JP3788297B2 - FREQUENCY ANALYZING APPARATUS, ABNORMALITY DETERMINING APPARATUS AND ABNORMALITY DETERMINING SYSTEM APPLYING FFT AFFECT - Google Patents

Description

更に、ＳＩＮ波形にこのＳｔａｂｌｅ Ｗｉｎｄｏｗ関数を掛けると、図１２に示すようになる。そして、このように窓関数を掛けた信号に対して、ＦＦＴ処理を行うことにより、周波数成分毎のパワースペクトルが生成できる。なお、ＦＦＴアルゴリズムは、既存の高速フーリエアルゴリズムを使用することができる。
【００４８】
本形態では、両端に減衰部を設けることにより、フレーム分割により切り出されたフレームデータを前後につなげたときの不連続性によるスペクトルの漏洩現象を防止する。
【００４９】
これにより、単発信号の発生タイミングのずれに伴う窓関数によりＦＦＴの周波数毎のスペクトル量が変化しない。つまり、安定検出部の範囲内で有れば、どの位置に瞬間的に単発の信号が発生したとしても、ＦＦＴ処理されて得られるスペクトル量は一定の値をとる。
【００５０】
さらに、本形態では、前後のフレームがフレームサイズの１／２分のデータ（時間軸領域）をオーバーラップさせたため、処理対象のフレームｉの両端の減衰部は、フレーム（ｉ−１）或いはフレーム（ｉ＋１）の安定検出部の領域とオーバーラップする。従って、単発信号が処理対象のフレームｉの減衰部の領域で発生した場合には、当該フレームｉに基づくＦＦＴ処理をすると、減衰されてしまいスペクトル量は小さくなり、消失してしまうおそれが高いが、前後いずれかのフレームの安定検出部に位置するので、係るフレームに対するＦＦＴ処理により所望のスペクトル量が生成される。
【００５１】
つまり、本形態によれば、たとえ短時間，瞬間的な単発信号がどこで発生しても、いずれかのフレームのＦＦＴ結果により検出でき、しかも、安定検出部のどの位置であっても生成されるスペクトル量でのばらつきはほとんどないので、安定したスペクトルを得ることができる。すなわち、検出対象となる単発信号の度合いを特定周波数帯のスペクトルとして安定して獲得することができる。
【００５２】
次に、ＦＦＴ処理をして得られたフレームｉについての周波数スペクトルデータから、周波数特徴量を演算する（ＳＴ６）。すなわち、抽出する特徴量としては、各種のものが用いることができるが、例えば、一定の周波数範囲単位毎のスペクトル量（平均，最大値など）を求めることができる。
【００５３】
そして、上記した処理対象のフレームｉに対するＦＦＴ処理並びにそれに基づく周波数特徴量演算を、フレーム１〜ｎまで順に実行する（ＳＴ７，８）。これにより、例えば特徴量を求める周波数単位を１０Ｈｚとすると、図１３に示すような結果が得られ、このデータがコンピュータプログラムメモリ上の２次元配列として保持される。
【００５４】
さらに、抽出する特徴量であるが、必ずしも全ての周波数帯に対して行う必要はなく、通常は検出対象の周波数帯域が分かっている場合が多いので、下限と上限で規定される所定の周波数範囲内について行う。そして、本実施の形態では、異常判定を行うものであるので、異常音の周波数１５ｋＨｚ程度を含む所定範囲（例えば、１４ｋＨｚから１６ｋＨｚや、１０ｋＨｚから２０ｋＨｚの範囲内など）について実行する。
【００５５】
上記した各処理を実行し、全てのフレームに対してＦＦＴ処理並びに周波数特徴量が求められたならば（ＳＴ７でＹｅｓ）、次に代表特徴量を求める（ＳＴ９）。つまり、上記周波数特徴量の中から、次工程での判定に使用する代表特徴量を抽出する。
【００５６】
具体的には、例えば、各周波数単位毎にｎ個のフレームの平均値を求めたり、ピーク値（スペクトル量の大きい物からｘ個分のデータの平均値）や、ボトム値（スペクトル量の小さい物からｙ個分のデータの平均値）や、変化量（前データとの差の絶対値を求め、大きい物からｚ個分のデータの平均値）などがある。もちろん、これに限るものではなく、他の数学・統計処理を使用して代表特徴量を求めても良い。
【００５７】
次いで、上記のようにして求めた代表特徴量に基づき異常の有無（良品／不良品）の判定を行う（ＳＴ１０）。この判定は、例えば代表特徴量を入力条件としてファジィ推論を実行することにより求めることができる。この判定結果は、良品／不良品の２種類を弁別するものでも良いし、判別がつかない（グレイ）という結果を含めた３種類とすることもできる。さらには、不良品（異常）の場合には、その種類まで特定する機能を備えても良い。
【００５８】
具体的な、ルールやメンバシップ関数などの記載は省略するが、例えば、特開平１１−１７３９０９号公報や、特開平１１−１７３９５６号公報並びに特開２００１−９１４１４号公報などに開示された技術を利用できる。
【００５９】
また、本発明は、短時間，単発に発生する信号でもＦＦＴ処理により検出可能とするものであるので、ＦＦＴ処理を中心に説明したが、上記した公報にも開示されているように、判定部における判定処理（ファジィ推論）は、ＦＦＴ処理にて生成されたスペクトルに基づく特徴量以外のものを用いてももちろん良い。
【００６０】
一例をあげると、データメモリ３に格納されている時間軸信号波形データを読み出し、ＦＦＴと同様にフレーム処理をし、所定の時間軸のデータを取得する。そして、前処理として、バンドパスフィルタを通過させて所望周波数（異常判定処理の場合には、例えば１０ｋＨｚから２０ｋＨｚ）の信号成分を抽出する。次いで、ピーク値演算や、ＲＭＳ演算などをし、フレーム毎の特徴量を求め、更に、全てのフレームに対する特徴量が求められたならば、そこから代表特徴量を求め、それをファジィ推論部に与える。
【００６１】
次に、本発明の効果を実証する実験結果について説明する。まず、極短時間に発生する異音のＦＦＴ演算結果による検証を行う。図１４に示すように検出対象となる短時間の異音（単発）を含む信号波形に対し、ＦＦＴ処理を行った。そして、フレームサイズは０．１０２４秒に対し、その中央部の約１／２の範囲（図中両方向矢印で示す領域）のデータに基づいて演算を行った（０．０５秒間を１１分割）。さらに、処理対象の信号波形は、図外の領域では、異音は発生せず、図示する振幅の小さい波形が続いている。
【００６２】
この状態において、ＦＦＴ処理するフレームの開始点をずらしていき、それぞれに対してハミング窓，ハニング窓，Ｓｔａｂｌｅ Ｗｉｎｄｏｗによる窓処理をしたものと、窓なしでＦＦＴ処理を行い、生成されたスペクトル量（最大値）を求めた。すなわち、処理対象の分析フレームに対する異音波形の位置ずれによる周波数成分（パワースペクトル：リニア）の影響を調べた。具体的には、図１４に示す状態（横軸フルスケール：０．１秒）を起点とし、そこから１０ｍ秒ずつずらした結果（中央の安定検出部となる１／２部分を１０分割した位置にずらした結果）、図１５に示すような実験結果が得られた。
【００６３】
この図から明らかなように、ハニング窓やハミング窓では、信号の発生位置のずれによりスペクトル量が大きく変化しているのに比べ、本発明のＳｔａｂｌｅＷｉｎｄｏｗを用いた窓処理を実行後のＦＦＴ結果では、安定したスペクトルが得られている。また、窓処理なしのものに比較すると、大きなスペクトル量が得られ、容易に検出できることが分かる。
【００６４】
次に、ＳＩＮ波形をＦＦＴ演算した結果でのＳｔａｂｌｅ Ｗｉｎｄｏｗの検証を行う。つまり、連続的に発生している信号をより安定して演算できるかを検証した。具体的には、５ｋＨｚのＳＩＮ波形を用い、上記と同様にハミング窓，ハニング窓，Ｓｔａｂｌｅ Ｗｉｎｄｏｗによる窓処理並びに窓処理なしのものに対し、フレームの開始位置をずらしながらＦＦＴ処理を実行した。
【００６５】
ここでは、図１５に示すＳＩＮ波形（５ｋＨｚ）のタイムレコード（フレーム）の開始位置を微妙にずらしたときの影響を検証するため、ずらす時間の単位は非常に短い時間に設定している。その結果、図１６に示すようになった。そして、偏差を見ると、図１７に示すような結果が得られた。また、ＳＩＮ波形に対するフレームの開始位置の位相（角度）ずれに対するパワースペクトル（リニア）の影響を示すと、図１８のようになる。
【００６６】
これらの図から明らかなように、連続周期的に発生しているＳＩＮ波形に関してみても、窓なしでは演算結果の偏差が大きく、漏えいによる影響があると判断できるが、本発明のＳｔａｂｌｅ Ｗｉｎｄｏｗでは、偏差が小さく、従来の窓関数に比べてもその精度は問題無いものと言える。
【００６７】
この検証結果から、本発明による処理を用いれば、極小時間のみ発生するような非周期的な信号と、ＳＩＮ波形のように周期的に発生する信号のいずれの場合も安定して検出できると言える。
【００６８】
次に、時間軸波形のシフト量をフレームサイズの１／２にしたことの効果の検証を行う。異常音のない信号波形と、約１秒間に一瞬だけ異常音成分（１６ｋＨｚ）が含まれている波形に対し、従来と同様に均等にフレームサイズに分割してＦＦＴを行ったところ、図１９に示す結果が得られた。これに対し、フレームサイズの１／２ずつずらしながら取得したフレームに対し、ＦＦＴを行ったところ、図２０に示す結果が得られた。各グラフは、異常音の成分である１６ｋＨｚのパワースペクトル量を表したものである。
【００６９】
図１９から明らかなように、従来の方法では、上記異常音成分が、４番目と５番目のフレームの境界部分で発生し、それぞれのフレームの両端のみに出現している。従って、異常があってもスペクトル量は小さく、検出するのは困難である。これに対し、図２０から明らかなように、１／２ずつずらした方法によれば、確実に異常成分を検出することができる。つまり、従来はフレームの両端部分にのみ現れるような成分はいかに大きな振幅成分であっても、十分な検出をすることができなかったが、本考案による処理により、より安定して検出することができると言える。
【００７０】
上記した実施の形態では、窓関数の安全検出部を１／２にし、シフト量をフレームサイズの１／２としたが、本発明はこれに限ることは無く、フレームサイズの半分以上でも良いし、半分以下でもよい。
【００７１】
つまり、シフト量をＰとすると、各フレームを構成するデータは、図２１に示すように、定義することができる。これにより、対象データとして分割する位置を任意に重複して取得することができる。
【００７２】
そして、シフト量を半分以上にする場合は、分析対象フレーム数が少なくなるのでＣＰＵによる合計処理時間は短くなるという利点を持つ。但し、安定検出部を１／２のままにすると、安定検出部に入らない時間帯が発生するため、対象データ中一瞬のみ発生する信号を安定検出することが困難になる。この場合でも、前後のフレームが一部重複する（フレームサイズよりもシフト量を短くする）ようにすれば、安定検出部に入らない時間帯の割合が少なくなるので、従来のものよりも性能は良い。
【００７３】
但し、好ましくは、窓関数の減衰量（安定検出部）の割合を適宜調整することである。一例を示すと、例えば、安定検出部：減衰部＝３：１、つまり、安定検出部がフレームサイズの３／４で、フレームサイズをＮにした場合の上記した窓関数を規定する式は、以下のようになる。

そして、ＳＩＮ波形に上記窓関数をかけた波形の一例を示すと、図２２のようになる。
【００７４】
そして、前述したようにシフトパラメータによるシフト量をＮ／２より大きくとる場合は、図２３に示すように、シフト量＝安定検出部のサイズとすることで、安定した検出を可能にする。なお、シフト量を大きく（安定検出部のサイズを大きく）すると、ＦＦＴ処理回数が削減されるので、全体の処理時間が短くなるので好ましいが、一定以上シフト量を大きくすると、それに伴ない減衰部が小さくなるので、窓無しの状態に近くなり、本来の窓関数の目的が損なわれてしまう。また、シフト量を１／２以下にすると、処理時間が長くなりすぎる。従って、Ｎ／２＜シフト量＜３Ｎ／４の範囲が実用的となる。
【００７５】
図２４は、上記した装置を実際のシステムに適用した一例を示している。同図に示すように、筐体１０内に図１に示すパソコン５（データメモリ３，ＣＰＵ４）が内蔵され、筐体１０の前面には、パソコン５のモニタ画面１１が位置される。また、筐体１０の下方には、ＰＬＣ１２が内蔵され、ＦＡ等の生産工場内に設置された装置制御盤１３と接続され、通信するようになっている。ＰＬＣ１２とパソコン５も接続されており、情報の送受ができるようになっている。つまり、ＰＬＣ１２は、生産工場内でのシステムの稼働状況が分かるので、例えば、測定対象の製品１５が所定位置に搬送されてきたことをパソコン５に通知し、パソコン５は係る通知をトリガとしてデータの取り込みや、ＦＦＴ処理等を行うことができる。
【００７６】
また、パソコン５には、マイクロフォン１６や加速度ピックアップ１７が接続されており、測定対象の製品１５から発生する音や振動を検出し、アナログ信号情報としてパソコン５に与える。
【００７７】
そして、パソコン５では、取得した波形信号に基づき、上記した処理を実行することにより良品／不良品の判定を行い、その結果を表示装置であるＰＴ１９に出力表示する。図に示すように、不良品の場合には、その原因まで表示するようにしている。
【００７８】
【発明の効果】
以上のように、この発明では、記憶部に格納された波形データに対してフレーム分割をしてフレームを生成するに際し、前後のフレームの一部をオーバーラップするようにし、さらに、ＦＦＴ処理をするに先立ち使用する窓関数を、両サイドに減衰部を設けるとともに、中央部分をフラットにしたため、短時間や不連続に発生する信号成分，過渡特性をＦＦＴアルゴリズムを応用した周波数解析装置でもって確実に検出することができる。よって、官能検査による異常判定を自動的に行うことができる。
【図面の簡単な説明】
【図１】従来のＦＦＴ処理をする際に実行するフレーム分割の一例を示す図である。
【図２】窓を用いないでＦＦＴ処理をした場合の一例を示す図である。
【図３】窓を用いないでＦＦＴ処理をして得られたパワースペクトルをデシベル表示した例を示す図である。
【図４】窓関数を用いてＦＦＴ処理をした場合の一例を示す図である。
【図５】窓関数を用いてＦＦＴ処理をして得られたパワースペクトルをデシベル表示した例を示す図である。
【図６】単発の異常音が発生した場合の波形信号に対し、窓関数を用いてＦＦＴ処理をした場合の一例を示す図である。
【図７】単発の異常音が発生した場合の波形信号に対し、窓関数を用いてＦＦＴ処理をした場合の一例を示す図である。
【図８】単発信号の発生タイミングのずれによるスペクトルの影響を示す実験結果である。
【図９】本発明の好適な一実施の形態を示すブロック図である。
【図１０】ＣＰＵの機能を説明するフローチャートである。
【図１１】フレーム分割処理を説明する図である。
【図１２】窓関数を説明する図である。
【図１３】各フレーム毎の周波数特徴量を求めた結果の一例を示す図である。
【図１４】本発明の効果を実証するために用いたサンプル波形信号の一例を示す図である。
【図１５】単発信号（短時間異音）の測定時間ずれに対するスペクトルの影響を示す実験結果である。
【図１６】ＳＩＮ波形の開始位置ずれによるパワースペクトルの影響を示す実験結果である。
【図１７】ＳＩＮ波形の開始位置ずれによるパワースペクトルの影響（偏差）を示す実験結果である。
【図１８】ＳＩＮ波形の位相ずれ（角度）によるパワースペクトルの影響を示す実験結果である。
【図１９】正常な音のみからなるサンプル波形と、瞬間的に発生した異常音を含むサンプル波形に対し従来方法によりスペクトルを求めた結果を示す図である。
【図２０】正常な音のみからなるサンプル波形と、瞬間的に発生した異常音を含むサンプル波形に対し、フレームを１／２シフトしながらスペクトルを求めた結果を示す図である。
【図２１】シフト量Ｐに対するフレームを構成するデータの関係を示す図である。
【図２２】ＳＩＮ波形に別の窓関数を掛けて得られた波形例を示す図である。
【図２３】シフト量と安定検出部（安定部）の適切な関係を示す図である。
【図２４】本発明に係る異常判定システムの好適な一実施の形態を示す図である。
【符号の説明】
１ アナログフィルタ
２ Ａ／Ｄボード
３ データメモリ
４ ＣＰＵ
５ パソコン
１０ 筐体
１１ モニタ画面
１２ ＰＬＣ
１３ 装置制御盤
１５ 製品
１６ マイクロフォン
１７ 加速度ピックアップ
１９ ＰＴ[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a frequency analysis apparatus, an abnormality determination apparatus, and an abnormality determination system that apply an FFT algorithm that measures an analog signal from a sensor such as vibration and sound and monitors a frequency component by an FFT (Fast Fourier Transform) algorithm. .
[0002]
[Prior art]
In automobiles and home appliances, a rotating device incorporating a motor is very often used. For example, taking an automobile as an example, rotating devices are mounted everywhere, such as power steering, power seats, and the like. When the rotating device is actually operated, a sound is generated with the rotation of the motor or the like.
[0003]
Some of these sounds are inevitably generated along with normal operations, and other sounds are generated due to defects. Examples of abnormal sounds associated with the failure include bearing abnormalities, internal abnormal contact, imbalance, and foreign matter contamination. More specifically, there are gear chipping, foreign object biting, spot flaws, and abnormal noise such that the rotating part and the fixed part inside the motor rub for a moment during rotation. Further, as sounds that people feel uncomfortable, for example, there are various sounds from 20 Hz to 20 kHz that humans can hear, for example, about 15 kHz. And when the sound of the predetermined frequency component is generated, it becomes an abnormal sound. Of course, abnormal sounds are not limited to this frequency.
[0004]
The sound associated with such defects is not only unpleasant, but may cause further malfunctions.In production factories, sensory inspections are usually conducted by inspectors to determine the presence or absence of abnormal sounds and to ensure quality assurance. Is going. Specifically, it is done by listening with the ear or touching it with the hand to check the vibration. The sensory test is defined by the sensory test term JIS Z8144.
[0005]
By the way, sensory tests that rely on the five senses of such inspectors not only require skilled skills, but also the results of judgments vary widely, such as individual differences and changes due to time. There was a problem that it was difficult and difficult to manage. Therefore, in order to solve such a problem, the present inventor measured vibration and sound of a product driving unit with a sensor as an apparatus for the purpose of automating a “sensory test” process, and applied an FFT algorithm to the analog signal. A method for inspecting and inspecting frequency components using a frequency analyzer has been developed (Japanese Patent Laid-Open Nos. 11-173906, 11-173909, 2001-91414).
[0006]
Briefly, a frequency analysis apparatus to which the FFT algorithm is applied can analyze a frequency domain of a time domain signal by a fast Fourier transform algorithm. On the other hand, the frequency range of abnormal sounds is also determined to some extent. Accordingly, since the component corresponding to the abnormal sound generation region can be extracted from the frequency components extracted by the analysis, the feature amount of the extracted component is obtained. Then, the presence / absence of the abnormality and the cause thereof are estimated from the feature amount using fuzzy inference.
[0007]
Furthermore, production line equipment and devices in the FA industry and the process industry are becoming faster and more complex. On the other hand, if an abnormality or failure occurs in this equipment / device, production and quality will be greatly affected. As one of the techniques of simple diagnosis and precise diagnosis for preventing this, frequency analysis using a frequency analysis apparatus to which the above-described FFT algorithm is applied has become an essential technique.
[0008]
[Problems to be solved by the invention]
According to the above-described invention, the basic configuration of an apparatus for automatically determining an abnormality has been made. And it was possible to perform some automatic judgment and data management based on it. However, when abnormal sound based on a defect is generated momentarily or only for a relatively short time, it is difficult to accurately recognize and determine that the abnormal sound is detected. .
[0009]
This is due to the following reason. That is, the frequency analysis apparatus and the fast Fourier transform algorithm applying the FFT algorithm perform the analysis based on the premise that “the input signal is a composite wave of a plurality of SIN / COS waveforms that are the same regardless of time”. On the other hand, “abnormal sounds that people feel unpleasant” as a target in the “sensory test” include discontinuous sounds and intermittent sounds that occur only in a very short time period. Therefore, the above assumption is not applicable, and accurate determination cannot be made.
[0010]
In addition, due to the development of mechatronics, devices that have become faster and more complex are not suitable for the above-mentioned assumptions because they are driven only for a very short time or the internal medium moves. Results are not obtained.
[0011]
To explain with a specific example, the FFT algorithm first reads a time signal stored in the data memory and performs time record processing. That is, in the FFT, processing is performed for each data block (time record) of N points of 2 n to generate a frequency spectrum of “(N / 2) +1” points.
[0012]
Then, as shown in FIG. 1, this time record is set so that adjacent time records are continuous. Actually, N points of data are extracted from a predetermined position of the read data in accordance with a preset sampling time, and set as a data group (frame 1) of the first time record. Similarly, data for N points from the (N + 1) th is extracted and set as a data group (frame 2) of the second time record.
[0013]
As described above, the FFT algorithm is based on the premise that the same data is continuously infinite before and after the cut time record. However, in reality, the phase cannot be set to 0 (amplitude is 0) at the boundary part (first and Nth data) of all time records. If connected, discontinuities will occur.
[0014]
Therefore, if the FFT processing is performed on the SIN waveform shown in FIG. 1 as it is, a leakage phenomenon occurs as shown in FIGS. That is, the waveform displayed in FIG. 2A is a time axis waveform to be processed, and is the same as that in FIG. 1 in this example. The waveform displayed in FIG. 2B is a power spectrum by FFT (the vertical axis is linear display). Originally, since it is a single SIN waveform, a spectrum is generated only for the frequency component of the SIN waveform. As is apparent from FIG. 3 in which the power spectrum is displayed in decibels, other frequency components are also generated. End up.
[0015]
Therefore, usually, window processing is performed on the signal waveform to be processed. That is, only the central part of the time record is processed by multiplying the signal waveform by a mountain-shaped function that makes both ends of the time record zero. As window function types, various window functions such as Hanning window, Hamming window, flat top window, and triangular window have been developed. As an example, the Hanning window is a function as shown in FIG. Since each window is multiplied by a small coefficient as it goes to both ends, the signal waveform of the central portion is conversely reversed in order to prevent the output in one time record from being lowered as it is. It is trying to amplify against. In each window function, the method of cutting the data at both ends differs from the amount of amplification at the center, but is common in that the center is amplified.
[0016]
When FFT is performed on the SIN waveform shown in FIG. 1 using a Hanning window, the time axis waveform multiplied by the window function is as shown in FIG. 4A, and the power spectrum of the waveform is linearly displayed. As shown in FIG. Then, when this power spectrum is displayed in decibels, it is as shown in FIG. As is apparent from FIG. 5, it can be confirmed that the side component is almost 0 dB or less and a near-linear spectrum is generated.
[0017]
Further, after generating the power spectrum, averaging processing (averaging) is performed. That is, after obtaining frequency axis arrangements for a plurality of time records in this way, a frequency analysis apparatus to which the FFT algorithm is applied performs statistical processing by averaging processing to obtain a stable result. Note that there are various types of averaging, such as RMS averaging and linear averaging (vector averaging). In any case, frequency spectrum data calculated from a plurality of time records can be obtained by mathematical processing. Averaging is performed to reduce non-reproducible signal components and noise.
[0018]
As described above, the frequency analysis apparatus using the FFT algorithm is devised so that a more stable result can be obtained by window processing or averaging processing. This is based on the premise that the target data is a continuous waveform in which SIN waves and COS waves are combined.
[0019]
By the way, when the signal component to be detected is a signal that occurs only for a moment such as “Kit” (a signal that lasts only a fraction of the time width of the time record) for the human ear. The detection accuracy is poor.
[0020]
That is, as an example of the signal that is generated only for a moment to be detected, there is an abnormal sound (single signal) when the automobile motor is driven as shown in FIG. The example shown in the figure is a time axis waveform (FIG. 6A) obtained by actually measuring a sound that a person feels uncomfortable with a microphone and a power spectrum by FFT (FIG. 6B). The main frequency of occurrence of the abnormal part A is about 15 kHz, and a person's ear feels a sound like a “ki!” Metal rubbed violently for a moment with discomfort.
[0021]
As can be seen from the figure, if one time record is about 0.1 seconds, the duration of abnormal sound generation is about 15 milliseconds. As described above, when the occurrence time of the abnormal portion with respect to the time record is much shorter, as shown in FIG. 6A, when the abnormal portion A is located at the center of the time record to be amplified, FIG. As shown in b), the power spectrum is also generated in the vicinity of 15 kHz and can be detected. However, as shown in FIG. 7A, when the occurrence position of the abnormal part is located at both ends of the time record, it is close to 0 by the window process, so the spectrum of the frequency to be detected (around 15 kHz). Disappears. (See Fig. 7 (b))
[0022]
Therefore, in the case as shown in FIG. 7, it is determined that there is no abnormality in the result of the FFT process, although the abnormal sound is originally generated. Alternatively, since a spectrum is generated slightly in the vicinity of 15 kHz, if the abnormality can be reliably detected by reducing the threshold value for determining the abnormality, it is determined that the normal one is abnormal. Therefore, it will be disposed of in vain. Therefore, not only the cost is increased but also waste of resources is caused, which is not preferable.
[0023]
Furthermore, even if the abnormal part does not disappear, the spectrum amount as a result of the FFT varies depending on the position where the abnormal part occurs in the time record. This is because, for a Hamming window or the like, the gain increases as it goes to the center as shown in FIG. Accordingly, when the occurrence position of the abnormal portion is at the center position of the time record, the spectrum amount becomes the largest, and the spectrum amount decreases as the distance from the center (approaching the end portion) increases. An example is shown in FIG. FIG. 8 shows a window process using a Hanning window when the start position of the time record is shifted by 10 msec for the same signal waveform (single abnormal sound is generated). Spectral quantities were obtained for those with and without window processing.
[0024]
As shown in this figure, even if an attempt is made to obtain the frequency spectrum of the same signal component using a frequency analysis device that applies the FFT algorithm, the spectrum is shifted from the generation timing of the same abnormal sound component due to the influence of the window function. It will change greatly depending on. That is, the size of the abnormal sound cannot be captured by the specific frequency spectrum amount of the FFT calculation result. This means that a frequency analysis device using a conventional FFT algorithm cannot stably detect a sound generated for a short time.
[0025]
As described above, the frequency analysis device applying the FFT algorithm is widely used for analyzing wave signals such as vibration, sound, heat conduction, etc. without being limited to abnormality determination. This is not appropriate for capturing transient characteristics.
[0026]
The present invention can reliably detect signal components and transient characteristics that occur in a short time or discontinuously with a frequency analysis apparatus applying an FFT algorithm, and can automatically perform abnormality determination by sensory inspection. An object of the present invention is to provide a frequency analysis device, an abnormality determination device, and an abnormality determination system to which an algorithm is applied.
[0027]
[Means for Solving the Problems]
The frequency analysis apparatus applying the FFT algorithm according to the present invention creates a frame by storing storage means for storing given waveform data and extracting data for a desired frame size from the waveform data stored in the storage means Frame generating means for performing the FFT processing on the basis of the data corrected by applying the window function to the frame data constituting the frame. The frame creation means has a function of extracting data so that a part of the preceding and succeeding frames overlaps on the time axis, and the window function includes an attenuation unit such that the amplitudes at both ends of the frame become zero. While having on both sides, the region of the center stability detecting portion is configured to be flat. The storage means corresponds to the data memory 3 in the embodiment. The frame creation means and the FFT processing means are realized by the CPU 4 in the embodiment. The former corresponds to the function of executing step 3, and the latter corresponds to the function of executing step 4.
[0028]
Since the attenuation parts are provided on both sides of the frame, the occurrence of spectrum leakage due to discontinuity when the frame data cut out by frame division is connected back and forth is suppressed, as with various conventional window functions. The
[0029]
Further, since the region of the stability detection unit is flattened, the amount of spectrum generated in the frame can take a constant value no matter where the signal of the stability detection unit appears for a short time.
[0030]
Furthermore, when a short-time signal appears in the attenuation part of a certain frame, if the FFT process is performed on that frame, the spectrum amount is attenuated and may not be detected. However, even in such a case, since a part of the preceding and succeeding frames is overlapped, it is highly likely that the frame is located in the stability detection unit of another adjacent frame, and can be detected by FFT processing based on the other frame.
[0031]
Then, the amount of overlap is ½ of the frame size, the region of the stability detection unit is ½ of the frame size, and the attenuation unit is ¼ of the frame size on both sides, respectively. Efficient and stable detection can be performed. Of course, the present invention is not limited to this, and the overlapping amount may be larger or smaller than ½ of the frame size. Similarly, the size of the stability detector may be larger or smaller than ½ of the frame size. Furthermore, it is preferable that the stability detectors of the front and rear frames do not overlap with each other, and that the attenuation unit is positioned at the adjacent stability detector, since this is efficient.
[0032]
That is, when the shift amount is set so as to satisfy such a condition (the stability detection unit is always located in the stability detection unit), all moments are put into the stabilization unit by the window (window function) of the present invention. Can do. In other words, it is because all the instants of the waveform have the merit that they can enter the stability detection unit of any frame. Moreover, since the stability detectors do not overlap, there is no waste. That is, an increase in the number of frames to be subjected to FFT processing is suppressed, and processing time is required. Of course, depending on the required use, the present invention is included even in a setting that does not satisfy the above-described conditions.
[0033]
An abnormality determination apparatus according to the present invention is an abnormality determination apparatus that determines an abnormality of a product based on measurement data measured by a measuring unit, and applies a frequency analysis apparatus to which the above FFT algorithm is applied and the FFT algorithm. A feature amount extraction unit that extracts a feature amount based on frequency spectrum information of each frame obtained by the frequency analysis device, and a determination unit that performs the abnormality determination based on the feature amount extracted by the feature amount extraction unit. Configured. The storage means stores the measurement data.
[0034]
Here, the measuring means can be realized by a microphone, an acceleration sensor (pickup), or other sensors. In addition, in the embodiment, the feature quantity extraction unit may be understood to include a function for obtaining the feature quantity of each frame in step 5 and a function for obtaining the representative feature quantity in step 7. In this case, the feature amount is a concept including two types of frequency feature amount and representative feature amount. Further, the function for obtaining the representative feature value may be included in the determination means side. Further, the concept of the representative feature value may be eliminated, and the determination process may be performed based on the feature value obtained for each frame.
[0035]
Furthermore, the abnormality determination system of the present invention includes the above-described abnormality determination device and a PLC, and the PLC can be linked to a production system that produces the product, and the abnormality determination device receives a signal from the PLC. Made to work based.
[0036]
According to this invention, the frequency analysis device applying the FFT algorithm can reliably detect even a short-time single-shot signal or a discontinuous signal, and the required spectral amount. Therefore, it is possible to reliably perform the abnormality determination based on the spectrum amount subjected to the FFT processing. Of course, it is not impeded that the actual abnormality determination is performed in consideration of the feature amount other than the spectrum amount.
[0037]
The above-described components of the present invention can be combined as much as possible. Each means constituting the apparatus according to the present invention can be realized by a dedicated hardware circuit, or can be realized by a programmed computer.
[0038]
DETAILED DESCRIPTION OF THE INVENTION
FIG. 9 shows a preferred embodiment of the present invention. As shown in the figure, a waveform signal (analog) to be processed is given to an A / D board 2 through an analog filter 1, converted into a digital signal therein, and stored in a data memory 3. Then, the CPU 4 reads out desired data stored in the data memory 3 and performs FFT processing, and performs abnormality diagnosis / determination from the result.
[0039]
As the data memory 3 and the CPU 4, those built in the main body of the personal computer 5 are used. The A / D board 2 is mounted as an expansion board for the personal computer 5. Further, the waveform signal to be processed may be an output of a vibration sensor or a microphone, current, power, or the like. The analog filter 1 removes noise and other unnecessary signal components from these input waveform signals (analog), and may be mounted separately on the personal computer 5 or on the sensor side described above. Also good.
[0040]
The function of the CPU 4 in the present invention is to execute the flowchart shown in FIG. That is, since the waveform signal converted into digital data by the A / D board 2 is stored in the data memory 3, the data memory 3 is accessed and data is extracted at a desired sampling period. Actually, by setting the sampling period in the A / D board 2 to an appropriate value, the data is already sampled at a desired period when data is stored in the data memory 3. In this case, the process of step 1 is not performed by the CPU 4.
[0041]
Next, data to be analyzed is extracted from the data stored in the data memory 3 (ST2). That is, as shown in FIG. 11A, a series of waveform data (digital) is read and stored in the buffer (start address 0) in order from the start data.
[0042]
Then, the extracted analysis target data is divided for each data of time records to be processed by FFT, that is, for each N points of data (hereinafter referred to as “frame”) (ST3). As a result, n frames are generated. At this time, as shown in FIG. 11B, assuming that the frame size is N (for example, 1024), the shift parameter is N / 2 (for example, 512). In other words, conventionally, the shift parameter is N, and the i-th frame and the (i + 1) -th frame are continuous without overlapping, but in this embodiment, the data of the previous and subsequent frames are taken so that the halves overlap. ing.
[0043]
Note that the amount of overlap (shift parameter) is N / 2 in the present embodiment so that half overlap, but the present invention is not limited to this, and may be half or more of the frame size, or half. The following is also acceptable.
[0044]
For each frame constituted by extracting data as described above, the FFT processing is performed on the i-th frame in order from i = 1 (ST4, 5). This FFT processing is basically executed by an algorithm generally performed conventionally, but the window function used in the window processing performed on the data of the frame to be processed is changed.
[0045]
That is, it is the same as in the prior art in that it is gradually attenuated so that the amplitude at both ends of the frame becomes 0, but the central portion is flat (in the conventional case, the amplitude increases toward the center even in the central portion). Window function (referred to as “Stable Window function”).
[0046]
Specifically, a quarter region on both ends of the frame is an attenuation unit, and a central half region is a stability detection unit. The left-side attenuation portion has a phase 0 to 90 ° COS waveform, and the right-side attenuation portion has a phase 90 ° to 180 ° COS waveform. In the region of the stability detection unit, flattening is as described above, but further, amplification is not performed (a coefficient to be applied to a signal to be processed is 1). That is, among the data constituting the frame, the data located in the stability detection unit uses the extracted data value as it is. In other words, it can be viewed as if there is no window.
[0047]
Then, the time record is weighted by the function of the following expression by multiplying the data constituting the processing target frame by the stable window function defined as described above.
When the number of target data is N, the following three equations are combined.

Further, when the sine waveform is multiplied by the stable window function, the result is as shown in FIG. A power spectrum for each frequency component can be generated by performing FFT processing on the signal thus multiplied by the window function. As the FFT algorithm, an existing fast Fourier algorithm can be used.
[0048]
In this embodiment, by providing attenuation portions at both ends, a spectrum leakage phenomenon due to discontinuity when frame data cut out by frame division is connected back and forth is prevented.
[0049]
As a result, the spectrum amount for each frequency of the FFT does not change due to the window function that accompanies the shift in the generation timing of the single signal. In other words, as long as it is within the range of the stability detection unit, the amount of spectrum obtained by the FFT processing takes a constant value regardless of where a single signal is instantaneously generated.
[0050]
Furthermore, in this embodiment, since the preceding and succeeding frames overlap the data (time axis region) corresponding to ½ of the frame size, the attenuation units at both ends of the frame i to be processed are the frame (i−1) or the frame It overlaps with the area of the stability detector (i + 1). Therefore, when a single signal is generated in the region of the attenuation part of the frame i to be processed, if FFT processing based on the frame i is performed, the signal is attenuated and the amount of spectrum becomes small, and there is a high possibility of disappearance. Since it is located in the stability detection part of either the previous or next frame, a desired spectral amount is generated by the FFT process for the frame.
[0051]
That is, according to the present embodiment, wherever a short-time, instantaneous single-shot signal occurs, it can be detected from the FFT result of any frame, and is generated at any position of the stability detection unit. Since there is almost no variation in the amount of spectrum, a stable spectrum can be obtained. That is, it is possible to stably acquire the degree of a single signal to be detected as a spectrum in a specific frequency band.
[0052]
Next, the frequency feature amount is calculated from the frequency spectrum data for the frame i obtained by the FFT process (ST6). That is, various types of feature quantities can be used, and for example, a spectrum quantity (average, maximum value, etc.) for each fixed frequency range unit can be obtained.
[0053]
Then, the FFT process for the frame i to be processed and the frequency feature amount calculation based on the FFT process are sequentially executed from the frames 1 to n (ST7, 8). Thus, for example, if the frequency unit for obtaining the feature value is 10 Hz, the result shown in FIG. 13 is obtained, and this data is held as a two-dimensional array on the computer program memory.
[0054]
Furthermore, although it is a feature quantity to be extracted, it is not always necessary to perform it for all frequency bands, and usually the frequency band to be detected is often known, so a predetermined frequency range defined by the lower and upper limits Do the inside. In this embodiment, since abnormality determination is performed, a predetermined range including an abnormal sound frequency of about 15 kHz (for example, within a range of 14 kHz to 16 kHz, 10 kHz to 20 kHz, etc.) is executed.
[0055]
If the above-described processes are executed and FFT processing and frequency feature amounts are obtained for all the frames (Yes in ST7), then representative feature amounts are obtained (ST9). That is, representative feature values used for determination in the next process are extracted from the frequency feature values.
[0056]
Specifically, for example, an average value of n frames is obtained for each frequency unit, a peak value (an average value of data for x pieces from a thing with a large spectral amount), or a bottom value (a small spectral amount). The average value of y data from an object) and the amount of change (the absolute value of the difference from the previous data is obtained, and the average value of z data from a large object). Of course, the present invention is not limited to this, and the representative feature amount may be obtained using other mathematical / statistical processing.
[0057]
Next, the presence / absence of an abnormality (non-defective product / defective product) is determined based on the representative feature value obtained as described above (ST10). This determination can be obtained, for example, by executing fuzzy inference using a representative feature amount as an input condition. The determination result may be a discrimination between two types of non-defective product / defective product, or may be three types including a result of being indistinguishable (gray). Furthermore, in the case of a defective product (abnormal), a function for specifying the type may be provided.
[0058]
Although specific descriptions of rules and membership functions are omitted, for example, the techniques disclosed in Japanese Patent Application Laid-Open No. 11-173909, Japanese Patent Application Laid-Open No. 11-173906, and Japanese Patent Application Laid-Open No. 2001-91414 are disclosed. Available.
[0059]
In addition, since the present invention can detect even a single occurrence of a signal by FFT processing for a short time, the description has focused on the FFT processing. However, as disclosed in the above publication, the determination unit Of course, the determination processing (fuzzy inference) in may use a feature other than the feature amount based on the spectrum generated by the FFT processing.
[0060]
As an example, the time axis signal waveform data stored in the data memory 3 is read out, frame processing is performed in the same manner as FFT, and data on a predetermined time axis is acquired. Then, as preprocessing, a signal component of a desired frequency (for example, 10 kHz to 20 kHz in the case of abnormality determination processing) is extracted by passing through a band pass filter. Next, peak value calculation, RMS calculation, etc. are performed to determine the feature amount for each frame. If the feature amounts for all the frames are determined, the representative feature amount is determined therefrom, and this is used as the fuzzy inference unit. give.
[0061]
Next, experimental results that demonstrate the effects of the present invention will be described. First, verification is performed based on the FFT calculation result of abnormal noise generated in an extremely short time. As shown in FIG. 14, FFT processing was performed on a signal waveform including a short time abnormal sound (single sound) to be detected. Then, for the frame size of 0.1024 seconds, calculation was performed based on data in the range of about 1/2 of the central portion (the region indicated by the double-headed arrow in the figure) (11 divisions for 0.05 seconds). Further, the signal waveform to be processed does not generate abnormal noise in the region outside the figure, and the waveform having a small amplitude shown in the figure continues.
[0062]
In this state, the start point of the frame to be subjected to FFT processing is shifted, and each of them is subjected to window processing by a Hamming window, Hanning window, and stable window, and FFT processing without a window, and the generated spectrum amount ( The maximum value was determined. That is, the influence of the frequency component (power spectrum: linear) due to the displacement of the abnormal sound waveform on the analysis frame to be processed was examined. Specifically, the result shown in FIG. 14 (horizontal axis full scale: 0.1 second) as a starting point, and the result of shifting by 10 milliseconds from that (position where the half portion serving as the center stability detection unit is divided into 10 parts) The result of the experiment shown in FIG. 15 was obtained.
[0063]
As is clear from this figure, in the Hanning window and the Hamming window, the amount of spectrum changes greatly due to the deviation of the signal generation position, and in the FFT result after executing the window processing using the stable window of the present invention, A stable spectrum is obtained. In addition, it can be seen that a large amount of spectrum is obtained and can be easily detected as compared with the case without window processing.
[0064]
Next, the verification of the stable window based on the result of the FFT operation on the SIN waveform is performed. In other words, it was verified whether a continuously generated signal can be calculated more stably. Specifically, using a SIN waveform of 5 kHz, FFT processing was performed while shifting the start position of the frame with respect to the window processing by the Hamming window, the Hanning window, the stable window and the window processing without window processing in the same manner as described above.
[0065]
Here, in order to verify the influence when the start position of the time record (frame) of the SIN waveform (5 kHz) shown in FIG. 15 is slightly shifted, the unit of shifting time is set to a very short time. As a result, it became as shown in FIG. When the deviation was observed, the result as shown in FIG. 17 was obtained. FIG. 18 shows the influence of the power spectrum (linear) on the phase (angle) shift of the frame start position with respect to the SIN waveform.
[0066]
As is clear from these figures, even with regard to the SIN waveform that occurs continuously periodically, it can be determined that there is a large deviation of the calculation result without a window and there is an influence due to leakage, but in the stable window of the present invention, The deviation is small, and it can be said that the accuracy is not a problem compared with the conventional window function.
[0067]
From this verification result, it can be said that if the processing according to the present invention is used, it is possible to stably detect both a non-periodic signal generated only in a minimum time and a signal generated periodically such as a SIN waveform. .
[0068]
Next, the effect of setting the shift amount of the time axis waveform to ½ of the frame size is verified. When a signal waveform without abnormal sound and a waveform containing an abnormal sound component (16 kHz) for about one second for only one second are divided into equal frame sizes as in the prior art, FFT is performed. The results shown are obtained. On the other hand, when the FFT was performed on the frames acquired while shifting by half the frame size, the result shown in FIG. 20 was obtained. Each graph represents the amount of power spectrum at 16 kHz, which is a component of abnormal sound.
[0069]
As is apparent from FIG. 19, in the conventional method, the abnormal sound component is generated at the boundary between the fourth and fifth frames and appears only at both ends of each frame. Therefore, even if there is an abnormality, the amount of spectrum is small and it is difficult to detect. On the other hand, as is apparent from FIG. 20, according to the method shifted by ½, the abnormal component can be reliably detected. In other words, in the past, a component that appeared only at both ends of the frame could not be detected sufficiently, no matter how large the amplitude component, but it can be detected more stably by the process according to the present invention. I can say that.
[0070]
In the embodiment described above, the window function safety detection unit is halved and the shift amount is halved of the frame size. However, the present invention is not limited to this, and may be half or more of the frame size. It may be less than half.
[0071]
That is, if the shift amount is P, the data constituting each frame can be defined as shown in FIG. Thereby, the position divided | segmented as object data can be acquired arbitrarily overlappingly.
[0072]
When the shift amount is more than half, the total processing time by the CPU is shortened because the number of frames to be analyzed is reduced. However, if the stability detection unit is left at 1/2, a time zone that does not enter the stability detection unit occurs, so that it is difficult to stably detect a signal that occurs only for a moment in the target data. Even in this case, if the previous and subsequent frames are partially overlapped (the shift amount is shorter than the frame size), the ratio of the time zone that does not enter the stability detection unit decreases, so the performance is better than the conventional one. good.
[0073]
However, preferably, the ratio of the attenuation amount (stability detection unit) of the window function is appropriately adjusted. As an example, for example, the stability detection unit: attenuation unit = 3: 1, that is, the equation that defines the above window function when the stability detection unit is 3/4 of the frame size and the frame size is N is: It becomes as follows.

An example of a waveform obtained by multiplying the SIN waveform by the window function is as shown in FIG.
[0074]
As described above, when the shift amount based on the shift parameter is larger than N / 2, as shown in FIG. 23, the shift amount is equal to the size of the stability detection unit, thereby enabling stable detection. It should be noted that increasing the shift amount (increasing the size of the stability detecting unit) reduces the number of FFT processings, which is preferable because the entire processing time is shortened. However, if the shift amount is increased beyond a certain level, the attenuation unit is accordingly increased. Becomes smaller, it becomes close to the state without a window, and the purpose of the original window function is lost. Further, if the shift amount is ½ or less, the processing time becomes too long. Therefore, the range of N / 2 <shift amount <3N / 4 is practical.
[0075]
FIG. 24 shows an example in which the above-described apparatus is applied to an actual system. As shown in the figure, the personal computer 5 (data memory 3, CPU 4) shown in FIG. 1 is built in the housing 10, and the monitor screen 11 of the personal computer 5 is located on the front surface of the housing 10. Further, a PLC 12 is built below the housing 10 and is connected to and communicates with a device control panel 13 installed in a production factory such as an FA. The PLC 12 and the personal computer 5 are also connected so that information can be transmitted and received. That is, the PLC 12 knows the operating status of the system in the production factory. For example, the PLC 12 notifies the personal computer 5 that the product 15 to be measured has been transported to a predetermined position, and the personal computer 5 uses the notification as a trigger for data. Capture, FFT processing, and the like.
[0076]
Further, a microphone 16 and an acceleration pickup 17 are connected to the personal computer 5, and sound and vibration generated from the measurement target product 15 are detected and given to the personal computer 5 as analog signal information.
[0077]
Then, the personal computer 5 performs the above-described processing based on the acquired waveform signal, thereby determining the non-defective product / defective product, and outputs and displays the result on the display device PT 19. As shown in the figure, in the case of a defective product, the cause is also displayed.
[0078]
【The invention's effect】
As described above, in the present invention, when generating a frame by dividing the waveform data stored in the storage unit, a part of the preceding and succeeding frames are overlapped, and the FFT process is further performed. The window function to be used is provided with an attenuation part on both sides and the center part is flattened, so that signal components and transient characteristics that occur in a short time or discontinuity can be reliably detected with a frequency analyzer that applies the FFT algorithm. Can be detected. Therefore, abnormality determination by a sensory test can be performed automatically.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating an example of frame division executed when performing conventional FFT processing.
FIG. 2 is a diagram illustrating an example when FFT processing is performed without using a window.
FIG. 3 is a diagram illustrating an example in which a power spectrum obtained by performing FFT processing without using a window is displayed in decibels.
FIG. 4 is a diagram illustrating an example when FFT processing is performed using a window function.
FIG. 5 is a diagram illustrating an example in which a power spectrum obtained by performing FFT processing using a window function is displayed in decibels.
FIG. 6 is a diagram illustrating an example in which FFT processing is performed on a waveform signal when a single abnormal sound occurs using a window function.
FIG. 7 is a diagram illustrating an example of a case where FFT processing is performed on a waveform signal when a single abnormal sound occurs using a window function.
FIG. 8 is an experimental result showing an influence of a spectrum due to a shift in generation timing of a single signal.
FIG. 9 is a block diagram showing a preferred embodiment of the present invention.
FIG. 10 is a flowchart illustrating functions of a CPU.
FIG. 11 is a diagram illustrating frame division processing.
FIG. 12 is a diagram illustrating a window function.
FIG. 13 is a diagram illustrating an example of a result of obtaining a frequency feature amount for each frame.
FIG. 14 is a diagram showing an example of a sample waveform signal used for demonstrating the effect of the present invention.
FIG. 15 is an experimental result showing an influence of a spectrum on a measurement time lag of a single signal (short-time abnormal sound).
FIG. 16 is an experimental result showing an influence of a power spectrum due to a start position shift of a SIN waveform.
FIG. 17 is an experimental result showing an influence (deviation) of a power spectrum due to a start position shift of a SIN waveform.
FIG. 18 is an experimental result showing an influence of a power spectrum due to a phase shift (angle) of a SIN waveform.
FIG. 19 is a diagram illustrating a result of obtaining a spectrum by a conventional method for a sample waveform including only normal sound and a sample waveform including instantaneously generated abnormal sound.
FIG. 20 is a diagram illustrating a result of obtaining a spectrum while shifting the frame by 1/2 for a sample waveform including only normal sound and a sample waveform including instantaneously generated abnormal sound.
FIG. 21 is a diagram illustrating a relationship of data constituting a frame with respect to a shift amount P;
FIG. 22 is a diagram illustrating a waveform example obtained by multiplying the SIN waveform by another window function.
FIG. 23 is a diagram illustrating an appropriate relationship between a shift amount and a stability detecting unit (stable unit).
FIG. 24 is a diagram showing a preferred embodiment of an abnormality determination system according to the present invention.
[Explanation of symbols]
1 Analog filter
2 A / D board
3 Data memory
4 CPU
5 PC
10 housing
11 Monitor screen
12 PLC
13 Device control panel
15 products
16 microphone
17 Accelerometer
19 PT

Claims (4)

Storage means for storing given waveform data;
Frame creating means for extracting data for a desired frame size from the waveform data stored in the storage means and creating a frame;
FFT processing means for executing FFT processing based on data corrected by applying a window function to the frame data constituting the frame,
The frame creation means has a function of extracting data so that a part of the preceding and following frames overlaps on the time axis,
Applying the FFT algorithm, wherein the window function has an attenuation part on both sides so that the amplitude at both ends of the frame is 0, and the center stability detection area is flat. Frequency analyzer. The amount of overlap is half the frame size,
2. The FFT algorithm according to claim 1, wherein the region of the stability detection unit is ½ of the frame size, and the attenuation unit is ¼ of the frame size on both sides, respectively. Frequency analysis device. An abnormality determination device that determines an abnormality of a product based on measurement data measured by a measurement means,
A frequency analysis apparatus to which the FFT algorithm according to claim 1 is applied;
A feature amount extracting means for extracting a feature amount based on the frequency spectrum information of each frame obtained by the frequency analysis device applying the FFT algorithm;
Determination means for performing the abnormality determination based on the feature amount extracted by the feature amount extraction means,
The abnormality determination device characterized in that the measurement data is stored in the storage means. An abnormality determination device according to claim 3 and a PLC,
The PLC can be linked to a production system for producing the product,
The abnormality determination system, wherein the abnormality determination device operates based on a signal from the PLC.

Images