JPH08136400A - Method and device for detecting element characteristic defect - Google Patents

Abstract

PURPOSE: To detect minute characteristic defects by measuring discontinuous points (kink) of a characteristic curve of differential efficiency (dL/dI) or the like of a laser diode or the like without any error. CONSTITUTION: Measured characteristic curve data of dL/dI (L: output light, I: drive current) of a laser diode are input to a model circuit 2 through a data input device 1. The model circuit 2 smoothes the characteristic curve data (a) by a remote two-point average method. Remote two-point average method model data (b) are input to a feature extraction circuit 3, and a difference between the characteristic curve data (a) which are input data and the remote two-point average method model data (b) is calculated in the circuit 3. The calculated result is output to a judgment circuit 4 as difference data (c). In the judgment circuit 4 the difference data (c) from the feature extraction circuit 3 are compared with discrimination threshold values (d), if the difference data (c) are smaller than the judged threshold values at all points, products are judged to be good and if not so, the products are judged to be inferior.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device characteristic defect detecting method and an apparatus therefor, and more particularly to a method for detecting discontinuity defects in a semiconductor device characteristic curve such as a laser diode and an apparatus therefor.

[0002]

2. Description of the Related Art The presence or absence of a discontinuity point of a characteristic curve is used as one of the means for judging the quality of a semiconductor element. For example, in the case of a laser diode, when a characteristic curve of differential efficiency is drawn, a discontinuous point called a kink appears in a defective product. A conventional method of detecting such a discontinuous point is to find a curve that approximates the characteristic curve, and if there is a large difference between the approximated curve and the characteristic curve based on the actual measurement data, The product was judged as a "defective product" by recognizing the product as the existence of the discontinuous point.

FIG. 8 is a block diagram of a conventional characteristic defect detecting device for detecting such a discontinuity of a characteristic curve. As shown in FIG. 1, a conventional characteristic defect detection device includes a data input device 1 to which measured characteristic curve data of an element characteristic is input and outputs the data, and a characteristic curve data a from the data input device 1. A least squares approximation circuit 18 for approximating the difference by a quadratic least squares method, a feature extraction circuit 19 for calculating difference data between the least squares processing data m 'from the least squares approximation circuit 18 and the characteristic curve data a,
It has a decision circuit 20 for comparing the feature extraction data p ′ from the feature extraction circuit with a preset decision threshold value d ″ to decide pass / fail.

Here, considering the characteristic curves of the actual device characteristics, the characteristics of the curves such as the curvature are various, and even when a non-defective characteristic curve is input, an approximation error may appear greatly in the least squares curve. . FIG. 9 shows a differential efficiency characteristic curve (I-dL / dI characteristic;
Represents a drive current, L represents a light output), and a curve based on the least squares processing data obtained by the above-described conventional technique. FIG. 10 is a graph in which the difference data between the least square approximation curve data (m ′) and the measured characteristic curve data (a) is plotted on the graph in the conventional technique.

In FIG. 9, a defect (where the characteristic curve is discontinuous) exists near the drive current If = 60 mA. However, as apparent from FIG. 10, there is a place where the difference data is large near If = 30 mA. I do. Therefore, If = 60m
Judgment threshold value d ″ so that a discontinuity near A can be detected
Is set, an element that is not originally defective is determined to be defective. To avoid this, the determination threshold value must be set high. For this reason, the conventional example has a disadvantage that a minute characteristic defect cannot be detected.

Accordingly, Japanese Patent Application Laid-Open No. Hei 4-236370 discloses a technique for keeping the defect detection error relatively small.
A characteristic defect detection apparatus has been proposed which applies a neural network to learn the characteristics of a characteristic curve. As shown in FIG. 11, the characteristic defect detection apparatus applying this neural network is a current drive unit 21 that drives a semiconductor laser and inputs a current to a differentiating circuit, and an optical signal measurement that measures the optical output of the semiconductor laser 22. A circuit 23, a differentiating circuit 24 for generating a differential output signal according to a drive current, an A / D converting circuit 25 for A / D converting the output of the differentiating circuit 24, and a microprocessor 26 are provided. Measuring means for measuring the optical output L of the semiconductor laser with respect to the current I, and (B) the drive current I
DL / dI with respect to, and A / D converted differential LI
(C) starting point searching means for obtaining a starting point at which the rate of change of the differential LI characteristic data obtained by the data generating means is maximum and light output is maximum; Means for generating a difference direction vector set from the differential LI characteristic data from the start point to the measurement end point, and (E) the difference direction vector on the feature detector using a step function having a threshold. Means for converting into a cell coordinate to generate a set of cell coordinate conversion vectors; (F) means for generating a cell coordinate code of the feature detector from the cell coordinate conversion vector; and (G) two sets of the cell coordinate codes. Means for generating an input layer unit matrix having components as components, and allocating component values of the input layer unit matrix to a feature vector having an input layer unit number to the neural network, H) counting the occurrence frequency information of the feature vector, and means for normalizing comprises means for obtaining (I) wherein blocking the input unit matrix for each row vector, the maximum frequency unit components in the block,
(J) means for locally connecting the input layer unit and the intermediate layer unit of the neural network having the maximum appearance frequency unit information of the feature vector, and (K) learning of the differential LI characteristic by learning the neural network. Means for classifying defects.

A characteristic defect detection apparatus to which this neural network is applied converts differential efficiency characteristic curve data of a laser diode into a set of cell coordinate transformation vectors on a characteristic detector, and converts a characteristic vector of a characteristic vector obtained from two sets of reaction cells. Learning is performed by connecting the input layer unit of the neural network having the maximum value of the occurrence frequency normalization information to the intermediate layer unit,
The maximum value of the output layer unit is output as a classification category for the presence or absence of a defect.

[0008]

In the method of square least square approximation shown in FIG. 8, since the difference between the actual characteristic curve and the approximated characteristic curve is large, the defect detection error is large and therefore the minute defect is small. Could not be detected.

On the other hand, in the conventional characteristic defect detection apparatus disclosed in Japanese Patent Laid-Open No. 4-236370, although the defect detection error is improved, the characteristics of the characteristic curve are learned because the neural net is applied. Had to let. Therefore, there is a problem that the number of man-hours for learning is required and that the defect detection ability considerably depends on the number of learning data to be learned and the characteristics of the learning data.

[0010] Further, the characteristic curve has a characteristic such as a curvature, which is affected by the kind and the individual manufacturing process, etc., even in a non-defective product, and the size and shape of the defective portion are various. There is a possibility that training data cannot be used. On the other hand, when a large number of varieties are used as learning data, it is difficult to detect a difference in quality characteristics. In addition, there is a drawback that the non-defective judgment standard (standard) is unclear.

[0011]

An element characteristic defect detecting method according to the present invention calculates difference data between measured characteristic curve data of element characteristics and smoothed data obtained by smoothing the characteristic curve data by a remote two-point averaging method. The pass / fail is determined based on whether or not there is a place where the difference data is larger than the determination threshold.

Further, another element characteristic defect detection method calculates vector approximation data obtained by vector approximation from measured characteristic curve data of element characteristics, classifies the type of characteristic curve according to the peak detection position of the characteristic curve, According to the classification, the section is divided into two sections, and model data obtained by approximating a characteristic curve for each divided section is obtained.
Further, the presence / absence of discontinuity of the measurement characteristic curve is detected based on the relationship between the trajectory of the vector approximation data and the trajectory of the modeled data, thereby determining the quality.

[0013]

Embodiments of the present invention will now be described with reference to the drawings. [First Embodiment] FIG. 1 is a block diagram of an element characteristic defect detecting apparatus according to a first embodiment of the present invention. In FIG. 1, measurement characteristic curve data of an element is input to a data input device 1, and is output from the data input device 1 to a modeling circuit 2. The modeling circuit 2 smoothes the characteristic curve data a by the remote two-point averaging method. Here, the remote two-point averaging method when the characteristic curves are considered in an XY coordinate system orthogonal to each other is defined by Expression (1).

Ys _j = {f (x _jk ) + f (x _{j + k} )} / 2 (1) Here, the j-th point (j = 1, 2,.
...,; j is an integer) Y coordinate y when x coordinate is x _j
j is represented by equation (2). Therefore, in Expression (1), f (x _jk ) is the Y coordinate data of the (j−k) point, and f (x _{j + k} ) is the Y coordinate of the (j + k) point.
It is coordinate data. k is a positive integer constant. Ys _{j on the} left side of Expression (1) gives the Y coordinate data of the j-th point after the remote 2-point averaging process. y _j = f (x _j ) (2)

For example, assuming that k = 20 (points), the Y coordinate data at the 100th point: y ₁₀₀ = 0.24
9, 80 point in the Y coordinate data: y ₈₀ = 0.205, 120 point in the Y coordinate data: y ₁₂₀ = 0.25
3 If, as the modeling data ys ₁₀₀ 100 point in the post remote two-point averaging process, the equation _{(1), ys 100 = (} y 80 + y 120) / 2 = (0.205 + 0.253) / 2 = 0.229.

Here, the modeled data ys _{100 at} the 100th point after the remote 2-point averaging process is Y at the 80th point.
It depends only on the coordinate data y ₈₀ and 120 point in the Y coordinate data y _120, a significant feature to be independent of the other data (especially 100 point in the Y coordinate data y _100).

In the equations (1) and (2), when the characteristic curve data is a differential efficiency characteristic curve (I-dL / dI characteristic) of the semiconductor laser diode, the X axis is the drive current I axis, The Y axis is the differential light output dL / dI axis.
The remote two-point averaging model data b obtained by the equation (1) is input to the feature extraction circuit 3, where the difference from the characteristic curve data a as input data is calculated, and the calculation result is the difference data c. Is output to the determination circuit 4.
The judgment circuit 4 compares the difference data c from the feature extraction circuit 3 with the judgment threshold value d, and if all of the points have the difference data c smaller than the judgment threshold value d, there is a "good" item and a point that becomes larger. If so, it is determined to be "defective".

FIG. 2 is a graph in which the characteristic curve data a and the modeled data b which are the processing results of the modeling circuit 2 are plotted. FIG. 3 shows the difference data c in the present invention on the coordinates. It is a graph plotted. here,
The processing results shown in FIGS. 2 and 3 are the results when k = 20. The value of k is a value determined from the characteristics of the input characteristic curve and the defective curve pattern. As is clear from FIG. 2 and FIG. 3, in the method according to this embodiment, the difference data c is remarkably large only at the defective portion, so that the quality judgment by the judgment threshold value d is easy.

[Second Embodiment] Next, a second embodiment of the present invention will be described in detail with reference to the drawings. FIG.
FIG. 4 is a block diagram of a device characteristic defect detecting apparatus according to a second embodiment of the present invention. In FIG. 4, differential efficiency characteristic curve data of a laser diode is input to a data input device 1 and output to a vector approximation circuit 5 and a modeling circuit 6.

The modeling circuit 6 includes a peak detection circuit 7 and
It is composed of a curve type classification circuit 8, section classification circuits 9, 10, a moving average circuit 11, a least squares circuit 12, and a section boundary moving average circuit 13. Peak detection circuit 7
The drive current value data f which takes the peak value of the characteristic curve data a of the differential efficiency is output to the curve type classification circuit 8.

The curve type classification circuit 8 determines which type the characteristic curve belongs to as shown in FIG. 5 based on the driving current value data f having a peak value. That is, if the drive current value data f having the peak value is not near the threshold current Ith, it is determined to be the type 1 type (TYPE 1), and if it is near the threshold current Ith, the type 2 type (TYPE).
Judge as 2). Then, according to the determination, the curve type classification signals g and g ′ are output to the section classification circuits 9 and 10.

If the type 1 type is determined, the section classification circuit 9 determines from the section from the threshold current Ith to the drive current value data f having the peak value and from the drive current value data f having the peak value. It is classified into sections up to the maximum drive current value Imax. The former section characteristic curve data h is smoothed by moving average processing in the moving average circuit 11 and then output to the section boundary moving average circuit 13. On the other hand, the latter section characteristic curve data i is approximated by the least square circuit 12 by the second least square processing, and then the section boundary moving average circuit 13
Is output to

If it is determined to be type 2 (TYPE 2), the section classification circuit 10 determines from the section from the peak current value to the peak current value + α and the peak current value + α. It is classified into sections up to the maximum drive current value Imax (α is a constant). The former section characteristic curve data j and the latter section characteristic curve data k are independently approximated by the least squares processing by the least squares circuit 12, and then output to the section boundary moving average circuit 13.

Here, the value of α is of type 2 (TYPE 2)
Is determined from the characteristics of the locus of the characteristic curve of the semiconductor laser diode, and the Y axis (differential light output dL / d) in the range from the vicinity of the peak current value to the maximum drive current value.
The value of I) is set near the boundary between the first half where the value rapidly changes and the second half where the value changes slowly.

The section boundary moving average circuit 13 to which the moving average circuit 11 for outputting the moving average processing data 1 and the least square circuit 12 for outputting the least square processing data m are connected,
In order to avoid the discontinuity of the boundary area of the section classified into two, the boundary area of the section is smoothed by moving average processing, and the smoothed data is used as the modeled data n to extract the feature extraction circuit B15 and the feature extraction circuit. Output to the circuit C16.

Here, a method of calculating the modeled data n will be described. Data y obtained by performing general moving average processing on the Y coordinate data given by equation (2)
m _j is shown in Formula (3). ym _j = {w _m · f (x _{j + m} )} / (2i + 1) (3) (where Σ is m, from −i to i).

In this embodiment, the constant of the equation (3) in the moving average circuit 11 is expressed as w _m = 1 (m = −i,..., I) i = 2. The constant of the equation (3) is set to w _m = 1 (m = −i,..., I) i = 20 (averaged at 101 points across the boundary in the section boundary moving average circuit 13). w _m and i
Is determined by the characteristics of the input characteristic curve. Modeling data n is generated by the moving average processing as described above.

The modeled data n is divided into two sections by the section classifying circuits 9 and 10, and each section is approximated by using an approximation technique properly and modeled, thereby becoming data with a small approximation error.

On the other hand, the vector approximation circuit 5 performs a vector approximation of the characteristic curve data a using, for example, a cone intersection method or the like, and uses the vector approximation data e as a feature extraction circuit A14.
Output to. The feature extraction circuit A14 calculates the presence or absence of a polarity change and the amount of angle variation for the vector approximation data e, and outputs the feature extraction data o to the determination circuit 17. The feature extraction circuit B15 calculates a difference between the vector approximation data e and the modeled data n, and outputs the feature extraction data p to the determination circuit 17. The feature extraction circuit C16 calculates the intersection angle and the vector length of the approximate vector having the intersection with respect to all intersections between the locus of the vector approximation data e and the modeling data n, and determines the feature extraction data q in the determination circuit 17 Output to In addition, various kinds of preset determination thresholds d ′ are input to the determination circuit 17.

The determination circuit 17 determines the feature extraction data o, p,
q, there is a change in the polarity of the approximate vector (feature extraction circuit A), the amount of angle variation between adjacent vectors is greater than a predetermined determination threshold (feature extraction circuit A), the locus of the approximate vector and the model The difference between the approximated vector curve and the modeled data curve is greater than the predetermined determination threshold value (feature extraction circuit B). If the vector length of the approximation vector including the intersection is longer than a predetermined determination threshold value (feature extraction circuit C), it is determined as "defective".

Next, the determination procedure of the determination circuit 17 will be described in more detail with reference to FIG. Since the vector approximation data e follows the characteristic curve data, it also follows, for example, a defect (where the curve is discontinuous). Therefore, it can be determined as a defect when the adjacent approximation vectors have the above-mentioned or feature as shown in FIG. Here, this determination is effective for relatively small-scale defects, but it is difficult to detect large-scale defects over a wide range because the change in vector angle of the vector approximation data e is small. In order to detect such a defect, the above determinations are made.

Since the modeled data n has a smooth trajectory ignoring the defect location, as shown in FIG. 6B, when there is a defect, the difference from the vector approximation data e that follows along the defect location is shown. ε increases. Therefore, it can be determined that a defect exists by comparing the difference ε with the determination threshold (determination method). However, when there is a step-like defect, as shown in FIG. 6B, the modeling curve (modeling data n) passes almost in the middle of the defect location, and the size of the defect is underestimated as compared with the actual case. there is a possibility. In such a defect, as shown in FIG. 6C, the intersection angle at the intersection between the modeling curve and the tracking vector increases, and the vector length of the tracking vector there increases, so that the above determination is made. Can be detected.

FIG. 7 shows a graph of the input characteristic curve data a and the vector approximation data e and modeled data n which are the processing data according to this embodiment. As is clear from FIG. 7, the modeled data n indicates an ideal trajectory ignoring the defect portion, and the vector approximation data e is a trajectory following the measured characteristic curve data a. Therefore, it can be seen that even the fine defect can be detected by the above-described determination method (1).

[0034]

As described above, the present invention provides (1)
Either approximating the curve of the characteristic curve data by the remote two-point averaging method and detecting a defect by comparing the approximated data with the original characteristic curve data, or (2) classifying the characteristic curve into two types and according to the classification In this method, the characteristic curve is vector-approximated, and the characteristic curve is vector-approximated, and the defect is detected from the trajectory of the vector-approached data or the comparison between the modeled data and the vector-approached data. It becomes possible to detect a defect with high sensitivity by making only the defective portion noticeable. Therefore, according to the present invention, it is possible to prevent an inconvenience in which a non-defective product is mistakenly determined to be defective, and it is possible to reliably detect a minute defect.

Further, since the defect detection error is not reduced by learning, there is no need for man-hours for that, and the defect detection capability depends on the number of learning data to be learned and the characteristics of the learning data. Is also resolved,
Highly accurate defect detection becomes possible.

[Brief description of drawings]

FIG. 1 is a block diagram of an element characteristic defect detection device according to a first embodiment of the present invention.

FIG. 2 is a characteristic curve diagram after a measurement characteristic curve and a process according to the first embodiment of the present invention are added.

FIG. 3 is a remote two-point averaging method difference data characteristic curve diagram for explaining the effect of the first embodiment of the present invention.

FIG. 4 is a block diagram of an element characteristic defect detection device according to a second embodiment of the present invention.

FIG. 5 is a diagram showing measurement characteristic curves by type for explaining the operation of the second embodiment of the present invention.

FIG. 6 is a tracking vector diagram and a modeled data curve diagram for explaining the operation of the second embodiment of the present invention.

FIG. 7 is a characteristic curve after a measurement characteristic curve and processing according to a second embodiment of the present invention are added.

FIG. 8 is a block diagram of a conventional example.

9A and 9B are measurement characteristic curves and characteristic curve diagrams after processing by a conventional example.

FIG. 10 is a difference data characteristic curve diagram for explaining a problem of the conventional example.

FIG. 11 is a block diagram of another conventional example.

[Explanation of symbols]

 1 Data input device 2 Modeling circuit (remote two-point averaging method) 3, 19 Feature extraction circuit (difference calculation circuit) 4, 17, 20 Judgment circuit 5 Vector approximation circuit 6 Modeling circuit 7 Peak detection circuit 8 Curve type classification circuit 9, 10 Section classification circuit 11 Moving average circuit 12 Least square circuit 13 Section boundary moving average circuit 14 Feature extraction circuit A (angle variation amount) 15 Feature extraction circuit B (difference calculation) 16 Feature extraction circuit C (crossing angle and vector length) 18 Least-squares approximation circuit 21 Current drive unit 22 Semiconductor laser 23 Optical signal measurement circuit 24 Differentiation circuit 25 A / D conversion circuit 26 Microprocessor a Characteristic curve data b Remote 2-point averaging method modeling data c Difference data d, d ', d "determination threshold value e vector approximation data f differential current peak value drive current value data g, 'Curve type classification signals h, i, j, k section curve data l moving average processing data m, m' least-squares process data n modeled data o, p, p ', q feature data

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] June 1, 1995

[Procedure Amendment 1]

[Document name to be amended] Statement

[Correction target item name] 0008

[Correction method] Change

[Correction content]

[0008]

In the second- order least-squares approximation shown in FIG. 8, the difference between the actual characteristic curve and the approximated characteristic curve is large, so that the defect detection error is large and the minute defect is small. Could not be detected.

Claims (6)

[Claims] 1. A method for calculating difference data between measured characteristic curve data of element characteristics and smoothed data obtained by smoothing the measured characteristic curve data by a remote two-point averaging method, wherein the difference data is larger than a determination threshold. An element characteristic defect detection method, characterized by judging whether or not there is a part. 2. Calculating vector approximation data obtained by vector approximation from measured characteristic curve data of an element characteristic, classifying a characteristic curve type according to a peak detection position of the measured characteristic curve, and dividing a section into two according to the classification. Then, modeled data obtained by approximating the characteristic curve by using an approximation method for each divided section, based on the locus of the vector approximate data, and the relationship between the locus of the vector approximate data and the locus of the modeled data An element characteristic defect detection method, comprising detecting the presence or absence of a discontinuity in a measurement characteristic curve, and judging the quality based on the discontinuity. 3. An approximation vector using the vector approximation data and the modeled data, wherein the vector angle of the approximation vector has a polarity change, and the angle variation between adjacent vectors is larger than a predetermined determination threshold value. The difference between the trajectory of the approximated vector and the modeled data curve is larger than a predetermined judgment threshold value. For the intersection between the approximate vector trajectory and the modeled data curve, the intersection angle is larger than the predetermined judgment threshold value. 3. The element characteristic defect according to claim 2, wherein if any one of the following conditions is satisfied: the vector length of the approximate vector including the intersection is longer than a predetermined determination threshold value, the determination is "defective". Detection method. 4. A data input device for outputting measured characteristic curve data of an element characteristic, a modeling circuit for smoothing the characteristic curve data output from the data input device by a remote two-point averaging method, and the modeling circuit A feature extraction circuit for calculating a difference between the modeled data output from the device and the characteristic curve data output from the data input device; and comparing the output data from the feature extraction circuit with a determination threshold to measure An element characteristic defect detection device, comprising: a determination circuit for determining the quality of the element. 5. A data input device for outputting measured characteristic curve data of element characteristics, a vector approximation circuit for performing vector approximation on characteristic curve data output from the data input device, and a characteristic curve output from the data input device. For the data, the type of the characteristic curve is classified according to the detection position of the peak, the section is divided into two according to the classification,
A modeling circuit that performs approximation modeling of a characteristic curve by properly using an approximation method for each section; an output data of the vector approximation circuit and an output data of the modeling circuit are input, and an approximation created by each output data A composite feature extraction circuit for extracting a characteristic of a change in a characteristic curve, and a determination circuit for comparing the feature extraction data output from the composite feature extraction circuit with a predetermined determination threshold value to determine the quality of the device under test. And a device characteristic defect detection device characterized by including: 6. A first feature extraction circuit, wherein the composite feature extraction circuit calculates presence / absence of a change in polarity of an approximate vector and an amount of angle variation using output data of the vector approximation circuit, and an output of the vector approximation circuit. A second feature extraction circuit that calculates a difference between data and output data of the modeling circuit, and an intersection between a locus of output data of the vector approximation circuit and a curve of output data of the modeling circuit.
A third feature extraction circuit for calculating a vector length of the approximate vector including the intersection angle and the intersection, and the determination circuit has a polarity change in the vector angle of the approximate vector; The vector angle variation is larger than a predetermined judgment threshold, the difference between the approximate vector locus and the modeled data curve is larger than a predetermined judgment threshold, the approximate vector locus and the modeled data curve If the intersection angle is larger than a predetermined determination threshold and the vector length of the approximate vector including the intersection is longer than a predetermined determination threshold, 6. The device characteristic defect detecting device according to claim 5, wherein the device characteristic is determined as "defective".