JP4253024B2 - Method and program for extracting features for inspection of non-defective and defective products from waveform pattern data - Google Patents

Description

  The present invention relates to a method and a program for extracting features for inspecting non-defective / defective products from spectral waveform pattern data.

  Products such as fluorescent tubes can be characterized by their spectral waveform patterns. That is, as shown in FIG. 3, the fluorescent tube has a large and small light energy for each wavelength, and the emission characteristics of the fluorescent tube are represented by the spectrum waveform. Such a light emission characteristic of the fluorescent tube is used as one of criteria for determining whether the product is a good product or a defective product. For example, it is determined by whether or not it is equivalent to a registered spectral waveform group of a good product. That is, in FIG. 3 (a) and FIG. 3 (b), it can be regarded as a waveform pattern having the same property in most places, but the <4> peak shape is as shown in FIG. 3 (a) and FIG. 3 (b). However, there is a difference, and there arises doubt about the identity of the spectrum waveform. In addition to fluorescent tubes, spectrum waveforms are used in the fields of chemicals, pharmaceuticals, and the like to determine whether a product is good or defective.

  Conventionally, the determination of non-defective / defective products as described above has been performed by humans. However, in recent years, it has been increasing that computers perform such operations. When causing a computer to execute the operation, it is important to accurately and sufficiently extract “features as numerical values” representing the characteristics from the spectrum waveform pattern. Features required for such work include (1) representing features related to unique features (eg, frequency, amplitude, etc.) of the target spectral waveform pattern data, and (2) the position of each peak. It is necessary to represent features relating to the height and shape of each peak. In addition, it is desirable that the computer can process the characteristics satisfying such necessary items in a short time.

  Features extracted from conventional spectral waveform pattern data include the horizontal axis position of each peak, the peak height, the half width of the peak, the area of the peak portion, and the like. In recent years, the concept of differential characteristics and integral characteristics has been proposed as means for extracting features of waveform patterns by simple processing (see, for example, Non-Patent Document 1), and application examples thereof have also been announced. (For example, refer nonpatent literature 2).

  Here, the differential characteristics and the integral characteristics will be briefly described with reference to FIG. Considering a state in which an appropriate number (5 in FIG. 7A) of horizontal lines (hereinafter referred to as “sample lines”) is drawn with respect to a certain range of waveforms, the differential characteristic is the difference between the sample line and the waveform. This is the number of intersections (in the example of FIG. 7A, the differential characteristic is 9), and the integral characteristic is the range where the waveform exists above the sample line (that is, the larger side in the y-axis direction). It is the sum of the lengths (illustrated by bold lines in FIG. 7B). In other words, the differential characteristic can be considered as an amount of change on the waveform sample line, and the integral characteristic can be considered as an abundance on the waveform sample line (hereinafter, “change amount” and “ The term “abundance”). The results of obtaining these characteristics for the sample lines <1> to <5> are shown in FIG. As a result, the waveform patterns in FIGS. 7A and 7B are expressed as ten numerical values shown in FIG. 7C.

Taichi Genichi, “Speech Pattern Recognition”, Quality Engineering, Quality Engineering Society, October 1995, Vol. 3, No. 5, p. 3-7 Shoichi Teshima et al., “Study on visual inspection technology applying Mahalanobis Taguchi system method”, Quality Engineering, Quality Engineering Society, October 1997, Vol. 5, No. 5, p. 38-45 Kamoshita et al., “Optimization of Multidimensional Information System Based on Mahalanobis Distance-In Case of Fire Alarm System”, Quality Engineering, Quality Engineering Society, June 1996, Vol. 4, No. 3, p. 54-68

  However, the conventional feature extraction method of spectral waveform pattern data is not sufficient to satisfy the above conditions (1) and (2). Depending on the problem to be solved, some pattern recognition means (for example, statistical Even if it is recognized by means, a neural network, etc., the target recognition result may be insufficient.

  For example, in the above feature quantities, that is, the horizontal axis position of each peak, peak height, peak half width, peak area, etc., the peak start position is clearly defined from the peak rising position, that is, the flat part or the non-peak part. It is difficult to define the data as follows, and the characteristic value fluctuates greatly depending on the definition method, which may cause a problem in reliability as data. Further, for example, information for capturing the difference in shape at the peak portion <4> in FIGS. 3A and 3B is insufficient, and it is determined whether the product is a good product or a defective product from the spectrum waveform pattern. Therefore, there was a problem that the information was insufficient.

  The present invention has been devised in view of such circumstances, and it is an object of the present invention to provide a feature extraction method and program for spectral waveform pattern data representing properties related to the shape and the like of each peak portion of a spectral waveform. It is said.

The method for extracting features for inspecting non-defective / defective products of the product from the spectral waveform pattern data according to claim 1 of the present invention specifies spectral waveform pattern data within a range of 1 or the spectral waveform pattern data A first step of dividing into a plurality of ranges in the horizontal axis direction defined by the wavelength direction of the second, a second step of respectively defining a plurality of sample lines for the spectral waveform pattern data in the specified or divided range; From the amount of change on the sample line defined by the number of locations where the spectral waveform pattern data and the predetermined sample line intersect , the first spectral waveform pattern on the sample line from the start position of the predetermined sample line A spectrum on the sample line defined by the length to the position where the data begins to exist Of the waveform waveform start point or the end point of the spectral waveform pattern on the sample line defined by the length from the start position of the predetermined sample line to the position where the spectral waveform pattern data ends on the sample line And a third step of extracting a part or all of it.

The program for extracting features for inspecting a non-defective product / defective product from the spectral waveform pattern data according to claim 2 of the present invention specifies spectral waveform pattern data within a range of 1 or the spectral waveform pattern data Dividing into a plurality of ranges in the horizontal axis direction defined by the direction of the wavelength, and a step of storing a plurality of sample lines for the spectral waveform pattern data in the specified or divided range and storing them in the storage device And the amount of change on the sample line defined by the number of points where the spectral waveform pattern data and the predetermined sample line intersect , and the spectrum on the sample line first from the start position of the predetermined sample line On the sample line defined by the length to the position where the waveform pattern data begins to exist A spectral waveform pattern on the sample line defined by a length from a start point of the spectral waveform pattern in the sample line or a position from the start position of the predetermined sample line to a position at which the spectral waveform pattern data ends on the sample line The computer is caused to execute a step of extracting a part or all of them and storing them in a storage device.

  According to the present invention, it is possible to extract the characteristics of a waveform pattern that represents properties related to the shape and the like of each peak portion of the spectrum waveform from the spectrum waveform pattern, thereby accurately inspecting non-defective / defective products. It becomes possible.

  Next, preferred embodiments of the present invention will be described in detail with reference to the drawings. The spectrum waveform pattern has a start point and an end point. In the present embodiment, a waveform pattern such as the spectral waveform of the fluorescent tube shown in FIG. 3 will be described as an example. First, the spectral waveform pattern is divided into a plurality of ranges (step B1). Various methods for setting the number of divisions and the range are conceivable, but the shape of the spectrum waveform pattern of a normal fluorescent tube is almost determined, and peaks of a constant shape appear at a substantially constant height at a plurality of wavelength locations. Therefore, the spectral waveform pattern is divided for each location where a peak exists. That is, as shown in FIG. 4, the range is divided into <1> to <7>.

  The division may be defined so that the ranges do not overlap as shown in FIG. 4 or may be defined so as to overlap. At this time, it is preferable to divide so that there is no undefined part over the entire range of the target wavelength, but it may be divided so that there is an undefined part. In some cases, such as when there is one peak, division is not necessary.

  Next, Lh sample lines are determined for the divided spectral waveform patterns (step B2). The number Lh of sample lines is determined according to how much detailed features are extracted from the spectrum waveform pattern, and may be uniform in the ranges <1> to <7>, or set for each range. Also good.

  Next, for each range, a part or all of the change amount, the existing amount, the minimum value of the existing amount, the maximum value of the existing amount, the start point of the waveform, or the end point of the waveform in the sample line L = 1 to Lh. Obtain (step B3). The amount of change and the amount of existence are as described above. Other features, that is, the minimum value of the existing amount, the maximum value of the existing amount, the start point of the waveform pattern, and the end point of the waveform pattern will be described with reference to FIG.

  In FIG. 5, for ease of explanation, the ranges <1> to <3> in FIG. 4 are shown as one range. First, the minimum value of the abundance is, for example, the length of the section cd in the sample line b in FIG. That is, in the sample line B, the locations where the waveform exists on the sample line are cd and ef, and the length of the minimum length portion cd is the minimum value of the existence amount. On the other hand, the maximum value of the abundance is the length of the section ef. The start point of the waveform pattern is the position where the waveform begins to exist first on the sample line, that is, the c point for the sample line B in FIG. 5A, and the c from the start position s of the sample line B Length to point. On the other hand, the end point of the waveform pattern is the position where the waveform ends on the sample line, that is, the f point for the sample line B in FIG. 5A, and the f point from the start position s of the sample line B. It becomes the length. The waveform start point pattern and the waveform pattern end point are the position a and the position b, respectively, for the sample line a in FIG. 5A, and the values are s′-a and s′-b, respectively. It becomes length.

  These features will be further described with reference to FIG. 6 which is drawn from the spectral waveform pattern range <4> of FIG. The characteristics of the spectral waveform pattern as shown in FIGS. 6A and 6B are shown in FIGS. 6 (c) and 6 (d) show the abundance, the waveform start point, and the number when the numbers 1, 2,... Are numbered in order from the bottom of the sample lines in FIGS. Each of the end points of the waveform is shown. In FIG. 6A and FIG. 6B, the peak position and the area of the peak portion are substantially the same, but the shape of the mountain is different. It can be seen that the different situation appears in each feature in the sample lines 6 to 9 in FIGS. The peak position is obtained as the start position and end position of the tenth sample line, and the area of the peak part is included in the abundance of each sample line. In other words, the feature shown here enables the difference in the shape of the peak portion to be expressed as a difference in numerical value. In FIGS. 6C and 6D, all the features after the sample line 11 are “0”. However, this is because the initial value of the feature is “0” and the sample line does not have a waveform. This is because the initial value is maintained. Although only three types of features are shown in FIGS. 6C and 6D, other features such as a change amount can be used depending on the purpose of detecting the difference between the non-defective product and the defective product.

  As described above, the features of the entire spectrum waveform pattern are obtained by obtaining the features for the entire range <1> to <7> in FIG. 4 and forming a series of feature groups as shown in FIG. The total number of features in this example is, for example, 20 × 6 = 120, assuming that the number of sample lines is 20 and there are 6 types of features in each sample line, such as the amount of change and the amount of existence. Therefore, when the same number of features are extracted in the range <1> to <7>, the total number of features is 120 × 7 = 840.

  The feature extraction method of the present invention may include a preprocessing step for leveling the spectrum waveform pattern data. As a means for leveling, for example, calculation of a moving average of data can be mentioned. When this preprocessing step is performed, the above-described steps B1 to B3 are performed on the normalized spectral waveform pattern data obtained by the preprocessing step.

  Furthermore, the feature extraction method of the present invention may include a step of extracting other features (for example, features conventionally used such as peak positions) (step B4). The features obtained in steps B3 and B4 are preferably input to the recognition system (step B5). In addition, as a recognition system, various systems, such as what applied statistical theory, what applied artificial intelligence theories, such as a neural network, are applicable.

  Next, a program for causing a computer to execute the above steps (that is, Step B1 to Step B3) will be described. A computer on which the program is executed includes a CPU (Central Processing Unit), a storage device such as a memory and a hard disk, an input device such as a keyboard, a display device, and an output device (all not shown) It may be of a general type having or a microchip type processing apparatus or the like.

  First, spectrum waveform pattern data input by the input device is stored in the memory. The spectral waveform pattern data stored in the memory is divided by the CPU according to the “number of divisions” input in advance from the input device, and the divided spectral waveform pattern data is stored in the memory (step B1).

  Next, with respect to the divided spectral waveform pattern data, one or a plurality of sample lines are determined in the CPU according to the “sample line number” input in advance, and the divided spectral waveform pattern data with the sample lines determined is stored in the memory. (Step B2).

  Next, with respect to the spectral waveform pattern data in which the sample line stored in the memory is defined, features such as the amount of change and the amount of existence are calculated and extracted by the CPU, and the features thus extracted are stored in the memory ( Step B3).

  Alternatively, the spectrum waveform pattern data stored in the memory may be leveled by the CPU, the leveled spectrum waveform pattern data may be temporarily stored in the memory, and Step B1 may be executed for the leveled spectrum waveform pattern data. The data obtained in step B3 is input to the recognition system from an output device or the like. In the above example, the data is described as being stored in the memory. However, when the amount of data is large, the data is stored in a mass storage device such as a hard disk.

  The present invention is not limited to the above-described embodiments, and various modifications can be made within the scope of the invention described in the claims, and these are also included in the scope of the present invention. Needless to say, it is something.

  For example, in the above-described embodiment, the sample line is shown as a horizontal straight line, but an inclined straight line (see FIG. 5B), an upwardly convex curve (see FIG. 5C), and downward. Arbitrary lines such as convex curves (not shown) and non-parallel lines may be used.

It is the flowchart which showed the structure of the characteristic extraction method of the spectrum waveform pattern which concerns on preferable embodiment of this invention. It is the flowchart which showed a part of procedure in the feature extraction method of FIG. 1 in detail. It is the graph which showed the spectrum waveform pattern. It is the graph which showed the spectrum waveform pattern divided | segmented into the several range. It is a graph for demonstrating the kind and characteristic of a sample line. (A), (b) are graphs showing spectral waveform patterns and sample lines in a certain range, (c), (d) are tables showing the characteristics in (a), (b), and (e) is a series. It is the conceptual diagram which showed the feature group. It is a chart for demonstrating the variation | change_quantity and abundance.

Claims (2)

A method for extracting features for inspecting non-defective / defective products from spectral waveform pattern data,
A first step of specifying spectral waveform pattern data in one range or dividing the spectral waveform pattern data into a plurality of ranges in a horizontal axis direction defined by a wavelength direction of the spectral waveform pattern data;
A second step of defining a plurality of sample lines for the spectral waveform pattern data in the specified or divided range;
The amount of change on the sample line defined by the number of locations where the spectral waveform pattern data and the predetermined sample line intersect ,
The starting point of the spectral waveform pattern on the sample line defined by the length from the determined starting position of the sample line to the position where spectral waveform pattern data first exists on the sample line;
Alternatively, a part or all of the end points of the spectrum waveform pattern on the sample line defined by the length from the start position of the determined sample line to the position where the spectrum waveform pattern data ends on the sample line is extracted. And a third step
A method comprising the steps of: A program that extracts features for inspection of non-defective and defective products from spectrum waveform pattern data,
Specifying the spectral waveform pattern data in one range, or dividing the spectral waveform pattern data into a plurality of ranges in a horizontal axis direction defined by a wavelength direction of the spectral waveform pattern data;
For spectral waveform pattern data in the above-mentioned range specified or divided, and storing in the storage unit defines several specimens lines double, respectively,
The amount of change on the sample line defined by the number of locations where the spectral waveform pattern data and the predetermined sample line intersect ,
The starting point of the spectral waveform pattern on the sample line defined by the length from the determined starting position of the sample line to the position where spectral waveform pattern data first exists on the sample line;
Alternatively, a part or all of the end points of the spectrum waveform pattern on the sample line defined by the length from the start position of the determined sample line to the position where the spectrum waveform pattern data ends on the sample line is extracted. And storing in the storage device,
A program that causes a computer to execute.

Images