supplied to the controlled system, and the positioning drive mechanism of the controlled system is driven in a state in which the influence of the periodic disturbance is compensated. A control method characterized by: In a control device for controlling the operation of a controlled system having a positioning drive mechanism ,
Comparison means for generating a position error signal indicating a position error in the operation of the controlled system by comparing a measurement signal output from the controlled system and indicating the operation state of the controlled system with a reference signal ;
A servo controller that drives a positioning drive mechanism of the controlled system and generates a correction signal that corrects a position error in the operation of the controlled system indicated by the position error signal generated by the comparison means;
The position error signal generated by the comparison means is input, and includes a plurality of hidden layer neuron units each outputting a predetermined function for the input, the plurality of hidden layer neuron units being arranged in pairs, A hidden layer configured such that the output of one neuron unit is input to the other neuron unit, and an output layer neuron unit that outputs a sum of weighted outputs of the plurality of hidden layer neuron units. An output layer, and by adjusting the weighting value when adding the outputs of the plurality of hidden layer neuron units in the output layer neuron unit, a periodic signal arbitrarily approximated to the input is output A neural network that
The influence of the periodic disturbance is learned based on the position error signal generated from the measurement signal indicating the operating state of the controlled system under the influence of the periodic disturbance and input to the neural network. Adjusting the weighting value when adding the outputs of the plurality of hidden layer neuron units in the output layer neuron unit to thereby generate a periodic signal for compensating for the influence of the periodic disturbance in the controlled system. A neural network learning mechanism for controlling the neural network to output from the output layer neuron unit of
Adding means for adding and synthesizing a correction signal generated by the servo controller and a periodic signal output from the output layer neuron unit;
A control signal obtained by adding and synthesizing the correction signal and the periodic signal by the adding means is supplied to the controlled system, and the positioning drive mechanism of the controlled system is driven in a state in which the influence of the periodic disturbance is compensated. A control device. Let n be the number of hidden layer neuron unit pairs, and i = 1,. . . , N, in each i pair, one of the hidden layer neuron units outputs a function iωv (h) of the corresponding input h, and the other of the pair outputs −iωv (h). 3. The control apparatus according to claim 2 , wherein v is a predetermined function and ω is a constant. The control device according to claim 3 , wherein the v (h) is a time integral of h. The control device according to claim 4 , wherein the hidden layer is configured by analog VLSI hardware. The control device according to claim 3 , wherein the v (h) is a discrete-time integration of h. The control device according to claim 6 , wherein the hidden layer neuron unit is configured by digital software. The neural network learning mechanism, each weight update, the weighting value, characterized in that it is configured to vary in an amount that is proportional constant multiplied by the instantaneous value of the output of the corresponding hidden layer claim 2 The control device described in 1. Control device according to claim 8, wherein the proportionality constant is characterized in that it is a function of the instantaneous value of the output and the error signal of the servo controller.   A positioning servo system comprising a controlled system having a positioning drive mechanism and a control device for controlling the operation of the controlled system,
  The control device includes:
  Comparison means for generating a position error signal indicating a position error in the operation of the controlled system by comparing a measurement signal output from the controlled system and indicating the operation state of the controlled system with a reference signal;
  A servo controller that drives a positioning drive mechanism of the controlled system and generates a correction signal that corrects a position error in the operation of the controlled system indicated by the position error signal generated by the comparison means;
  The position error signal generated by the comparison means is input, and includes a plurality of hidden layer neuron units each outputting a predetermined function for the input, the plurality of hidden layer neuron units being arranged in pairs, A hidden layer configured such that the output of one neuron unit is input to the other neuron unit, and an output layer neuron unit that outputs a sum of weighted outputs of the plurality of hidden layer neuron units. An output layer, and by adjusting the weighting value when adding the outputs of the plurality of hidden layer neuron units in the output layer neuron unit, a periodic signal arbitrarily approximated to the input is output A neural network that
  Based on the position error signal generated by the comparison means from the measurement signal indicating the operating state of the controlled system under the influence of periodic disturbance and input to the neural network, the plurality of output layer neuron units By adjusting the weighting value when adding the output of the hidden layer neuron unit, to output a periodic signal from the output layer neuron unit of the neural network to compensate for the influence of the periodic disturbance in the controlled system, A neural network learning mechanism for learning the influence of the periodic disturbance and controlling the neural network;
  Adding means for adding and synthesizing a correction signal generated by the servo controller and a periodic signal output from the output layer neuron unit;
  A control signal obtained by adding and synthesizing the correction signal and the periodic signal by the adding means is supplied to the controlled system, and the positioning drive mechanism of the controlled system is driven in a state in which the influence of the periodic disturbance is compensated. Do
  A positioning servo system characterized by that.

Images