JP2005000389A - Myoelectric pattern discriminating method and device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reduce quantization errors by using a quantization method in which quantization characteristics are modified suitably depending on the polarization in the distribution. <P>SOLUTION: A device is equipped with a controller, which discriminates differences of myoelectric patterns by quantizing values measured by a plurality of myoelectric sensors into discrete values, and decides motions of a subject to be controlled based on results of the discrimination. The quantization for discriminating the myoelectric patterns is implemented by using the quantization method in which the quantization characteristics are modified suitably depending on the distribution characteristics of the patterns. Typical μ-law quantization and median cut quantization are used for the method in which the quantization characteristics are modified suitably depending on the distribution characteristics of the patterns. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

再構成可能ハードウェアとしては、Ｐｒｏｇｒａｍｍａｂｌｅ Ｌｏｇｉｃ Ａｒｒａｙ （以下ではＰＬＡと略す）が用いられている。ＰＬＡは、図２に示す例のように、ＡＮＤアレイ（接続している入力信号の論理積を作る）とＯＲアレイ（接続しているＡＮＤアレイの出力信号の論理和を作る）からなる。この図では、黒丸と白丸が、それぞれ、入力と出力間の接続を決定するスイッチを示しており（黒丸が接続を意味する）、このスイッチのＯＮ／ＯＦＦをコンフィグレーションビット列で指定することによって、任意の論理回路を構成できる。
【００２０】
この進化型チップを用いることにより、筋電の個人差や経時変化に適応できるだけでなく、ＬＳＩ単体で適応的な回路合成が可能であるため、外部のＰＣなどを必要としない。このため、例えば外出先などで筋電特性が変化した場合でも、その場で識別回路を再構成することができる。
【００２１】
［筋電パターンの測定］
前述したように、我々が開発中の筋電義手では、前腕周りの２箇所から２ｃｈの筋電を測定し、それらのパターン識別により義手の動作を決定する。しかしながら、筋電パターンの識別精度は測定位置に強く依存するにもかかわらず、現状では、試行錯誤によって最適な測定位置を探す以外に、有効な方式は開発されていない。
【００２２】
本発明は、最適な筋電測定位置を探すことが目的ではないので、前腕周りのいくつかの位置から測定した筋電パターンの識別を行い、最も識別精度の良い測定位置を選択する。具体的には、図３に示す屈筋側の３ヶ所（図中２、４、６）と伸筋側（図中１、３、５）の３ヶ所から筋電を測定し、屈筋側と伸筋側から１ヶ所づつの測定位置を選択する９通りの組合せで、筋電パターンの識別を行った。なお、図中、ｕｌｎａ は尺骨を、ｒａｄｉｕｓ は橈骨をそれぞれ表している。
【００２３】
図４は、この筋電の測定に用いる筋電測定装置を示す模式図である。本装置は、筋電センサ （図中４個） 、参照電極 （図中２個） 、増幅器ボックス （Ｄｅｌｓｙｓ製：２個）、ＡＤ変換ＰＣカード （ＲＡＴＯＣシステムズ製：１個）、１台のノートＰＣからなる。即ち、２個の筋電センサ、１個の参照電極、１個の増幅器が１セットとなり、これが２セットある。１セットで２ｃｈの信号を測定することができるので、２セットで合計４ｃｈの信号を同時に測定可能である。増幅器ボックは、［１００、１０００、１００００］倍に増幅可能であるが、ここでは、増幅率を１００００倍に設定した。
【００２４】
ＡＤ変換ＰＣカードは、合計４ｃｈのアナログ信号をデジタル信号に変換し、ＰＣに取り込み可能である。そこで本装置では、筋電センサ、参照電極、増幅器ボックスを２組づつ用い、合計４ｃｈの筋電信号を同時に測定する。つまり、６つの測定位置（図３中の１〜６）から、表２に示す５通りの組合せで４つの測定位置を選び、４ｃｈづつ同時に筋電を測定する。
【表２】

【００２５】
測定された信号は、それぞれ１０００ＨｚでＡＤ変換してＰＣに取り込み、ＰＣ上で平滑化を行う。ここでは１０００個づつのデータの絶対値の平均を求めることによって平滑化を行った。平滑化処理は、一般的にはアナログフィルタを用いて行うが、情報を失わないためにアナログフィルタによる平滑化は行わなかったので、ＰＣ上でソフトウェアによる平滑化を行った。市販の筋電義手で用いられている筋電センサには、内部にアナログローパスフィルタを持つものがあり、図示の構成においても、同様のセンサを用いることが可能である。これによって、平滑化処理における遅延は無視できる程度におさえることができる。
【００２６】
各測定において、被験者には６つの動作（前腕の回内 （ｆｏｒｅａｒｍ ｐｒｏｎａｔｉｏｎ）、回外 （ｆｏｒｅａｒｍ ｓｕｐｉｎａｔｉｏｎ）、手首の掌屈 （ｗｒｉｓｔ ｆｌｅｘｉｏｎ）、背屈 （ｗｒｉｓｔｅｘｔｅｎｓｉｏｎ）、手先の開 （ｈａｎｄ ｏｐｅｎ）、閉 （ｈａｎｄ ｃｌｏｓｅ））を意図して筋肉を収縮してもらい、動作毎２０パターン測定する。つまり、表２の筋電測定位置の組合せ毎に、２０パターン×６動作＝１２０パターン測定するので、電極位置１と２の組からは合計４８０パターン、それ以外の電極位置の組みからは合計２４０パターンづつ測定する。
【００２７】
パターン識別では、これらの筋電パターンから、各動作ランダムに２０パターンづつ（合計１２０パターン）選択し、それらを、識別回路の合成に用いる。さらに、残りの筋電パターンから動作毎２０パターン（合計１２０パターン）を選び、合成した識別回路の評価に用いる。以下では、回路合成に用いる筋電パターンをトレーニングパターン、回路の評価に用いるパターンをテストパターンと呼ぶ。
被験者としては、男性２人、女性３人に参加してもらい、以下では、順に、被験者Ａ、Ｂ、Ｃ、Ｄ、Ｅと呼ぶことにする。被験者Ａは複数回われわれの実験に参加し、システムの仕組みをよく理解している（以下、熟練者と呼ぶ）。残りの４人は、初めてわれわれの実験に参加した（以下、初心者と呼ぶ）。
【００２８】
［量子化エラーの原因］
筋電パターンの量子化では、分布をいくつかの領域に分割し、各領域に分布する筋電パターンを、対応する一つの整数値に量子化する。我々のこれまでの実験では、筋電パターンを０〜７の整数値に量子化しているので、ここでもこの方式を用いる場合について説明する。つまり、図５に示すように、０〜１に正規化した筋電パターンの分布領域を０．１２５（量子化幅）づつ８分割し、各領域に分布する筋電パターンを対応する整数値に量子化（線形量子化）する。
【００２９】
この量子化において、図５の各領域に分布する、つまり同じ離散パターンに量子化される筋電パターンの分布密度を、図６に示す。この図から、筋電パターンは、値の小さい領域に極端に偏って分布しているのがわかる。
そこで、最も顕著な例として領域（０，０）に分布する４１個の筋電パターン［（０，０）に量子化される筋電パターン］が、どの動作時に発生したものかを調べる。すると、１９パターンは回内 （ｆｏｒｅａｒｍ ｐｒｏｎａｔｉｏｎ） 動作時、２０パターンは回外 （ｆｏｒｅａｒｍ ｓｕｐｉｎａｔｉｏｎ）、２パターンは手先を開く （ｈａｎｄ ｏｐｅｎ） 動作時となっており、異なる動作時に発生した筋電パターンが、同じ離散的なパターン（０，０）に量子化されている。この場合、いかなるパターン識別器を用いてもこれらの筋電パターンを識別することはできず、このために、高精度なパターン識別が困難である。
【００３０】
以下の説明では、異なる動作時に発生したにもかかわらず、同じ領域に分布するために同じ離散パターンに量子化される量子化エラーをコリジョン（ｃｏｌｌｉｓｉｏｎ）と呼び、本装置では、コリジョン数を小さくすることによって、識別精度を改善する。そこで、まず、被験者毎のコリジョンの数を、表３に示す。これは、全ての領域におけるコリジョンの合計であり、各電極位置の組合せのうち、最もコリジョンの少ないものを示した。
【表３】

このコリジョンの原因の一つは、図６に示すように偏って分布している筋電パターンを、線形量子化を用いて量子化しているためである。そこで、非線形な量子化法として代表的な、次に示す２つの手法を用いて、この問題を解決する。つまり、音声信号の量子化に用いられるμ−ｌａｗ量子化と、画像データの量子化に用いられるｍｅｄｉａｎ ｃｕｔ量子化である。
【００３１】
［μ−ｌａｗ対数量子化器］
μ−ｌａｗ量子化器は、信号の分布が指数分布と見なせる場合に、その逆特性である対数変換をしてから線形に量子化する方式である。図７の例に代表されるように、筋電パターンの分布は指数特性を示すことが多く、μ−ｌａｗ量子化が有効であると期待できる。
筋電パターンが指数分布するのは、複数動作時に発生する筋電パターンを用いるためである。つまり、筋電パターンを２つのセンサで測定した場合、従来、センサ近くの筋収縮から得られる強い筋電信号だけでは、市販の筋電義手のように３動作以上の識別が困難である。そこで、センサから遠い筋収縮時に発生する筋電信号も用いて動作識別を行うが、体内を伝搬する時に信号が減衰して値の小さい領域に分布する。このため、値の小さい領域に複数の動作時に発生する筋電パターンが密集し、指数特性を示すのである。
【００３２】
具体的には、以下の式によって対数変換した後、量子化を行う。

ここで、ｅ：入力信号、ｖ：対数変換出力、Ｖ：入力の最大値、μ：変換比であり、変換比μを変化させることによって、変換特性を変えることができる。つまり、図７に示すように、μの値を小さくすることにより、ほぼ線形の変換となり、逆にμ値を大きくすると、非線形性が大きくなる。
【００３３】
表４に、μ−ｌａｗ量子化を用いた場合のコリジョンの数を示す。μ値としては図７に示した７つの値を用い、最もコリジョンの少ないものを示した。この結果から、μ−ｌａｗ量子化を用いることによって、コリジョンを少なくすることができるのがわかる。被験者Ｂでコリジョン数が減少しなかったのは、分布が指数特性を示さないためと考えられ、次に示す ｍｅｄｉａｎ ｃｕｔ量子化が有効であると期待できる。
【表４】

Ｍｅｄｉａｎ Ｃｕｔ 量子化は、フルカラー画像 （２４ビット） を低階調 （例：８ビット、２５６階調） ディスプレーに表示する際、画像の品質を落とさずに現画像を近似表示するために提案された手法である。そこでは、まず、分布のヒストグラムを作成し、分布の広がりの中央値 （ｍｅｄｉａｎ） で２つの領域に分割する。このヒストグラム作成では、ある程度高階調 （例：１０ビット、１０２４階調） で量子化し、ヒストグラムを作成する。この量子化を以下では第１段階の量子化と呼び、以下に述べる低階調への量子化を第２段階の量子化と呼ぶ。
【００３４】
次に、分割された領域を、さらにそれぞれの中央値で分ける操作を繰り返すことによって、分布を量子化する。つまり、分布密度の大きいところは細かく分割し、分布密度の小さいところは粗く分割する。このため、μ−ｌａｗ量子化は分布が指数特性を示す場合に有効であったのに対し、Ｍｅｄｉａｎ Ｃｕｔは任意の分布特性に対して有効である。
【００３５】
表５に、ｍｅｄｉａｎ ｃｕｔ量子化を用いた場合のコリジョンの数を示す。ここでは、第２段階の量子化が８階調であるので、第１段階の量子化を２５６階調 （８ビット）とした。この結果から、ｍｅｄｉａｎ ｃｕｔ 量子化を用いることによって、コリジョンを少なくすることができるのがわかる。さらに、μ−ｌａｗの場合にコリジョン数を改善できなかった被験者Ｂでも効果を確認できた。
【表５】

【００３６】
【実施例】
ここでは、μ−ｌａｗ量子化とｍｅｄｉａｎ ｃｕｔ量子化を用いる効果を確認するために、筋電パターンを量子化し、パターン識別を行う。進化型チップによるパターン識別回路の合成では、トレーニングパターンに対する識別率を評価値としてＧＡを適用する。そこで、ＧＡの評価処理を１，０００，０００回行った時点で回路合成を終了し、最も評価値の大きい回路の、テストパターンに対する識別率を用いて、本発明手法の効果を確認する。本実験では、進化型チップのソフトウェアエミュレータを用いた。
【００３７】
μ−ｌａｗ量子化においては、図７に示した７通りのμ値で実験を行い、テストパターンに対する識別率が最も良かったμ値を採用した。ｍｅｄｉａｎ ｃｕｔ量子化においては、第１段階の量子化幅として［５、６、７、８、９、１０、１１］ビットを用い、最も識別率の高いものを採用した。それぞれ［３２、６４、１２８、２５６、５１２、１０２４、２０４８］階調である。表６に採用したμ値（μ−ｌａｗ）と第１段階の量子化幅（ｍｅｄｉａｎ ｃｕｔ）、および、各パターン識別実験で採用した筋電測定位置（最も識別率の高かった筋電測定位置）を示してある。
【表６】

【００３８】
図８に、合成した回路のテストパターンに対する識別率（乱数の初期値を変えて１０試行した平均）のグラフを示す。折れ線Ａ〜Ｅは被験者ＩＤを表す。まず、被験者Ａ（熟練者）においては、線形量子化では８２．３％に留まっていたのに対し、μ−ｌａｗ量子化では９７．８％、ｍｅｄｉａｎ ｃｕｔ 量子化では９５．８％と、実用的なパターン識別率を達成した。つまり、リハビリトレーニングによって操作方法に習熟することによって、思い通りに多機能義手を操作できるようになると期待できる。一方、被験者Ｂ〜Ｅ（初心者）でも、線形量子化を用いた場合に対して、大幅に識別精度が改善しており、提案手法を用いて量子化エラー（コリジョン）を少なくすることによって、識別精度を改善できることが実証できた。
【００３９】
次に、本手法を用いた義手の実用化を考慮し、提案した２つの方式を比較する。図８から、これらの方式ではほぼ同程度の識別精度を達成できており、識別精度だけを考えると、どちらの方式を用いても高精度な動作決定が可能である。しかしながら、μ−ｌａｗ量子化では、μ値に応じて量子化特性が大きく変わるため、適切な値を設定しなければ、高精度な識別が困難であると考えられる。
【００４０】
そこで、表７に、７つのμ値を用いた場合と、ｍｅｄｉａｎ ｃｕｔ 量子化において７通りの第１段階の量子化幅を用いた場合の、パターン識別率の標準偏差を示す。ｍｅｄｉａｎ ｃｕｔ 量子化では、標準偏差が小さい、つまり第１段階の量子化幅を変化させても大きく識別率が変化しないのに対し、μ−ｌａｗ量子化では、標準偏差が大きく、適切なμ設定が重要であることがわかる。この筋電義手の実用化のためには、現場の医師や義肢装具士によって、これらの設定を行う必要があるため、適切なμの設定が必要なμ−ｌａｗ量子化よりも、ｍｅｄｉａｎ ｃｕｔ量子化の方が適切である。
【表７】

【００４１】
【発明の効果】
本発明によれば、分布の偏りに応じて適応的に量子化特性を変更可能な量子化法を用いることにより、量子化エラーを減少させることができる。
これらの量子化を用いることによって、５人の被験者から測定した筋電パターンに対してパターン識別を行い、実用的な筋電パターン識別率（μ−ｌａｗ：９７．８％、 ｍｅｄｉａｎ ｃｕｔ：９５．８％）を達成した。さらに、本発明を具体化した義手の実用化を考慮し、これらの方式の比較を行った結果、現場の医師や義肢装具士の負担の少ないｍｅｄｉａｎ ｃｕｔ量子化が、実用化に適していることが明らかになった。
【図面の簡単な説明】
【図１】前腕筋電義手の基本コンセプトを示す図である。
【図２】Ｐｒｏｇｒａｍｍａｂｌｅ Ｌｏｇｉｃ Ａｒｒａｙ （ＰＬＡ）を示す図である。
【図３】前腕の断面図と筋電測定位置を示す図である。
【図４】筋電測定装置を例示する図である。
【図５】筋電パターンの分布例を示す図である。
【図６】図５に示した筋電パターンの分布密度を示す図である。
【図７】筋電パターンの指数特性分布を示す図である。
【図８】論理回路による筋電パターンの識別率を示すグラフである。
【図９】パターン識別を用いた多機能筋電義手の一般的制御方式を示す図である。[0001]
BACKGROUND OF THE INVENTION
The present invention discriminates myoelectric pattern differences by quantizing the values measured using a plurality of myoelectric sensors into discrete values, and determines the operation of the control target based on the identification result. The present invention relates to an electric pattern identification method and apparatus.
[0002]
[Prior art]
Prosthetic hands (myoelectric prostheses) that can be operated by muscle action potentials (myoelectrics) have been studied since the 1960s, and many have already been put into practical use. For example, in the United States and Europe, it is paid in units of thousands per year and is positioned as one of the standard prosthetic hands.
However, many existing myoelectric prostheses can be operated by myoelectricity with only one function (eg, opening and closing of the hand), and are often inconvenient in daily life. For this reason, in recent years, expectation for a multi-functional myoelectric prosthesis capable of operating a plurality of functions has increased. For example, three points to be improved for existing prosthetic hands include (1) addition of fingertip function, (2) addition of wrist function, and (3) natural appearance.
[0003]
In order to operate a commercially available single-function myoelectric prosthesis, basically, two myoelectric sensors are brought into contact with the skin surface at the cut end, and a 2ch myoelectric signal is measured. Therefore, when each signal intensity exceeds a predetermined threshold value, a corresponding operation is selected. For example, in the case of a forearm myoelectric prosthesis, 2ch myoelectricity is measured by a sensor in contact with the surface of the extensor and flexor muscles remaining at the stump, and the strength of the electromyogram measured on the flexor side exceeds the threshold value. When the extensor side exceeds the threshold, the hand is opened.
[0004]
On the other hand, in order to operate more functions by myoelectricity, three methods have been conventionally used. That is, (1) increase the number of thresholds, (2) increase the myoelectric measurement site, and (3) use the myoelectric pattern identification.
First, in the method of increasing the number of thresholds, for example, in an example in which a 2ch myoelectric signal is measured from the surface of the extensor and flexor muscles remaining at the stump, two threshold values are set for each signal. Therefore, the two operations are identified by utilizing the fact that the time from when the first threshold is exceeded to when the second threshold is exceeded differs between when the force is applied slowly and when the force is applied quickly. For example, if the extensor side is slowly applied, the hand is unscrewed, and if the force is applied quickly, the hand is opened. In the case of the flexor side, the hand is turned around when it is slow, and the hand is closed when it is quick. However, with this method, it is difficult to increase the number of thresholds further.
[0005]
Next, in the method of increasing the number of measurement sites, the same number of electrodes as the operation to be operated are brought into contact with the skin surface on the same number of muscles (groups). Therefore, when each muscle (group) is contracted independently, the measured myoelectric strength exceeds the threshold value and the corresponding action is selected. However, in this method, since it is necessary to move the prosthetic hand by muscle contraction that is completely different from the operation image before cutting, it takes time for training of the user, and in recent years it has hardly been used. Also, considering that the area where electrodes can be mounted is limited and that myoelectric sensors are very expensive, it is not practical to increase the number of electrodes.
[0006]
On the other hand, in the method using pattern identification as shown in FIG. 9, the muscle is contracted with the same operation image as before the cutting, and the difference in the electromyographic pattern generated at that time is identified. Simple. In FIG. 9, the measurement of the myoelectric pattern is performed using a plurality of myoelectric sensors. The difference in the measured myoelectric pattern is identified by the controller, and based on this result, the operation of the controlled object is determined.
In early studies of this method, statistical pattern identification methods were used, but in recent years, nonlinear pattern identification methods such as neural networks and logic circuits have been used due to myoelectric non-linearity ( Non-patent document 1).
[0007]
The present inventor is developing a controller capable of operating a plurality of functions with the aim of putting the multi-functional myoelectric prosthesis into practical use. That is, the controller determines the action of the prosthetic hand by the myoelectric pattern identification, but conventionally, only the improvement of the identification accuracy has been emphasized. However, considering that it is always worn and carried, light weight and downsizing are issues that cannot be ignored. We are aiming to realize a compact controller by using myoelectric pattern identification by logic circuits.
[0008]
In order to realize this controller, it is necessary to roughly divide two problems. First of all, in the design of the pattern discriminator, the operation specifications necessary for synthesizing the discriminating circuit cannot be determined in advance due to individual differences in electromyography and changes over time. Therefore, we solved this problem by using an LSI called an evolution chip that can synthesize and reconfigure logic circuits adaptively to changes in operation specifications, etc., for myoelectric pattern identification (Non-Patent Documents) 1 and 2).
[0009]
Second, in pattern identification by a logic circuit, myoelectricity is quantized into discrete values, which are encoded into 0 and 1 bit strings to be input signals to the identification circuit (logic circuit). Therefore, an efficient quantization method is indispensable in the controller design. For example, two myoelectric patterns that must be identified as different patterns may be quantized into the same discrete myoelectric pattern, which cannot be distinguished using any discriminator. In other words, in order to realize highly accurate myoelectric pattern identification, it is necessary to minimize errors that occur during quantization (quantization errors).
[0010]
[Non-Patent Document 1]
Kajitani et al. , “An Evolveable Hardware Chip and Its Application as a Multi-Function Prosthetic Hand Controller”, Proceedings of the World Antienton National Conferencing. 182-187, 1999.
[Non-Patent Document 2]
Kajitani et al. , A gate-level EHW chip: Implementing GA operations and reconfigurable hardware in a single LSI ”, Evolveable Systems 12: No. 78, No. 8 in the system, From Biology to Hardware.
[Non-Patent Document 3]
B. Smith, “Instantaneous Companding of Quantized Signals”, The Bell System Technical Journal, May, 1957.
[Non-Patent Document 4]
P. Heckbert, "Color Image quantization for frame buffer display", Computer Graphics, Vol. 16, no. 3, 1982.
[0011]
[Problems to be solved by the invention]
Conventionally, linear quantization that uniformly quantizes the distribution has been used for the quantization process. However, as will be described later, there are many quantization errors due to uneven distribution of myoelectric patterns in linear quantization. It became clear that it occurred.
An object of the present invention is to reduce a quantization error by using a quantization method that can adaptively change a quantization characteristic in accordance with a distribution bias.
[0012]
Typical μ-law quantization (see Non-Patent Document 3) and median cut quantization (see Non-Patent Document 4) are used as a method capable of adaptively changing the quantization characteristic according to the pattern distribution characteristics. be able to.
In μ-law quantization, since the distribution of myoelectric patterns often exhibits exponential characteristics, quantization is performed after logarithmic transformation, which is the inverse characteristic. This logarithmic conversion characteristic can be adaptively changed by a variable called μ value. In media cut quantization, depending on the distribution characteristics of the myoelectric pattern, fine quantization is performed where the distribution density is large and coarse quantization is performed where the distribution density is small, so that the quantization error can be reduced.
[0013]
[Means for Solving the Problems]
The present invention uses a quantization method that can adaptively change the quantization characteristics according to the distribution bias in order to reduce the quantization error. Typical μ-law quantization and median cut quantization can be used as a method capable of adaptively changing the quantization characteristic in accordance with the pattern distribution characteristic.
In μ-law quantization, since the distribution of myoelectric patterns often exhibits exponential characteristics, quantization is performed after logarithmic transformation, which is the inverse characteristic. This logarithmic conversion characteristic can be adaptively changed by a variable called μ value. In media cut quantization, depending on the distribution characteristics of the myoelectric pattern, fine quantization is performed where the distribution density is large and coarse quantization is performed where the distribution density is small, so that the quantization error can be reduced.
The myoelectric pattern identification method and apparatus of the present invention includes a controller that identifies a difference in myoelectric patterns by quantizing values measured using a plurality of myoelectric sensors into discrete values, and this identification result The operation of the control target is determined based on the above. The quantization for identifying the myoelectric pattern is characterized in that it is performed using a quantization method that can adaptively change the quantization characteristic in accordance with the distribution characteristic of the pattern.
[0014]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, a system configuration of a multifunctional myoelectric prosthesis embodying the present invention and an evolution chip used as a pattern discriminator will be described based on examples. Basically, as described above with reference to FIG. 9, the present invention determines the operation of a multifunctional myoelectric prosthetic hand or the like as a control target by identifying a myoelectric pattern. This multi-functional myoelectric hand uses a pattern discriminator based on a logic circuit in order to achieve compactness.
[0015]
[System configuration]
FIG. 1 is a diagram showing a basic concept of a forearm myoelectric prosthetic hand, which includes a multi-function hand, a controller, a battery, two myoelectric sensors (two), and a socket. Since the myoelectric sensor (myoelectric electrode) is mounted inside the socket, the cut end is inserted into the socket to contact the skin surface and measure 2ch myoelectricity. Therefore, the measured 2ch myoelectric pattern is identified by the controller, and the operation of the multifunction hand is determined.
The controller has a built-in pattern discriminator based on a logic circuit, and determines the action of the prosthetic hand based on myoelectric pattern discrimination. As this myoelectric pattern discriminator, we use an LSI called the evolution chip described below.
[0016]
[Evolved chip]
In contrast to conventional hardware, the circuit configuration is fixed depending on the design process, whereas the evolved chip has the ability to change its circuit configuration dynamically and autonomously, and can change specifications and operating environments. Designed to adapt to changes in
The evolution type chip includes hardware that can change the circuit configuration any number of times (reconfigurable hardware) and a circuit that adaptively changes the circuit configuration. The circuit configuration of the reconfigurable hardware can be changed any number of times by downloading a software bit string called a configuration bit string. For adaptive rewriting of a circuit configuration, a circuit that executes a search technique called a genetic algorithm (hereinafter abbreviated as GA) at high speed is used (see Non-Patent Document 2).
[0017]
GA is a search method inspired by the evolution of living organisms, and searches for an optimal solution in parallel by encoding a plurality of solution candidates into bit strings of 0 and 1, respectively. In other words, these bit strings are regarded as chromosomes, and a new chromosome (solution candidate) is created by performing an operation called crossover or mutation. Therefore, in order to evaluate how close each chromosome is to the solution, an evaluation value is obtained by a predetermined evaluation function, and based on this evaluation value, an object far from the solution is deleted and an object close to the solution is left. Search for a solution by repeating.
[0018]
In an evolutionary chip, a configuration bit string that specifies the circuit configuration of reconfigurable hardware is regarded as a chromosome. Therefore, by applying GA to evolve the chromosome, if the evaluation function is appropriately set, a configuration bit string that specifies an optimum circuit configuration can be obtained.
This evolved chip consists of (1) GA operation circuit, (2) Reconfigurable hardware, (3) Chromosome memory (chromosome memory), (4) Training pattern memory (for training pattern) Memory), (5) It consists of input / output interfaces.
[0019]
Table 1 shows the GA operation and parameter settings. The reason why the number of individuals is as small as 32 is that a technique that enables efficient search with a small number of individuals is used (see Non-Patent Document 2).
[Table 1]

A programmable logic array (hereinafter abbreviated as PLA) is used as the reconfigurable hardware. As shown in FIG. 2, the PLA is composed of an AND array (which produces a logical product of connected input signals) and an OR array (which produces a logical sum of output signals of connected AND arrays). In this figure, black circles and white circles indicate switches that determine the connection between input and output, respectively (black circles indicate connection), and by specifying ON / OFF of this switch in the configuration bit string, Arbitrary logic circuits can be configured.
[0020]
By using this evolved chip, not only can it be adapted to individual differences in myoelectricity and changes over time, but it is also possible to perform adaptive circuit synthesis with an LSI alone, so no external PC or the like is required. For this reason, for example, even when the myoelectric characteristics change when going out, the identification circuit can be reconfigured on the spot.
[0021]
[Measurement of myoelectric pattern]
As described above, the myoelectric prosthetic hand that we are developing measures 2ch myoelectricity from two places around the forearm and determines the motion of the prosthetic hand by identifying these patterns. However, despite the fact that the myoelectric pattern identification accuracy strongly depends on the measurement position, at present, no effective method has been developed other than searching for the optimal measurement position by trial and error.
[0022]
The purpose of the present invention is not to search for an optimal myoelectric measurement position, so the myoelectric pattern measured from several positions around the forearm is identified, and the measurement position with the highest identification accuracy is selected. Specifically, myoelectricity is measured from the three positions on the flexor side (2, 4, 6 in the figure) and the extensor side (1, 3, 5 in the figure) shown in FIG. The myoelectric pattern was identified by nine combinations of selecting one measurement position from the muscle side. In the figure, ulna represents the ulna, and radius represents the rib.
[0023]
FIG. 4 is a schematic diagram showing a myoelectric measurement device used for the measurement of myoelectricity. This device consists of myoelectric sensors (4 in the figure), reference electrodes (2 in the figure), amplifier box (manufactured by Delsys: 2), AD conversion PC card (manufactured by RATOC Systems: 1), one notebook It consists of a PC. That is, two myoelectric sensors, one reference electrode, and one amplifier form one set, and there are two sets. Since 2 sets of signals can be measured with one set, a total of 4 channels of signals can be measured simultaneously with 2 sets. The amplifier box can amplify [100, 1000, 10000] times. Here, the amplification factor is set to 10000 times.
[0024]
The AD conversion PC card can convert a total of 4 channels of analog signals into digital signals and take them into a PC. Therefore, in this device, two sets of myoelectric sensors, reference electrodes, and amplifier boxes are used, and a total of 4 channels of myoelectric signals are measured simultaneously. That is, four measurement positions are selected from the six measurement positions (1 to 6 in FIG. 3) in five combinations shown in Table 2, and the myoelectricity is measured simultaneously for each of four channels.
[Table 2]

[0025]
The measured signals are each AD converted at 1000 Hz, taken into a PC, and smoothed on the PC. Here, smoothing was performed by obtaining an average of absolute values of 1000 pieces of data. The smoothing process is generally performed using an analog filter, but smoothing by software was performed on a PC because smoothing by an analog filter was not performed in order not to lose information. Some myoelectric sensors used in commercially available myoelectric prostheses have an analog low-pass filter inside, and the same sensor can be used in the illustrated configuration. Thereby, the delay in the smoothing process can be suppressed to a negligible level.
[0026]
In each measurement, the subject had six movements (forearm pronation, forearm supination, wrist flexion, wrist extension, hand open, The muscle is contracted for the purpose of hand closing, and 20 patterns are measured for each movement. That is, 20 patterns × 6 motions = 120 patterns are measured for each combination of EMG measurement positions in Table 2, so that a total of 480 patterns are obtained from the combination of electrode positions 1 and 2, and a total of 240 are obtained from the combinations of other electrode positions. Measure by pattern.
[0027]
In pattern identification, 20 patterns are randomly selected from these myoelectric patterns (120 patterns in total) and used for synthesis of an identification circuit. Further, 20 patterns (total 120 patterns) are selected from the remaining myoelectric patterns and used for evaluation of the synthesized identification circuit. Hereinafter, the myoelectric pattern used for circuit synthesis is called a training pattern, and the pattern used for circuit evaluation is called a test pattern.
As subjects, two men and three women participate, and in the following, subjects A, B, C, D, and E will be called in order. Subject A participates in our experiments multiple times and understands how the system works (hereinafter referred to as an expert). The remaining four people participated in our experiment for the first time (hereinafter referred to as beginners).
[0028]
[Cause of quantization error]
In the quantization of the myoelectric pattern, the distribution is divided into several regions, and the myoelectric pattern distributed in each region is quantized into one corresponding integer value. In our previous experiments, the myoelectric pattern is quantized to an integer value of 0 to 7, so the case where this method is used will be described here. That is, as shown in FIG. 5, the distribution region of the myoelectric pattern normalized to 0 to 1 is divided into eight by 0.125 (quantization width), and the myoelectric pattern distributed in each region is set to a corresponding integer value. Quantize (linear quantization).
[0029]
FIG. 6 shows the distribution density of the myoelectric pattern distributed in each region of FIG. 5 in this quantization, that is, quantized to the same discrete pattern. From this figure, it can be seen that the myoelectric pattern is extremely biased and distributed in a region having a small value.
Therefore, as the most prominent example, it is examined at which operation the 41 myoelectric patterns distributed in the region (0,0) [myoelectric patterns quantized to (0,0)] occurred. Then, 19 patterns are at the time of foreground movement, 20 patterns are at the time of forehead (spinning), 2 patterns are at the time of hand open, and the myoelectric patterns generated at the time of different actions are Are quantized to the same discrete pattern (0,0). In this case, these myoelectric patterns cannot be identified using any pattern discriminator, which makes it difficult to identify patterns with high accuracy.
[0030]
In the following description, a quantization error that is quantized into the same discrete pattern because it is distributed in the same region even though it occurs during different operations is referred to as collision, and this apparatus reduces the number of collisions. This improves the identification accuracy. First, Table 3 shows the number of collisions for each subject. This is the total number of collisions in all regions, and shows the least collision among combinations of electrode positions.
[Table 3]

One of the causes of the collision is that the myoelectric pattern that is unevenly distributed as shown in FIG. 6 is quantized using linear quantization. Therefore, this problem is solved by using the following two methods that are typical as nonlinear quantization methods. That is, μ-law quantization used for quantization of an audio signal and median cut quantization used for quantization of image data.
[0031]
[Μ-law logarithmic quantizer]
The μ-law quantizer is a method of performing linear transformation after logarithmic transformation, which is the inverse characteristic, when a signal distribution can be regarded as an exponential distribution. As represented by the example in FIG. 7, the distribution of myoelectric patterns often exhibits exponential characteristics, and it can be expected that μ-law quantization is effective.
The reason why the myoelectric pattern is exponentially distributed is that the myoelectric pattern generated during a plurality of operations is used. That is, when the myoelectric pattern is measured by two sensors, conventionally, it is difficult to identify three or more movements using only a strong myoelectric signal obtained from muscle contraction near the sensor like a commercially available myoelectric hand. Therefore, although the motion identification is performed also using the myoelectric signal generated when the muscle contracts far from the sensor, the signal is attenuated when propagating through the body and distributed in a region having a small value. For this reason, myoelectric patterns generated at the time of a plurality of operations are concentrated in an area having a small value, and exhibit an exponential characteristic.
[0032]
Specifically, quantization is performed after logarithmic transformation by the following equation.

Here, e is the input signal, v is the logarithmic conversion output, V is the maximum value of the input, and μ is the conversion ratio, and the conversion characteristics can be changed by changing the conversion ratio μ. That is, as shown in FIG. 7, by reducing the value of μ, the conversion becomes almost linear, and conversely, increasing the μ value increases the nonlinearity.
[0033]
Table 4 shows the number of collisions when μ-law quantization is used. As the μ value, the seven values shown in FIG. 7 were used, and the one with the least collision was shown. From this result, it can be seen that collision can be reduced by using μ-law quantization. The reason why the number of collisions did not decrease in subject B is thought to be because the distribution does not show exponential characteristics, and it can be expected that the following media cut quantization is effective.
[Table 4]

Median Cut quantization was proposed to approximate the current image without degrading the image quality when displaying a full color image (24 bits) on a low gradation (eg 8 bits, 256 gradations) display. It is a technique. First, a distribution histogram is created and divided into two regions based on the median of the spread of the distribution. In this histogram creation, a histogram is created by quantizing with some high gradation (eg, 10 bits, 1024 gradations). This quantization is hereinafter referred to as first-stage quantization, and the quantization to the low gradation described below is referred to as second-stage quantization.
[0034]
Next, the distribution is quantized by repeating the operation of dividing the divided areas by their median values. That is, a portion with a high distribution density is divided finely, and a portion with a low distribution density is divided roughly. For this reason, μ-law quantization is effective when the distribution exhibits an exponential characteristic, whereas Median Cut is effective for an arbitrary distribution characteristic.
[0035]
Table 5 shows the number of collisions when media cut quantization is used. Here, since the second stage quantization is 8 gradations, the first stage quantization is 256 gradations (8 bits). From this result, it can be seen that the collision can be reduced by using median cut quantization. Furthermore, in the case of μ-law, the effect could be confirmed even in the subject B who could not improve the collision number.
[Table 5]

[0036]
【Example】
Here, in order to confirm the effect of using μ-law quantization and median cut quantization, the myoelectric pattern is quantized and pattern identification is performed. In the synthesis of the pattern identification circuit by the evolution type chip, GA is applied with the identification rate for the training pattern as an evaluation value. Therefore, when the GA evaluation process is performed 1,000,000 times, the circuit synthesis is completed, and the effect of the method of the present invention is confirmed using the discrimination rate of the circuit having the largest evaluation value with respect to the test pattern. In this experiment, an evolution chip software emulator was used.
[0037]
In the μ-law quantization, an experiment was performed with the seven μ values shown in FIG. 7, and the μ value with the best discrimination rate for the test pattern was adopted. In media cut quantization, [5, 6, 7, 8, 9, 10, 11] bits were used as the first-stage quantization width, and the one with the highest discrimination rate was adopted. [32, 64, 128, 256, 512, 1024, 2048] gradations, respectively. The μ value (μ-law) and the first stage quantization width adopted in Table 6 and the myoelectric measurement position adopted in each pattern discrimination experiment (myoelectric measurement position with the highest discrimination rate). Is shown.
[Table 6]

[0038]
FIG. 8 shows a graph of the identification rate (average of 10 trials by changing the initial value of the random number) against the test pattern of the synthesized circuit. Lines A to E represent subject IDs. First, for subject A (expert), linear quantization remained at 82.3%, but μ-law quantization was 97.8%, and median cut quantization was 95.8%. Pattern recognition rate was achieved. In other words, you can expect to be able to operate your multi-functional prosthesis as you wish by familiarizing yourself with the operation methods through rehabilitation training. On the other hand, subjects B to E (beginners) also have significantly improved discrimination accuracy compared to the case of using linear quantization, and the discrimination is reduced by reducing the quantization error (collision) using the proposed method. It was proved that the accuracy could be improved.
[0039]
Next, considering the practical application of the prosthetic hand using this method, the two proposed methods are compared. From FIG. 8, these methods can achieve almost the same level of identification accuracy, and considering only the identification accuracy, it is possible to determine the operation with high accuracy using either method. However, in the μ-law quantization, the quantization characteristic changes greatly according to the μ value, and therefore it is considered difficult to identify with high accuracy unless an appropriate value is set.
[0040]
Therefore, Table 7 shows the standard deviation of the pattern identification rate when seven μ values are used and when seven first-stage quantization widths are used in median cut quantization. In media cut quantization, the standard deviation is small, that is, the identification rate does not change greatly even if the quantization width of the first stage is changed, whereas in μ-law quantization, the standard deviation is large and an appropriate μ setting is set. Is important. In order to put this myoelectric prosthesis into practical use, it is necessary to make these settings by a doctor on the spot or a prosthetic limb prosthesis. Therefore, the median cut quantum rather than the μ-law quantization that requires an appropriate μ setting. Is more appropriate.
[Table 7]

[0041]
【The invention's effect】
According to the present invention, it is possible to reduce quantization errors by using a quantization method that can adaptively change quantization characteristics in accordance with the distribution bias.
By using these quantizations, pattern discrimination is performed on myoelectric patterns measured from five subjects, and a practical myoelectric pattern discrimination rate (μ-law: 97.8%, median cut: 95. 8%) was achieved. Furthermore, considering the practical application of the prosthetic hand that embodies the present invention, as a result of comparing these methods, media cut quantization, which is less burdensome on the spot doctors and prosthetic orthoses, is suitable for practical use. Became clear.
[Brief description of the drawings]
FIG. 1 is a diagram showing a basic concept of a forearm myoelectric hand.
FIG. 2 is a diagram showing a programmable logic array (PLA).
FIG. 3 is a cross-sectional view of a forearm and a diagram showing a myoelectric measurement position.
FIG. 4 is a diagram illustrating a myoelectric measurement device.
FIG. 5 is a diagram showing an example of distribution of myoelectric patterns.
6 is a diagram showing the distribution density of the myoelectric pattern shown in FIG. 5. FIG.
FIG. 7 is a diagram showing an exponential characteristic distribution of a myoelectric pattern.
FIG. 8 is a graph showing a myoelectric pattern identification rate by a logic circuit.
FIG. 9 is a diagram illustrating a general control method of a multi-function myoelectric hand using pattern identification.

Claims (5)

A controller for identifying a difference in myoelectric pattern by quantizing values measured using a plurality of myoelectric sensors into discrete values, and a muscle for determining the operation of the control target based on the identification result In the electric pattern identification method,
Quantization for identifying the myoelectric pattern is performed using a quantization method capable of adaptively changing the quantization characteristic according to the distribution characteristic of the pattern. The myoelectric pattern identification method according to claim 1, wherein the quantization method capable of adaptively changing the quantization characteristic is a μ-law quantization method or a median cut quantization method. The myoelectric pattern identification method according to claim 1, wherein the control target is a multi-functional myoelectric hand, and the controller is configured to be capable of operating a plurality of functions. The myoelectric pattern identification method according to claim 1, wherein the controller performs myoelectric pattern identification using an evolution type chip that is capable of adaptively synthesizing and reconfiguring a logic circuit in response to a change in operation specifications. A controller for identifying a difference in myoelectric pattern by quantizing values measured using a plurality of myoelectric sensors into discrete values, and a muscle for determining the operation of the control target based on the identification result In the electric pattern identification device,
Quantization for identifying the myoelectric pattern is performed using a quantization method that can adaptively change the quantization characteristic according to the distribution characteristic of the pattern.

Images