JP2006038478A - Quality determination method and device for specimen - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a quality determination device detecting/determining a microdefect with a reduced determination error as to quality of the specimen. <P>SOLUTION: This quality determination device is provided with: an oscillation means for oscillating the specimen by a jet flow of a fluid; a vibration sensor for detecting a vibration waveform of the specimen oscillated by the oscillation means; a waveform cutting-out means for cutting out a sample waveform from the vibration waveform detected by the vibration sensor; a characteristic amount calculating means for calculating a characteristic amount in the vibration of the specimen, based on the sample wave form cut out by the waveform cutting-out means; a determination means for inputting the characteristic amount calculated by the characteristic amount calculating means, into a boundary learning type neural network constructed by inputting the characteristic amount of a nondefective sample of the specimen, and for finding a difference between the characteristic amount of the specimen and the characteristic amount of the nondefective sample, so as to determine whether the error is within a range of a threshold value or not; and a display means for displaying a determination result determined by the determination means. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a method and apparatus for determining the quality of an object, and more particularly to a method and apparatus suitable for determining the quality of a part that is produced by pressing such as the body of an automobile or the exterior of an electrical appliance.

Products produced by press working, such as automobile bodies and exteriors of electrical appliances, have locations where stress is concentrated due to the processing stage and the processing shape. In such a place, defects such as cracks and constriction are likely to occur. Conventionally, manual visual inspection has been the mainstream, but inspections are different because of differences in pass / fail judgments by inspectors, inspection errors due to manual operation, noise, etc. are not suitable for the environment when performing inspections, etc. There is a need for automation.
Therefore, an examination method as shown in Patent Document 1 has been proposed as a method for determining the quality of a subject.

  The inspection method disclosed in Patent Document 1 is as follows. First, the subject is hammered with an impulse hammer, and the vibration waveform of the subject obtained by the hammering is collected. Next, the collected vibration waveform is taken out. The extracted vibration waveform is subjected to a fast Fourier transform process to obtain a frequency spectrum of the subject. Next, the peak frequency of the frequency spectrum of the subject obtained as described above is compared with the peak frequencies of the frequency spectrum of a plurality of non-defective products that have been measured in advance. The quality of the subject is determined based on whether or not it exists within the distribution of the peak frequency.

According to such an inspection method, it is possible to keep the determination standard constant because there is no need for human intervention in the inspection of the subject. In addition, the automation of the inspection process eliminates so-called inadvertent mistakes. Furthermore, it is not necessary for the worker to work in a noisy environment.
JP-A-9-171008

  According to the inspection method disclosed in Patent Document 1, it is possible to perform an inspection without any manual intervention. However, the subject may vary depending on the shape and production accuracy of the subject. In such a case, the statistical distribution for each peak frequency of the reference non-defective product becomes complicated, and there is a possibility that a determination error may occur. In addition, since the fast Fourier transform process is not suitable for analyzing short-time phenomena or transient phenomena, it may not be possible to deal with defects that are different from normal defects in the subject. In addition, in hammering with an impulse hammer, the range in which the subject can be vibrated is limited, and it is considered difficult to excite high-frequency vibrations that can be used to detect the presence of minute flaws. . Further, the hammering by the impulse hammer may give a secondary defect to the subject depending on the shape and structure of the subject.

  In the present invention, the above-mentioned problems are solved, the determination error concerning the quality of the subject is small, the quality can be judged even for an unexpected defect, and a secondary defect is given when inspecting the subject. It is an object of the present invention to provide a quality determination method and apparatus.

  In order to achieve the above object, it is considered that a determination means capable of recognizing a difference between a non-defective vibration waveform and other vibration waveforms among vibrations generated when the subject is vibrated is necessary. It is done. Further, as a means for vibrating the subject, the frequency band to be vibrated can be changed according to the material, shape, strength, etc. of the subject, and there is no contact between the solid and the subject. It is considered preferable to do so.

Therefore, the quality determination method of the subject according to the present invention is a method for determining whether or not the subject is a non-defective product, wherein the non-defective sample of the subject is vibrated by a jet of fluid, Detecting the vibration of the non-defective sample, cutting out the sample waveform from the vibration waveform of the non-defective sample obtained by vibration detection, calculating a linear prediction coefficient based on the sample waveform, and using the linear prediction coefficient as a feature amount for learning, A boundary learning type neural network is constructed by using a plurality of feature values of the non-defective sample, and the feature quantity of the subject to be judged is determined in the boundary learning type neural network in the same manner as the feature quantity of the non-defective sample. The difference between the feature quantity of the subject and the feature quantity of the non-defective product sample that is the target of the pass / fail judgment by the output of the boundary learning neural network is input, As a numerical value boss was in the range, and judging the quality of the subject according to whether the number is a value within the range of the threshold value defined by the boundary learning neural network.
In the above-described method for determining whether or not a subject is good or bad, the excitation may be performed so that the subject generates vibrations in a high frequency band of 20 kHz or more.

  An object pass / fail determination apparatus according to the present invention for realizing the above method is an apparatus for determining whether or not an object is a non-defective product, and an excitation means for vibrating the object with a jet of fluid. A vibration sensor for detecting a vibration waveform of the subject vibrated by the vibration means, a waveform cut-out means for cutting a sample waveform from the vibration waveform detected by the vibration sensor, and the wave cut-out means Feature amount calculation means for calculating a feature amount with respect to the vibration of the subject based on the sample waveform, and a boundary learning type in which the feature amount calculated by the feature amount calculation means is constructed by inputting the feature amount of a non-defective sample of the subject A determination unit that inputs to the neural network and obtains a difference between the feature amount of the subject and the feature amount of the non-defective sample and determines whether the difference is within a threshold range; Display means for displaying a determination result by the determining means, comprising the.

In the apparatus for determining pass / fail of a subject, the vibration means has a nozzle for ejecting fluid, and the subject is subjected to vibration in a high frequency region of 20 kHz or more. It is good to be configured so that the speed can be changed.
Furthermore, a plurality of vibration sensors may be provided for one subject.

  According to the method and apparatus as described above, since the linear prediction coefficient is used as the feature amount of the vibration generated by the subject, the change in the feature amount is possible even for a minute defect that does not appear as a change in amplitude in the vibration frequency of the subject. Can be captured. For this reason, the inspection accuracy can be improved. In addition, a boundary learning type neural network is constructed and output as a numerical value within the set range of the feature quantity of the subject to be judged as good or bad, and the difference between the feature quantity of the subject and the feature quantity of the non-defective sample, Since the quality of the subject is determined based on the numerical value, the feature value of the vibration waveform that is generated when the subject subject to pass / fail determination is vibrated and the feature value of the vibration waveform that is generated when the non-defective product is vibrated It is clear how much the difference is. For this reason, even if an unexpected defect occurs in the subject, it can be determined as “defective” without misidentifying it. Further, by adopting a configuration in which a fluid is used when vibrating the subject, unlike in the case of hammering the subject, concentrated stress may be generated at the place of vibration of the subject and secondary defects may occur. There is no. Further, the subject can be vibrated in a high frequency region of 20 kHz or more by changing the nozzle diameter of the vibration means, the ejection speed (ejection pressure), or the like. For this reason, signs of minute scratches appearing in the vibration can also be represented in the waveform. Here, since a linear prediction coefficient is used as a feature amount, a temporally short vibration change appearing in a waveform in a high frequency region can be captured, so that the inspection accuracy is high.

  DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments according to a method and an apparatus for determining quality of a subject of the present invention will be described below with reference to the drawings. The following embodiment is a partial form according to the present invention, and the object quality determination method and apparatus of the present invention includes various forms as long as the main part is not changed.

  First, a quality determination device for a subject according to the present invention will be described with reference to FIG. The subject pass / fail determination apparatus 50 according to the present embodiment includes a vibration unit 10 for vibrating the subject 100 (see FIG. 2), and vibration of the subject 100 renewed by the vibration unit 10. The vibration sensor 20 to be measured, the filter 30 that removes the vibration waveform in the unnecessary frequency band from the vibration waveform obtained by the vibration sensor 20, and increases the SN ratio (Signal to Noise ratio), and filtered by the filter 30. The basic configuration is a computer 40 that analyzes the obtained waveform and determines the quality of the subject.

  In the object pass / fail determination apparatus 50 based on the above-described configuration, the vibration means 10 may be configured to vibrate the subject 100 with a jet of fluid. The fluid may be a liquid or a gas, but in the present embodiment, it will be described as a gas that is easy to handle, for example, air. The configuration of the vibration means 10 may be anything as shown in FIG. That is, the nozzle 11 is provided near the subject 100 and ejects compressed air to the subject 100, and the compressed air is ejected from the nozzle 11 in accordance with the vibration time and vibration pattern of the subject 100. This is a configuration including a solenoid valve 12 for electrical control and a tank 13 for storing compressed air for excitation supplied to the nozzle 11. The compressed air supply source 14 stored in the tank 13 may be configured to include a small compressor as a part of the vibration means 10 or to be directly supplied from a large compressor such as a factory. In the vibration means 10 configured as described above, the solenoid valve 12 is electromagnetically opened and closed in accordance with a control signal from the outside (for example, the computer 40). The supply of the compressed air to the nozzle 11 may be performed continuously or intermittently according to the control signal. Further, in the vibration means 10 having the above-described configuration, by changing the diameter of the nozzle 11, the distance to the subject, the pressure / jetting speed of the compressed air, etc., the intensity of vibration and the vibration applied to the subject are changed. The frequency can be changed. For this reason, it is possible to vibrate the subject 100 to a high frequency band of 20 kHz or higher where it is easy to find minute scratches.

  The vibration sensor 20 may be provided near the inspection location of the subject 100 (a location where defects such as cracks and constrictions are likely to occur) or in the vicinity of the inspection location. Regardless of contact type. Further, when actually inspecting the subject 100, it is desirable to provide the vibration sensors 20 at a plurality of locations of the subject 100. Thereby, flaw detection can be performed at a plurality of locations by a single vibration inspection.

  Next, details of the configuration of the computer 40 will be described. The computer 40 is first provided with A / D conversion means 41 for converting the analog signal filtered by the filter 30 into a digital signal that can be electrically analyzed. Connected to the output side of the A / D conversion means 41 is a waveform cutout means 42 that cuts out a waveform for calculating a feature value, which will be described later, from the signal after digital conversion.

  The waveform cutout means 42 cuts out a waveform in a certain time band for calculating a feature amount described later from vibration waveforms (signals) input from the vibration sensor 20. By performing this waveform cut-out, the calculation for calculating the feature amount can be simplified.

  Filter means 43 is connected to the output side of the waveform cutout means 42. The filter unit 43 is a filtering unit configured to pass only a signal in a frequency band of vibration generated from the subject 100 having defects such as cracks. Although the filter means 43 is described as filtering the signal after digital conversion, it may be connected to the input side upstream of the A / D conversion means 41 as an analog filter like the filter 30 described above. .

  A linear prediction coefficient calculation unit (feature amount calculation unit) 44 is connected to the output side of the filter unit 43. The linear prediction coefficient calculation unit 44 calculates a linear prediction coefficient that is a feature amount of the signal that has passed through the filter unit 43. As a calculation method, a method based on the maximum entropy method can be cited.

ここで、ｐはＳ（ｎ）以前の時刻におけるモデルの次数を示し、ε（ｎ）は予測値の予測誤差を示す。このような数式１に用いられる係数ａ_iが線形予測係数である。 Here, the linear prediction coefficient is one of feature quantities used in signal processing, and has a feature that can store the frequency information of a signal in a compressed form. For example, when the measured value of a signal at a certain time n is S (n), the predicted value of S (n) is obtained as a measurement interval (sampling interval) Δt by the calculation of Equation 1.

Here, p represents the order of the model at a time before S (n), and ε (n) represents the prediction error of the predicted value. The coefficient a _i used in Equation 1 is a linear prediction coefficient.

  The above Equation 1 shows that a signal at an arbitrary time can be expressed as a weighted sum of the previous signals, and is called an autoregressive model. Further, the power spectrum of the waveform signal obtained by the vibration of the subject can be calculated based on the linear prediction coefficient. For this reason, it can be said that there is a correspondence between the power spectrum of the signal and the linear prediction coefficient. Therefore, if there is a difference in the power spectrum, a difference is observed in the linear prediction coefficient, and the linear prediction coefficient can be used as a feature amount of the detection signal (vibration waveform). In addition, since the power spectrum can be normally calculated with a small number of linear prediction coefficients, the number of feature amounts may be smaller than when other feature amounts are used. Usually, in order to identify the characteristics of a signal, the number of linear prediction coefficients is about 20 and reaches a level at which self-diagnosis is possible.

  The linear prediction coefficient obtained by the maximum entropy method does not change the coefficient value even when the model order p is changed. When calculating a spectrum using the linear prediction coefficient, a high-accuracy spectrum for temporally short data There is a feature that can be obtained.

  The output side of the linear prediction coefficient calculation unit 44 that calculates the linear prediction coefficient described above is connected to a determination unit 45 that determines the quality of the subject. The determination unit 45 is a feature for learning a linear prediction coefficient calculated through the linear prediction coefficient calculation unit 44 from a vibration waveform obtained when a good sample (hereinafter simply referred to as a non-defective sample) of a subject is vibrated. It is a boundary learning type neural network constructed by inputting a plurality of feature quantities obtained from a plurality of non-defective samples. In the boundary learning type neural network, the distance between the feature quantity of the vibration waveform obtained by the excitation of the non-defective product and the feature quantity of the vibration waveform obtained by the excitation of the subject to be judged as good or bad is determined. The difference can be output as a numerical value within a set range (for example, 0 to 1). That is, when the feature amount of a good product is defined as 1, a determination result (numerical value) close to 1 is output if the subject is a good product and close to 0 if the subject is defective. Whether the subject is good or bad depends on whether the output result is within the boundary (threshold value) defined by the learning of the boundary learning neural network. That is, according to the above setting, the difference obtained by subtracting the threshold value is obtained from the numerical value output as the determination result of the feature amount of the object, and if the difference is positive, it is determined as a non-defective product. The

  Regarding a method for determining the quality of a subject using a boundary learning neural network, a plurality of learning input patterns are input to an input unit group, and an output pattern of an output unit group corresponding to the input learning input pattern is determined in advance. Compared with the pattern, an error between the two is obtained, and based on this error, the threshold value of the unit constituting each unit group and the weighting factor of the coupling element connecting each unit group (in this embodiment, the linear prediction coefficient) ) Is repeated to learn the input pattern for learning, and the output pattern of the output unit group obtained by inputting an arbitrary pattern to the input unit group is compared with the output pattern after learning. Based on the pattern matching method. In this method, an average value m and a standard deviation σ of feature values are calculated from learning data, and data defining part or all of the learning data as m ± aσ (a is a real value) is used as boundary data. The neural network is learned by adding the boundary data created in this way to the learning data. In such a method, how to determine the real value a may be obtained with reference to knowledge of normal distribution. Thereby, a neural network having a desirable pattern recognition capability is constructed. In addition, according to the boundary learning type neural network used in the present invention, it is sufficient to have only non-defective data as learning data.

  Next, the operation of the subject quality determination device 50 will be described. The vibration sensor 20 attached to the subject 100 (not shown) measures the vibration of the subject 100 that is vibrated by the vibration means 10. 3 and 4 show the measurement results of the vibration sensor 20. Here, FIG. 3 shows a vibration waveform when the subject is a non-defective product, and FIG. 4 shows a vibration waveform when the subject 100 has a defect such as a crack or a crack. In the waveforms shown in FIG. 3 and FIG. 4, a large difference cannot be seen in determining whether the subject 100 is good or bad.

  The vibration waveform detected as described above passes through the filter 30, the SN ratio is increased, is input into the computer 40, and is digitally converted by the A / D conversion means. From the digitally converted signal, a signal in a time zone for calculating a feature amount is cut out by the waveform cutout unit, and is input to the filter unit 43 and filtered.

  The filtered signal is input to the linear prediction coefficient calculation means 44, and the above-described linear prediction coefficient is calculated. The calculated linear prediction coefficients are shown in FIGS. 5 and 6 show linear prediction coefficients for 1 second from the time when 0.5 seconds have elapsed from the top of the waveforms of FIGS. 3 and 4. 5 shows linear prediction coefficients when the subject is a non-defective product, and FIG. 6 shows linear prediction coefficients when the subject is a defective product. Comparing FIG. 5 and FIG. 6, it can be seen that there is a difference in each linear prediction coefficient.

Then, the linear prediction coefficient calculated as described above is input to the determination unit 45, and the difference between the feature quantity of the subject 100 (the obtained linear prediction coefficient) by the boundary learning type neural network and the feature quantity of the non-defective product learned is obtained. Calculated as a numerical value within the set range. FIG. 7 shows the result of determining whether the subject 100 is good or defective. In the boundary learning type network, the feature quantity of the non-defective product is set to 1, and the distance between the feature quantity of the subject is determined, and a value between 0 and 1 is output. When the feature amount is close to a non-defective product, a value close to 1 is output, and when the feature amount is far from the non-defective product, a value close to 0 is output. For this reason, when the value from 0 to 1 output by the boundary learning type neural network is lower than the threshold value determined by the construction of the boundary learning type neural network, it is determined as bad, and when the value is higher than the threshold value, it is good. Can be determined.
The determination result obtained as described above is output to a monitor (not shown) of the control display means 46 provided in the computer 40.

  The following effects can be obtained by using a linear prediction coefficient as a feature quantity of vibration obtained by the vibration sensor 20. That is, the linear prediction coefficient can express the characteristics of the waveform by changing the time information within the analysis window length to a small amount of information such as the order. In addition, it is possible to capture a temporal change in the vibration waveform when a subject made possible by air excitation is vibrated in a high frequency region. For this reason, comparing with other detection methods, for example, a method that uses the maximum amplitude value of the original signal as an evaluation criterion, and a method that calculates the frequency spectrum of the original signal by fast Fourier transform and uses it as a feature amount, a smaller defect And unexpected defects can be detected. The linear prediction coefficient can be calculated relatively quickly. In addition, since the calculation of the neural network using the linear prediction coefficient can be performed at high speed, the quality of the subject can be determined in real time.

  Furthermore, since the linear prediction coefficient reflects the frequency spectrum of the signal, defects such as minute cracks and cracks that cannot be confirmed as amplitude changes can be detected in the vibration waveforms shown in FIGS. The accuracy of pass / fail judgment is improved.

Further, since the output of the boundary learning type neural network takes a value from 0 to 1, the degree of the defect can be determined based on the magnitude of the output value.
In addition, since the vibration means for vibrating the subject is an air vibration using a nozzle, there is no possibility of damaging the subject during vibration.

  In the embodiment, it is described that the vibration means uses a jet flow from the nozzle. However, the pass / fail judgment method of the subject of the present invention is also used for vibration by other vibration means such as hammering. be able to.

It is a block diagram which shows the quality determination apparatus of the subject which concerns on this invention. It is a block diagram which shows the structural example of the vibration means used for the quality determination apparatus of the subject of this invention. It is a figure which shows the vibration waveform which can be detected when a good quality object is vibrated. It is a figure which shows the vibration waveform which can be detected when a defective subject is vibrated. It is a figure which shows the linear prediction coefficient computed from the vibration waveform of the quality object. It is a figure which shows the linear prediction coefficient computed from the vibration waveform of the inferior goods subject. It is a figure which shows the quality of the subject determined by the boundary learning type neural network.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ......... Excitation means, 11 ......... Nozzle, 12 ......... Solenoid valve, 13 ......... Tank, 14 ......... Compressed air supply source, 20 ......... Vibration sensor, 30 ......... Filter , 40... Computer 41... A / D conversion means 42... Waveform cutting means 43 43 Filter means 44 Linear prediction coefficient calculating means 45 45 Determination means 46 ... …… Control display means, 100 ... …… Subject.

Claims (5)

  A method for determining whether or not a subject is a non-defective product, wherein a non-defective sample of the subject is vibrated by a jet of fluid, and vibrations of the good sample due to the vibration are detected and obtained by vibration detection. A sample waveform is cut out from the vibration waveform of a non-defective sample, a linear prediction coefficient is calculated based on the sample waveform, the linear prediction coefficient is used as a feature quantity for learning, and a boundary learning type neural network using a plurality of feature quantities of the non-defective sample is used. A network is constructed, and the feature quantity of the subject to be judged is passed to the boundary learning type neural network in the same manner as the feature quantity of the non-defective sample, and the pass / fail result is output by the boundary learning type neural network. The difference between the feature amount of the subject to be determined and the feature amount of the non-defective sample is obtained as a numerical value within the set range, and the numerical value is Quality determination method of the subject and judging the quality of the subject depending on whether a value in the range of determined threshold by field learning neural network.   2. The method of determining pass / fail of a subject according to claim 1, wherein the excitation causes the subject to generate vibration in a high frequency band of 20 kHz or more.   An apparatus for determining whether or not a subject is a non-defective product, the vibration means for exciting the subject with a jet of fluid, and vibration for detecting a vibration waveform of the subject vibrated by the vibration means A sensor, a waveform cutout unit that cuts out a sample waveform from the vibration waveform detected by the vibration sensor, and a feature amount calculation unit that calculates a feature amount with respect to the vibration of the subject based on the sample waveform cut out by the waveform cutout unit The feature amount calculated by the feature amount calculation means is input to a boundary learning type neural network constructed by inputting the feature amount of the non-defective sample of the subject, and the feature amount of the subject and the feature amount of the non-defective sample are calculated. A determination means for determining a difference and determining whether the difference is within a threshold range; and a display means for displaying a determination result by the determination means. Diagnosis device of a subject in.   The vibration means has a nozzle that ejects fluid, and is configured to be capable of changing the nozzle diameter and fluid ejection speed in order to give the subject vibration in a high frequency region of 20 kHz or more. The object pass / fail determination apparatus according to claim 3. The object pass / fail determination apparatus according to claim 3 or 4, wherein a plurality of the vibration sensors are provided for one object.

Images