JP2008008708A - Judgment knowledge creation system, judgment knowledge creation method, program, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a judgment knowledge creation system which creates judgment knowledge in a short period of time and facilitates an explanation of a derivation process of the judgment knowledge. <P>SOLUTION: The judgment knowledge creation system 20 determines: a characteristic amount suitable for an object to be inspected; and a parameter for calculating the characteristic amount, in an inspection system which inspects whether the object to be inspected is a non-defective item or a defective item on the basis of the characteristic amount which is obtained by performing characteristic amount extraction processing on acquired waveform data. The judgment knowledge creation system 20 comprises: a frequency parameter determination section 30 which determines a frequency parameter indicating an optimum frequency range in which an abnormal waveform component occurs as the parameter from the waveform data of the non-defective item and the defective item provided in advance; and a characteristic amount optimization section 40 which determines an effective characteristic amount on the basis of the waveform component of the optimum frequency range indicated by the frequency parameter determined by the frequency parameter determination section 30. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention provides the inspection object in an inspection apparatus that inspects whether the inspection object is a non-defective product or a defective product based on a feature amount obtained by performing feature extraction processing on the acquired waveform data. The present invention relates to a judgment knowledge creating apparatus for obtaining effective features that are algorithms suitable for the above.

  Conventionally, product inspection is performed in order to guarantee the quality of the product. In the abnormal sound inspection among the inspections of products, conventionally, a non-defective product or a defective product has been determined by a so-called sensory inspection in which a skilled inspector hears and determines the sound when the product is driven. However, the sensory test cannot be performed quantitatively and stably because it has a high personality and requires a high level of expertise.

  Therefore, for the purpose of labor saving to reduce costs and improving the reliability of products by improving the accuracy and variation of inspections, sensory inspections that had so far only relied on the intuition and experience of skilled inspectors In recent years, an abnormal sound measurement system that automatically performs the mechanism has been developed.

  Such an abnormal sound measurement system measures the sound and vibration of a product to be inspected, and uses the knowledge for inspection (abnormal sound or An algorithm that emphasizes abnormal vibrations) is required. The inspection knowledge is individually created based on the sound and vibration of the target product, environmental noise, and time change.

  Here, the creation of inspection knowledge, specifically, the selection of parameters used for features (algorithms for calculating feature values) has conventionally been performed by an analyst (who creates inspection knowledge for an abnormal sound inspection system). ). However, the number of combinations of parameters is enormous.

  For example, the sampling frequency is 40 kHz, the analysis time is 1 second, the frame width (unit time) is 0.1 seconds, and the frame shift width is 0.05 seconds. The frequency band that humans can hear is 10 Hz to 20 kHz. However, since the signal processing algorithm requires twice the resolution, the sampling frequency is 40 kHz. In addition, one frame is 0.1 seconds in which at least one wavelength of the minimum frequency 10 Hz that can be heard by humans enters.

The upper limit frequency and the lower limit frequency (for example, 3000 to 4000 Hz), which are parameters that determine the range of the frequency filter, can take values in increments of 10 Hz of 10 to 20 kHz. Therefore, the total number of combinations of the upper limit frequency and the lower limit frequency is about 2 million ( ₂₀₄₈ C ₂ ).

  In the range of the frequency filter, there are about 60,000 features (algorithms) including the total number of combinations of the parameters. The feature amount obtained by feature extraction (obtaining the feature amount by the feature (algorithm)) is, for example, the value of the waveform data s (t) in a certain range of time t (specified by the lower limit parameter TL and the upper limit parameter TH). Is the number of peaks PN (Peak Numbers) exceeding the threshold (specified by the parameter PT). In addition, there is a peak value PV (Peak Value) located at a certain rank (specified by the parameter RP) in a certain range of time t (specified by the lower limit parameter TL and the upper limit parameter TH).

Furthermore, it is necessary to select a time zone (frame) in which an abnormal component occurs. Since the total number of frames is 20 (= 1 / 0.05), there are 200 ( ₂₀ C ₂ ) possible time zones. Furthermore, the abnormal component has a stationary component and a shock component, and the stationary abnormal component can be confirmed by paying attention to the value of the feature value of each frame. This can be confirmed by paying attention to the amount of change before and after the feature amount. Therefore, it is conceivable that the time zone is selected by paying attention to the value of each frame and the time zone is selected by paying attention to the value of the amount of change before and after each frame. 200 × 2).

  The product of the total number of these combinations is about 48 trillion. Therefore, in the past, analysts have selected from such enormous combinations using advanced waveform analysis expertise. Moreover, the abnormal sound (or abnormal vibration) of the target product is very fine, and there are many things that cannot be understood unless a certain frequency band or time zone is emphasized. Therefore, a very large man-hour is required, which hinders automation of inspection work.

  Therefore, in Patent Document 1 and Non-Patent Documents 1 and 2, each type of feature amount is an individual, and various parameters (frequency parameter indicating a frequency band, calculation for calculating the feature amount) are extracted. A knowledge creation support device that searches for an optimal parameter using an optimization technique such as a genetic algorithm using a parameter as a gene and a combination of parameters as an individual is disclosed.

  The outline of the procedure of the parameter search method in this knowledge creation support apparatus will be described with reference to the flowchart shown in FIG.

  First, the knowledge creation support apparatus reads waveform data in which a non-defective product (OK) / defective product (NG) is discriminated in advance by a skilled inspector (S101). Each waveform data is associated with a determination result (OK / NG).

  Next, the knowledge creation support apparatus sets search conditions according to the user input (S102). Various algorithms can be applied to the parameter search. Search conditions when using a genetic algorithm include the following.

  Search conditions that define the operation of the genetic algorithm include the number of individuals, the crossover rate, the mutation rate, and the number of generations. The number of individuals is the number of individuals (solution candidates) used for the search. The crossover rate is the probability of crossing an individual. Mutation rate is the probability of mutating a gene in an individual. The number of generations is the number of generations to which the genetic algorithm is applied.

  Further, as a search condition for defining a search method, there is a selection of importance on the degree of separation or importance on the number of separations.

  In addition, the search conditions that define the end of the search include when the evaluation value indicating the degree of separation or the number of separations exceeds a certain value (exceeding the evaluation value), or when the number of generations with the same evaluation value exceeds a certain number of generations. It can be assumed that the search condition is satisfied when there is at least one point (exceeding the number of times) and at least one is included.

  Subsequently, the knowledge creation support device randomly generates a plurality of initial individuals (S103). As described above, the term “individual” refers to various feature quantities, such as the feature quantities PN, PV, and the like, as disclosed in JP-A-11-173909.

  Next, the value of each individual (feature value) is calculated (S104).

  Thereafter, the knowledge creation support apparatus totals the calculation results for each individual into OK products and NG products, and calculates an evaluation value for each individual (S105).

  That is, first, an evaluation value for each feature amount is obtained using Expression (1). Here, the coefficient α is NG when the average of the sensor data (hereinafter referred to as OK product) that is OK in the inspection result file is smaller than the average of the sensor data (hereinafter referred to as NG product) that is NG in the inspection result file. This is a coefficient for increasing the value when the product is detected at a high value. On the other hand, the count β is a coefficient for adding points when the OK product group and the NG product group are completely separated. Formula (1) is an example, and other formulas may be used.

Vn = α × β × (NGAven−OKAven) / OKσn (1)
here,
Vn = 0: OKσ = 0
α = 100: OKσ> 0, AND, (NGAven−OKAven)> 0
−20: OKσ> 0, AND, (NGAven−OKAven) ≦ 0
β = 2: (NGMinn−OKMaxn)> 0
β = 1: (NGMinn−OKMaxn) ≦ 0
However, OKAven: average value of feature value n of OK product OKσn: variance of feature value n of OK product NGven: average value of feature value n of NG product NGσn: variance of feature value n of NG product OKMinn, OKMaxn: feature of OK product Maximum value and minimum value of quantity n NGMinn, NGMaxn: Maximum value and minimum value of NG product feature quantity n: 0 to number of feature quantities -1.

  The final evaluation value Val is obtained as follows. That is, for the evaluation value Val, the expression (2) and the expression (2) ′ are properly used according to the search method designated by the user.

Val = Vm (2)
Vm: MAX (Vn: n = 0 to feature quantity −1)
Val = (w * Vm + (1-w) * Vk) / 1 (2) ′
Vm: MAX (Vn: n = 0 to feature quantity −1)
Vk: MAX (Vn: n = 0 to feature quantity −1 and m is excluded)
w: 0.0 to 1.0 (weight set by the user)
When the search method has priority on the degree of separation, Expression (2) is used, and when the search method has priority on the number of separations, Expression (2) ′ is used. Expression (2) ′ is an expression using the upper two evaluation values Vn, but may be a weighted average of arbitrary upper evaluation values.

  Next, the knowledge creation support apparatus confirms whether the evaluation value calculated based on each of the above expressions satisfies the search end condition that defines the search end (S106). The search condition that defines the end of the search is set by executing S101.

  When the search end condition is not satisfied (No in S106), the process returns to S103, a new individual is generated by crossover or mutation, an individual with a low evaluation value is replaced with a new individual, and S104 to S106 are repeated.

On the other hand, when the search end condition is satisfied (Yes in S106), the knowledge creation support apparatus determines the optimal solution (one parameter solution having the maximum evaluation value) that satisfies the search end condition (S107), and from the optimal solution, Non-defective / defective product discrimination knowledge (discrimination algorithm) is created (S108).
Japanese Patent Application Laid-Open No. 2004-279111 (released on October 7, 2004) OMRON Technics Vol. 43 No. 1 pp. 99-105 (2003) OMRON Technics Vol. 44 no. 1 pp. 48-53 (2004)

  Many customers who intend to introduce a new abnormal sound inspection system have conducted sensory inspections by the ears of skilled inspectors. In such a case, since the abnormal sound inspection system is used as a replacement for the sensory inspection, it is necessary to explain to the customer the consistency between the inspection standard of the skilled inspector so far and the discrimination knowledge of good / defective products. is there. Therefore, if the derivation process of the optimum solution cannot be explained, the customer's trust cannot be obtained.

  However, an optimization method such as a genetic algorithm performs a search at random based only on the evaluation function, and thus the derivation process cannot be rationally explained. In addition, when the search space is complex, it may take a very long time to derive an optimal solution.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to provide a determination knowledge creation device and a determination that can make determination knowledge in a short time and easily explain the derivation process of the determination knowledge. It is to realize a knowledge creation method, a program, and a recording medium.

  In order to solve the above-described problem, the determination knowledge creating apparatus according to the present invention is based on a feature amount obtained by performing a feature extraction process on acquired waveform data, and the inspection object is a non-defective product or a defective product. In the inspection apparatus for inspecting whether or not, a determination knowledge creating apparatus for obtaining an effective feature, which is an algorithm suitable for the inspection object, from abnormal waveform data given in advance and defective waveform data, an abnormal waveform A frequency parameter determining means for determining a frequency parameter indicating an optimal frequency band in which the component is generated, and the effective feature based on the waveform component in the optimal frequency band indicated by the frequency parameter determined by the frequency parameter determining means. Characteristic optimization means to be obtained.

  In order to solve the above-described problem, the determination knowledge creating method according to the present invention is based on the feature amount obtained by performing the feature extraction process on the acquired waveform data, and the inspection object is a non-defective product or a defective product. A determination knowledge generation method for a determination knowledge generation device for obtaining an effective feature that is an algorithm suitable for the inspection object in an inspection device for inspecting whether the waveform is a good product waveform and a defective product waveform that are given in advance A frequency parameter determining step for determining a frequency parameter indicating an optimum frequency band in which an abnormal waveform component is generated from data, and a feature optimization for obtaining the effective feature based on the waveform component in the optimum frequency band indicated by the frequency parameter. Including the step of converting.

  Here, the “feature” (algorithm) is an arithmetic algorithm for the waveform component. The “feature amount” is a value obtained from the waveform component by the above feature. “Feature extraction” means calculating the feature amount using the feature.

  As described above, the total number of feature amount calculation methods (that is, the number of “features” (algorithms)) as a part of the determination knowledge is considered in consideration of a combination of parameters for calculating feature amounts. The number is about 48 trillion. Among them, the number of combinations having the largest number is a frequency parameter indicating a frequency band used for the calculation of the feature amount, which is about 2 million.

  Conventionally, the optimum solution is searched omnidirectionally from about 48 trillion combinations. According to the above configuration, first, the optimum frequency band in which the abnormal waveform component is generated is determined as the frequency parameter. Thereafter, an effective feature (the effective feature may include a calculation parameter other than the frequency parameter) is obtained. Here, as a method of obtaining the effective feature, a known optimization method can be used.

  Therefore, the time for searching for one or several optimal frequency parameters from 2 million ways can be further shortened compared to the time for searching from 48 trillion ways. In addition, since the frequency parameter is determined first, the time for obtaining the optimum effective feature in the band indicated by the frequency parameter can be shortened. Further, the total time for searching for the frequency parameter and the time for finding the effective feature is also shorter than the time for searching from 48 trillion ways. In this way, by obtaining effective features step by step, an optimal solution of effective features as judgment knowledge can be output in a short time.

  In addition, compared with the case where the search is brute force from 48 trillion ways, it becomes easier to explain to the customer the derivation process in which the frequency parameter is determined first and the effective feature is obtained from the determined frequency parameter. You can get credit for.

  In particular, the frequency parameter is an important factor that determines whether or not an abnormal waveform component is captured. At the stage of determining the frequency parameter, the customer can confirm that the frequency parameter is optimal by comparing the non-defective / defective product waveform data with the determined frequency parameter. As a result, the customer's confidence in the process of deriving effective features can be further obtained.

  Further, in the determination knowledge creating device according to the present invention, the frequency parameter determining means extracts a plurality of unit frequency band waveform components from the non-defective waveform data and the defective waveform data, and based on each waveform component A first calculation unit that calculates a first value, and a separation state between a first value corresponding to good waveform data and a first value corresponding to defective waveform data for each unit frequency band. Second frequency calculation means for calculating a second value to be indicated, and determining the frequency parameter based on the second value.

  Here, in a certain unit frequency band, the first value corresponding to the non-defective product waveform data and the first value corresponding to the defective waveform data are separated from each other in the unit frequency band. It means that the waveform component of the non-defective product and the waveform component of the non-defective product are different, that is, an abnormal waveform component is generated in the waveform component of the defective product.

  Therefore, according to the above configuration, the second value indicating that the first value corresponding to the non-defective product waveform data and the first value corresponding to the defective waveform data are further separated. The unit frequency band possessed, that is, the unit frequency band in which the abnormal waveform component is generated can be determined as the optimum frequency band.

  Furthermore, in the determination knowledge creating apparatus according to the present invention, it is preferable that the first calculation means calculates a value indicating the intensity of the waveform component as the first value.

  As for the abnormal waveform component (stationary abnormal waveform component) generated regardless of time, the strength that is not found in a non-defective product is added to the defective product. Therefore, according to the above configuration, by setting the first value as a value indicating the intensity of the waveform component, the frequency band in which the stationary abnormal waveform component is generated can be determined as the optimum frequency band.

  The value indicating the intensity of the waveform component includes an average value or a median value of the intensity.

  Furthermore, in the determination knowledge creating apparatus according to the present invention, it is preferable that the first calculation means calculates a value indicating a variation in intensity for each unit time in the waveform component as the first value.

  The abnormal waveform component (impact abnormal waveform component) that occurs suddenly or periodically has a unit time intensity varying between different unit times. That is, assuming that one unit time is one frame, the strength of the frame including the shocking abnormal waveform component is larger than the strength of the frame not including the shocking abnormal waveform component. Therefore, according to the above configuration, the frequency band in which the shocking abnormal waveform component is generated can be determined as the optimum frequency band by using the strength of the frame for calculating the first value.

  The first value includes a variance value of intensity per unit time in the waveform component, a kurtosis degree of the waveform component, or a maximum value of the waveform component. When the strength of the frame including the impact abnormal waveform component is sufficiently larger than the strength of the frame including the impact abnormal waveform component, the dispersion value can be used. Further, when the maximum value of the waveform component data including the shocking abnormal waveform component is sufficiently larger than the maximum value of the waveform component data not including the shocking abnormal waveform component, the maximum value is set to the second value. It can be used as an indicator. In addition, when the number of frames including the shocking abnormal waveform component is sufficiently smaller than the total number of frames including the shocking abnormal waveform component, the kurtosis can be used as the second index. . The degree of impact can be evaluated by using the kurtosis degree. Also, by using the maximum value, it is possible to catch a single abnormality.

  Furthermore, in the determination knowledge creating apparatus according to the present invention, the second value for each unit frequency band is divided into a first value corresponding to non-defective waveform data and a first value corresponding to defective waveform data. When the rank when arranged in the order of further separation is the separation rank, the frequency parameter determining means preferably determines a plurality of unit frequency bands higher in the separation rank as the optimum frequency band.

  According to the above configuration, a plurality of unit frequency bands in which the first value corresponding to the non-defective product waveform data and the first value corresponding to the non-defective product waveform data are further separated are determined as the optimum frequency bands. can do. The frequency of the abnormal waveform component varies depending on the failure mode. Therefore, even when abnormal waveform components corresponding to a plurality of types of failure modes are generated, the frequency band in which each abnormal waveform component is generated can be determined as the optimum frequency band.

  Furthermore, in the determination knowledge creating apparatus according to the present invention, the second value for each unit frequency band is divided into a first value corresponding to non-defective waveform data and a first value corresponding to defective waveform data. When the order when arranged in the order of separation is the separation order, the frequency parameter determination means includes a selection means for selecting a plurality of unit frequency bands higher in the separation order as candidate frequency bands, FFT (Fast Fourier Transform) intensity graph corresponding to waveform data, graph creation means for creating an FFT intensity graph corresponding to the waveform data of the defective product, and a plurality of candidates selected by the FFT intensity graph and the selection means The frequency band is displayed on the display device, and the optimum frequency band determining means for determining the optimum frequency band from the plurality of candidate frequency bands according to the user input It is preferably provided and.

  According to the above configuration, the user can select which candidate frequency band is the optimum frequency band while comparing the displayed FFT intensity graph with the candidate frequency band. By checking the FFT intensity graph, the user can exclude the candidate frequency band selected, for example, due to the influence of abnormal noise or outliers from the optimum frequency band.

  Furthermore, in the determination knowledge creating device according to the present invention, the selecting means collectively includes the adjacent unit frequency bands when there are adjacent unit frequency bands in the plurality of unit frequency bands higher in the separation order. One candidate frequency band is preferable.

  In the inspection apparatus, it is desired to determine a plurality of failure modes by a single inspection. However, the frequency band in which the abnormal waveform component is generated differs depending on the failure mode. In addition, there is a limit to the number of frequency bands that can be inspected at once by the inspection apparatus (for example, 5).

  However, according to the above configuration, adjacent unit frequency bands are grouped into one candidate frequency band. As a result, by selecting this candidate frequency band, a plurality of adjacent unit frequency bands in which abnormal waveform components are generated can be determined as one optimum frequency band. The single optimum frequency band includes a plurality of abnormal waveform components, and when these abnormal waveform components are due to different failure modes, the plurality of failure modes can be inspected at a time.

  The determination knowledge creation device may be realized by a computer. In this case, a program for causing the determination knowledge creation device to be realized by a computer by operating the computer as each of the means, and a computer recording the same A readable recording medium falls within the scope of the present invention.

  As described above, the determination knowledge creating apparatus according to the present invention determines the frequency parameter indicating the optimum frequency band in which the abnormal waveform component is generated from the waveform data of the non-defective product and the waveform data of the defective product given in advance. Parameter determining means, and feature optimizing means for obtaining the effective feature based on the waveform component of the optimum frequency band indicated by the frequency parameter determined by the frequency parameter determining means. Thereby, judgment knowledge can be created in a short time, and the derivation process of the judgment knowledge can be easily explained.

  Prior to the description of an embodiment relating to the determination knowledge creating apparatus of the present invention, an inspection apparatus that performs pass / fail determination of a target product using determination rules (determination knowledge) set by the apparatus will be described.

  FIG. 2 is a block diagram showing the configuration of the inspection apparatus 10. As shown in FIG. 2, the inspection apparatus 10 includes a waveform acquisition unit 11, a feature extraction unit 12, a determination unit 13, an inspection result output unit 14, a feature amount calculation parameter storage unit 15, and a fuzzy rule storage unit. 16.

  For each product, the waveform acquisition unit 11 acquires waveform data from sound and vibration when the product is driven. Also, waveform data is acquired from vibrations during product driving. Here, it is assumed that the sound when the product is driven is obtained from the sound sensor 3 such as a microphone. The sound sensor 3 is arranged so as to be in contact with or close to the product, and collects sound generated when the product is driven. Further, the vibration at the time of driving the product is obtained from the acceleration sensor 3, for example. The acceleration sensor 3 is arranged so as to be in contact with or close to the product, and collects vibrations generated when the product is driven. The collected sound and vibration are amplified by an amplifier (not shown), and sent to the waveform acquisition unit 11 through the A / D converter 5.

  The feature amount calculation parameter storage unit 15 is a calculation algorithm for calculating a feature amount from waveform data, and stores effective features suitable for the waveform data. The feature quantity calculation parameter storage unit 15 can calculate from the frequency parameter indicating the upper limit frequency and the lower limit frequency of the frequency band to be filtered from the waveform data and the waveform data filtered by the frequency parameter as the effective feature. And feature amount identification information for identifying a feature amount used for quality determination. The feature amount is an element (measurement item) that is indicated by a value obtained from the waveform data and makes it possible to distinguish between a case where the product is a good product and a case where the product is a defective product. The feature amount is a physical characteristic used for determining whether or not a product is a non-defective product, and can be regarded as a characteristic (quality characteristic) that well represents the quality of the product.

  The feature quantity includes “a feature quantity with parameters” that requires some parameters to calculate the feature quantities and “feature quantities without parameters” that do not require parameters. Examples of the feature quantity without parameters include an average value and a variance value of the intensity of waveform data.

  On the other hand, as an example of the feature quantity having parameters, the value of the waveform data s (t) in a certain range of time t (specified by the lower limit parameter TL and the upper limit parameter TH) as shown in FIG. The number PN (Peak Numbers) of peaks exceeding (specified by PT) and the value PV of a peak located at a certain rank (specified by parameter RP) in a certain range of time t (specified by lower limit parameter TL and upper limit parameter TH) ( Peak Value). When the feature quantity calculation parameter storage unit 15 stores the feature quantity identification information of “feature quantity with parameters”, the values of various parameters for calculating the feature quantity (in the case of the feature quantity PN, the parameter TL, (TH and PT values) are also stored.

  The feature extraction unit 12 calculates the feature amount indicated by the feature amount identification information stored in the feature amount calculation parameter storage unit 15 from the waveform data acquired by the waveform acquisition unit 11. However, when the feature quantity indicated by the feature quantity identification information is a feature quantity with a parameter, the feature extraction unit 12 uses the values of various parameters stored in the feature quantity calculation parameter storage unit 15 to calculate the feature quantity. calculate.

  The fuzzy rule storage unit 16 stores a fuzzy rule used when the determination unit 13 performs pass / fail determination processing.

  The determination unit 13 performs fuzzy inference on the feature amount calculated by the feature extraction unit 12 in accordance with the fuzzy rules stored in the fuzzy rule storage unit 16, and performs pass / fail determination by comparison with an allowable range (threshold). .

  The inspection result output unit 14 displays, for example, information including a result determined by the determination unit 13 on a display device (not shown), that is, information on the inspection result of the product, or a storage device (not shown). Or store.

  As the specific configuration of the inspection apparatus 10, various types of conventionally known ones can be applied, and thus detailed description thereof is omitted. The determination knowledge creating apparatus according to the present invention includes a frequency parameter stored in the above-described feature amount calculation parameter storage unit 15, feature amount identification information, values of various parameters in the case of a feature amount with parameters, and a fuzzy rule storage unit 16. The fuzzy rule stored in is created.

(Configuration of judgment knowledge creation device)
FIG. 4 is a block diagram showing a schematic configuration of the determination knowledge creation device 20 of the present embodiment. As shown in FIG. 4, the determination knowledge creation device 20 includes a waveform database 21, an inspection result database 22, and a pass / fail determination algorithm generation unit 23. The determination knowledge creating apparatus 20 is connected with an input device 1 such as a keyboard and a mouse and a display device 2 as a user interface.

  As shown in FIG. 5, the waveform database 21 stores a plurality of waveform data obtained by the sensor 3 for sound and vibration generated when a sample product or the like is operated. The waveform database 21 stores a waveform ID for identifying the waveform data in association with each waveform data.

  The inspection result database 22 stores inspection result information in which the waveform ID is associated with the non-defective / defective product determination result for the waveform data corresponding to the waveform ID.

  The determination of non-defective / defective is performed in advance by a skilled inspector. For example, at the time of waveform data acquisition, an inspector listens to sound actually generated from the target product and makes a judgment, or the stored waveform data is reproduced, and a skilled inspector listens to the reproduced sound to make a judgment.

  The pass / fail judgment algorithm generation unit 23 is stored in the above-described feature amount calculation parameter storage unit 15 based on the waveform data stored in the waveform database 21 and the inspection result information stored in the inspection result database 22. The feature amount identification information and the values of various parameters (limited to the feature amount having parameters) and the fuzzy rules stored in the fuzzy rule storage unit 16 are created.

  The pass / fail judgment algorithm generation unit 23 includes a frequency parameter determination unit (frequency parameter determination unit) 30, a feature optimization unit (feature optimization unit) 40, and a rule creation unit 50.

  Based on the waveform data stored in the waveform database 21 and the inspection result information stored in the inspection result database 22, the frequency parameter determination unit 30 generates a frequency band (optimum) in which the abnormal waveform component is most frequently generated. Frequency band) is determined, and the determined upper limit frequency and lower limit frequency of the optimum frequency band are stored as frequency parameters in the feature amount calculation parameter storage unit 15 of the inspection apparatus 10. The detailed configuration of the frequency parameter determination unit 30 will be described later.

  The feature optimization unit 40 searches for an effective feature that is an arithmetic algorithm for calculating a feature amount (optimum feature amount) that can best separate a non-defective product and a defective product in the frequency band determined by the frequency parameter determination unit 30. To do. Further, when determining the feature quantity with parameters as the optimum feature quantity, the feature optimization unit 40 also obtains a computation parameter for computing the feature quantity as an effective feature.

  As described above, the combination of all parameters including frequency parameters and calculation parameters necessary for calculating the feature amount is an enormous amount of about 48 trillion. However, in this embodiment, the frequency parameter determination unit 30 first determines the frequency parameter. The total number of combinations of frequency parameters is about 2 million. For this reason, the total number of combinations of parameters other than the frequency parameters is within a range in which the total number can be evaluated. Therefore, the feature optimizing unit 40 evaluates all the “features (algorithms)” (in the case of a feature quantity with a parameter, including the value of the operation parameter) so that the non-defective product / defective product can be best separated (Hereinafter referred to as a feature value determination evaluation value) is calculated, and a feature (algorithm) having the largest feature value determination evaluation value is output to the rule creation unit 50 as an optimal solution (effective feature).

  Then, the feature optimization unit 40 stores the obtained effective feature in the feature amount calculation parameter storage unit 15. Specifically, the feature optimizing unit 40 accesses the feature amount calculation parameter storage unit 15 of the inspection apparatus 10, and feature amount identification information for identifying the feature amount calculated by the optimal solution, and the feature amount with parameters In the case of, the value of the calculation parameter is stored.

  The rule creation unit 50 uses the frequency parameter determined by the frequency parameter determination unit 30 and the feature value of the optimum solution obtained by the feature optimization unit 40 (in the case of a feature value with parameters, the value of the operation parameter). Based on the above, a fuzzy rule is created. Since the feature values and parameters used for creating the fuzzy rule have already been determined, the “IF THEN method” rule can be easily created by a known method. A membership function can also be created based on this. Then, the rule creation unit 50 accesses the fuzzy rule storage unit 16 of the inspection apparatus 10 and stores the created fuzzy rule.

(Configuration of frequency parameter determination unit)
Next, the detailed configuration of the frequency parameter determination unit 30 will be described. FIG. 1 is a block diagram illustrating a configuration of the frequency parameter determination unit 30. As shown in FIG. 1, the frequency parameter determination unit 30 includes a search condition setting unit 31, a waveform data division processing unit 32, an index calculation unit (first calculation means) 33, and an evaluation value calculation unit (second calculation). Means) 34, a candidate band selection section (selection means) 35, an FFT intensity graph creation section (FFT intensity graph creation means) 36, and an optimum band setting section (feature optimization means) 37.

  The waveform data division processing unit 32 extracts a plurality of frame data of a predetermined unit time shifted by the frame shift width from each waveform data stored in the waveform database 21, and performs a filtering process on the frame data. Is divided into waveform component data for each unit frequency band. At this time, the waveform data division processing unit 32 adds the waveform ID associated with the waveform data before division to each waveform component data.

  As shown in FIG. 6, the index calculation unit 33 reads the intensity of each waveform data from the waveform component data of the non-defective product and the defective product divided by the waveform data division processing unit 32 for each unit frequency band. Based on the intensity, an index (a first index and a second index to be described later) for evaluating the presence or absence of an abnormal waveform component is calculated.

  The abnormal waveform component is a waveform component that does not appear in the waveform component data of the non-defective product but appears in the waveform component data of the defective product. There are two types of abnormal waveform components, a stationary abnormal waveform component and an impact abnormal waveform component.

  The stationary abnormal waveform component is an intensity added to the intensity (power) of non-defective waveform component data regardless of the time direction. For example, it is a sound (boo sound) that is constantly generated.

  On the other hand, the shocking abnormal waveform component is one in which the intensity of unit time varies between different unit times. Here, one unit time is defined as one frame. In this case, the strength of the frame including the shocking abnormal waveform component is larger than the strength of the frame not including the shocking abnormal waveform component. For example, a shocking abnormal waveform component is a sound (kick sound) generated every certain period.

  Therefore, the index calculation unit 33 evaluates, for each unit of frequency data, a first index for evaluating the presence / absence of a stationary abnormal waveform component and the presence / absence of a shocking abnormal waveform component for each unit frequency band. A second index is calculated. As will be described later, the first index is an average value, and the second index is a variance value. Then, as illustrated in FIG. 9, the index calculation unit 33 generates index data in which the waveform ID, the unit frequency band, the first index, and the second index are associated with each other.

  The search condition setting unit 31 sets the above-described unit time, frame shift width, unit frequency band, first index, and second index as frequency parameter search conditions for determining the optimum frequency band. The search condition setting unit 31 sets the frequency parameter search condition in accordance with the user input. If there is no input from the user, the search condition setting unit 31 sets the default condition.

  The default unit time is 0.1 second, and the default frame shift width is 0.05 seconds. This is because the time required for at least one wave having the lowest frequency of 10 Hz that can be heard by humans is 0.1 second. The search condition setting unit 31 can set a predetermined multiple of the rotation cycle as a unit time according to a user input, for example, when the target product is rotationally driven such as an engine.

  The 1/3 octave band is defined as the default unit frequency band. The search condition setting unit 31 can set a 1 / n octave band (n ≠ 3), an equally spaced band for each predetermined frequency, and the like as a unit frequency band according to a user input.

  As a default of the first index, an average value of the intensity (power) of the waveform component data is defined. This is because the strength of the waveform component data of the defective product in which the stationary abnormal waveform component is generated is larger than the strength of the waveform component data of the non-defective product. As shown in FIG. 7, for example, when a stationary abnormal waveform component is generated at 8000 to 10000 Hz, in a unit frequency band of 8000 to 10000 Hz, it is less than the average value in the time direction in the intensity of the waveform component data of good products. The average value in the time direction in the intensity of the waveform component data of the non-defective product becomes large.

  The search condition setting unit 31 can set the median value of the intensity of the waveform component data as the first index according to the user input. In the case of the median value, the influence of noise and outliers is reduced.

  As a second index, when the intensity of the frame including the shocking abnormal waveform component is sufficiently larger than the intensity of the frame including the shocking abnormal waveform component, the intensity per unit time in the waveform component data, A dispersion value indicating variation can be used. As shown in FIG. 8, when an abnormal waveform component with impact is generated at 8000 to 10000 Hz, for example, the unit of the waveform component data of the defective product is more than the variance value of the intensity per unit time of the waveform component data of the non-defective product. The dispersion value of the intensity for each time increases.

  In addition, the kurtosis and the maximum value of the waveform component data can be set as the second index. When the maximum value of the intensity of the waveform component data including the abnormal abnormal waveform component is sufficiently larger than the maximum value of the intensity of the waveform component data not including the abnormal abnormal waveform component, the maximum value is used as the second index. Can be used. In addition, when the number of frames including the shocking abnormal waveform component is sufficiently smaller than the total number of frames including the shocking abnormal waveform component, the kurtosis can be used as the second index. . The degree of impact can be evaluated by using the kurtosis degree. Also, by using the maximum value, a single abnormality can be captured.

  The search condition setting unit 31 sets the second index according to the user input. Here, it is assumed that a variance value of “intensity per unit time” is set as the second index.

  The evaluation value calculation unit 34 is based on the inspection result information stored in the inspection result database 22 and the value of the first index calculated by the index calculation unit 33 for the non-defective waveform data for each unit frequency band. Then, an evaluation value indicating the degree of separation between the value of the first index calculated by the index calculation unit 33 with respect to the waveform data of the defective product (hereinafter referred to as a frequency determination evaluation value) is calculated. Similarly, the evaluation value calculation unit 34 uses the second index calculated by the index calculation unit 33 for the non-defective waveform data for each unit frequency band based on the inspection result information stored in the inspection result database 22. And an evaluation value for frequency determination indicating the degree of separation of the second index value calculated by the index calculating unit 33 with respect to the waveform data of the defective product. Then, as shown in FIG. 11, the evaluation value calculation unit 34 is an evaluation value in which a unit frequency band, a frequency determination evaluation value for the first index, and a frequency determination evaluation value for the second index are associated with each other. Generate data.

  In the present embodiment, the evaluation value calculation unit 34 calculates the SN ratio of the desired characteristic indicated by the following formula (3) as the frequency determination evaluation value.

  When the abnormal signal component is included, the first index (or the second index) calculated from the waveform data of the defective product is calculated from the waveform data of the non-defective product. The value increases with distance from the distribution of the first index (or second index). The SN ratio of the desired large characteristic has an advantage that it is a highly accurate value even if the number of defective waveform data is small. Depending on the production line, almost all manufactured products are non-defective products, and there is a low probability that defective products will occur. In such a case, when creating the determination knowledge, the number of waveform data of defective samples is smaller than the number of waveform data of non-defective samples. However, even in such a case, by using the S / N ratio shown in Expression (3), the first index (or the second index) calculated by the index calculation unit 33 on the non-defective waveform data can be obtained. The degree of separation between the first index (or the second index) calculated by the index calculation unit 33 with respect to the waveform data of the defective product can be accurately evaluated.

  FIG. 10 is a diagram showing the relationship between the magnitude of the SN ratio and the first index (or second index) calculated from the non-defective / defective product waveform data. In the unit frequency band in which the abnormal waveform component is included in the waveform component data of the defective product, the first index (or the second index) calculated from the waveform data of the defective product is the first index calculated from the waveform data of the non-defective product. (Or the second index) will depart from the distribution. At this time, as shown in FIG. 10, the SN ratio increases. On the other hand, in the unit frequency band in which the abnormal waveform component is not included in the waveform component data of the defective product, the first index (or the second index) calculated from the waveform data of the defective product is the first calculated from the waveform data of the non-defective product. It will be included in the distribution of one index (or second index). At this time, as shown in FIG. 10, the SN ratio becomes small.

  Note that the frequency determination evaluation value is not limited to the desired signal-to-noise ratio represented by Equation (3). In addition to this, an average value difference, a correlation ratio indicating a distribution difference, a Cullback library information amount, an F value, or the like may be used.

  The candidate band selection unit 35 is expected to generate a stationary abnormal waveform component based on the evaluation value for frequency determination calculated by the evaluation value calculation unit 34 (here, the SN ratio of the desired size characteristic). Frequency band candidates and frequency band candidates that are expected to generate shocking abnormal waveform components.

  A specific selection method of the candidate band selection unit 35 in the present embodiment will be described with reference to FIG. First, the candidate band selecting unit 35 arranges the SN ratios of the desired characteristics of the first index and the second index of each unit frequency band in descending order of values (hereinafter, referred to as separation order) (in FIG. 12, (The numbers in the circles indicate the order of increasing SN ratio). And the candidate zone | band selection part 35 makes the unit frequency zone corresponding to the S / N ratio of the isolation | separation ranking 1st as a 1st group. Here, an index (first index or second index) corresponding to the first SN ratio in the separation order is set as the index of the first group. Next, the candidate band selection unit 35 has the same index (first index or second index) corresponding to the S / N ratio of the second highest separation order as the first group index and the second highest separation order. When the unit frequency band corresponding to the SN ratio is adjacent to the unit frequency band included in the first group, the unit frequency band corresponding to the second highest SN ratio is included in the first group, and otherwise The unit frequency band corresponding to the S / N ratio of the second highest separation order is set as the second group. Thereafter, the candidate band selection unit 35 performs grouping of the unit frequency bands in the order of separation order. That is, the candidate band selecting unit 35 has the same index (first index or second index) corresponding to the SN ratio of the nth separation rank as the index of any group of the Nth group that has been created, and When the unit frequency band corresponding to the SN ratio of the nth separation order is adjacent to the unit frequency band included in the group, the unit frequency band corresponding to the SN ratio of the nth separation order is included in the group. Otherwise, the unit frequency band corresponding to the SN ratio of the nth separation order is set as the (N + 1) th group. In this manner, the candidate band selection unit 35 creates up to the sixth group, and selects the first to fifth groups as frequency band candidates (candidate frequency bands) where an abnormal waveform component is expected to occur. To do.

  When the group index is the first index, the group is set as a candidate frequency band of the stationary abnormal waveform component, and when the group index is the second index, the group is defined as the impact abnormal waveform component. Candidate frequency band.

  Note that depending on the target product, there may be a bias toward abnormal waveform components that are either stationary or shocking. Even in such a case, by selecting the candidate frequency band by the above method, the candidate frequency band is also biased to either stationary or impact depending on the bias of the abnormal waveform component.

  The FFT intensity graph creation unit 36 performs FFT analysis on each waveform data, totals the FFT analysis results of waveform data whose inspection result information corresponds to “non-defective product”, and waveform data whose inspection result information corresponds to “defective product”. The FFT analysis results are summarized. Then, as shown in FIG. 13, the FFT intensity graph creating unit 36 is configured with a time axis and a frequency axis for each of the inspection performance information “non-defective product” / “defective product” based on the tabulated FFT analysis results. A two-dimensional FFT intensity graph is created. In FIG. 13, the intensity is represented by shading (or color).

  The optimum band setting unit 37 is a frequency band (optimum) in which the feature amount and its parameters are optimized by the feature optimization unit 40 from among the candidate frequency bands of the first to fifth groups selected by the candidate band selection unit 35. Frequency band).

  In the present embodiment, the optimum band setting unit 37 includes the candidate frequency bands of the first to fifth groups selected by the candidate band selecting unit 35 and the inspection result information “non-defective” / An FFT intensity graph corresponding to each “defective product” is displayed on the display device 2. Then, the optimum band setting unit 37 determines the candidate frequency band designated by the user as the optimum frequency band from among the candidate frequency bands of the first to fifth groups.

  The optimum band setting unit 37 may determine a plurality of bands as the optimum frequency band. This is because different types of abnormal waveform components (stationary abnormal waveform components and impact abnormal waveform components) do not always occur in the same frequency band.

  The user can specify the optimum frequency band by comparing the FFT intensity graph with the candidate frequency band.

(Flow of judgment knowledge creation process)
Next, the flow of determination knowledge creation processing in the determination knowledge creation device 20 according to the present embodiment will be described with reference to the flowchart shown in FIG.

  First, the frequency parameter determination unit 30 and the feature optimization unit 40 of the pass / fail determination algorithm generation unit 23 read waveform data from the waveform database 21 and inspection result information from the inspection result database 22 (S1).

  Next, the search condition setting unit 31 and the feature optimization unit 40 of the frequency parameter determination unit 30 set search conditions for the optimal solution according to the user input to the input device 1 (S2).

  Specifically, the search condition setting unit 31 of the frequency parameter determination unit 30 sets a unit time, a frame shift width, a unit frequency band, a first index, and a second index, which are frequency parameter search conditions, according to a user instruction. . However, the search condition setting unit 31 sets the frequency parameter search condition to default when there is no user instruction. Here, it is assumed that the search condition setting unit 31 performs default setting.

  Further, the feature optimization unit 40 sets either separation priority or separation number priority as a search method according to a user instruction.

  Next, the frequency parameter determination unit 30 determines a frequency parameter indicating the upper and lower limits of the optimum frequency band in which the abnormal waveform component is generated (S3). Details of the process of S3 will be described later.

  Subsequently, the feature optimization unit 40 performs filtering processing on each waveform data stored in the waveform database 21 in the optimal frequency band determined by the frequency parameter determination unit 30 to generate post-filtering data. Then, the feature optimization unit 40 calculates a plurality of predetermined feature amounts based on each filtered data (S4).

  Here, the feature optimization unit 40 calculates the feature amount for each combination of the parameters for the feature amount with parameters. That is, for the feature quantity M having m parameter combinations, the feature optimization unit 40 calculates the feature quantities M-1 to M-m corresponding to each parameter combination.

  Next, the feature optimizing unit 40 performs grouping of non-defective products / defective products based on the inspection performance information stored in the inspection performance database 22 for each feature amount, and the separation degree or separation of the non-defective products / defective products. An evaluation value for determining the feature quantity indicating the number is calculated (S5).

  A known technique may be used as a method for calculating the feature value determination evaluation value. That is, the feature optimizing unit 40 obtains an evaluation value for each feature amount using the formula (1) described in the background art section, and when the search method “separation priority” is set in S2, When the evaluation value Val for determining the feature quantity is calculated according to the formula (2) described in the technology column and the search method “separation number priority” is set, the equation (2) ′ described in the background technology column is used. Then, an evaluation value Val for determining the feature amount is calculated.

  Then, the feature optimizing unit 40 sets the feature (algorithm) for calculating the feature amount having the highest feature amount determination evaluation value as the optimum solution (effective feature), and when the feature amount and the parameter are present, The value of the calculation parameter is output to the rule creation unit 50 (S6).

  Thereafter, the rule creation unit 50 creates a fuzzy rule based on the feature amount determined by the feature optimization unit 40 (in the case of a feature amount with a parameter, the value of the operation parameter). It is stored in the fuzzy rule storage unit 16 (S7). Thereby, the process is terminated.

(Flow of determining optimal frequency band)
Next, the flow of the optimum frequency band determination process in the frequency parameter determination unit 30 will be described with reference to the flowchart shown in FIG.

  First, the waveform data division processing unit 32 extracts a plurality of frame data having a unit time and a frame shift width set in S2 from each waveform data stored in the waveform database 21. Further, the waveform data division processing unit 32 performs a filtering process on each frame data, and divides it into waveform component data for each unit frequency band set in S2 (S11).

  Next, as shown in FIG. 6, the index calculation unit 33, for each waveform data, for each unit frequency band, the first index (here, the average value in the time direction) and the second index (here, the time direction) Is calculated (S12). Then, as illustrated in FIG. 9, the index calculation unit 33 generates index data in which the waveform ID, the unit frequency band, the first index, and the second index are associated with each other.

  Subsequently, the evaluation value calculation unit 34 refers to the inspection result information in the inspection result database 22 and divides each waveform ID into either a non-defective product or a defective product. Then, the evaluation value calculation unit 34 calculates the average value calculated from the non-defective waveform component data and the average calculated from the defective waveform component data for each unit frequency band based on the index data shown in FIG. The SN ratio of the desired characteristic indicating the degree of separation from the value is calculated. Similarly, for each unit frequency band, the evaluation value calculation unit 34 is a telescopic characteristic indicating the degree of separation between the dispersion value calculated from the non-defective waveform component data and the dispersion value calculated from the defective waveform component data. Is calculated (S13). Then, as illustrated in FIG. 11, the evaluation value calculation unit 34 generates evaluation value data in which the unit frequency band, the SN ratio with respect to the first index, and the SN ratio with respect to the second index are associated with each other.

  Thereafter, the candidate band selecting unit 35 selects five candidate frequency bands as shown in FIG. 12 according to the magnitude of the SN ratio calculated in S13 (S14). Since the details of the selection method have been described above, the description thereof is omitted here.

  Next, the FFT intensity graph creation unit 36 obtains an FFT intensity graph based on each waveform data. Then, the FFT intensity graph creation unit 36 aggregates the FFT intensity graphs corresponding to the non-defective waveform data, and generates one FFT intensity graph. Similarly, the FFT intensity graph creating unit 36 aggregates FFT intensity graphs corresponding to each waveform data of defective products for each type of defective mode, and generates one FFT intensity graph (S15).

  Next, the optimum band setting unit 37 displays the candidate frequency band selected by the candidate band selecting unit 35 and the FFT intensity graph created by the FFT intensity graph creating unit 36 on the display device 2, and re-candidate the candidate frequency band. A selection instruction is accepted (S16). At this time, the user can check whether the candidate frequency band is a band including an abnormal waveform component by comparing the FFT intensity graph with the candidate frequency band. When it is determined that reselection is necessary, the user may input a reselection instruction to the input device 1. On the other hand, when it is determined that reselection is not necessary, that is, there is a frequency band that appropriately includes the abnormal waveform component in the candidate frequency band, the user may select and input the frequency band as the optimum frequency band.

  Thereafter, the optimum band setting unit 37 confirms whether or not there is an instruction to reselect candidate frequency bands (S17). When a reselection instruction is input (Yes in S17), the search condition setting unit 31 displays a screen for prompting re-input of the frequency parameter search condition on the display device 2, and the frequency parameter search condition is reset according to the user input. Setting is performed (S18).

  On the other hand, when there is no reselection instruction (No in S17), the optimum band setting unit 37 determines the candidate frequency band designated by the user input as the optimum frequency band, and the frequency parameter indicating the upper and lower limits of the optimum frequency band Is stored in the feature amount calculation parameter storage unit 15 of the inspection apparatus 10 (S19).

(Modification)
In the above description, it is assumed that the inspection result information includes a non-defective / defective product determination result. However, the present invention is not limited to this, and the inspection result information may include class-specific information indicating the type (class) of a defect (abnormality) when the product is a defective product.

  In this case, the FFT intensity graph creation unit 36 may compile the FFT analysis results of the waveform data for each class and create an FFT intensity graph. Thereby, the user can confirm which class of abnormality is occurring at which frequency.

  In the above description, the feature optimization unit 40 calculates, for all features (algorithms), feature value determination evaluation values that can best separate non-defective / defective products, and the feature value determination evaluation value is the highest. A large feature is output to the rule creation unit 50 as an optimal solution (effective feature). However, the present invention is not limited to this, and the feature optimization unit 40 may use GA (genetic algorithm), NN (neural network), SVM (support vector machine), or the like as a method for searching for an optimal solution.

  In the above description, the optimum band setting unit 37 determines a band designated from the candidate frequency bands as the optimum frequency band according to the user input. However, the present invention is not limited to this, and the optimum band setting unit 37 may determine all candidate frequency bands as optimum frequency bands. In this case, a user operation can be omitted.

  In the above description, the candidate band selection unit 35 selects five candidate frequency bands in accordance with the method shown in FIG. However, the selection method is not limited to this. For example, the candidate band selection unit 35 may set only the unit frequency band having the first rank of the feature value determination evaluation value as the candidate frequency band. Or the candidate band selection part 35 is good also considering the unit frequency band to which the order of the evaluation value for feature-value determination to the 1st to n-th rank as a candidate frequency band.

  As described above, the determination knowledge creating apparatus 20 according to the present embodiment determines whether the inspection target is a good product or a defective product based on the feature value obtained by performing the feature value extraction process on the acquired waveform data. In the inspection apparatus 10 for inspecting whether or not there is, an effective feature that is an algorithm suitable for the inspection object is obtained. Then, the determination knowledge creating device 20 uses a frequency parameter determination unit (frequency parameter) that determines a frequency parameter indicating an optimum frequency band in which an abnormal waveform component is generated from waveform data of a good product and a defective product given in advance. And a feature optimizing unit (feature optimizing unit) 40 for obtaining the effective feature based on the waveform component in the optimum frequency band indicated by the frequency parameter determined by the frequency parameter determining unit 30.

  As described above, the calculation algorithm (“feature”) for calculating the feature amount is an enormous number of about 48 trillion when considering the combination of parameters for calculating the feature amount. Among them, the number of combinations having the largest number is a frequency parameter indicating a frequency band from which a feature amount is extracted, and there are about 2 million kinds.

  Conventionally, an optimal solution (effective feature) has been searched out of about 48 trillion combinations, but according to the above configuration, first, the optimal frequency at which an abnormal waveform component is generated A band is determined as a frequency parameter, and then an effective feature is obtained.

  Therefore, the time for searching for one or several optimal frequency parameters from 2 million ways can be further shortened compared to the time for searching from 48 trillion ways. In addition, since the frequency parameter is determined first, the time required to obtain the effective feature in the band indicated by the frequency parameter is short. Further, the total time for searching for the frequency parameter and the time for finding the effective feature is also shorter than the time for searching from 48 trillion ways. In this way, by obtaining effective features step by step, it is possible to output an optimum solution of feature amounts as judgment knowledge in a short time.

  In addition, compared with the case where the search is brute force from 48 trillion ways, it becomes easier to explain to the customer the derivation process in which the frequency parameter is determined first and the effective feature is obtained from the determined frequency parameter. You can get credit for.

  In particular, the frequency parameter is an important factor that determines whether or not an abnormal waveform component is captured. At the stage of determining the frequency parameter, the customer can confirm that the frequency parameter is optimal by comparing the non-defective / defective product waveform data with the determined frequency parameter. As a result, it is possible to further obtain customer confidence in the process of deriving the feature amount.

  Further, the frequency parameter determination unit 30 extracts waveform components of a plurality of unit frequency bands from the non-defective product waveform data and the defective product waveform data, and indexes indicating the characteristics of the respective waveform components (first index and second index: The index calculation unit (first calculation means) 33 for calculating the first value), and the separation state between the index corresponding to the non-defective waveform data and the index corresponding to the waveform data of the defective product for each unit frequency band. An evaluation value calculation unit (second calculation unit) 34 that calculates an evaluation value for frequency determination (second value) to be shown, and a candidate band selection unit (selection unit) that determines a frequency parameter based on the evaluation value for frequency determination 35, an FFT intensity creating unit (FFT intensity graph creating unit) 36 and an optimum band setting unit (optimum frequency band determining unit) 37.

  As described above, the index corresponding to the non-defective product waveform data and the index corresponding to the non-defective product waveform data are separated in a certain unit frequency band. This means that an abnormal waveform component is generated.

  Therefore, a unit frequency band with an evaluation value for frequency determination indicating that the index corresponding to the waveform data of the non-defective product and the index corresponding to the waveform data of the defective product are further separated, that is, an abnormal waveform component is generated. The unit frequency band being used can be determined as the optimum frequency band.

  The index calculator 33 calculates a first index (for example, an average value or a median value) indicating the intensity of the waveform component. The strength of a stationary abnormal waveform component not added to a non-defective product is added to a defective product. Therefore, based on the frequency determination evaluation value based on the first index, the candidate band selection unit 35 and the optimum band setting unit 37 can determine the frequency band in which the stationary abnormal waveform component is generated as the optimum frequency band. .

  In addition, the index calculation unit 33 calculates a second index (for example, a variance value, a kurtosis degree, a maximum value) indicating the variation in the time direction of the intensity of the waveform component. The abnormal waveform component of impact property is a thing in which strength present in non-defective products varies with time in defective products. Therefore, based on the frequency determination evaluation value based on the second index, the candidate band selection unit 35 and the optimal band setting unit 37 can determine the frequency band in which the impact abnormal waveform component is generated as the optimal frequency band. .

  Also, when the frequency determination evaluation values for each unit frequency band are arranged in the order in which the index corresponding to the non-defective product waveform data and the index corresponding to the non-defective product waveform data are separated, the separation order is used. The candidate band selection unit 35 sets a plurality of unit frequency bands higher in the separation order as candidate frequency bands, and the optimum band setting unit 37 determines the candidate frequency band as the optimum frequency band.

  Thereby, a plurality of unit frequency bands in which the index corresponding to the non-defective product waveform data and the index corresponding to the defective waveform data are further separated can be determined as the optimum frequency band. The frequency of the abnormal waveform component varies depending on the failure mode. Therefore, even when abnormal waveform components corresponding to a plurality of types of failure modes are generated, the frequency band in which each abnormal waveform component is generated can be determined as the optimum frequency band.

  Furthermore, the determination knowledge creation device 20 includes a candidate band selection unit (selection unit) 35 that selects a plurality of unit frequency bands of higher separation order as candidate frequency bands, an FFT intensity graph corresponding to non-defective waveform data, and An FFT intensity graph creating unit (FFT intensity graph creating means) 36 for creating an FFT intensity graph corresponding to the waveform data of the defective product, and a plurality of candidate frequency bands selected by the FFT intensity graph and the candidate band selecting unit 35 are displayed. An optimum band setting unit (optimum frequency band determining means) 37 that is displayed on the apparatus 2 and determines an optimum frequency band from among the plurality of candidate frequency bands according to a user input is provided.

  According to the above configuration, the user can select which candidate frequency band is the optimum frequency band while comparing the displayed FFT intensity graph with the candidate frequency band. By checking the FFT intensity graph, the user can exclude the candidate frequency band selected, for example, due to the influence of abnormal noise or outliers from the optimum frequency band.

  Further, in the determination knowledge creating apparatus 20, when there are adjacent unit frequency bands in the plurality of unit frequency bands higher in the separation order, the candidate band selecting unit 35 collects these adjacent unit frequency bands as 1 One candidate frequency band.

  In the inspection apparatus, it is desired to determine a plurality of failure modes by a single inspection. However, the frequency band in which the abnormal waveform component is generated differs depending on the failure mode. In addition, there is a limit to the number of frequency bands that can be inspected at once by the inspection apparatus (for example, 5).

  However, according to the above configuration, adjacent unit frequency bands are grouped into one candidate frequency band. As a result, by selecting this candidate frequency band, a plurality of adjacent unit frequency bands in which abnormal waveform components are generated can be determined as one optimum frequency band. The single optimum frequency band includes a plurality of abnormal waveform components, and when these abnormal waveform components are due to different failure modes, the plurality of failure modes can be inspected at a time.

  The present invention is not limited to the above-described embodiments, and various modifications can be made within the scope of the claims. That is, embodiments obtained by combining technical means appropriately modified within the scope of the claims are also included in the technical scope of the present invention.

  Finally, each block of the determination knowledge creation device 20 and the pass / fail determination algorithm generation unit 23 may be configured by hardware logic, or may be realized by software using a CPU as follows.

  That is, the determination knowledge creating device 20 includes a CPU (central processing unit) that executes instructions of a control program that realizes each function, a ROM (read only memory) that stores the program, and a RAM (random access memory) that expands the program. ), A storage device (recording medium) such as a memory for storing the program and various data. An object of the present invention is a recording medium in which a program code (execution format program, intermediate code program, source program) of a control program of the determination knowledge creating apparatus 20 which is software for realizing the above-described functions is recorded so as to be readable by a computer. Can also be achieved by reading the program code recorded on the recording medium and executing it by the computer (or CPU or MPU).

  Examples of the recording medium include tapes such as magnetic tapes and cassette tapes, magnetic disks such as floppy (registered trademark) disks / hard disks, and disks including optical disks such as CD-ROM / MO / MD / DVD / CD-R. Card system such as IC card, IC card (including memory card) / optical card, or semiconductor memory system such as mask ROM / EPROM / EEPROM / flash ROM.

  Further, the determination knowledge creating device 20 may be configured to be connectable to a communication network, and the program code may be supplied via the communication network. The communication network is not particularly limited. For example, the Internet, intranet, extranet, LAN, ISDN, VAN, CATV communication network, virtual private network, telephone line network, mobile communication network, satellite communication. A net or the like is available. Also, the transmission medium constituting the communication network is not particularly limited. For example, even in the case of wired such as IEEE 1394, USB, power line carrier, cable TV line, telephone line, ADSL line, etc., infrared rays such as IrDA and remote control, Bluetooth ( (Registered trademark), 802.11 wireless, HDR, mobile phone network, satellite line, terrestrial digital network, and the like can also be used. The present invention can also be realized in the form of a computer data signal embedded in a carrier wave in which the program code is embodied by electronic transmission.

  The present invention can be applied to an apparatus that creates determination knowledge in an inspection apparatus that performs determination of a non-defective product / defective product by using sound and vibration generated from a product such as an engine.

It is a block diagram which shows the structure of the frequency parameter determination part with which the determination knowledge preparation apparatus which concerns on this embodiment is provided. It is a block diagram which shows the structure of the inspection apparatus which concerns on this embodiment. It is a figure which shows the example of a calculation of the feature-value with a parameter. It is a block diagram which shows schematic structure of the determination knowledge preparation apparatus which concerns on this embodiment. It is a figure which shows an example of the waveform data which a waveform database memorize | stores. It is a figure which shows the correspondence of the intensity value of waveform component data, and a 1st, 2nd parameter | index. It is a figure which shows the difference of the average value of the waveform intensity | strength of a non-defective product / defective product when the stationary abnormal waveform component has generate | occur | produced. It is a figure which shows the difference of the dispersion | distribution value of the waveform intensity | strength of a non-defective product / defective product when the abnormal waveform component of impact property has generate | occur | produced. It is a figure which shows the data structure of the parameter | index data which an parameter | index calculating part produces | generates. It is a figure which shows the relationship between the magnitude | size of SN ratio, and the 1st parameter | index and the 2nd parameter | index calculated from the waveform data of a good article / defective article. It is a figure which shows the data structure of the evaluation value data which an evaluation value calculating part produces | generates. It is a figure which shows the selection method of the candidate of the frequency band where it is estimated that the abnormal waveform component has generate | occur | produced. It is a figure which shows an example of a FFT intensity | strength graph. It is a flowchart which shows the flow of the creation process of the judgment knowledge in this embodiment. It is a flowchart which shows the flow of the determination process of an optimal frequency band (frequency parameter). It is a flowchart which shows the flow of the creation processing of the conventional judgment knowledge.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Input device 2 Display apparatus 10 Inspection apparatus 15 Feature-value calculation parameter memory | storage part 16 Fuzzy rule memory | storage part 20 Judgment knowledge creation apparatus 21 Waveform database 22 Inspection performance database 30 Frequency parameter determination part (frequency parameter determination means)
31 Search condition setting unit 32 Waveform data division processing unit 33 Index calculation unit (first calculation means)
34 Evaluation Value Calculation Unit (Second Calculation Unit)
35 Candidate band selection section (selection means)
36 FFT strength graph creation unit (FFT strength graph creation means)
37 Optimum band setting unit (optimum frequency band determining means)
40 Feature optimization unit (feature optimization means)
50 Rule creation department

Claims (12)

An algorithm suitable for the inspection object in an inspection apparatus that inspects whether the inspection object is a non-defective product or a defective product based on the feature amount obtained by performing feature extraction processing on the acquired waveform data It is a judgment knowledge creation device for obtaining an effective feature,
A frequency parameter determining means for determining a frequency parameter indicating an optimum frequency band in which an abnormal waveform component is generated from the waveform data of a good product and a defective product given in advance;
A determination knowledge creating apparatus comprising: a feature optimization unit that obtains the effective feature based on a waveform component in an optimum frequency band indicated by the frequency parameter determined by the frequency parameter determination unit. The frequency parameter determining means includes
First calculation means for extracting a waveform component of a plurality of unit frequency bands from the waveform data of the non-defective product and the waveform data of the defective product, and calculating a first value based on each waveform component;
Second calculation means for calculating, for each unit frequency band, a second value indicating a separation state between a first value corresponding to non-defective product waveform data and a first value corresponding to defective product waveform data; With
The determination knowledge creating apparatus according to claim 1, wherein the frequency parameter is determined based on the second value.   The determination knowledge creating apparatus according to claim 2, wherein the first calculation means calculates a value indicating the intensity of the waveform component as the first value.   The determination knowledge creating apparatus according to claim 2, wherein the first calculation unit calculates an average value or a median value of the intensities of the waveform components as the first value.   The determination knowledge creating apparatus according to claim 2, wherein the first calculation unit calculates a value indicating variation in intensity for each unit time in the waveform component as the first value.   The first calculation means calculates, as the first value, a variance value of intensity per unit time in the waveform component, a kurtosis degree of the waveform component, or a maximum value of the waveform component. The determination knowledge creating apparatus according to claim 2. The order when the second value for each unit frequency band is arranged in the order in which the first value corresponding to the non-defective waveform data and the first value corresponding to the defective waveform data are further separated. When using separation order
3. The determination knowledge creating apparatus according to claim 2, wherein the frequency parameter determining means determines a plurality of unit frequency bands higher in the separation order as the optimum frequency band. The order when the second value for each unit frequency band is arranged in the order in which the first value corresponding to the non-defective waveform data and the first value corresponding to the defective waveform data are further separated. When using separation order
The frequency parameter determining means includes
Selection means for selecting a plurality of unit frequency bands higher in the separation order as candidate frequency bands;
An FFT intensity graph corresponding to the waveform data of the non-defective product, and a graph creating means for creating an FFT intensity graph corresponding to the waveform data of the defective product;
The FFT intensity graph and a plurality of candidate frequency bands selected by the selecting means are displayed on a display device, and an optimum frequency band determining means for determining an optimum frequency band from the plurality of candidate frequency bands according to a user input The determination knowledge creating apparatus according to claim 2, further comprising:   In the case where there are adjacent unit frequency bands in the plurality of unit frequency bands higher in the separation order, the selecting means collectively sets the adjacent unit frequency bands into one candidate frequency band. The determination knowledge creating apparatus according to claim 8. An algorithm suitable for the inspection object in an inspection apparatus that inspects whether the inspection object is a non-defective product or a defective product based on the feature amount obtained by performing feature extraction processing on the acquired waveform data A determination knowledge creation method of a determination knowledge creation device for obtaining an effective feature is,
A frequency parameter determining step for determining a frequency parameter indicating an optimum frequency band in which an abnormal waveform component is generated from waveform data of a good product and a defective product given in advance;
And a feature optimization step for obtaining the effective feature based on a waveform component in an optimum frequency band indicated by the frequency parameter.   The program for functioning as said each means in the determination knowledge preparation apparatus of any one of Claims 1-9.   A computer-readable recording medium on which the program according to claim 11 is recorded.

Images