JP3977782B2 - Time-series pattern generation apparatus and time-series pattern generation method - Google Patents

Description

を用いて、後述する各時系列パターン発生モデル４０９（途中段階としての時系列パターン）に対する１セグメントデータ４０３の尤度（適合度）をそれぞれ計算する。
【００７３】
ここで、（式１）において、ｌ_ｉ、Δｈ_ｉ、ｙ_ｉ（ｔ）、μ_ｌｉ、μ_Δｈｉ、σ_ｌｉ及びσ_Δｈｉは、時系列パターン発生モデル４０９のパラメータである。より詳しくは、ｌ_ｉは、長さ、Δｈ_ｉは高さ、ｙ_ｉ（ｔ）は形状データを示す。また、後述から明らかになるように、μ_ｌｉ及σ_ｌｉは、長さの平均及び標準偏差、μ_Δｈｉ及びσ_Δｈｉは、高さの平均及び標準偏差を示す。一方、ｌ'_ｉ、Δｈ'_ｉ、ｙ'_ｉ（ｔ）は、１セグメントデータ４０３のパラメータである。また、ｇ_{μ , σ}（ｘ）は平均をμ、標準偏差をσとした正規分布関数を表す。
【００７４】
（式１）に示すように、この尤度関数は、１セグメントデータ４０３と各時系列パターン発生モデルとの適合度を、長さ、高さ及び形状の観点から求める。具体的には、尤度関数は、Σ内に３つの乗算項を有しており、一番左の乗算項が長さの適合度、２番目の乗算項が高さの適合度、３番目の乗算項が形状の適合度に対応する。形状の適合度は、１セグメントデータ４０３の形状データと時系列パターン発生モデル４０９の形状データをそれぞれ正規化した状態で（長さと高さの影響を排除した基準となるサイズに修正した状態で）求めている。従って、（式１）によれば、長さと高さが多少異なっていても、形状が類似していれば、高い適合性を示す尤度が算出される。但し、１セグメントデータ４０３の長さあるいは高さが時系列パターン発生モデル４０９とがあまりに大きく異なる場合は、形状が類似していても、一番左の乗算項と２番目の乗算項によって、低い適合性を示す尤度が算出される。
【００７５】
モデル尤度計算部４０４は、（式１）を用いて算出した尤度のうち最大尤度と２番目の尤度を尤度データ４０８として、図４に示すように、１セグメントデータ４０３とともに、次の尤度評価部４０５に送出する。
【００７６】
尤度評価部４０５は、モデル尤度計算部４０４で算出された最大尤度（maxQ_i）と２番目の尤度（secondQ_i）を元に、次の（式２）に示す尤度比算出式
【数２】

によって算出される尤度比Ｒを計算する。
【００７７】
尤度評価部４０５は、算出した尤度比R=maxQ_i/secondQ_iを、入力された類似度パラメータ１０６（１セグメントデータ４０３と時系列パターン発生モデルとの類似度を判別する際の判別基準値）とを比較する。尤度評価部４０５は、尤度比R=maxQ_i/secondQ_iが類似度パラメータ１０６以下であると判断した場合は、この１セグメントデータに類似する時系列パターン発生モデル４０９は存在しないと判断し、この１セグメントデータ４０３を新規モデル作成部４０７に送出する。
【００７８】
新規モデル作成部４０７は、受け取った１セグメントデータ４０３の長さ、高さ及び形状データと、長さおよび高さの平均（ここでは初期値として１セグメントデータの長さと高さそのものを用いる）と、長さ及び高さの標準偏差（ここでは初期値として０または既に得られている標準偏差の平均の３０％など）とからなる、時系列パターン発生モデル４０９を作成する。新規モデル作成部４０７は、作成した時系列パターン発生モデル４０９を図示しないバッファに記憶するとともに、継続指示データ４１２をセグメント取出継続評価部４０１に送出して、セグメント継続評価部４０１に処理を移行する。
【００７９】
一方、尤度評価部４０５は、算出した尤度比R=maxQ_i/secondQ_iが類似度パラメータ１０６よりも大きいと判断した場合は、この１セグメントデータ４０３は、最大尤度maxQ_iを持つ時系列パターン発生モデル４０９と類似していると判断する。尤度評価部４０５は、１セグメントデータ４０３と類似情報４１０（いずれの時系列パターン発生モデルに類似しているかを示す情報）とをモデル再学習部４０６に送出して、モデル再学習部４０６へ処理を移行する。
【００８０】
モデル再学習部４０６は、類似情報４１０に指定された時系列パターン発生モデルを図示しないバッファから取り出し、以下の（式３）に示す平均・標準偏差更新関数に従って、この時系列パターン発生モデルのパラメータを学習（更新）する。
【数３】

【００８１】
（式３）に示す通り、モデル再学習部４０６は、形状の学習を行わず、長さの平均と標準偏差、及び、高さの平均と標準偏差のみを更新する。モデル再学習部４０６は、更新後の時系列パターン発生モデル４０９を用いて、図示しないバッファ内の対応する時系列パターン発生モデル４０９を置き換える。一方、モデル再学習部４０６は、継続指示データ４１２をセグメント取出継続評価部４０１に送出して、セグメント取出継続評価部４０１に処理を移行する。
【００８２】
以下、本モデル群生成部３０６による処理動作について、セグメント長が１の生成時系列パターン３０７（図７（ａ）参照）を生成する場合を例に説明する。
【００８３】
まず、図４に示すように、セグメント取出継続評価部４０１は、セグメントデータ３０５（歩行時のもの）内に取り出すべき１セグメントデータ４０３が存在するか否かを判断する。ここでは、セグメント取出継続評価部４０１は、取り出すべき１セグメントデータ４０３は存在すると判断し、１セグメント取出部４０２に処理を移行する。
【００８４】
１セグメント取出部４０２は、セグメントデータ３０５から１セグメントデータ４０３（図６における形状ｙ１の区間参照）を取り出し、モデル尤度計算部４０４に送出する。
【００８５】
モデル尤度計算部４０４は、時系列パターン発生モデル４０９が図示しないバッファ内にまだ存在しないので（尤度比Ｒを算出するためには少なくとも２つの時系列パターン発生モデルが必要）、尤度を算出することなく、受け取った１セグメントデータ４０３をそのまま尤度評価部４０５に送出する。
【００８６】
尤度評価部４０５も、受け取った１セグメントデータ４０３をそのまま新規モデル作成部４０７に送出する。
【００８７】
新規モデル作成部４０７は、受け取った１セグメントデータ４０３を用いて、形状データと、長さ及び高さと、長さおよび高さの平均及び標準偏差とを第１の時系列パターン発生モデル４０９として生成し、図示しないバッファ内に記憶する。また、新規モデル作成部４０７は、セグメント取出継続評価部４０１に継続指示データ４１２を送出する。
【００８８】
継続指示データ４１２を受け取ったセグメント取出継続評価部４０１は、セグメントデータ３０５（図６参照）内にまだ取り出すべき１セグメントデータ４０３が存在するか否かを判断する。ここでは、セグメント取出継続評価部４０１は、取り出すべき１セグメントデータ４０３はまだ存在すると判断し、１セグメント取出部４０２に処理を移行する。
【００８９】
１セグメント取出部４０２は、セグメントデータ３０５から次の１セグメントデータ４０３（図６における形状ｙ２の区間参照）を取り出して、モデル尤度計算部４０４に送出する。
【００９０】
モデル尤度計算部４０４は、上述同様、１セグメントデータ４０３をそのまま尤度評価部４０５に送出する。
【００９１】
尤度評価部４０５も、受け取った１セグメントデータ４０３をそのまま新規モデル作成部４０７に送出する。
【００９２】
新規モデル作成部４０７は、上述と同様にして、受け取った１セグメント４０３を用いて、第２の時系列パターン発生モデル４０９を生成して、図示しないバッファ内に記憶するとともに、セグメント取出継続評価部４０１に継続指示データ４１２を送出する。
【００９３】
継続指示データ４１２を受け取ったセグメント取出継続評価部４０１は、セグメントデータ３０５（図６参照）内にまだ取り出すべき１セグメントデータが存在するか否かを判断する。ここでは、セグメント取出継続評価部４０１は、取り出すべき１セグメントデータ４０３はまだ存在すると判断し、１セグメント取出部４０２に処理を移行する。
【００９４】
１セグメント取出部４０２は、セグメントデータ３０５内からさらに次の１セグメントデータ４０３（図６における形状ｙ３の区間参照）を取り出して、モデル尤度計算部４０４に送出する。
【００９５】
モデル尤度算出部４０４は、上述した図示しないバッファ内における第１の時系列パターン発生モデル４０９と、１セグメントデータとを尤度関数（式１）に入力して第１の時系列パターン発生モデル４０９に対する尤度を求める。また、モデル尤度算出部４０４は、上述した第２の時系列パターン発生モデル４０９と１セグメントデータとを尤度関数（式１）に入力して第２の時系列パターン発生モデル４０９に対する尤度を求める。モデル尤度算出部４０４は、算出した第１及び第２の時系列パターン発生モデル４０９に対する尤度（尤度データ４０８）と、１セグメント取出部４０２から受け取った１セグメントデータ４０３とを、尤度評価部４０５に送出する。
【００９６】
尤度評価部４０５は、（式２）に従って、尤度データ４０８を用いて、尤度比Ｒを算出し、算出した尤度比Ｒを、入力された類似度パラメータ１０６と比較する。尤度評価部４０５は、ここでは、尤度比Ｒは類似度パラメータ１０６以下であると判断し（第１及び第２の時系列パターン発生モデル４０９のいずれにも類似していないと判断し）、１セグメントデータ４０３を新規モデル作成部４０７に送出する。
【００９７】
新規モデル作成部４０７は、上述と同様にして、受け取った１セグメントデータ４０３を用いて、第３の時系列モデル発生モデル４０９を生成して図示しないバッファ内に記憶するとともに、セグメント取出継続評価部４０１に継続指示データ４１２を送出する。
【００９８】
継続指示データ４１２を受け取ったセグメント取出継続評価部４０１は、セグメントデータ３０５（図６参照）内にまだ取り出すべき１セグメントデータ４０３が存在するか否かを判断する。ここでは、セグメント取出継続評価部４０１は、取り出すべき１セグメントデータ４０３はまだ存在すると判断し、１セグメント取出部４０２に処理を移行する。
【００９９】
１セグメント取出部４０２は、セグメントデータ３０５内からさらに次の１セグメントデータ４０３（図６における形状ｙ４の区間参照）を取り出し、モデル尤度計算部４０４に送出する。
【０１００】
モデル尤度算出部４０４は、上述と同様に、第１、第２および第３の時系列パターン発生モデルに対する１セグメントデータ４０３の尤度を、尤度関数（式１）を用いて算出する。モデル尤度算出部４０４は、算出した３つの尤度のうち１番目及び２番目に大きいもの（尤度データ４０８）と、１セグメント取出部４０２から受け取った１セグメントデータ４０３とを、尤度評価部４０５に送出する。
【０１０１】
尤度評価部４０５は、受け取った尤度データ４０８を用いて、（式２）に従って、尤度比Ｒを算出する。尤度評価部４０５は、算出した尤度比Ｒと類似度パラメータ１０６とを比較する。尤度評価部４０５は、ここでは尤度比Ｒは類似度パラメータ１０６より大きいと判断する。尤度評価部４０５は、最大尤度を持つ時系列パターン発生モデル（ここでは第１の時系列パターン発生モデルとする）と１セグメントデータ４０３とは類似していると判断する。従って、尤度評価部４０５は、１セグメントデータ４０３と類似情報４１０とをモデル再学習部４０６に送出する。
【０１０２】
モデル再学習部４０６は、類似情報４１０に基づいて第１の時系列パターン発生モデル４０９を特定し、この第１の時系列パターン発生モデル４０９と１セグメントデータ４０３とを用いて、上述の（式３）に従って、第１の時系列パターン発生モデル４０９のパラメータを学習（更新）する。上述したように、形状の学習は行わず、長さと高さの平均及び標準偏差のみを更新する。パラメータを更新したモデル再学習部４０６は、セグメント取出継続評価部４０１に継続指示データ４１２を送出する。
【０１０３】
以上に説明した処理を、セグメントデータ３０５内の残りの他の１セグメントデータ４０３（図６における形状ｙ５の区間参照）および走行時における各１セグメントデータ４０３（図６における形状ｙ６〜ｙ１０の区間参照）についても行い、最終的に、パターン長が１である５つの時系列パターン発生モデル４０９（図７（ａ）参照）が取得される（図示しないバッファ内に格納される）。そして、セグメント取出継続評価部４０１は、セグメントデータ３０５内に取り出すべき１セグメントデータが存在しないと判断したら、上で取得された５つの時系列パターン発生モデル４０９をそれぞれ時系列パターンとして確定し、これらを生成時系列パターン３０７として出力する（図７（ａ）参照）。上述から分かるように、生成時系列パターン３０７を構成する各時系列パターンは、新規に時系列パターン発生モデルを生成する元となった１セグメントデータの長さ、高さ、形状と、上述した、長さ及び高さの平均及び標準偏差とを有する。なお、図７、図８、図９、図１１及び図１２に示される各時系列パターンに付された長さｌ、高さΔｈ及び形状ｙは、ここでは、この１セグメントデータの長さ、高さ、形状とする。
【０１０４】
以上では、パターン長が１の場合におけるモデル群生成部３０６の処理動作について説明したが、パターン長が２，３・・・の場合も同様にして行われる。
【０１０５】
次に、図１に示すように、時系列パターン評価部１０５について説明する。
【０１０６】
図２は、時系列パターン評価部１０５の構成を詳細に示す図である。
【０１０７】
上述したように、時系列パターン評価部１０５は、出力時系列パターン１０４が所望の基準を満たすか否かを判断し、満たす場合は、後述するようにして最終時系列パターン１０７を出力し、一方、満たさない場合は、類似度パラメータ１０６を補正する。
【０１０８】
図２に示すように、この時系列パターン評価部１０５は、出現確率計算部２０１、時系列パターン評価値計算部２０３、時系列パターン評価値評価部２０５および類似度パラメータ補正部２０６を備える。
【０１０９】
出現確率計算部２０１は、出力時系列パターン１０４を構成する各時系列パターン（図９参照）が、各時系列の状態（例えば歩行および走行）において出現する頻度（出現頻度）をそれぞれ調べる。そして、出現確率計算部２０１は、調べた結果を元に、各時系列パターンが、各状態において出現する確率（出現確率）を計算し、算出結果を出現確率２０２として時系列パターン評価値計算部２０３に送出する。
【０１１０】
図１０は、図９の各時系列パターン１〜５，１２，４５（出力時系列パターン１０４）について算出した各状態における出現頻度（歩行時頻度及び走行時頻度）、各状態における出現確率２０２（歩行時確率及び走行時確率）及び時系列パターン評価値２０４（時系列パターン個別評価値の集合）を示した図表である。
【０１１１】
図１０に示すように、各時系列パターン１〜５、１２、４５（図９参照）が、歩行時及び走行時のどちらに頻繁に現れるのかが分かる。例えば、時系列パターン１は、歩行時頻度が２、走行時頻度が０で、従って、歩行時確率は１、走行時確率は０である。
【０１１２】
時系列パターン評価値計算部２０３は、出現確率計算部２０１から受け取った出現確率２０２を用いて、各時系列パターンを評価するための時系列パターン個別評価値をそれぞれ算出し、算出した時系列パターン個別評価値を時系列パターン評価値２０４として時系列パターン評価値評価部２０５に送出する。ここで、時系列パターン個別評価値は、各時系列パターンの各状態への偏り具合状況を数値化して表したものである。
【０１１３】
図１０に示す各時系列パターン個別評価値は、各時系列パターンの平均情報量（エントロピ）を元に算出したものである。この評価では、各時系列パターンは、平均情報量の値が低いほど各状態に偏った良い時系列パターンとなる。例えば、時系列パターン個別評価値が０のパターン１，２，４，５，１２，４５は、歩行時および走行時のいずれかにのみ出現する好ましい時系列パターンといえる。これに対し、時系列パターン個別評価値が０．１５０５１５のパターン３は、歩行時および走行時にそれぞれ同等の確率で出現しており、好ましくない時系列パターンであるといえる。
【０１１４】
時系列パターン評価値評価部２０５は、時系列パターン評価値計算部２０３から受け取った時系列パターン評価値２０４と入力された出力時系列パターン１０４とを用いて、出力時系列パターン１０４の評価を行う。
【０１１５】
より詳しくは、時系列パターン評価値評価部２０５は、時系列パターン評価値２０４を用いて出力時系列パターン１０４全体が所望の基準を満たすか否か評価するための時系列パターン全体評価値を算出する。
【０１１６】
時系列パターン評価値評価部２０５は、算出した時系列パターン全体評価値が所定の基準内にあれば、時系列パターン個別評価値が所定の基準外にある時系列パターンを除去したものを最終時系列パターン１０７として出力する。
【０１１７】
一方、時系列パターン評価値評価部２０５は、算出した時系列パターン全体評価値が所定の基準外ならば、時系列パターン評価値２０４を類似度パラメータ補正部２０６へ送出する。
【０１１８】
以下、出力時系列パターン１０４の評価について具体的に説明する。
【０１１９】
まず、図９〜図１１を用いて、時系列パターン全体評価値が所定の基準内になる場合について説明する。
【０１２０】
時系列パターン評価値評価部２０５は、まず、時系列パターン評価値２０４の平均を算出し、これを時系列パターン全体評価値とする。時系列パターン評価値評価部２０５は、算出した時系列パターン全体評価値と、ユーザ等により設定された全体設定許容値0.11（ユーザにとって納得のいく出力時系列パターン１０４の水準を示し、好適には例えば0.15以下とする）とを比較する。つまり、時系列パターン評価値評価部２０５は、図１０の時系列パターン評価値２０４の平均値0.022（=0.150515/7）と、全体設定許容値0.11とを比較する。平均値0.022は、全体設定許容値0.11よりも小さいので、時系列パターン評価値評価部２０５は、時系列パターン全体評価値は所定の基準内にあると判断する。時系列パターン評価値評価部２０５は、出力時系列パターン１０４を構成する各時系列パターン（図９参照）のうち、個別設定許容値（ユーザにとって納得のいく、各時系列パターンを評価する水準であり、ここでは全体設定許容値と同じ値0.11を有する）よりも高い時系列パターン個別評価値を有する時系列パターン（図１０参照）が存在するか否かを判断する。そして、時系列パターン評価値評価部２０５は、そのような時系列パターンが存在すると判断した場合は、図９と図１１とを比較して分かるように、その時系列パターン（図１０の時系列パターン３）を出力時系列パターン１０４から除去する。そして、時系列パターン評価値評価部２０５は、除去後の出力時系列パターン１０４（図１１参照）を最終時系列パターン１０７として出力する（図２参照）。
【０１２１】
上述では、全体設定許容値及び個別設定許容値としてそれぞれ同じ値を用いたが、それぞれ異なる値を設定してもよい。
【０１２２】
次に、図１２および図１３を用いて、時系列パターン評価値２０４から算出された時系列パターン全体評価値が所定の基準外になる場合について説明する。
【０１２３】
図１２は、上述とは異なる値（上述よりも緩い値）を有する類似度パラメータ１０６によって時系列パターン生成部１０３（図１参照）において生成された出力時系列パターン１０４例を示す図である。この出力時系列パターン１０４は、上述と同様に、図５の時系列データ１０２から生成されたものである。
【０１２４】
図１２に示すように、図９では異なる時系列パターンとして認識されていた時系列パターン１２，４５（図９参照）が、ここでは同一の時系列パターン１２として認識されている。同様に、図９では異なる時系列パターンとして認識されていた時系列パターン１，４が、ここでは同一のパターン１として、また、図９では異なる時系列パターンとして認識されていた時系列パターン２，５が、ここでは、同一の時系列パターン２として認識されている。
【０１２５】
図１３は、図１２の各時系列パターンに対して上述と同様にして算出した各状態における出現頻度、各状態における出現確率２０２及び時系列パターン評価値２０４を示した図表である。
【０１２６】
ここで、出力時系列パターン１０４（図１２参照）の評価について説明すると以下の通りである。
【０１２７】
上述と同様に、まず、時系列パターン評価値評価部２０５（図２参照）は、時系列パターン評価値２０４の平均（時系列パターン全体評価値）と、全体設定許容値0.11（所定の基準）とを比較する。時系列パターン評価値２０４の平均（時系列パターン全体評価値）は、図１３から分かるように、0.150515であり、全体設定許容値0.11よりも大きい。従って、時系列パターン評価値評価部２０５は、時系列パターン全体評価値は所定の基準外にあると判断する。つまり、時系列パターン評価値評価部２０５は、図１２の出力時系列パターン１０４は、目的を満たしていないと判断し、図２に示すように、時系列パターン評価値２０４を類似度パラメータ補正部２０６に送出する。
【０１２８】
類似度パラメータ補正部２０６は、時系列パターン評価値評価部２０５から受け取った時系列パターン評価値２０４を用いて類似度パラメータ２０６値を補正する（より厳しいものにする）。補正の例としては、時系列パターン評価値２０４の平均の前回との差分（初期値は例えば０とする）と、類似度パラメータ１０６の前回との差分（初期値は０とする）との比を定数倍し、今回の類似度パラメータ１０６からこれを減ずるなどがある。
【０１２９】
類似度パラメータ補正部２０６は、以上のようにして算出した補正後の類似度パラメータ１０６を、図１に示すように、時系列パターン生成部１０３に送出する。
【０１３０】
以上に説明したように、本実施の形態によれば、時系列データから、類似度パラメータで示される程度に相互に類似しない時系列パターンを、形状が複雑なものも含めて、パターン長を固定することなく、生成することができる。
【０１３１】
また、時系列データに加えて、時系列の状態も用いるようにしたので、各状態について、できるだけ他の状態には出現しない時系列パターンを生成することが可能となる。
【０１３２】
また、目的の基準に沿った時系列パターンが得られるまで、類似度パラメータを何度でも補正して、最終的に、時系列の状態全てに対して固有の時系列パターンを得ることができる。この際、効率よく類似パラメータを補正することで、より早く目的の基準に沿った時系列パターンを生成することが可能になる。
【０１３３】
また、生成された時系列パターンの各状態における出現確率を算出するようにしたので、時系列パターン個別評価値として、例えば平均情報量を用いることが可能となる。
【０１３４】
また、出現頻度が低い時系列パターンについては削除するようにしたので有効な時系列パターンの生成が可能となる。
【０１３５】
また、削除後残存時系列パターン内の各時系列パターンに対応する時系列データを含むように、パターン長増加後におけるセグメントデータの切出しを行うようにしたので、あまり重要でないセグメントデータの切出しを無視して、効率のよいセグメントデータの切出しが可能になる。
【０１３６】
また、セグメント情報（長さ及び高さ）のみではなく、正規化した形状をも用いて、時系列パターン発生モデルに対する尤度を計算するようにしたので、セグメントデータの形状が複雑な形状を保持していても、適正な類否判別が可能となる。つまり、セグメントデータが複雑な形状を保持していても、相互に類似しない時系列パターンを、長さと高さの曖昧さを吸収して、時系列データから効率よく生成することができる。
【０１３７】
また、得られた時系列パターンが目的に合うかどうかを評価し、目的に沿わない時系列パターンは除去するようにしたので、目的に沿った時系列パターンのみを確実に抽出することができる。
【０１３８】
また、各時系列の状態に偏った時系列パターンを抽出できるので、時系列パターンによるクラス分類も行うことができる。
【０１３９】
また、時系列パターン全体評価値（時系列パターン個別評価値の平均）と全体設定許容値（所定の基準）とを比較して出力時系列パターンを評価するようにしたので、ユーザにとって理解しやすい閾値設定を行うことが可能になる。
【０１４０】
また、各状態における時系列パターンの出現確率や、時系列パターン全体評価値、時系列パターン個別評価値等を画面上等に表示することで、ユーザが時系列パターンを自ら選択することもできる。
【０１４１】
【発明の効果】
本発明により、各状態に対して他の状態と区別することができる固有の時系列パターンを各状態から得られた時系列データに基づき自動的に生成できる。
【図面の簡単な説明】
【図１】本発明の一実施形態に関わる時系列パターン生成装置の構成例を示す図。
【図２】時系列パターン生成装置における時系列パターン評価値評価部の構成例を示す図。
【図３】時系列パターン生成装置における時系列パターン生成部の構成例を示す図。
【図４】時系列パターン生成部におけるモデル群生成部の構成例を示す図。
【図５】状態データおよび時系列データを説明するための図。
【図６】セグメント情報、セグメントデータおよび１セグメントデータを説明するための図。
【図７】生成時系列パターンの例を示す図。
【図８】削除後残存時系列パターンの例（パターン長が２の場合）を示す図。
【図９】出力時系列パターンの例を示す図。
【図１０】出現頻度、出現確率および時系列パターン個別評価値の例を示す図表。
【図１１】最終時系列パターンの例を示す図。
【図１２】途中段階での出力時系列パターンの例を示す図。
【図１３】途中段階での出現確率および時系列パターン個別評価値の例を示す図表。
【符号の説明】
１０１ 状態データ
１０２ 時系列データ
１０３ 時系列パターン生成部
１０４ 出力時系列パターン
１０５ 時系列パターン評価部
１０６ 類似度パラメータ
１０７ 最終時系列パターン
２０１ 出現確率計算部
２０２ 出現確率
２０３ 時系列パターン評価値計算部
２０４ 時系列パターン評価値
２０５ 時系列パターン評価値評価部
２０６ 類似度パラメータ補正部
３０１ セグメント化部
３０２ セメント情報
３０３ セグメント切出し部
３０５ セグメントデータ
３０６ モデル群生成部
３０７ 生成時系列パターン
３０８ 時系列パターン削除部
３０９ 削除後残存時系列パターン
３１１ 継続評価部
３１２ 継続処理部
４０１ セグメント取出継続評価部
４０２ １セグメント取り出し部
４０３ １セグメントデータ
４０４ モデル尤度計算部
４０５ 尤度評価部
４０６ モデル再学習部
４０７ 新規モデル作成部[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a time-series pattern generation apparatus and a time-series pattern generation method for generating a time-series pattern specific to each state from time-series data corresponding to a plurality of states to be observed, and time series data that are not similar to each other. The present invention relates to a time series pattern generation method for generating a series pattern.
[0002]
[Prior art]
With the improvement of sensing technology and processor technology, sensors have become familiar with the reduction in cost. For example, it has become possible to attach a sensor to the foot and acquire the operation as time-series data.
[0003]
Therefore, it is becoming possible to attach a sensor to a person to acquire time series data, and to generate a characteristic time series pattern that appears when the person performs some action from the acquired time series data. Yes. In addition, in order to identify what kind of action the person is doing, the time series data obtained from each action (time series state) is used to generate a time series pattern specific to each action. Has also become required.
[0004]
However, in general, the time series pattern to be acquired often has a huge amount, and there is a problem that it takes a lot of labor to search for the time series pattern by hand.
[0005]
Also, there is no known method for generating a time series pattern specific to each operation using the above-described latter, that is, time series data obtained from each state.
[0006]
On the other hand, with respect to the former, that is, to generate a time-series pattern characteristic of this state using time-series data obtained from a certain state, as a method for realizing this, for example, time-series clustering and rule analysis There is a non-patent document 2 extended in time series.
[0007]
However, in any case, it is necessary to divide the time series data into segments (segmentation) according to an appropriate standard for processing, but in the former time series clustering, the number of segments (number of segments), that is, the pattern length is constant. There is a drawback that it is necessary to be (fixed).
[0008]
On the other hand, the latter rule analysis has the advantage that the pattern length is automatically determined and the time series pattern can be generated efficiently, but when generating a time series pattern with a long pattern length, the time series with a short pattern length A long time series pattern is expressed by combining patterns. For this reason, there exists a fault which each time series pattern tends to be averaged. This is because a short time-series pattern is simple in shape, and thus there are many approximate time-series patterns in many cases. Therefore, this method has a drawback that it cannot handle time-series patterns having a complicated shape when the pattern length becomes long.
[0009]
By the way, when generating a time series pattern, it is preferable that a lot of similar time series patterns are not generated. Therefore, it is necessary to determine similarity of the extracted time series data. In the similarity determination, there is a request that it is the same if the shape is similar even if the length and height are different (to absorb the ambiguity of the length and height). There was a drawback that only the length ambiguity could be absorbed. On the other hand, there is a method using a time series pattern generation model (Non-Patent Document 1) as a method that can absorb both ambiguity in length and height. In this method, the shape of the time series is not retained, and is regarded as a straight line. For this reason, there is a disadvantage that complicated shapes cannot be distinguished from each other.
[0010]
[Patent Document 1]
Japanese Patent Laid-Open No. 11-326542
[Patent Document 2]
JP-A-9-34719
[Non-Patent Document 1]
Xianping Ge and Padhraic Smyth, Deformable Markov Model Templates for Time-Series Pattern Matching, ACM SIGKDD (KDD-2000), pp.81-90, 2000
[Non-Patent Document 2]
Rakesh Agrawal, Tomasz Imielinski, Arun N. Swami, Mining Association Rules between Sets of Items in Large Databases, SIGMOD Conference
1993, pp.207-216, 1993
[0011]
[Problems to be solved by the invention]
As described above, conventionally, there has been a problem in that a time series pattern that can be distinguished from other time series states cannot be automatically generated for each time series state. In addition, when the pattern length is increased, there is a problem that a complicated shape cannot be handled. Furthermore, there is a problem that it is impossible to determine the similarity of a complicated shape that absorbs the ambiguity of length and height.
[0012]
The present invention has been made in consideration of the above-described circumstances, and each time-series state can be distinguished from other time-series states, and a time-series pattern unique to each time-series state is automatically generated. An object of the present invention is to provide a time-series pattern generation device and a time-series pattern generation method that can be generated easily. It is another object of the present invention to provide a time-series pattern generation method that can handle a complicated shape even when the pattern length increases. Furthermore, an object of the present invention is to provide a time-series pattern generation method that makes it possible to determine the similarity of complex shapes, thereby taking into account complex shapes in addition to length and height.
[0013]
  The time-series pattern generation device as one aspect of the present invention is
  A time-series pattern generation device used for generating a time-series pattern specific to each state from time-series data corresponding to a plurality of states to be observed,
  A data cutout unit that divides each time series data into several segments according to a certain standard, and cuts out data from each time series data with a specific segment length,
  A time series pattern corresponding to each of the plurality of data groups obtained by dividing each extracted data into a plurality of data groups having the same similarity using a similarity criterion for determining the similarity between data A model group generation unit for generating
  A time series pattern deleting unit that deletes a time series pattern corresponding to a data group composed of a number of data equal to or less than a deletion threshold;
  The remaining time series pattern is stored, and when the number of the remaining time series patterns is larger than a threshold value, the segment length is set in the data extraction unit.Instruct to enlargeA continuous evaluation department;With
  The data cutout unit cuts out data so as to partially include data corresponding to the time-series pattern remaining in the previous segment length.
  It is characterized by.
[0014]
  A time-series pattern generation method as one aspect of the present invention includes:
  A time series pattern generation method used to generate a time series pattern specific to each state from time series data corresponding to a plurality of states to be observed,
  A data extraction step of dividing each time series data into several segments according to a certain standard, and cutting out data from each time series data with a specific segment length,
  When each cut out data is divided into a plurality of data groups having the same similarity relationship by using a similarity determination criterion for determining the similarity between data, and corresponding to each of the plurality of data groups A model group generation step for generating a sequence pattern;
  A time series pattern deletion step of deleting a time series pattern corresponding to a data group consisting of a number of data equal to or less than a deletion threshold;
  The remaining time series pattern is stored, and when the number of the remaining time series patterns is larger than a threshold, the segment length is set.EnlargeA continuous evaluation step for determining to perform the data cutting step,
  In the data extraction step, the data is extracted so as to partially include data corresponding to the time series pattern remaining in the previous segment length.
It is characterized by.
[0015]
  Claim14In the method, the time series pattern evaluation value evaluation step stops outputting each time series pattern when there is a state that is not associated with any of the time series patterns.
  The similarity determination criterion correction step corrects the similarity determination criterion based on the overall evaluation value of the time series pattern.
[0017]
DETAILED DESCRIPTION OF THE INVENTION
First, features of the present embodiment will be briefly described.
[0018]
FIG. 5 is a diagram showing time-series data 102 composed of time-series data obtained during walking and time-series data obtained during travel. This time-series data 102 is obtained from, for example, an acceleration sensor attached to a biological foot.
[0019]
In the figure, state data (label) 101 including a walking label and a traveling label is shown. The state data 101 is used to specify in what state the time-series data 102 has been acquired, as will be described later.
[0020]
In the present embodiment, a time series pattern specific to each state is generated using such time series data of each state (eg, walking, running).
[0021]
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
[0022]
FIG. 1 is a diagram illustrating a configuration example of a time-series pattern generation apparatus according to an embodiment of the present invention.
[0023]
As illustrated in FIG. 1, the time-series pattern generation device includes a time-series pattern generation unit 103 and a time-series pattern evaluation unit 105.
[0024]
The time-series pattern generation unit 103 uses the input state data 101 (see FIG. 5) and time-series data 102 (see FIG. 5) to have various pattern lengths that are not similar to each other (the pattern length is unfixed). Generate a time series pattern. Generation of time series patterns that are not similar to each other is performed using a similarity parameter 106 (similarity criterion) described later. The time series pattern generation unit 103 outputs the generated time series pattern group as an output time series pattern 104.
[0025]
FIG. 9 to be described later is a diagram showing an example of a time series pattern generated from the time series data 102 of FIG. In FIG. 9, a time series pattern with a pattern length of 1 (upper stage) and a time series pattern with a pattern length of 2 (lower stage) are shown.
[0026]
The time series pattern evaluation unit 105 uses the output time series pattern 104 output from the time series pattern generation unit 103 to evaluate whether or not the output time series pattern 104 satisfies a desired criterion.
[0027]
More specifically, first, the time series pattern evaluation unit 105 individually evaluates the time series pattern overall evaluation value for evaluating the entire output time series pattern 104 and each time series pattern constituting the output time series pattern 104. The time series pattern individual evaluation value for this is calculated using a method described later.
[0028]
Next, the time-series pattern evaluation unit 105 determines whether or not the time-series pattern overall evaluation value is within a predetermined standard, and if it is determined that it is within the predetermined standard, next, It is determined whether or not the time series pattern individual evaluation value is within a predetermined standard. When the time-series pattern evaluation unit 105 determines that there is a time-series pattern individual evaluation value that is outside the predetermined reference, the time-series pattern evaluation unit 105 deletes the time-series pattern having the time-series pattern individual evaluation value from the output time-series pattern 104. The time series pattern evaluation unit 105 outputs the output time series pattern 104 in this state as a final time series pattern 107 as shown in FIG.
[0029]
On the other hand, when the time series pattern evaluation unit 105 determines that the overall time series pattern evaluation value is outside the predetermined reference, the above-described similarity is determined using the overall time series pattern evaluation value or each time series pattern individual evaluation value. The parameter 106 is corrected in a more severe direction. Then, as shown in FIG. 1, the time-series pattern evaluation unit 105 sends the corrected similarity parameter 106 to the time-series pattern generation unit 104 and again causes the time-series pattern generation unit 103 to correct the similarity after correction. The output time series pattern 104 is output by the parameter 106. The time series pattern evaluation unit 105 repeats the correction of the similarity parameter 106 until the time series pattern overall evaluation value falls within a predetermined standard.
[0030]
The final time series pattern 107 output by the time series pattern evaluation unit 105 is a time series pattern finally obtained by this method (for example, a time series pattern unique to walking, a time series pattern unique to running). An example of the final time series pattern 107 is shown in FIG. This is obtained by removing the time series pattern (pattern 3) in which the time series pattern individual evaluation value is determined to be outside the predetermined reference from the output time series pattern 104 of FIG. 9 as described above. As will be apparent from the description below, the patterns 1, 2, and 12 are time series patterns unique to walking, and the other patterns 4, 5, and 45 are time series patterns unique to running. Note that the time series pattern evaluation unit 105 outputs only some of the top ranks (for example, those having a high appearance frequency) when there are a large number of time series patterns whose time series pattern individual evaluation values are within a predetermined standard. Anyway.
[0031]
Next, the time series pattern generation unit 103 and the time series pattern evaluation unit 105 will be described in more detail.
[0032]
First, the time series pattern generation unit 103 will be described.
[0033]
FIG. 3 is a diagram showing in detail the configuration of the time series pattern generation unit 103.
[0034]
As described above, the time series pattern generation unit 103 generates the output time series pattern 104 using the time series data 102 and the state data 101.
[0035]
As shown in FIG. 3, the time series pattern generation unit 103 includes a segmentation unit 301, a segment cutout unit 303, a model group generation unit 306, a time series pattern deletion unit 308, a continuation evaluation unit 311, and a continuation processing unit 312. Prepare.
[0036]
The segmentation unit 301 divides the input time series data 102 into a plurality of sections (segments) on a predetermined basis (segmentation), and for each section, a section size (length in the case of a one-dimensional time series), Calculate the height of the section. The segmenting unit 301 outputs the section size and the section height for each section as segment information 302. As an example of segmentation, there is piecewise linear approximation (in the case of a one-dimensional time series, broken line).
[0037]
FIG. 6 is a diagram showing a state in which time-series data (see FIG. 5) during walking and running are segmented.
[0038]
In FIG. 6, the segment information 302 described above is all data of (length l, height Δh) = (l1, Δh1), (l2, Δh2),..., (L10, Δh10). Here, heights Δh1, Δh2,... Δh10 (not shown) are the differences between the maximum height and the minimum height in each corresponding section. As the height of the section, for example, a difference in height between the beginning and the end of the section may be taken.
[0039]
As shown in FIG. 3, the segment cutout unit 303 adds segment data 305 (state data is not shown) to each piece of data constituting the segment information 302 described above by adding shape data of each section and input state data. ) Is generated. In the example of FIG. 6, the segment data 305 includes (length l, height Δh, shape y) = (l1, Δh1, y1), (l2, Δh2, y2),..., (L10, Δh10, y10). ) Of all data (state data not shown). The shape data y1, y2... Includes the length and height of the corresponding section, but the length and height data are obtained separately from the shape data because they are used in calculations described later. . Each piece of data (length l, height Δh, shape y) (state data not shown) constituting the segment data 305 is referred to as one segment data 403.
[0040]
The model group generation unit 306 uses the segment data 305 generated by the segment cutout unit 303 and the similarity parameter 106 described later, and is not similar to each other, and a time-series pattern based on a specific pattern length (number of segments) (for each state) Including frequency of appearance in time series data). The model group generation unit 306 outputs the generated time series pattern group as a generation time series pattern 307.
[0041]
FIGS. 7A and 7B are diagrams showing an example of a generated time series pattern 307 generated from the segment data 305 (see FIG. 6).
[0042]
More specifically, FIG. 7A is a diagram illustrating an example of a generated time series pattern 307 generated when the pattern length is 1. As shown in the figure, here, five time series patterns 1 to 5 are generated. Here, for example, “Reference: y1, y4” of the time-series pattern 1 indicates that it corresponds to the shapes y1, y4 in FIG. That is, the pattern 1 is a time series pattern generated from the shapes y1 and y4.
[0043]
On the other hand, FIG. 7B is a diagram showing an example of a generation time series pattern when the pattern length is 2. As shown in the figure, when the pattern length is 2, six time-series patterns 12, 23, 31, 45, 53, and 34 are generated.
[0044]
As shown in FIG. 3, the time series pattern deletion unit 308 deletes time series patterns whose appearance frequency is equal to or less than the deletion threshold among the time series patterns in the generated time series pattern 307 and deletes the remaining time series patterns. The remaining time series pattern 309 is output. This is because the time series pattern with a low appearance frequency is deleted to increase the calculation efficiency in the continuation evaluation unit 311 described later.
[0045]
FIG. 8 is a diagram illustrating an example of a post-deletion remaining time series pattern 309 obtained when the deletion threshold is set to 1 with respect to the generation time series pattern 307 (see FIG. 7B) having a pattern length of 2. .
[0046]
As can be seen by comparing FIG. 7B and FIG. 8, time series patterns 23, 31, 53, and 34 having an appearance frequency of 1 are deleted from the six time series patterns shown in FIG. 7B. Then, the patterns 12 and 45 whose appearance frequency is 2 are output as the remaining time series pattern 309 after deletion.
[0047]
The continuous evaluation unit 311 examines the number of time series patterns in the post-deletion remaining time series pattern 309 output by the time series pattern deletion unit 308. If the continuous evaluation unit 311 determines that the number of time-series patterns in the remaining time-series pattern 309 after deletion is not 0, that is, if it is determined that there is a possibility of generating a time-series pattern having a larger pattern length, After storing the remaining time series pattern 309 after deletion, the processing is shifted to the continuation processing unit 312.
[0048]
That is, the continuation processing unit 312 generates the segment data 305 having an increased pattern length (for example, increasing the pattern length from 1 to 2 or from 2 to 3) using the time-series data 102. At this time, the segment data 305 is acquired in a state including a portion of the time series data corresponding to the time series pattern in the remaining time series pattern 309 after deletion. This is to achieve efficient processing by suppressing the extraction of data that seems to be less important.
[0049]
On the other hand, if the continuation evaluation unit 311 determines that the number of time series patterns in the remaining time series pattern 309 after deletion is 0, that is, it is determined that there is no possibility of generating a time series pattern having a larger pattern length. In this case, the post-deletion remaining time series pattern 309 for each pattern length is output as the output time series pattern 104.
[0050]
FIG. 9 briefly described above is a diagram showing an example of an output time series pattern 104 obtained from the time series data 102 shown in FIG.
[0051]
Hereinafter, the processing operation by the above-described time series pattern generation unit 103 will be described in detail by taking as an example the case of generating the output time series pattern 104 of FIG. 9 from the time series data 102 of FIG.
[0052]
First, as shown in FIG. 3, the segmenting unit 301 segments the time-series data 102 (see FIG. 5) (see FIG. 6) to segment information 302 (length l1, height Δh) (length l2, Height Δh2)... Is generated and sent to the segment cutout unit 303.
[0053]
The segment cutout unit 303 uses the segment information 302 to generate segment data 305 (length 11, height Δh 1, shape y 1) (length 12, height Δh 2, shape y 2) (state data is not shown). )) Is generated and sent to the model group generation unit 306.
[0054]
The model group generation unit 306 uses the segment data 305 (length 11, height Δh 1, shape y 1) (length 12, height Δh 2, shape y 2) (state data not shown). A generated time-series pattern 307 (see FIG. 7A) consisting of time-similar patterns having a length of 1 and not similar to each other is generated and sent to the time-series pattern deleting unit 308.
[0055]
The time-series pattern deletion unit 308 deletes the time-series pattern having an appearance frequency of 1 or less from the received generated time-series pattern 307 and sends it to the continuous evaluation unit 311 as the remaining time-series pattern 309 after deletion (upper part of FIG. 9). reference).
[0056]
The continuous evaluation unit 311 determines the number of time-series patterns in the remaining time-series pattern 309 after deletion. In this case, since the number of patterns is not 0 (the number of patterns is 5 as shown in the upper part of FIG. 9), the post-deletion remaining time series pattern 309 is stored internally, and the post-deletion residual time series pattern 309 is continuously processed. To the unit 312.
[0057]
The continuation processing unit 312 generates segment data 305 having a pattern length of 2 using the received post-deletion remaining time-series pattern 309 and the time-series data 102 and sends the segment data 305 to the model group generation unit 306.
[0058]
Using the received segment data 305, the model group generation unit 306 generates a generation time series pattern 307 having a pattern length of 2 (see FIG. 7B) and sends it to the time series pattern deletion unit 308.
[0059]
The time series pattern deletion unit 308 deletes the time series pattern having an appearance frequency of 1 or less from the received generated time series pattern 307 (see the lower part of FIG. 9), and continues the evaluation unit 311 as the remaining time series pattern 309 after deletion. To send.
[0060]
The continuous evaluation unit 311 determines the number of time series patterns in the post-deletion remaining time series pattern 309. In this case, since the number of patterns is not 0 (the number of patterns is 2 as shown in the lower part of FIG. 9), the post-deletion remaining time series pattern 309 is stored, and the post-deletion remaining time series pattern 309 is stored in the continuation processing unit 312. Send it out.
[0061]
The continuation processing unit 312 generates segment data 305 with a pattern length of 3 using the received post-deletion remaining time-series pattern 309 and time-series data 102 and sends the segment data 305 to the model group generation unit 306.
[0062]
Using the received segment data 305, the model group generation unit 306 generates a generation time series pattern 307 having a pattern length of 3 (not shown) and sends it to the time series pattern deletion unit 308.
[0063]
The time-series pattern deleting unit 308 deletes the time-series pattern having an appearance frequency of 1 or less from the received generated time-series pattern 307 as a post-deletion remaining time-series pattern 309 (not shown) and sends it to the continuous evaluation unit 311. To do.
[0064]
The continuous evaluation unit 311 determines the number of time series patterns in the post-deletion remaining time series pattern 309. Since the number of time series patterns is 0 (not shown), the continuation evaluation unit 311 outputs the post-deletion remaining time series pattern 309 (see FIG. 9) having the previously stored pattern lengths 1 and 2 as the output time series pattern 104. Output as.
[0065]
Here, the model group generation unit 306 (see FIG. 3) constituting the time series pattern generation unit 103 will be described in more detail.
[0066]
FIG. 4 is a diagram illustrating a detailed configuration example of the model group generation unit 306.
[0067]
As described above, the model group generation unit 306 includes the segment data 305 (length 11, height Δh 1, shape y 1) (length 12, height Δh 2, shape y 2) (state data not shown). The generation time series pattern 307 is generated from the above.
[0068]
As shown in FIG. 4, the model group generation unit 306 includes a segment extraction continuation evaluation unit 401, a one segment extraction unit 402, a model likelihood calculation unit 404, a likelihood evaluation unit 405, a model relearning unit 406, and a new model. A creation unit 407 is provided.
[0069]
The segment extraction continuation evaluation unit 401 determines whether all the one segment data 403 in the segment data 305 (each data constituting the segment data 305 as described above) has been extracted by the next one segment extraction unit 402. . When the segment extraction continuation evaluation unit 401 determines that one segment data 403 cannot be extracted any more, that is, if all the one segment data 403 in the segment data 305 has been extracted, A series pattern (see FIGS. 7A and 7B) is taken out from a buffer (not shown) and output as a generation time series pattern 307.
[0070]
On the other hand, when the segment extraction continuation evaluation unit 401 determines that the one segment data 403 to be extracted still exists in the segment data 305, the process proceeds to the one segment extraction unit 402.
[0071]
The one segment extraction unit 402 extracts the one segment data 403 from the segment data 305 and sends it to the model likelihood calculation unit 404.
[0072]
The model likelihood calculation unit 404 is, for example, a likelihood function shown in the following (Equation 1).
[Expression 1]

Are used to calculate the likelihood (fitness) of the one segment data 403 for each time series pattern generation model 409 (time series pattern as an intermediate stage) to be described later.
[0073]
Here, in (Equation 1), l_i, Δh_i, Y_i(T), μ_li, Μ_Δhi, Σ_liAnd σ_ΔhiAre parameters of the time-series pattern generation model 409. More specifically, l_iIs the length, Δh_iIs height, y_i(T) shows shape data. Also, as will become clear from the following, μ_liAnd σ_liIs the mean and standard deviation of the length, μ_ΔhiAnd σ_ΔhiIndicates the mean and standard deviation of the height. On the other hand, l '_i, Δh ′_i, Y '_i(T) is a parameter of the one-segment data 403. G_{μ , σ}(X) represents a normal distribution function in which the average is μ and the standard deviation is σ.
[0074]
As shown in (Expression 1), this likelihood function obtains the degree of fit between the one-segment data 403 and each time-series pattern generation model from the viewpoints of length, height, and shape. Specifically, the likelihood function has three multiplication terms in Σ, the leftmost multiplication term is the fitness of length, the second multiplication term is the fitness of height, the third Corresponds to the conformity of the shape. The conformity of the shape is obtained by normalizing the shape data of the one-segment data 403 and the shape data of the time-series pattern generation model 409 (in a state in which the size is corrected to a reference size excluding the influence of the length and height). Seeking. Therefore, according to (Equation 1), even if the length and height are somewhat different, if the shapes are similar, the likelihood indicating high suitability is calculated. However, when the length or height of the one-segment data 403 is very different from the time-series pattern generation model 409, even if the shapes are similar, the length is low depending on the leftmost multiplication term and the second multiplication term. A likelihood indicating suitability is calculated.
[0075]
The model likelihood calculation unit 404 uses the maximum likelihood and the second likelihood of the likelihoods calculated using (Equation 1) as the likelihood data 408, as shown in FIG. It is sent to the next likelihood evaluation unit 405.
[0076]
The likelihood evaluating unit 405 is configured to calculate the maximum likelihood (maxQ calculated by the model likelihood calculating unit 404)._i) And second likelihood (secondQ_i) Based on the likelihood ratio calculation formula shown in the following (Formula 2)
[Expression 2]

The likelihood ratio R calculated by is calculated.
[0077]
The likelihood evaluating unit 405 calculates the calculated likelihood ratio R = maxQ_i/ secondQ_iIs compared with the input similarity parameter 106 (a determination reference value for determining the similarity between the one-segment data 403 and the time-series pattern generation model). Likelihood evaluation section 405 uses likelihood ratio R = maxQ._i/ secondQ_iIs determined to be equal to or less than the similarity parameter 106, it is determined that there is no time-series pattern generation model 409 similar to the one-segment data, and the one-segment data 403 is sent to the new model creation unit 407.
[0078]
The new model creation unit 407 receives the length, height and shape data of the received one-segment data 403, and the average of the length and height (here, the length and height of the one-segment data are used as initial values) A time-series pattern generation model 409 is created which includes standard deviations of length and height (here, 0 as an initial value or 30% of the average of standard deviations already obtained). The new model creation unit 407 stores the created time-series pattern generation model 409 in a buffer (not shown), and sends the continuation instruction data 412 to the segment extraction continuation evaluation unit 401, and shifts the processing to the segment continuation evaluation unit 401. .
[0079]
On the other hand, the likelihood evaluation unit 405 calculates the calculated likelihood ratio R = maxQ._i/ secondQ_iIs determined to be larger than the similarity parameter 106, the one-segment data 403 is obtained from the maximum likelihood maxQ_iIs determined to be similar to the time-series pattern generation model 409. Likelihood evaluation unit 405 sends 1-segment data 403 and similar information 410 (information indicating which time-series pattern generation model is similar) to model re-learning unit 406, to model re-learning unit 406. Migrate processing.
[0080]
The model re-learning unit 406 takes out the time-series pattern generation model specified in the similar information 410 from a buffer (not shown), and sets the parameters of this time-series pattern generation model according to the average / standard deviation update function shown in (Equation 3) below. Learn (update).
[Equation 3]

[0081]
As shown in (Expression 3), the model re-learning unit 406 does not learn the shape, and updates only the average length and standard deviation, and the average height and standard deviation. The model relearning unit 406 uses the updated time series pattern generation model 409 to replace the corresponding time series pattern generation model 409 in a buffer (not shown). On the other hand, the model re-learning unit 406 sends the continuation instruction data 412 to the segment extraction continuation evaluation unit 401 and shifts the processing to the segment extraction continuation evaluation unit 401.
[0082]
Hereinafter, the processing operation by the model group generation unit 306 will be described by taking as an example a case where a generation time series pattern 307 (see FIG. 7A) having a segment length of 1 is generated.
[0083]
First, as shown in FIG. 4, the segment extraction continuation evaluation unit 401 determines whether or not one segment data 403 to be extracted exists in the segment data 305 (the one at the time of walking). Here, the segment extraction continuation evaluation unit 401 determines that there is one segment data 403 to be extracted, and shifts the processing to the one segment extraction unit 402.
[0084]
The 1-segment extraction unit 402 extracts 1-segment data 403 (see the section of the shape y1 in FIG. 6) from the segment data 305 and sends it to the model likelihood calculation unit 404.
[0085]
Since the model likelihood calculation unit 404 does not yet have the time series pattern generation model 409 in the buffer (not shown) (at least two time series pattern generation models are required to calculate the likelihood ratio R), the likelihood is calculated. Without calculation, the received one-segment data 403 is sent to the likelihood evaluation unit 405 as it is.
[0086]
The likelihood evaluating unit 405 also sends the received one segment data 403 to the new model creating unit 407 as it is.
[0087]
The new model creation unit 407 uses the received 1-segment data 403 to generate shape data, length and height, and average and standard deviation of the length and height as the first time-series pattern generation model 409. And stored in a buffer (not shown). Also, the new model creation unit 407 sends continuation instruction data 412 to the segment extraction continuation evaluation unit 401.
[0088]
The segment extraction continuation evaluation unit 401 that has received the continuation instruction data 412 determines whether there is still one segment data 403 to be extracted in the segment data 305 (see FIG. 6). Here, the segment extraction continuation evaluation unit 401 determines that there is still one segment data 403 to be extracted, and shifts the processing to the one segment extraction unit 402.
[0089]
The 1-segment extraction unit 402 extracts the next 1-segment data 403 (see the section of the shape y2 in FIG. 6) from the segment data 305 and sends it to the model likelihood calculation unit 404.
[0090]
The model likelihood calculation unit 404 sends the one segment data 403 to the likelihood evaluation unit 405 as it is, as described above.
[0091]
The likelihood evaluating unit 405 also sends the received one segment data 403 to the new model creating unit 407 as it is.
[0092]
The new model creation unit 407 generates a second time-series pattern generation model 409 using the received one segment 403 in the same manner as described above, stores it in a buffer (not shown), and extracts the segment extraction continuation evaluation unit. The continuation instruction data 412 is sent to 401.
[0093]
The segment extraction continuation evaluation unit 401 that has received the continuation instruction data 412 determines whether there is still one segment data to be extracted in the segment data 305 (see FIG. 6). Here, the segment extraction continuation evaluation unit 401 determines that there is still one segment data 403 to be extracted, and shifts the processing to the one segment extraction unit 402.
[0094]
The 1-segment extraction unit 402 further extracts the next 1-segment data 403 (see the section of the shape y3 in FIG. 6) from the segment data 305 and sends it to the model likelihood calculation unit 404.
[0095]
The model likelihood calculating unit 404 inputs the first time-series pattern generation model 409 and the one-segment data in the above-described buffer (not shown) to the likelihood function (Equation 1), and inputs the first time-series pattern generation model. The likelihood for 409 is obtained. Further, the model likelihood calculating unit 404 inputs the above-described second time series pattern generation model 409 and one segment data into the likelihood function (Equation 1), and the likelihood for the second time series pattern generation model 409. Ask for. The model likelihood calculation unit 404 calculates the likelihood (likelihood data 408) for the calculated first and second time-series pattern generation models 409 and the one-segment data 403 received from the one-segment extraction unit 402 as the likelihood. The data is sent to the evaluation unit 405.
[0096]
The likelihood evaluation unit 405 calculates a likelihood ratio R using the likelihood data 408 according to (Equation 2), and compares the calculated likelihood ratio R with the input similarity parameter 106. Here, the likelihood evaluation unit 405 determines that the likelihood ratio R is equal to or less than the similarity parameter 106 (determines that it is not similar to any of the first and second time-series pattern generation models 409). 1 segment data 403 is sent to the new model creation unit 407.
[0097]
The new model creation unit 407 generates a third time series model generation model 409 using the received 1-segment data 403 and stores it in a buffer (not shown) in the same manner as described above. The continuation instruction data 412 is sent to 401.
[0098]
The segment extraction continuation evaluation unit 401 that has received the continuation instruction data 412 determines whether there is still one segment data 403 to be extracted in the segment data 305 (see FIG. 6). Here, the segment extraction continuation evaluation unit 401 determines that there is still one segment data 403 to be extracted, and shifts the processing to the one segment extraction unit 402.
[0099]
The 1-segment extraction unit 402 further extracts the next 1-segment data 403 (see the section of the shape y4 in FIG. 6) from the segment data 305 and sends it to the model likelihood calculation unit 404.
[0100]
The model likelihood calculation unit 404 calculates the likelihood of the one segment data 403 for the first, second, and third time series pattern generation models using the likelihood function (Equation 1), as described above. The model likelihood calculating unit 404 evaluates the likelihood of the first and second largest (likelihood data 408) of the calculated three likelihoods and the one segment data 403 received from the one segment extracting unit 402. To the unit 405.
[0101]
The likelihood evaluating unit 405 calculates a likelihood ratio R according to (Equation 2) using the received likelihood data 408. The likelihood evaluation unit 405 compares the calculated likelihood ratio R with the similarity parameter 106. Here, the likelihood evaluation unit 405 determines that the likelihood ratio R is larger than the similarity parameter 106. The likelihood evaluation unit 405 determines that the time series pattern generation model having the maximum likelihood (here, the first time series pattern generation model) is similar to the one segment data 403. Therefore, the likelihood evaluation unit 405 sends the one-segment data 403 and the similar information 410 to the model relearning unit 406.
[0102]
The model re-learning unit 406 identifies the first time-series pattern generation model 409 based on the similar information 410, and uses the first time-series pattern generation model 409 and the one-segment data 403, to According to 3), the parameters of the first time-series pattern generation model 409 are learned (updated). As described above, shape learning is not performed, and only the average of length and height and the standard deviation are updated. The model relearning unit 406 that has updated the parameters sends the continuation instruction data 412 to the segment extraction continuation evaluation unit 401.
[0103]
The processing described above is performed by using the remaining one segment data 403 in the segment data 305 (refer to the section of the shape y5 in FIG. 6) and the one segment data 403 at the time of traveling (refer to the sections of the shapes y6 to y10 in FIG. 6). ), And finally, five time-series pattern generation models 409 (see FIG. 7A) having a pattern length of 1 are acquired (stored in a buffer (not shown)). When the segment extraction continuation evaluation unit 401 determines that there is no one segment data to be extracted from the segment data 305, the five time-series pattern generation models 409 acquired above are determined as time-series patterns, respectively. Is generated as a generated time series pattern 307 (see FIG. 7A). As can be seen from the above, each time-series pattern constituting the generated time-series pattern 307 includes the length, height, and shape of the one-segment data from which the time-series pattern generation model is newly generated, With mean and standard deviation of length and height. The length l, the height Δh, and the shape y given to each time series pattern shown in FIGS. 7, 8, 9, 11, and 12 are the length of this one segment data, Height and shape.
[0104]
The processing operation of the model group generation unit 306 when the pattern length is 1 has been described above, but the same operation is performed when the pattern length is 2, 3,.
[0105]
Next, as shown in FIG. 1, the time series pattern evaluation unit 105 will be described.
[0106]
FIG. 2 is a diagram showing in detail the configuration of the time series pattern evaluation unit 105.
[0107]
As described above, the time series pattern evaluation unit 105 determines whether or not the output time series pattern 104 satisfies a desired criterion, and if so, outputs the final time series pattern 107 as described later, If not satisfied, the similarity parameter 106 is corrected.
[0108]
As illustrated in FIG. 2, the time series pattern evaluation unit 105 includes an appearance probability calculation unit 201, a time series pattern evaluation value calculation unit 203, a time series pattern evaluation value evaluation unit 205, and a similarity parameter correction unit 206.
[0109]
The appearance probability calculation unit 201 examines the frequency (appearance frequency) at which each time series pattern (see FIG. 9) constituting the output time series pattern 104 appears in each time series state (for example, walking and running). Then, the appearance probability calculation unit 201 calculates the probability (appearance probability) that each time series pattern appears in each state based on the result of the examination, and sets the calculation result as the appearance probability 202 to calculate the time series pattern evaluation value To 203.
[0110]
FIG. 10 shows the appearance frequency (walking frequency and running frequency) in each state calculated for each time series pattern 1 to 5, 12, 45 (output time series pattern 104) in FIG. 10 is a chart showing a walking probability and a running probability) and a time-series pattern evaluation value 204 (a set of time-series pattern individual evaluation values).
[0111]
As shown in FIG. 10, it can be seen whether each time-series pattern 1-5, 12, 45 (see FIG. 9) appears frequently during walking or running. For example, the time series pattern 1 has a walking frequency of 2 and a running frequency of 0, and thus the walking probability is 1 and the running probability is 0.
[0112]
The time series pattern evaluation value calculation unit 203 uses the appearance probability 202 received from the appearance probability calculation unit 201 to calculate each time series pattern individual evaluation value for evaluating each time series pattern, and calculates the calculated time series pattern. The individual evaluation value is sent to the time series pattern evaluation value evaluation unit 205 as the time series pattern evaluation value 204. Here, the time-series pattern individual evaluation value is a numerical representation of the state of bias of each time-series pattern to each state.
[0113]
Each time series pattern individual evaluation value shown in FIG. 10 is calculated based on the average information amount (entropy) of each time series pattern. In this evaluation, each time series pattern becomes a good time series pattern that is biased to each state as the value of the average information amount is lower. For example, the patterns 1, 2, 4, 5, 12, and 45 having the time series pattern individual evaluation value of 0 can be said to be preferable time series patterns that appear only when walking or running. On the other hand, the pattern 3 having the time series pattern individual evaluation value of 0.150515 appears at the same probability during walking and running, and can be said to be an unfavorable time series pattern.
[0114]
The time series pattern evaluation value evaluation unit 205 evaluates the output time series pattern 104 using the time series pattern evaluation value 204 received from the time series pattern evaluation value calculation unit 203 and the input output time series pattern 104. .
[0115]
More specifically, the time series pattern evaluation value evaluation unit 205 uses the time series pattern evaluation value 204 to calculate a time series pattern overall evaluation value for evaluating whether or not the entire output time series pattern 104 satisfies a desired criterion. To do.
[0116]
The time series pattern evaluation value evaluation unit 205 removes the time series pattern in which the time series pattern individual evaluation value is outside the predetermined reference if the calculated time series pattern overall evaluation value is within the predetermined reference. Output as a sequence pattern 107.
[0117]
On the other hand, the time series pattern evaluation value evaluation unit 205 sends the time series pattern evaluation value 204 to the similarity parameter correction unit 206 if the calculated time series pattern overall evaluation value is outside a predetermined reference.
[0118]
Hereinafter, the evaluation of the output time series pattern 104 will be specifically described.
[0119]
First, the case where the entire time series pattern evaluation value falls within a predetermined standard will be described with reference to FIGS.
[0120]
First, the time-series pattern evaluation value evaluation unit 205 calculates an average of the time-series pattern evaluation values 204 and sets this as the time-series pattern overall evaluation value. The time-series pattern evaluation value evaluation unit 205 indicates the calculated time-series pattern overall evaluation value and the overall setting allowable value 0.11 set by the user or the like (shows the level of the output time-series pattern 104 that is satisfactory to the user, preferably For example, 0.15 or less). That is, the time series pattern evaluation value evaluation unit 205 compares the average value 0.022 (= 0.150515 / 7) of the time series pattern evaluation value 204 of FIG. 10 with the overall setting allowable value 0.11. Since the average value 0.022 is smaller than the overall setting allowable value 0.11, the time series pattern evaluation value evaluation unit 205 determines that the time series pattern overall evaluation value is within a predetermined standard. The time-series pattern evaluation value evaluation unit 205 is an individual setting allowable value (at a level for evaluating each time-series pattern that is satisfactory to the user among the time-series patterns constituting the output time-series pattern 104 (see FIG. 9). Yes, here, it is determined whether or not there is a time series pattern (see FIG. 10) having a time series pattern individual evaluation value higher than (the same value 0.11 as the overall setting allowable value). When the time-series pattern evaluation value evaluating unit 205 determines that such a time-series pattern exists, the time-series pattern (the time-series pattern in FIG. 10) can be understood by comparing FIG. 9 and FIG. 3) is removed from the output time series pattern 104. Then, the time series pattern evaluation value evaluation unit 205 outputs the output time series pattern 104 after removal (see FIG. 11) as the final time series pattern 107 (see FIG. 2).
[0121]
In the above description, the same value is used as the overall setting allowable value and the individual setting allowable value, but different values may be set.
[0122]
Next, a case where the time series pattern overall evaluation value calculated from the time series pattern evaluation value 204 is outside a predetermined standard will be described with reference to FIGS. 12 and 13.
[0123]
FIG. 12 is a diagram illustrating an example of the output time series pattern 104 generated in the time series pattern generation unit 103 (see FIG. 1) by the similarity parameter 106 having a value different from the above (a value looser than the above). This output time series pattern 104 is generated from the time series data 102 of FIG. 5 as described above.
[0124]
As shown in FIG. 12, the time series patterns 12 and 45 (see FIG. 9) recognized as different time series patterns in FIG. 9 are recognized as the same time series pattern 12 here. Similarly, time series patterns 1 and 4 recognized as different time series patterns in FIG. 9 are the same pattern 1 here, and time series patterns 2 and 2 recognized as different time series patterns in FIG. 5 are recognized as the same time-series pattern 2 here.
[0125]
FIG. 13 is a chart showing the appearance frequency in each state, the appearance probability 202 in each state, and the time series pattern evaluation value 204 calculated in the same manner as described above for each time series pattern in FIG.
[0126]
Here, the evaluation of the output time series pattern 104 (see FIG. 12) will be described as follows.
[0127]
Similar to the above, first, the time-series pattern evaluation value evaluation unit 205 (see FIG. 2) first calculates the average of the time-series pattern evaluation values 204 (time-series pattern overall evaluation value) and the overall setting allowable value 0.11 (predetermined standard). And compare. As can be seen from FIG. 13, the average of the time series pattern evaluation values 204 (time series pattern overall evaluation value) is 0.150515, which is larger than the overall setting allowable value 0.11. Therefore, the time-series pattern evaluation value evaluation unit 205 determines that the overall time-series pattern evaluation value is outside a predetermined standard. That is, the time series pattern evaluation value evaluation unit 205 determines that the output time series pattern 104 in FIG. 12 does not satisfy the purpose, and converts the time series pattern evaluation value 204 to the similarity parameter correction unit as shown in FIG. To 206.
[0128]
The similarity parameter correction unit 206 corrects the similarity parameter 206 value (makes it stricter) using the time series pattern evaluation value 204 received from the time series pattern evaluation value evaluation unit 205. As an example of correction, the ratio of the difference between the average of the time-series pattern evaluation values 204 (the initial value is 0, for example) and the difference between the previous time of the similarity parameter 106 (the initial value is 0) Is multiplied by a constant, and this is subtracted from the similarity parameter 106 of this time.
[0129]
The similarity parameter correction unit 206 sends the corrected similarity parameter 106 calculated as described above to the time-series pattern generation unit 103 as shown in FIG.
[0130]
As described above, according to the present embodiment, the time length of the time series patterns that are not similar to each other as much as indicated by the similarity parameter is fixed from the time series data, including those having complicated shapes. Can be generated without.
[0131]
In addition to the time-series data, the time-series state is also used, so that it is possible to generate a time-series pattern that does not appear in other states as much as possible.
[0132]
Further, the similarity parameter can be corrected any number of times until a time series pattern that meets the target standard is obtained, and finally, a unique time series pattern can be obtained for all the time series states. At this time, it is possible to generate a time series pattern according to a target criterion earlier by efficiently correcting the similar parameters.
[0133]
Moreover, since the appearance probability in each state of the generated time series pattern is calculated, for example, an average information amount can be used as the time series pattern individual evaluation value.
[0134]
Further, since a time series pattern with a low appearance frequency is deleted, an effective time series pattern can be generated.
[0135]
In addition, segment data is extracted after the pattern length has been increased so that time series data corresponding to each time series pattern in the remaining time series pattern after deletion is included, so segmentation of less important segment data is ignored. Thus, efficient segment data can be extracted.
[0136]
In addition, not only segment information (length and height) but also the normalized shape is used to calculate the likelihood for the time-series pattern generation model, so the shape of the segment data retains a complex shape Even if this is done, it is possible to determine the appropriate similarity. That is, even if the segment data has a complicated shape, time series patterns that are not similar to each other can be efficiently generated from the time series data by absorbing the ambiguity of length and height.
[0137]
In addition, it is evaluated whether or not the obtained time series pattern meets the purpose, and the time series pattern that does not meet the purpose is removed, so that only the time series pattern that meets the purpose can be reliably extracted.
[0138]
Further, since a time series pattern biased to each time series state can be extracted, class classification based on the time series pattern can also be performed.
[0139]
In addition, since the time series pattern overall evaluation value (average of time series pattern individual evaluation values) is compared with the overall set allowable value (predetermined standard), the output time series pattern is evaluated, which is easy for the user to understand. It is possible to set a threshold value.
[0140]
In addition, by displaying the appearance probability of the time series pattern in each state, the entire time series pattern evaluation value, the time series pattern individual evaluation value, and the like on the screen or the like, the user can select the time series pattern himself.
[0141]
【The invention's effect】
According to the present invention, a unique time series pattern that can be distinguished from other states for each state can be automatically generated based on time series data obtained from each state.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating a configuration example of a time-series pattern generation apparatus according to an embodiment of the present invention.
FIG. 2 is a diagram illustrating a configuration example of a time series pattern evaluation value evaluation unit in the time series pattern generation device.
FIG. 3 is a diagram illustrating a configuration example of a time series pattern generation unit in the time series pattern generation apparatus.
FIG. 4 is a diagram illustrating a configuration example of a model group generation unit in a time series pattern generation unit.
FIG. 5 is a diagram for explaining state data and time-series data.
FIG. 6 is a diagram for explaining segment information, segment data, and one-segment data.
FIG. 7 is a diagram showing an example of a generated time series pattern.
FIG. 8 is a diagram showing an example of a time series pattern remaining after deletion (when the pattern length is 2);
FIG. 9 is a diagram showing an example of an output time series pattern.
FIG. 10 is a chart showing an example of appearance frequency, appearance probability, and time-series pattern individual evaluation value.
FIG. 11 is a diagram showing an example of a final time series pattern.
FIG. 12 is a diagram showing an example of an output time-series pattern at an intermediate stage.
FIG. 13 is a chart showing an example of an appearance probability and a time-series pattern individual evaluation value at an intermediate stage.
[Explanation of symbols]
101 Status data
102 Time series data
103 Time-series pattern generator
104 Output time series pattern
105 Time-series pattern evaluation unit
106 Similarity parameter
107 Final time series pattern
201 Appearance probability calculator
202 Appearance probability
203 Time series pattern evaluation value calculator
204 Time series pattern evaluation value
205 Time-series pattern evaluation value evaluation unit
206 Similarity parameter correction unit
301 Segmentation section
302 Cement information
303 Segment cutout
305 segment data
306 Model group generator
307 Generation time series pattern
308 Time series pattern deletion part
309 Time series pattern remaining after deletion
311 Continuous Evaluation Department
312 Continuation processing section
401 Segment extraction continuation evaluation department
402 1 segment take-out part
403 1 segment data
404 Model likelihood calculator
405 Likelihood evaluation unit
406 Model re-learning section
407 New model creation department

Claims (20)

A time-series pattern generation device used for generating a time-series pattern specific to each state from time-series data corresponding to a plurality of states to be observed,
A data cutout unit that divides each time series data into several segments according to a certain standard, and cuts out data from each time series data with a specific segment length,
A time series pattern corresponding to each of the plurality of data groups obtained by dividing each extracted data into a plurality of data groups having the same similarity using a similarity criterion for determining the similarity between data A model group generation unit for generating
A time series pattern deleting unit that deletes a time series pattern corresponding to a data group composed of a number of data equal to or less than a deletion threshold;
A continuation evaluation unit that stores the remaining time series pattern, and instructs the data extraction unit to increase the segment length when the number of the remaining time series patterns is greater than a threshold ,
The data cutout unit cuts out data so as to partially include data corresponding to the time-series pattern remaining in the previous segment length.
A time-series pattern generating apparatus characterized by the above . The model group generation unit includes:
A new model creation unit that creates a time series pattern generation model that is a model for generating a time series pattern,
A model relearning unit for correcting the time-series pattern generation model;
A model likelihood calculation unit that calculates the likelihood of each time-series pattern generation model of the data cut out by the data cut-out unit from the viewpoint of the length, height, and shape of the data;
Evaluate whether or not a time series pattern generation model similar to the data exists from each calculated likelihood, and if not, create a time series pattern generation model using the data in the new model creation unit A likelihood evaluation unit that instructs the model re-learning unit to correct a time-series pattern generation model that is most similar to the data, if present, with the data;
The time-series pattern generating apparatus according to claim 1 , comprising: The likelihood evaluation unit has the highest likelihood when the ratio of the highest likelihood to the next highest likelihood among the calculated likelihoods satisfies the value of the similarity parameter given in advance. 3. The time series pattern generation apparatus according to claim 2 , wherein it is determined that the data is most similar to a series pattern generation model. The new model creation unit calculates the length, height and shape of the data as creation of the time series pattern generation model,
The model relearning unit calculates the average and standard deviation of the length and height as a modification of the time-series pattern generation model,
The model likelihood calculation unit includes the length, height, and shape of each time series pattern generation model, the average and standard deviation of the length and height, and the length, height, and shape of the data. The time series pattern generation apparatus according to claim 2 , wherein a likelihood for each time series pattern generation model is calculated. Evaluating whether to adopt the set of each time series pattern based on the appearance frequency in each time series data of each time series pattern stored by the continuation evaluation unit,
When adopting the set, specify each time series pattern to one of the states based on the appearance frequency in each time series data of each time series pattern included in the set, each specified state Output time series pattern,
If the set is not adopted, the model group generation unit is instructed to correct the similarity determination criterion in a stricter direction.
The time series pattern generation device according to claim 1, further comprising a time series pattern evaluation unit. The time-series pattern evaluation unit
An appearance probability calculating unit for calculating an appearance probability of each time series pattern appearing in each time series data;
Using each occurrence probability calculated for each time series pattern, a time series pattern evaluation value calculation unit that calculates an entire time series pattern evaluation value for evaluating whether or not to adopt each time series pattern set. When,
When the time series pattern overall evaluation value satisfies a pre-set overall setting allowable value, a time series pattern evaluation value evaluation unit that identifies each time series pattern as having the highest appearance probability,
When the time series pattern overall evaluation value does not satisfy the overall setting allowable value, the similarity determination criterion is corrected based on the time series pattern overall evaluation value, and the corrected similarity determination criterion is designated in the model group generation unit A similarity determination reference correction unit,
The time-series pattern generation apparatus according to claim 5 , comprising: The time-series pattern evaluation value calculation unit calculates a time-series pattern individual evaluation value for evaluating a bias of appearance of each time-series pattern with respect to each time-series data for each time-series pattern. Calculate the time series pattern overall evaluation value from the time series pattern individual evaluation value,
The time-series pattern evaluation value evaluation unit inspects whether each time-series pattern individual evaluation value satisfies a predetermined individual setting allowable value when the entire time-series pattern evaluation value satisfies the total setting allowable value. The time series pattern having the time series pattern individual evaluation value that does not satisfy the individual setting allowable value is removed, and the time series pattern having the time series pattern individual evaluation value that satisfies the individual setting allowable value has the highest appearance probability. Specific to the state,
The similarity determination criterion correction unit, when the time series pattern overall evaluation value does not satisfy the overall setting allowable value, based on at least one of the time series pattern individual evaluation value and the time series pattern overall evaluation value Correct the degree criterion,
The time-series pattern generation apparatus according to claim 6 . A time series pattern generation method used to generate a time series pattern specific to each state from time series data corresponding to a plurality of states to be observed,
A data extraction step of dividing each time series data into several segments according to a certain standard, and cutting out data from each time series data with a specific segment length,
When each cut out data is divided into a plurality of data groups having the same similarity relationship by using a similarity determination criterion for determining the similarity between data, and corresponding to each of the plurality of data groups A model group generation step for generating a sequence pattern;
A time series pattern deletion step of deleting a time series pattern corresponding to a data group consisting of a number of data equal to or less than a deletion threshold;
Storing a remaining time series pattern, and when the number of the remaining time series patterns is larger than a threshold, a continuous evaluation step for determining to perform the data extraction step by increasing the segment length, and
In the data extraction step, the data is extracted so as to partially include data corresponding to the time series pattern remaining in the previous segment length.
A time-series pattern generation method characterized by that . The model group generation step includes:
A new model creation step for creating a time series pattern generation model to be a model for generating a time series pattern;
A model relearning step for correcting the time-series pattern generation model;
A model likelihood calculation step for calculating the likelihood of each time-series pattern generation model of the data extracted by the data extraction step from the viewpoint of the length, height, and shape of the data;
Evaluate whether or not there is a time series pattern generation model similar to the data from each calculated likelihood, and if not, create a time series pattern generation model from the data by the new model creation step A likelihood evaluation step of correcting a time-series pattern generation model most similar to the data based on the data in the model relearning step,
The time series pattern generation method according to claim 8 , comprising: When the likelihood evaluation step has the highest likelihood when the ratio of the highest likelihood to the next highest likelihood among the calculated likelihoods satisfies the value of the similarity parameter given in advance. The time series pattern generation method according to claim 9 , wherein it is determined that the data is most similar to a series pattern generation model. The new model creation step calculates the length, height and shape of the data as the creation of the time series pattern generation model,
The model relearning step calculates the average and standard deviation of the length and height as a modification of the time series pattern generation model,
The model likelihood calculation step includes the length, height, and shape of each time series pattern generation model, the average and standard deviation of the length and height, and the length, height, and shape of the data. The method for generating a time series pattern according to claim 9 or 10 , wherein a likelihood for each time series pattern generation model is calculated. Evaluating whether to adopt the set of each time series pattern based on the appearance frequency in each time series data of each time series pattern stored by the continuous evaluation step,
When adopting the set, specify each time series pattern to one of the states based on the appearance frequency in each time series data of each time series pattern included in the set, each specified state Output time series pattern,
If the set is not adopted, a time series pattern evaluation step for correcting the similarity determination criterion in a stricter direction,
The time series pattern generation method according to claim 8 , further comprising: The time series pattern evaluation step includes:
An appearance probability calculating step for calculating an appearance probability of each time series pattern appearing in each time series data;
A time series pattern evaluation value calculating step for calculating an entire time series pattern evaluation value for evaluating whether or not to adopt each time series pattern set from each calculated appearance probability;
When the time series pattern overall evaluation value satisfies a pre-set overall setting allowable value, each time series pattern is specified as a state having the highest appearance probability, and each time series pattern in which the state is specified is output. A pattern evaluation value evaluation step;
When the time-series pattern overall evaluation value does not satisfy the overall setting allowable value, the similarity determination criterion correction step for correcting the similarity determination criterion based on the time-series pattern overall evaluation value;
The time series pattern generation method according to claim 12 , further comprising: In the time series pattern evaluation value evaluation step, when there is a state that is not associated with any of the time series patterns, the output of each time series pattern is stopped,
The time series pattern generation method according to claim 13 , wherein the similarity determination criterion correction step corrects the similarity determination criterion based on the overall evaluation value of the time series pattern. The time series pattern evaluation value calculating step includes, for each time series pattern, a maximum appearance probability among the appearance probabilities in each time series data of the time series pattern and all appearance probabilities excluding the maximum appearance probability. And calculating the average information amount using the maximum appearance probability and the sum of the appearance probabilities, and using the average information amount obtained from each time series pattern, The time series pattern generation method according to claim 13 , wherein the time series pattern generation method is calculated. The time series pattern evaluation value calculation step calculates a time series pattern individual evaluation value for evaluating a bias of appearance of each time series pattern with respect to each time series data for each time series pattern, Calculate the time series pattern overall evaluation value from the time series pattern individual evaluation value,
In the time series pattern evaluation value evaluation step, when the time series pattern overall evaluation value satisfies the overall setting allowable value, it is checked whether or not each time series pattern individual evaluation value satisfies a predetermined individual setting allowable value. The time series pattern having the time series pattern individual evaluation value that does not satisfy the individual setting allowable value is removed, and the time series pattern having the time series pattern individual evaluation value that satisfies the individual setting allowable value has the highest appearance probability. Specific to the state,
The similarity determination reference correction step uses the time series pattern individual evaluation value and the time series pattern overall evaluation value when the time series pattern overall evaluation value does not satisfy the overall setting allowable value. Correct the similarity criterion,
The time-series pattern generation method according to claim 13 . The time series pattern evaluation value calculation step obtains a maximum appearance probability among the appearance probabilities in each time series data of the time series pattern and a sum of all appearance probabilities excluding the maximum appearance probability, using the occurrence probability between the sum of the occurrence probability to calculate the average amount of information, when according to the average amount of information to claim 16, characterized in that to obtain a time series pattern individual evaluation value of the time series pattern Series pattern generation method. The time series pattern evaluation value according to claim 17 , wherein the time series pattern evaluation value calculation step obtains the time series pattern overall evaluation value by averaging the time series pattern individual evaluation values of the time series patterns. Generation method. The occurrence probability, the time series pattern overall evaluation value, and according to any one of the time series pattern to claims 16, characterized in that further comprising an output step of outputting at least one of the individual evaluation value 18 Time series pattern generation method. An appearance probability calculating step of calculating an appearance probability of each time series pattern stored in the continuous evaluation step appearing in each time series data;
Using each occurrence probability calculated for each time-series pattern, calculate a time-series pattern individual evaluation value for evaluating the bias of appearance of each time-series pattern with respect to each time-series data. Further, the time series pattern evaluation value for calculating the time series pattern overall evaluation value for evaluating whether or not to adopt the set of each time series pattern stored in the continuous evaluation step from the time series pattern individual evaluation value. A calculation step;
When the time series pattern overall evaluation value satisfies a predetermined overall setting allowable value, the time series pattern individual evaluation value is inspected to satisfy a predetermined individual setting allowable value, and the individual setting allowable value The time series pattern having a time series pattern individual evaluation value that does not satisfy the time series pattern is determined, and the time series pattern having a time series pattern individual evaluation value that satisfies the individual setting allowable value is identified as having the highest appearance probability. A pattern evaluation value evaluation step;
When the overall evaluation value of the time series pattern does not satisfy the overall setting allowable value, the similarity parameter is based on the difference between the average of the time series pattern individual evaluation values and the previous difference of the similarity parameter. A similarity criterion correction step for correcting the image in a stricter direction,
The time series pattern generation method according to claim 10 , comprising:

Images