CN102362282A - Signal classification method and signal classification device - Google Patents
Signal classification method and signal classification device Download PDFInfo
- Publication number
- CN102362282A CN102362282A CN2010800120253A CN201080012025A CN102362282A CN 102362282 A CN102362282 A CN 102362282A CN 2010800120253 A CN2010800120253 A CN 2010800120253A CN 201080012025 A CN201080012025 A CN 201080012025A CN 102362282 A CN102362282 A CN 102362282A
- Authority
- CN
- China
- Prior art keywords
- mentioned
- recognizer
- study
- data
- inspection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims description 48
- 238000000605 extraction Methods 0.000 claims abstract description 75
- 230000008859 change Effects 0.000 claims abstract description 35
- 230000002159 abnormal effect Effects 0.000 claims abstract description 6
- 238000007689 inspection Methods 0.000 claims description 127
- 230000007246 mechanism Effects 0.000 claims description 40
- 230000001174 ascending effect Effects 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 5
- 239000000284 extract Substances 0.000 claims description 5
- 230000004044 response Effects 0.000 abstract 1
- 238000012545 processing Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 230000015572 biosynthetic process Effects 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000002860 competitive effect Effects 0.000 description 2
- 239000012467 final product Substances 0.000 description 2
- 230000008676 import Effects 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000002203 pretreatment Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 230000037221 weight management Effects 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000012797 qualification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2433—Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
- G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/78—Detection of presence or absence of voice signals
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Signal Processing (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
During learning, when a learning signal that includes normal samples and abnormal samples is inputted to a feature value extraction unit (2), the learning signal is short-time Fourier transformed and learning data is extracted. For each combination of time and frequency, a classifier creation unit (6) uses decision results from a learning decision unit (4) to create a classifier that minimizes an erroneous-decision rate as computed by a computation unit (5). From among the classifiers, which were created for each time/frequency combination, a classifier selection unit (7) selects the classifier with the lowest erroneous-decision rate and computes the reliability. In response to decision results from the selected classifier, a weighting instruction unit (31) instructs a weight setting change unit (30) to change the weights of the learning data. During examination, an examination decision unit (8) uses the plurality of classifiers selected during learning to decide whether or not the examination subject is in a normal state.
Description
Technical field
The present invention relates to discern the signal recognition method and the signal identification device of the state of checking object.
Background technology
All the time, the known signal identification device (for example with reference to patent documentation 1,2) that has use competitive learning type neural network to discern the state of inspection object.
Signal identification device in the past does, when study with a plurality of study with data input competitive learning type neural network, make the dendrogram of having set classification, this classification representes to check the state of object.
The signal identification device in the past of when study, making dendrogram does; When inspection, will import dendrogram with data, neuronic weight coefficient data of the output layer on the dendrogram and inspection will be calculated with the Euclidean distance between the data based on the inspection of the state of checking object.Then, signal identification device in the past, with inspection uses data qualification to Euclidean distance as in the neuronic classification of the output layer of minimum, discern the state of checking object with the classification of data according to the inspection of having classified.According to signal identification device in the past, also can automatically learn even without professional knowledge.
The prior art document
Patent documentation
Patent documentation 1: TOHKEMY 2004-354111 communique
Patent documentation 2: TOHKEMY 2008-040683 communique
But; Signal identification device in the past; Though also can automatically learn even without professional knowledge; But when inspection, need calculate Euclidean distance in the whole output layer neurons on dendrogram, and obtain the minimum value of Euclidean distance, have elongated problem of supervision time.
Summary of the invention
The present invention carries out in view of the above problems, and its purpose is to provide signal recognition method and signal identification device that also can automatically learn and can shorten the supervision time even without professional knowledge.
According to first technical scheme of the present invention, a kind of signal recognition method is provided, use to possess the signal identification device of a plurality of recognizers of the state of identification inspection object respectively; Set each recognizer judgment standard study and will contrast with the judgment standard of data and each recognizer based on the inspection of the state of above-mentioned inspection object and discern the inspection of the state of this inspection object; This signal recognition method is characterised in that, when study, input comprises that a plurality of study of normal signal and abnormal signal use electric signal; Use predefined method for distilling to learn with extracting characteristic quantity the electric signal from each; This characteristic quantity use data as study, to each above-mentioned study with data setting or change weight, to each predefined extraction scope; Each study is set at threshold value successively with each key element value of the extraction scope of data; Magnitude relationship to each key element value and above-mentioned threshold value compares, and in the change judgment standard, carries out each study thus with the key element value of the extraction scope of data and the contrast of this judgment standard, judges as each study and uses the study of the extraction source of data to use electric signal whether to be normal signal; When each study is set or changes with the above-mentioned weight of data; To each above-mentioned judgment standard, the disconnected above-mentioned study of judge by accident is calculated as the erroneous judgement rate of breaking with the summation of the above-mentioned weight of data, to each said extracted scope; Use the disconnected rate of above-mentioned erroneous judgement and make the recognizer candidate for judgment standard hour; When above-mentioned weight is set or changes, to select the recognizer candidate of each said extracted scope made the disconnected rate of above-mentioned erroneous judgement be minimum recognizer candidate as above-mentioned recognizer, each above-mentioned recognizer is used break rate and calculate fiduciary level of above-mentioned erroneous judgement; When above-mentioned recognizer is selected, makes by this recognizer and judge the above-mentioned weight increase of disconnected above-mentioned study by accident with data; When inspection; Electric signal is used in the inspection of the state of the above-mentioned inspection object of input expression; Use said extracted method with extracting characteristic quantity the electric signal, is used data with this characteristic quantity as above-mentioned inspection, from this inspection in each recognizer of the above-mentioned condition for identification of decision; The judgment standard of above-mentioned inspection with data and above-mentioned recognizer contrasted; Judging whether above-mentioned inspection object is normal condition, is that finally being judged as above-mentioned inspection object is normal condition under the situation more than the summation of fiduciary level of the recognizer that to be judged as above-mentioned inspection object be ERST in the summation of the fiduciary level of the recognizer that to be judged as above-mentioned inspection object be normal condition.
In addition, above-mentioned inspection also can be the one dimension waveform with electric signal.
And then, also can be with above-mentioned each study with the key element value of the extraction scope of data according to ascending order or descending sort, each the key element value after the arrangement is set at threshold value successively.
Here, so-called one dimension waveform is meant the waveform of the regulation variable (physical quantity) of expression inspection object with respect to the time.
In addition, the said extracted method also can be a Short Time Fourier Transform.
In addition, the said extracted method also can be a continuous wavelet transform.
According to a second technical aspect of the present invention; A kind of signal identification device is provided; Possess a plurality of recognizers of the state of identification inspection object respectively; Set each recognizer judgment standard study and will contrast with the judgment standard of data and each recognizer based on the inspection of the state of above-mentioned inspection object and discern the inspection of the state of this inspection object; This signal identification device is characterised in that to possess: Characteristic Extraction mechanism is transfused to a plurality of study that comprise normal signal and abnormal signal and uses electric signal when study; Electric signal is used in the inspection that when inspection, is transfused to the state of the above-mentioned inspection object of expression, uses predefined method for distilling from this electric signal, to extract characteristic quantity; Weight setting change mechanism will use data as study from each study respectively with the characteristic quantity that extracts the electric signal by above-mentioned Characteristic Extraction mechanism, should learn with data setting or will change weight each; Decision mechanism during study; When study; To each predefined extraction scope, each study is set at threshold value successively with each key element value of the extraction scope of data, the magnitude relationship of each key element value and above-mentioned threshold value is compared; In the change judgment standard, carry out each study thus with the key element value of the said extracted scope of data and the contrast of this judgment standard, judge as each study and use the study of the extraction source of data to use electric signal whether to be normal signal; Calculation mechanism; When setting by above-mentioned weight setting change mechanism or each study of change during with the above-mentioned weight of data; To each above-mentioned judgment standard, the disconnected above-mentioned study of decision mechanism erroneous judgement is calculated as the erroneous judgement rate of breaking with the summation of the above-mentioned weight of data in the time of will be by above-mentioned study; Recognizer is made mechanism, to each said extracted scope, uses the disconnected rate of the above-mentioned erroneous judgement that is calculated by aforementioned calculation mechanism and makes the recognizer candidate for judgment standard hour; Recognizer selection mechanism; When above-mentioned weight is set or changes; From make the recognizer candidate that mechanism makes each said extracted scope by above-mentioned recognizer, select the disconnected rate of above-mentioned erroneous judgement be minimum recognizer candidate as above-mentioned recognizer, each above-mentioned recognizer is used the disconnected rate of above-mentioned erroneous judgement and calculates fiduciary level; Weight indicating mechanism when selecting above-mentioned recognizer by above-mentioned recognizer selection mechanism, changes mechanism to above-mentioned weight setting and indicates, so that judge the above-mentioned weight increase of disconnected above-mentioned study with data by accident by this recognizer; And when inspection decision mechanism; When inspection; To use data with the characteristic quantity that extracts the electric signal as above-mentioned inspection from above-mentioned inspection by above-mentioned Characteristic Extraction mechanism; In each recognizer of the above-mentioned condition for identification of decision, the judgment standard of above-mentioned inspection with data and above-mentioned recognizer contrasted, judge whether above-mentioned inspection object is normal condition; Be that finally being judged as above-mentioned inspection object is normal condition under the situation more than the summation of fiduciary level of the recognizer that to be judged as above-mentioned inspection object be ERST in the summation of the fiduciary level of the recognizer that to be judged as above-mentioned inspection object be normal condition.
Here, so-called one dimension waveform is meant the waveform of the variable (physical quantity) of the regulation of representing the inspection object with respect to the time.
According to first and second technical scheme of the present invention; When selecting the disconnected rate of erroneous judgement to be the recognizer of minimum; Change so that increase, reduce with the weight of data by the weight of the electric signal of the normal judgement of selecting last time of recognizer by the disconnected study of the recognizer erroneous judgement of selecting last time; When weight is set or change, recognizer is selected, also can when learning, automatically carry out best setting even without professional knowledge thus.In addition, according to the invention of technical scheme 1,6, when inspection, only the judged result of each recognizer is carried out comprehensively getting final product, so compare and to shorten the supervision time with the situation of having used neural network.
In addition, first and second technical scheme according to the present invention can consider respectively the size of the fiduciary level of each recognizer of calculating independently, and whether final judgement inspection object is normal condition.
And then first and second technical scheme according to the present invention through each key element value is set at threshold value successively, can judge easily that each study is normal signal or abnormal signal with the extraction source of data, so can calculate the disconnected rate of erroneous judgement at short notice.
In addition, in Short Time Fourier Transform, constitute and to comprise whole characteristics that the one dimension waveform has and have from the recognizer of the judgment standard of the key element value of wherein only using needs, so need not prepare a plurality of feature extraction filtrators in advance.As a result, compare with signal identification device in the past, can dwindle needs amount of memory.In addition, according to this embodiment,,, can dwindle program size so compare and to shorten learning time with signal identification device in the past because the hunting zone is defined to the characteristic quantity after the Short Time Fourier Transform.
In addition, in continuous wavelet transform, constitute and to comprise whole characteristics that the one dimension waveform has and have from the recognizer of the judgment standard of the key element value of wherein only using needs, so need not prepare a plurality of Characteristic Extraction filtrators in advance.As a result, compare with signal identification device in the past, can dwindle needs amount of memory.In addition, according to the invention of technical scheme 3,,, can dwindle program size so compare and to shorten learning time with signal identification device in the past because the hunting zone is defined to the characteristic quantity behind the continuous wavelet transform.
Description of drawings
Fig. 1 is the block diagram of formation of the signal identification device of expression embodiment 1.
Fig. 2 is the process flow diagram of learning method in the signal recognition method of expression embodiment 1.
Fig. 3 is the process flow diagram of the recognizer method for making in the learning method of signal recognition method of expression embodiment 1.
Fig. 4 is the process flow diagram of inspection method in the signal recognition method of expression embodiment 1.
Fig. 5 is the process flow diagram of learning method in the signal recognition method of expression embodiment 2.
Embodiment
(embodiment 1)
At first, the formation to the signal identification device of embodiment 1 describes.The signal identification device of this embodiment uses the AdaBoost method to discern the state of inspection object (the inspection object is normal condition or ERST).The signal identification device of this embodiment the time uses pre-prepd a plurality of (m) study with electric signal (below be called " signal is used in study ") in study, set a plurality of a little less than recognizers (below be called " recognizer ").The both sides of the electric signal when a plurality of study use signal to comprise the electric signal of inspection object during as normal condition (below be called " normal sample ") and check object as ERST (below be called " exceptional sample ").Use signal for each study, for other learn with signal discern and allocate in advance independent sample sequence number i (i=0 ..., m-1).Each recognizer does, when from inspection object input checking during with electric signal (below be called " signal use in inspection "), judgement checks whether object is normal condition.
As shown in Figure 1, judging part (decision mechanism during study) 4, calculating part (calculation mechanism) 5, recognizer preparing department (recognizer is made mechanism) 6, recognizer selection portion (recognizer selection mechanism) 7, judging part (decision mechanism during inspection) 8 and efferent 9 when checking when signal identification device possesses signal input part 1, Characteristic Extraction portion (Characteristic Extraction mechanism) 2, weight management department 3, study.In addition, inspection object (not shown) mainly is device or the equipment etc. that comprise slewing, but is not limited to above-mentioned situation.
For Characteristic Extraction portion 2, import the normal sample and the exceptional sample of one dimension waveform with signal as study.Signal is used in each study that 2 pairs in Characteristic Extraction portion is imported, to the pre-treatment of learning to remove denoising with signal.Then, Characteristic Extraction portion 2 uses 1,2 pairs of study of formula to carry out Short Time Fourier Transform with signal, and (b, set f) is extracted as characteristic quantity with the key element value F (i) of the combination of time (the center time of window function) b and frequency f.Signal is used in study to sample sequence number i, and (b, set f) is used data as the study of sample sequence number i to the key element value F (i) of the time b frequency f that will be extracted by Characteristic Extraction portion 2.For example, with the key element value representation of sample sequence number i=2, time b3 frequency f 2 be key element value F (2) (b3, f2).The key element value F (i) of formula 1 expression sample sequence number i time b frequency f (b, f), the window function Rd (t-b) of formula 2 representation formulas 1.Window function Rd (t-b) is to be the Gaussian window at center with time b.X (t) is an electric signal.
[formula 1]
[formula 2]
On the other hand, in when inspection, to Characteristic Extraction portion 2, as inspection with signal from vibration transducer 10 or microphone 11 inputs electric signal based on the simulation of the state of checking object (not shown).After input checking was with signal, Characteristic Extraction portion 2 was same during with study, and inspection is carried out Short Time Fourier Transform with signal, and (b, set f) is extracted as characteristic quantity with the key element value F of time b frequency f.(b, set f) is used data as inspection with the key element value F of time b frequency f.For example, the key element value representation of time b2 frequency f 5 is that (b2 f5) (with reference to formula 1,2, omits i) to F.
Weight management department 3 possesses: weight setting change portion (weight setting change mechanism) 30, set and change weights W (i) with data (signal is used in study) each study when study; With weight indication portion (weight indicating mechanism) 31, the weights W (i) of each study with data changed in 30 indications of weight setting change portion.The study of W (i) expression sample sequence number i is with the weight of data.For example, the study of sample sequence number i=1 is shown W (1) with the weight table of data, the study of sample sequence number i=m-1 is shown W (m-1) with the weight table of data.
30 pairs of each weights W of weight setting change portion (i) are set and are changed, so that each study becomes 1 (∑ W (i)=1) with the summation ∑ W (i) of the weights W (i) of data.Each study is set with the initial value of the weights W (i) of data equably.Use under the data situation individual as m in study, each study is set to 1/m with the weights W (i) of data.
31 pairs of weight setting changes portion of weight indication portion 30 indicates; So that whenever as after when selecting recognizer DEC (n) by recognizer selection portion 7 stating; Make by weights W (i) increase of the disconnected study of recognizer DEC (n) erroneous judgement, the study that judged rightly by recognizer DEC (n) is reduced with the weights W (i) of data with data.
Judging part 4 is to the combination of each time b frequency f during study, and (b, f), (b f) arranges with ascending order with the key element value F (i) that takes out with the key element value F (i) of take-off time b frequency f the data from global learning.The key element value F (c) that is arranged according to from small to large sequence list be shown F (0), F (1) ..., F (m-1).Then, judging part 4 is set at threshold value δ successively with each key element value F (c) during study, and each key element value F (c) and threshold value δ are compared.If key element value F (c) is less than threshold value δ, to be judged as the study that comprises key element value F (c) are normal samples with the extraction source of data to judging part 4 when then learning.If key element value F (c) is more than the threshold value δ, to be judged as the study that comprises key element value F (c) are exceptional samples with the extraction source of data to judging part 4 when then learning.In addition, judging part 4 can (b be with ascending order but with descending sort f) with key element value F (i) yet during study.
Calculating part 5 does; When setting by the combination of 30 pairs of whole time b frequency f of weight setting change portion or changing weights W (i); To each threshold value δ, will be calculated as the erroneous judgement rate ε that breaks with the summation ∑ W (i) of the weights W (i) of data by the disconnected study of when study judging part 4 erroneous judgements.So-called erroneous judgement is disconnected to be meant, the study that comprises key element value F (c) uses the extraction source of data be originally normal sample but be judged as exceptional sample or the said extracted source originally as exceptional sample but be judged as normal sample.Then, calculating part 5 the disconnected rate ε of the erroneous judgement that calculates greater than 0.5 situation under, (the disconnected rate ε of 1-erroneous judgement) is made as the disconnected rate ε of new erroneous judgement.
Recognizer preparing department 6 does, to the combination of each time b frequency f, use the disconnected rate ε of the erroneous judgement that calculates by calculating part 5 for judgment standard hour make recognizer (weak recognizer candidate) DEC (b, f).The content of above-mentioned judgment standard does; Disconnected rate ε is under the situation below 0.5 in the erroneous judgement that is calculated by calculating part 5; If key element value F (c) is less than threshold value δ; Then being judged as the study that comprises key element value F (c) is normal samples with the extraction source of data, if key element value F (c) is more than the threshold value δ, then being judged as the said extracted source is exceptional sample.On the other hand; The disconnected rate ε of the erroneous judgement that calculates by calculating part 5 greater than 0.5, (the disconnected rate ε of 1-erroneous judgement) become under the situation of the disconnected rate ε of new erroneous judgement; The content of above-mentioned judgment standard does, is exceptional samples if key element value F (c), then is judged as the study that comprises key element value F (c) less than threshold value δ with the extraction source of data; If key element value F (c) is more than the threshold value δ, then being judged as the said extracted source is normal sample.
Recognizer selection portion 7 does, from a plurality of recognizer DEC of the combination of each time b frequency f being made by recognizer preparing department 6 (b, f) among, selecting the disconnected rate ε of erroneous judgement be that (b is f) as recognizer (check and use recognizer) DEC (n) for minimum recognizer DEC.N is the recognizer sequence number that is used to discern each recognizer DEC (n).For example, the 3rd selecteed recognizer of DEC (3) expression.Recognizer selection portion 7 does, for the recognizer DEC (n) that selects, with the disconnected rate ε substitution formula 3 of the erroneous judgement of recognizer DEC (n) and calculate fiduciary level α (n), with fiduciary level α (n) substitution formula 4 and design factor β (n) (β (n)＜1).Recognizer selection portion 7 is that selection recognizer (n) is all selected n recognizer DEC (n) altogether when weights W (i) is set or changes.A plurality of recognizer DEC (n) arrange with selecting sequence, are stored in storage part 70.
[formula 3]
[formula 4]
β(n)＝exp(-α(n))
Will be by Characteristic Extraction portion 2 from checking that (b f) compares with threshold value δ, judges whether the inspection object is normal condition with the key element value F (i) that extracts the signal.Being judged as the inspection object at recognizer DEC (n) is under the situation of normal condition, and judging part 8 will add that value behind this fiduciary level α (n) of recognizer DEC (n) is as the summation S of new fiduciary level to the summation S of fiduciary level so far during inspection.Being judged as the inspection object at recognizer DEC (n) is under the situation of ERST, and judging part 8 will deduct value behind this fiduciary level α (n) of recognizer DEC (n) from the summation S of fiduciary level so far as the summation S of new fiduciary level during inspection.After judging part 8 has carried out above-mentioned judgement during inspection in whole recognizer DEC (n), be under the situation 0 or more, finally be judged as and check that object is a normal condition in the summation of fiduciary level.The summation S of fiduciary level less than 0 situation under, finally to be judged as the inspection object be ERST to judging part 8 during inspection.The final judged result of judging part 8 is to efferent 9 outputs during inspection.
Efferent 9 does, when from inspection during the final judged result of judging part 8 inputs, according to ring alarm tone or show the warning picture of the final judged result of being imported.In addition, efferent 9 does, under the situation that is connected with external unit (not shown), and also can be with exporting to external unit corresponding to the information of final judged result.
Then, the signal recognition method to this embodiment describes.At first, use Fig. 2 that learning method is described.
At first, learn to be imported into Characteristic Extraction portion 2 (S1) for m that comprises normal sample and exceptional sample with signal (waveform sample is used in study).Sample sequence number i is set to initial value 0 (S2).Then, Characteristic Extraction portion 2 will learn to carry out Short Time Fourier Transform with signal, and (b, the study of set f) is with data (S3) as key element value F (i) in extraction.Characteristic Extraction portion 2 judges whether from whole (m) study with having extracted study the signal with data (S4).Not extracting study with under the data conditions with signal from global learning, be set at i=i+1 (S5), and repeating step S3.Extracting study with signal with under the data conditions from global learning, weight setting change portion 30 is set at initial value (1/m) (S6) with each study with the weights W (i) of data.Then, recognizer sequence number n is set to initial value 1 (S7), and time b is set to initial value 0 (S8), and frequency f is set to initial value 0 (S9).
Then, the recognizer of recognizer preparing department 6 Production Time b frequency f (weak recognizer candidate) DEC (b, f) (S10).At this moment, calculate recognizer DEC (b, the rate ε that breaks of erroneous judgement f).Then, recognizer selection portion 7 judges (whether b, the disconnected rate ε of erroneous judgement f) are minimum (S11) for this recognizer DEC in the disconnected rate ε of erroneous judgement so far.Under the disconnected rate ε of this erroneous judgement be minimum situation, (b, the fiduciary level α (n) during f) as recognizer DEC (n) (S12) with this recognizer DEC for 7 calculating of recognizer selection portion.Then, recognizer selection portion 7 is updated to the recognizer DEC (n) of recognizer sequence number n this recognizer DEC (b, f) (S13).Recognizer selection portion 7 judge whether time b to whole frequency f carried out from step S10 to step S13 processing (first handle) (S14).Whole frequency f are not carried out be set at f=f+1 (S15) under the situation of first processing, repeat from step S10 to step S13 at time b.Whole frequency f have been carried out under first situation about handling at time b, recognizer selection portion 7 judge whether to whole time b carried out from step S9 to step S14 processing (second handles) (S16).Under the situation of whole time b not being carried out second processing, be set at b=b+1 (S17), repeat from step S9 to step S14.
Under the situation of All Time b having been carried out second processing, 30 pairs of each study of weight setting change portion are changed (S18) with the weights W (i) of data.31 pairs of weight setting changes portion of weight indication portion 30 indicates, so that the study that judges rightly for the recognizer DEC (n) by final updated multiply by factor beta (n) (β (n)＜1) and weights W (i) is reduced with data.On the other hand, 31 pairs of weight setting changes portion of weight indication portion 30 indicates, so that will with data weights W (i) increased by the disconnected study of erroneous judgement.
Recognizer selection portion 7 judges whether to have selected all (n), and recognizer DEC (n) (S19).Under the situation of non-selected whole recognizer DEC (n), be set at n=n+1 (S20), repeat from step S8 to step S18.Under the situation of having selected whole recognizer DEC (n), accomplish study.
Then, (b, method for making f) describes to the recognizer DEC of the time b frequency f of step S10 in the above-mentioned learning method to use Fig. 3.
At first, judging part 4 is for the combination of each time b frequency f during study, and (b f) arranges (S21) with ascending order with the key element value F (i) of data with global learning.The key element value F (c) that is arranged be followed successively by from small to large F (0), F (1) ..., F (m-1).Then, judging part 4 is set at initial value F (0) (S22) with threshold value δ during study.The recognizer DEC of judging part 4 making threshold value δ during study (b, f).Recognizer DEC (b to threshold value δ; F) set following judgment standard (S23): each key element value F (c) and threshold value δ are compared respectively; To comprise less than the study of the key element value F (c) of threshold value δ and be judged as normal sample with the extraction source of data, the study that will comprise the above key element value F (c) of threshold value δ is judged as exceptional sample with the extraction source of data.
Then, calculating part 5 uses to comprise and is calculated erroneous judgement by the study of the disconnected key element value F (c) of erroneous judgement with the weights W (i) of data and break that rate ε=∑ W (i) (S24) in step S23.Then, recognizer preparing department 6 judges that whether the disconnected rate ε of erroneous judgement is greater than 0.5 (S25).The disconnected rate ε of erroneous judgement greater than 0.5 situation under, recognizer preparing department 6 makes judgment standard opposite, and (the disconnected rate ε of 1-erroneous judgement) is made as the disconnected rate ε (S26) of new erroneous judgement.Promptly; Recognizer preparing department 6 change judgment standards; Be judged as exceptional sample so that will comprise less than the study of the key element value F (c) of threshold value δ with the extraction source of data, the study that will comprise the above key element value F (c) of threshold value δ is judged as normal sample with the extraction source of data.Perhaps, disconnected rate ε is under the situation below 0.5 in erroneous judgement, and recognizer preparing department 6 judges whether the disconnected rate ε of this erroneous judgement is minimum (S27) among the disconnected rate of erroneous judgement so far.Disconnected rate ε is that (b, judgment standard f) upgrades (S28) to 6 couples of recognizer DEC of recognizer preparing department under the situation of minimum in this erroneous judgement.
Then, recognizer preparing department 6 judges whether whole threshold value δ have been carried out handling (step S23 is to step S28) (S29).Under situation about whole threshold value δ not being handled, be set at c=c+1 (S30), repeat from step S23 to step S28.Whole threshold value δ are being carried out under the situation about handling, recognizer preparing department 6 confirms recognizer DEC (b, f) (S31) of time b frequency f.(b f) has by the judgment standard of final updated the recognizer DEC that confirms.
Then, use Fig. 4 that the inspection method of the signal recognition method of this embodiment is described.At first, will check with signal and be input to Characteristic Extraction portion 2 (S41) from signal input part 1.Characteristic Extraction portion 2 will check with signal and carry out Short Time Fourier Transform that (b, the inspection of set f) is with data (S42) as key element value F in extraction.
Judging part 8 is set at initial value 0 (S43) with the summation S of fiduciary level during inspection, and recognizer sequence number n is set at initial value 1 (S44).Then, judging part 8 is implemented based on the inspection of recognizer DEC (n) judgement (S45) with data during inspection.Be judged as by recognizer DEC (n) under the situation that the inspection object is a normal condition (S46), judging part 8 makes (the summation S of new fiduciary level)=(the summation S of fiduciary level so far)+(the fiduciary level α (n) of this recognizer DEC (n)) (S47) during inspection.Be judged as by recognizer DEC (n) under the situation that the inspection object is an ERST (S46), judging part 8 makes (the summation S of new fiduciary level)=(the summation S of fiduciary level so far)-(the fiduciary level α (n) of this recognizer DEC (n)) (S48) during inspection.Judging part 8 judges whether all implementing the judgement (S49) of inspection with data among the recognizer DEC (n) during inspection., do not repeat from step S45 to step S48 all implementing to be set at n=n+1 (S50) under the situation of inspection with the judgement of data among the recognizer DEC (n).All implementing under the situation of inspection with the judgement of data among the recognizer DEC (n), whether the summation S of judging part 8 judge reliabilities is (S51) more than 0 during inspection.At the summation S of fiduciary level is under the situation more than 0, and judging part 8 finally is judged as and checks that object is normal condition (S52) during inspection.The summation S of fiduciary level less than 0 situation under, finally to be judged as the inspection object be ERST (S53) to judging part 8 during inspection.The final judged result of judging part 8 is outputed to efferent 9 during inspection.
More than; According to this embodiment; To judge the recognizer DEC (b of disconnected rate ε by accident for minimum; When f) being chosen as recognizer DEC (n), change, when weights W (i) is set or changes, just select recognizer DEC (n) so that the study that is increased, makes the normal judgement of being selected by last time of recognizer DEC (n) by the disconnected study of recognizer DEC (n) erroneous judgement of selecting last time with the weights W (i) of data reduces with the weights W (i) of data; Even without professional knowledge, when study, also can automatically carry out best setting thus.Particularly, can consider respectively the size of the fiduciary level of each recognizer of calculating independently, and to checking whether object is that normal condition is finally judged.
In addition,, when inspection, the judged result of each recognizer DEC (n) is carried out comprehensively getting final product,, can shorten the supervision time so compare with the situation of having used neural network according to this embodiment.
And then; According to this embodiment, in when study judging part 4, through each key element value F (c) is set at threshold value δ successively; Can judge easily that each study is normal sample or exceptional sample with the extraction source of data, so can calculate the disconnected rate ε of erroneous judgement at short notice.
Yet signal identification device in the past is that when extracting characteristic quantity, the Characteristic Extraction filtrator that automatic selection is suitable for checking from a plurality of Characteristic Extraction filtrators is so need to prepare in advance whole Characteristic Extraction filtrators that use of supposing.But, the change of inspection object also comprise the object that may occur from now on and exist countless, so be difficult to prepare in advance the Characteristic Extraction filtrators that all supposition are used.In addition,, select the parameter combinations of optimal setting for a plurality of combinations of parameters that each Characteristic Extraction filtrator is had are all searched for, also need a large amount of processing times even prepared a plurality of Characteristic Extraction filtrators in advance.
On the other hand; According to this embodiment; In Short Time Fourier Transform; Constitute and to comprise whole characteristics that the one dimension waveform has and to have that (b, the recognizer DEC (n) of judgment standard f) is not so need prepare a plurality of Characteristic Extraction filtrators in advance from the key element value F (i) that wherein only used needs.As a result, compare with signal identification device in the past, can dwindle needs amount of memory.In addition,,,, can shorten learning time, can dwindle program size so compare with signal identification device in the past because the hunting zone is defined to the characteristic quantity after the Short Time Fourier Transform according to this embodiment.
(embodiment 2)
The difference of the signal identification device of the signal identification device of embodiment 2 and embodiment 1 is that electric signal is not carried out Short Time Fourier Transform and carries out continuous wavelet transform.
The Characteristic Extraction portion 2 of this embodiment does; After the study that comprises normal sample and exceptional sample for each has been carried out with signal removing the pre-treatment of denoising with signal, to study; Use formula 5,6 will learn to carry out continuous wavelet transform with signal; (b, set a) is extracted as characteristic quantity with the key element value Y (i) of the combination of time b and parameter a.The key element value Y (i) of formula 5 expression sample sequence number i time b parameter a (b, a), formula 6 expression female small echo ψ (t).X (t) is an electric signal.
[formula 5]
[formula 6]
On the other hand, in when inspection, Characteristic Extraction portion 2 is same during with study, also will check with signal and carry out continuous wavelet transform, and (b, data are used in the inspection of set a) as the key element value Y of time b parameter a in extraction.
Then, the signal recognition method to this embodiment describes.At first, use Fig. 5 that learning method is described.
At first, m the study that comprises normal sample and exceptional sample is transfused to Characteristic Extraction portion 2 (S61) with signal.Sample sequence number i is set to initial value 0 (S62).Then, Characteristic Extraction portion 2 will learn to carry out continuous wavelet transform with signal, and (b, the study of set a) is with data (S63) as key element value Y (i) in extraction.Characteristic Extraction portion 2 judges whether from whole (m) study with having extracted study the signal with data (S64).Not extracting study with under the data conditions with signal, be set at i=i+1 (S65), repeating step S63 from global learning.Extracting study with signal with under the data conditions from global learning, weight setting change portion 30 is set at initial value (1/m) (S66) with each study with the weights W (i) of data.Then, recognizer sequence number n is set to initial value 1 (S67), and time b is set to initial value 0 (S68), and parameter a is set to initial value 0 (S69).
Then, recognizer (weak recognizer candidate) DEC (b, a) (S70) during recognizer preparing department 6 Production Time b parameter a.At this moment, calculate recognizer DEC (b, the rate ε that breaks of erroneous judgement a).Then, recognizer selection portion 7 judges (whether b, the disconnected rate ε of erroneous judgement a) are minimum (S71) for this recognizer DEC among the disconnected rate ε of erroneous judgement so far.Under the disconnected rate ε of this erroneous judgement be minimum situation, (b, the fiduciary level α (n) during a) as recognizer DEC (n) (S72) with this recognizer DEC for 7 calculating of recognizer selection portion.Then, recognizer selection portion 7 is updated to the recognizer DEC (n) of recognizer sequence number n this recognizer DEC (b, a) (S73).Recognizer selection portion 7 judge whether time b to whole parameter a carried out from step S70 to step S73 processing (the 3rd handle) (S74).Whole parameter a are not carried out be set at a=a+1 (S75) under the situation of the 3rd processing, repeat from step S70 to step S73 at time b.Whole parameter a have been carried out under the 3rd situation about handling at time b, recognizer selection portion 7 judge whether to All Time b carried out from step S69 to step S74 processing (the reason) everywhere (S76).All Time b not being carried out the everywhere under the situation of reason, be set at b=b+1 (S77), repeat from step S69 to step S74.
All Time b carried out the everywhere under the situation of reason, 30 pairs of each study of weight setting change portion are changed (S78) with the weights W (i) of data.31 pairs of weight setting changes portion of weight indication portion 30 indicates, so that the study that the recognizer DEC (n) by final updated is judged rightly multiply by factor beta (n) (β (n)＜1) and weights W (i) is reduced with data.On the other hand, 31 pairs of weight setting changes portion of weight indication portion 30 indicates, so that will with data weights W (i) increased by the disconnected study of erroneous judgement.
Recognizer selection portion 7 judges whether to have selected all (n), and recognizer DEC (n) (S79).Under the situation of non-selected whole recognizer DEC (n), be set at n=n+1 (S80), repeat from step S68 to step S78.Under the situation of having selected whole recognizer DEC (n), accomplish study.
Then, (b, method for making a) describes to the recognizer DEC of the time b parameter a of step S70 in the above-mentioned learning method to use Fig. 3.
At first, judging part 4 is for the combination of each time b parameter a during study, and (b a) arranges (S21) with ascending order with the key element value Y (i) of data with global learning.Key element value Y (c) after the arrangement be followed successively by from small to large Y (0), Y (1) ..., Y (m-1).Then, judging part 4 is set at initial value Y (0) (S22) with threshold value δ during study.The recognizer DEC of judging part 4 making threshold value δ during study (b, a).Recognizer DEC (b for threshold value δ; A) set following judgment standard (S23): each key element value Y (c) and threshold value δ are compared respectively; To comprise less than the study of the key element value Y (c) of threshold value δ and be judged as normal sample with the extraction source of data, the study that will comprise the above key element value Y (c) of threshold value δ is judged as exceptional sample with the extraction source of data.
Then, calculating part 5 uses to comprise and is calculated erroneous judgement by the study of the disconnected key element value Y (c) of erroneous judgement with the weights W (i) of data and break that rate ε=∑ W (i) (S24) in step S23.Then, recognizer preparing department 6 judges that whether the disconnected rate ε of erroneous judgement is greater than 0.5 (S25).The disconnected rate ε of erroneous judgement greater than 0.5 situation under, recognizer preparing department 6 makes judgment standard opposite, and (the disconnected rate ε of 1-erroneous judgement) is made as the disconnected rate ε (S26) of new erroneous judgement.That is, recognizer preparing department 6 change judgment standards are judged as exceptional sample will comprise less than the study of the key element value Y (c) of threshold value δ with the extraction source of data, and the study that will comprise the key element value Y (c) more than the threshold value δ is judged as normal sample with the extraction source of data.Then, be under the situation below 0.5 perhaps at the disconnected rate ε of erroneous judgement, recognizer preparing department 6 judges whether the disconnected rate ε of this erroneous judgement is minimum (S27) in the disconnected rate of erroneous judgement so far.Disconnected rate ε is that (b, judgment standard a) upgrades (S28) to 6 couples of recognizer DEC of recognizer preparing department under the situation of minimum in this erroneous judgement.
Then, recognizer preparing department 6 judges whether whole threshold value δ have been carried out handling (step S23 is to step S28) (S29).Under situation about whole threshold value δ not being handled, be set at c=c+1 (S30), repeat from step S23 to step S28.Whole threshold value δ are being carried out under the situation about handling recognizer DEC (b, a) (S31) when recognizer preparing department 6 confirms time b parameter a.(b a) has by the judgment standard of final updated the recognizer DEC that confirms.
Then, use Fig. 4 that the inspection method of the signal recognition method of this embodiment is described.At first, check with signal from 2 (S41) of signal input part 1 input feature vector amount extraction portion.Characteristic Extraction portion 2 will check with signal and carry out continuous wavelet transform that (b, the inspection of set a) is with data (S42) as key element value Y in extraction.About after action, with the same (S43～S53) of embodiment 1.
More than; According to this embodiment, in continuous wavelet transform, formation comprises whole characteristics that the one dimension waveform has and has from wherein only having used key element value Y (the i) (b of needs; The recognizer DEC (n) of judgment standard a) is not so need prepare a plurality of Characteristic Extraction filtrators in advance.As a result, compare with signal identification device in the past, can dwindle needs amount of memory.In addition,,,, can shorten learning time, can dwindle program size so compare with signal identification device in the past because the hunting zone is defined to the characteristic quantity behind the continuous wavelet transform according to this embodiment.
More than, with reference to accompanying drawing preferred implementation of the present invention is illustrated, but the present invention is not limited to said example.So long as those skilled in the art can expect various change examples or revise example that these also belong to technical scope of the present invention certainly certainly in the technical scope that the scope of patent request is put down in writing.
Claims (7)
1. signal recognition method; Use possesses the signal identification device of a plurality of recognizers of the state of identification inspection object respectively; Set each recognizer judgment standard study and will contrast with the judgment standard of data and each recognizer based on the inspection of the state of above-mentioned inspection object and discern the inspection of the state of this inspection object; This signal recognition method is characterised in that
When study,
Input comprises that a plurality of study of normal signal and abnormal signal use electric signal, uses predefined method for distilling to learn with extracting characteristic quantity the electric signal from each, with this characteristic quantity as learning to use data,
To each above-mentioned study with data setting or the change weight,
To each predefined extraction scope; Each study is set at threshold value successively with each key element value of the extraction scope of data; Magnitude relationship to each key element value and above-mentioned threshold value compares; In the change judgment standard, carry out each study thus with the key element value of the extraction scope of data and the contrast of this judgment standard, judge as each study and use the study of the extraction source of data to use electric signal whether to be normal signal
When each study is set or changes with the above-mentioned weight of data, to each above-mentioned judgment standard, institute judge the above-mentioned study of breaking and be calculated as the rate of breaking of judging by accident with the summation of the above-mentioned weight of data,
To each said extracted scope, use the disconnected rate of above-mentioned erroneous judgement and make the recognizer candidate for judgment standard hour,
When above-mentioned weight is set or changes; From to recognizer candidate that to select the disconnected rate of above-mentioned erroneous judgement the recognizer candidate of each said extracted scope made be minimum as above-mentioned recognizer; Each above-mentioned recognizer is used the disconnected rate of above-mentioned erroneous judgement and calculates fiduciary level; When above-mentioned recognizer is selected, makes by this recognizer and judge the above-mentioned weight increase of disconnected above-mentioned study by accident with data
When inspection,
Electric signal is used in the inspection of the state of the above-mentioned inspection object of input expression, and use said extracted method with extracting characteristic quantity the electric signal, is used data with this characteristic quantity as above-mentioned inspection from this inspection,
In each recognizer of the above-mentioned condition for identification of decision; The judgment standard of above-mentioned inspection with data and above-mentioned recognizer contrasted; Judge whether above-mentioned inspection object is normal condition; Be that finally being judged as above-mentioned inspection object is normal condition under the situation more than the summation of fiduciary level of the recognizer that to be judged as above-mentioned inspection object be ERST in the summation of the fiduciary level of the recognizer that to be judged as above-mentioned inspection object be normal condition.
2. the signal recognition method of putting down in writing like claim 1 is characterized in that,
Above-mentioned inspection uses electric signal to be the one dimension waveform.
3. the signal recognition method of putting down in writing like claim 1 is characterized in that,
With above-mentioned each study with the key element value of the extraction scope of data according to ascending order or descending sort, each the key element value after the arrangement is set at threshold value successively.
4. the signal recognition method of putting down in writing like claim 1 is characterized in that,
The said extracted method is a Short Time Fourier Transform.
5. the signal recognition method of putting down in writing like claim 1 is characterized in that,
The said extracted method is a continuous wavelet transform.
6. signal identification device; Possess a plurality of recognizers of the state of identification inspection object respectively; Set each recognizer judgment standard study and will contrast with the judgment standard of data and each recognizer based on the inspection of the state of above-mentioned inspection object and discern the inspection of the state of this inspection object; This signal identification device is characterised in that to possess:
Characteristic Extraction mechanism; When study, be transfused to a plurality of study that comprise normal signal and abnormal signal and use electric signal; Electric signal is used in the inspection that when inspection, is transfused to the state of the above-mentioned inspection object of expression, uses predefined method for distilling from this electric signal, to extract characteristic quantity;
Weight setting change mechanism will use data as study from each study respectively with the characteristic quantity that extracts the electric signal by above-mentioned Characteristic Extraction mechanism, should learn with data setting or will change weight each;
Decision mechanism during study; When study; To each predefined extraction scope, each study is set at threshold value successively with each key element value of the extraction scope of data, the magnitude relationship of each key element value and above-mentioned threshold value is compared; In the change judgment standard, carry out each study thus with the key element value of the said extracted scope of data and the contrast of this judgment standard, judge as each study and use the study of the extraction source of data to use electric signal whether to be normal signal;
Calculation mechanism; When setting by above-mentioned weight setting change mechanism or each study of change during with the above-mentioned weight of data; To each above-mentioned judgment standard, the disconnected above-mentioned study of decision mechanism erroneous judgement is calculated as the erroneous judgement rate of breaking with the summation of the above-mentioned weight of data in the time of will be by above-mentioned study;
Recognizer is made mechanism, to each said extracted scope, uses the disconnected rate of the above-mentioned erroneous judgement that is calculated by aforementioned calculation mechanism and makes the recognizer candidate for judgment standard hour;
Recognizer selection mechanism; When above-mentioned weight is set or changes; From make the recognizer candidate that mechanism makes each said extracted scope by above-mentioned recognizer, select the disconnected rate of above-mentioned erroneous judgement be minimum recognizer candidate as above-mentioned recognizer, each above-mentioned recognizer is used the disconnected rate of above-mentioned erroneous judgement and calculates fiduciary level;
Weight indicating mechanism when selecting above-mentioned recognizer by above-mentioned recognizer selection mechanism, changes mechanism to above-mentioned weight setting and indicates, so that judge the above-mentioned weight increase of disconnected above-mentioned study with data by accident by this recognizer; And
Decision mechanism during inspection; When inspection; To use data with the characteristic quantity that extracts the electric signal as above-mentioned inspection from above-mentioned inspection by above-mentioned Characteristic Extraction mechanism; In each recognizer of the above-mentioned condition for identification of decision, the judgment standard of above-mentioned inspection with data and above-mentioned recognizer contrasted, judge whether above-mentioned inspection object is normal condition; Be that finally being judged as above-mentioned inspection object is normal condition under the situation more than the summation of fiduciary level of the recognizer that to be judged as above-mentioned inspection object be ERST in the summation of the fiduciary level of the recognizer that to be judged as above-mentioned inspection object be normal condition.
7. the signal identification device of putting down in writing like claim 6 is characterized in that,
With above-mentioned each study with the key element value of the extraction scope of data according to ascending order or descending sort, each the key element value after the arrangement is set at threshold value successively.
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2009077688A JP2010231455A (en) | 2009-03-26 | 2009-03-26 | Method and device for signal identification |
| JP077688/2009 | 2009-03-26 | ||
| PCT/IB2010/000579 WO2010116222A1 (en) | 2009-03-26 | 2010-03-18 | Signal classification method and signal classification device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN102362282A true CN102362282A (en) | 2012-02-22 |
Family
ID=42935684
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2010800120253A Pending CN102362282A (en) | 2009-03-26 | 2010-03-18 | Signal classification method and signal classification device |
Country Status (4)
| Country | Link |
|---|---|
| JP (1) | JP2010231455A (en) |
| KR (1) | KR20110122748A (en) |
| CN (1) | CN102362282A (en) |
| WO (1) | WO2010116222A1 (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107389545A (en) * | 2016-05-17 | 2017-11-24 | 柳光龙 | Centring means for detection object |
| CN108836574A (en) * | 2018-06-20 | 2018-11-20 | 广州智能装备研究院有限公司 | It is a kind of to utilize neck vibrator work intelligent sounding system and its vocal technique |
| CN109961421A (en) * | 2017-12-25 | 2019-07-02 | 欧姆龙株式会社 | Data generating device, data creation method and data generate recording medium |
| CN110199233A (en) * | 2017-01-27 | 2019-09-03 | 株式会社爱霓威亚 | State judges system |
| CN111758117A (en) * | 2018-03-14 | 2020-10-09 | 欧姆龙株式会社 | Inspection system, recognition system, and learning data generation device |
| CN112067324A (en) * | 2019-06-11 | 2020-12-11 | 株式会社日立制作所 | Automatic inspection system |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP7175216B2 (en) * | 2019-02-15 | 2022-11-18 | ルネサスエレクトロニクス株式会社 | Anomaly detection device, anomaly detection system, anomaly detection method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2004234175A (en) * | 2003-01-29 | 2004-08-19 | Matsushita Electric Ind Co Ltd | Contents retrieval device and program therefor |
| JP2006031387A (en) * | 2004-07-15 | 2006-02-02 | Yamaha Motor Co Ltd | Image recognition device, image recognition method, image recognition program, and recording medium to which the image recognition program is recorded |
| JP2006154961A (en) * | 2004-11-25 | 2006-06-15 | Sumitomo Electric Ind Ltd | Traffic sound identification device, traffic sound decision program for making computer function as traffic sound identification device, recording medium and traffic sound decision method |
| US20080077543A1 (en) * | 2006-09-06 | 2008-03-27 | Takumi Kobayashi | Recognizing apparatus and recognizing method |
| US20080219565A1 (en) * | 2007-03-06 | 2008-09-11 | Kabushiki Kaisha Toshiba | Training device and pattern recognizing device |
| US20080240579A1 (en) * | 2007-03-30 | 2008-10-02 | Kabushiki Kaisha Toshiba | Video discrimination method and video discrimination apparatus |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP3340357B2 (en) * | 1997-08-25 | 2002-11-05 | 三菱電機株式会社 | Information processing device management system and information processing device management method |
-
2009
- 2009-03-26 JP JP2009077688A patent/JP2010231455A/en active Pending
-
2010
- 2010-03-18 KR KR1020117022207A patent/KR20110122748A/en not_active Application Discontinuation
- 2010-03-18 CN CN2010800120253A patent/CN102362282A/en active Pending
- 2010-03-18 WO PCT/IB2010/000579 patent/WO2010116222A1/en active Application Filing
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2004234175A (en) * | 2003-01-29 | 2004-08-19 | Matsushita Electric Ind Co Ltd | Contents retrieval device and program therefor |
| JP2006031387A (en) * | 2004-07-15 | 2006-02-02 | Yamaha Motor Co Ltd | Image recognition device, image recognition method, image recognition program, and recording medium to which the image recognition program is recorded |
| JP2006154961A (en) * | 2004-11-25 | 2006-06-15 | Sumitomo Electric Ind Ltd | Traffic sound identification device, traffic sound decision program for making computer function as traffic sound identification device, recording medium and traffic sound decision method |
| US20080077543A1 (en) * | 2006-09-06 | 2008-03-27 | Takumi Kobayashi | Recognizing apparatus and recognizing method |
| US20080219565A1 (en) * | 2007-03-06 | 2008-09-11 | Kabushiki Kaisha Toshiba | Training device and pattern recognizing device |
| US20080240579A1 (en) * | 2007-03-30 | 2008-10-02 | Kabushiki Kaisha Toshiba | Video discrimination method and video discrimination apparatus |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107389545A (en) * | 2016-05-17 | 2017-11-24 | 柳光龙 | Centring means for detection object |
| CN107389545B (en) * | 2016-05-17 | 2020-10-16 | 伸龙有限公司 | Centering device for detecting object |
| CN110199233A (en) * | 2017-01-27 | 2019-09-03 | 株式会社爱霓威亚 | State judges system |
| CN109961421A (en) * | 2017-12-25 | 2019-07-02 | 欧姆龙株式会社 | Data generating device, data creation method and data generate recording medium |
| CN109961421B (en) * | 2017-12-25 | 2023-11-24 | 欧姆龙株式会社 | Data generating device, data generating method, and data generating recording medium |
| CN111758117A (en) * | 2018-03-14 | 2020-10-09 | 欧姆龙株式会社 | Inspection system, recognition system, and learning data generation device |
| CN111758117B (en) * | 2018-03-14 | 2023-10-27 | 欧姆龙株式会社 | Inspection system, recognition system, and learning data generation device |
| CN108836574A (en) * | 2018-06-20 | 2018-11-20 | 广州智能装备研究院有限公司 | It is a kind of to utilize neck vibrator work intelligent sounding system and its vocal technique |
| CN112067324A (en) * | 2019-06-11 | 2020-12-11 | 株式会社日立制作所 | Automatic inspection system |
Also Published As
| Publication number | Publication date |
|---|---|
| JP2010231455A (en) | 2010-10-14 |
| WO2010116222A1 (en) | 2010-10-14 |
| KR20110122748A (en) | 2011-11-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111181939B (en) | Network intrusion detection method and device based on ensemble learning | |
| CN102362282A (en) | Signal classification method and signal classification device | |
| CN103400577B (en) | The acoustic model method for building up of multilingual speech recognition and device | |
| EP2323080B1 (en) | Sample attribution class estimation method, sample attribution class estimation program, and attribution class estimation apparatus | |
| CN106570513A (en) | Fault diagnosis method and apparatus for big data network system | |
| CN112735383A (en) | Voice signal processing method, device, equipment and storage medium | |
| KR102198273B1 (en) | Machine learning based voice data analysis method, device and program | |
| EP2028645A1 (en) | Method and system of optimal selection strategy for statistical classifications in dialog systems | |
| CN104573013A (en) | Category weight combined integrated learning classifying method | |
| CN103617429A (en) | Sorting method and system for active learning | |
| CN103761965B (en) | A kind of sorting technique of instrument signal | |
| CN107180084A (en) | Word library updating method and device | |
| CN106601230A (en) | Logistics sorting place name speech recognition method, system and logistics sorting system based on continuous Gaussian mixture HMM | |
| CN110455512B (en) | Rotary mechanical multi-integration fault diagnosis method based on depth self-encoder DAE | |
| KR20070094690A (en) | Speech signal classification system and method thereof | |
| WO2017217412A1 (en) | Signal processing device, signal processing method, and computer-readable recording medium | |
| US7233692B2 (en) | Method and computer program product for identifying output classes with multi-modal dispersion in feature space and incorporating multi-modal structure into a pattern recognition system | |
| CN107798351A (en) | A kind of personal identification method and system based on deep learning neutral net | |
| CN107357895A (en) | A kind of processing method of the text representation based on bag of words | |
| JP4848492B2 (en) | Signal identification method and signal identification apparatus | |
| US7181062B2 (en) | Modular classification architecture for a pattern recognition application | |
| CN112906672B (en) | Rail defect identification method and system | |
| Rosales-Pérez et al. | Infant cry classification using genetic selection of a fuzzy model | |
| WO2017217396A1 (en) | Signal processing device, signal processing method, and computer-readable recording medium | |
| CN116153336B (en) | Synthetic voice detection method based on multi-domain information fusion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20120222 |