JP7226230B2 - Thermal resistance prediction method and property inspection machine - Google Patents

Description

The present invention relates to a thermal resistance prediction method and a characteristic tester.

2. Description of the Related Art Conventionally, there is known a technique for evaluating the remaining life of an inverter that controls a motor based on the thermal resistance of a semiconductor element included in the inverter (see, for example, Patent Document 1). According to the technique of Patent Document 1, the thermal resistance of a semiconductor element is measured based on the voltage drop of the semiconductor element under predetermined conditions.

In addition, conventionally, in order to keep the temperature of the light emitting element below the heat resistance temperature, there is a known technique of starting to suppress the current to the light emitting element when the temperature of the light emitting element reaches a set current suppression start temperature ( For example, see Patent Document 2). According to the technique of Patent Document 1, the current suppression start temperature is set based on the thermal resistance of the light emitting element, the thermal resistance of the circuit side, the heat resistant temperature of the light emitting element, and the like.

In this way, it is important to understand the thermal resistance of a semiconductor element in order to suppress deterioration and failure of the semiconductor element due to heat and to ensure reliability. However, a characteristic inspection machine used for characteristic inspection of light-emitting diodes (LEDs) usually does not have a function of directly measuring thermal resistance. For this reason, a method called the _ΔVF method, which can indirectly measure the thermal resistance, is generally used.

The _ΔVF method is a method of calculating the thermal resistance based on the temperature-voltage characteristics of the LED and the voltage difference before and after the current is applied to the LED.

Japanese Patent Application Laid-Open No. 2002-119043 JP 2014-172520 A

However, measuring the temperature-voltage characteristics of LEDs for individual LED packages is not realistic because it requires a huge amount of man-hours. Also, when calculating the thermal resistance using the temperature-voltage characteristics of a specific LED, the characteristics of individual LEDs vary, so the inspection criteria must be excessively strict. Even individuals that should not be tested are judged to be defective.

An object of the present invention is a thermal resistance prediction method that can be implemented using a characteristic tester that does not have a function to directly measure thermal resistance, and can accurately obtain the thermal resistance of a light emitting element in a small number of steps. , and to provide a characteristic inspection machine used for implementing the thermal resistance prediction method.

One aspect of the present invention provides the following thermal resistance prediction methods [1] to [7] and a characteristic tester [8] in order to achieve the above object.

[1] A thermal resistance prediction method for predicting the thermal resistance of a first light emitting element using a classifier created using a plurality of second light emitting elements, the first light emitting element constituting a package performing voltage drop measurements on the device under a plurality of conditions in which at least one of the magnitude and direction of the current flowing through the first light emitting device is different to obtain a first voltage measurement value; and estimating the thermal resistance of the first light emitting device from the first voltage measurements using a classifier, wherein the classifier predicts the plurality of second light emitting devices forming a package. On the other hand, performing voltage drop measurement under a plurality of conditions in which at least one of the magnitude and direction of the current flowing through the second light emitting element is different, obtaining a second voltage measurement value ; A step of measuring the thermal resistance of the second light emitting element using a resistance measuring instrument to obtain a thermal resistance measurement value, and using the second voltage measurement value as an explanatory variable and the thermal resistance measurement value as an objective variable. performing machine learning to derive a causal relationship between the second voltage measurement and the thermal resistance measurement. a classifier for classifying and predicting the thermal resistance of the first light emitting element from values, wherein the plurality of conditions of the voltage drop measurement for obtaining the first voltage measurement is performed by the machine learning in the thermal resistance prediction method , wherein the results of narrowing down the explanatory variables are the same as the plurality of conditions of the voltage drop measurement for obtaining the second voltage measurement value finally used as the explanatory variable. .
[2] The second voltage measurement is the forward voltage V _F1 of the second light emitting element under a measurement current that does not affect the junction temperature of the second light emitting element; _ΔVF , which is the difference between the forward voltages _VF2 of the second light-emitting element immediately after the measurement current is applied after the junction temperature is raised by applying a current of a predetermined magnitude larger than the current; The thermal resistance prediction method according to [1] above.
[3] The thermal resistance prediction method according to [1] or [2] above, including the step of deleting Null values and measurement abnormal values from the second voltage measurement value before the machine learning.
[4] The thermal resistance prediction method according to any one of [1] to [3] above, including the step of normalizing the second voltage measurement value before the machine learning.
[5] In the machine learning, the importance of the explanatory variables is calculated by first machine learning using a random forest as an algorithm, and the explanatory variables are narrowed down based on the importance [1] above. The thermal resistance prediction method according to any one of [4].
[6] The thermal resistance prediction method according to [5] above, wherein in the machine learning, after the first machine learning, a second machine learning using a support vector machine as an algorithm is performed.
[7] The thermal resistance prediction method according to [6] above, wherein Linear is used as the kernel function of the support vector machine.
[8] Machine learning for classifying and predicting the thermal resistance of the light-emitting element by performing voltage drop measurement on the light-emitting element and acquiring a voltage measurement value, and the voltage measurement value. and a processing unit that refers to the classifier stored in the storage unit and classifies and predicts the thermal resistance of the light emitting element from the voltage measurement value acquired from the measurement unit. and an output unit that acquires the estimated thermal resistance from the processing unit and outputs it to the outside.

According to the present invention, a thermal resistance prediction method that can be implemented using a characteristic tester that does not have a function of directly measuring thermal resistance and can accurately obtain the thermal resistance of a light emitting element in a small number of steps. , and a property inspection machine used to implement the thermal resistance prediction method.

FIG. 1 is a flow chart showing the flow of a thermal resistance prediction method according to an embodiment of the present invention. FIG. 2 is a block diagram showing an example configuration of a property checker when a classifier is stored in the property checker. FIG. 3 is a flow chart showing the flow of creating a classifier according to the embodiment of the present invention. 4(a)-(c) are graphs relating to _ΔVF measurements. FIG. 5 is a flowchart showing an example of the flow of machine learning. FIGS. 6A and 6B are graphs showing examples of the relationship between predicted values and measured values of thermal resistance when explanatory variables are not normalized and when explanatory variables are normalized. FIG. 7 shows an example of decision trees formed in a random forest. FIG. 8 is a conceptual diagram for explaining SVM. FIGS. 9A and 9B are graphs showing examples of relationships between predicted values and measured values of thermal resistance when Linear and rbf are used as SVM kernel functions, respectively. Further, FIG. 9C is a graph showing an example of a case where machine learning by random forest is performed as a comparative example. FIGS. 10(a) and 10(b) are graphs showing examples of relationships between predicted values and measured values of thermal resistance before and after hyperparameter optimization, respectively.

[Embodiment]
(Thermal resistance prediction method)
FIG. 1 is a flow chart showing the flow of a thermal resistance prediction method according to an embodiment of the present invention. The thermal resistance prediction method will be described below with reference to FIG.

First, voltage drop measurement is performed under a plurality of conditions for an LED included in an LED package whose thermal resistance R _th is to be predicted (step S1). Hereinafter, the LED included in the LED package whose thermal resistance is to be predicted is referred to as the first light emitting element, and the voltage drop measurement value obtained by this measurement is referred to as the first voltage measurement value _Vm1 .

The measurement of the first voltage measurement _Vm1 in step S1 can be performed by an LED characterization machine. The first voltage measurement value V _m1 is obtained by voltage drop measurement under the same conditions as the second voltage measurement value V _m2 finally used as an explanatory variable in machine learning for creating a classifier described later. be done.

In a light-emitting element flip-chip bonded to a wiring board or the like in a package, the thermal resistance is closely related to the bonding area of the bonding portion by AuSn, solder, or the like, and greatly affects the life and characteristics of the element. Therefore, it can be said that the thermal resistance prediction method according to the present embodiment is particularly useful when the first light emitting element is a light emitting element flip-chip bonded in a package.

In addition, when the LED package including the first light emitting element is an LED package used in applications where the management of thermal resistance is important, such as an LED package for headlights, which is used at a high current, It can be said that the thermal resistance prediction method based on the morphology is particularly useful. The form of the LED package is not particularly limited.

A classifier is then used to predict the thermal resistance of the first light emitting element from the first voltage measurement _Vm1 (step S2). This classifier is a classifier for predicting the thermal resistance R _th from the voltage drop measurements obtained by machine learning. Creation of classifiers by machine learning will be described later. Also, the classifier may be regression type or classification type.

The classifier is stored and used in, for example, an LED characteristic inspection machine, a server on a computer network, or the like.

Next, the value of thermal resistance predicted by the classifier is output (step S3).

FIG. 2 is a block diagram showing a configuration example of the characteristic inspection machine 10 when the classifier α is stored in the characteristic inspection machine 10. As shown in FIG.

The characteristic tester 10 includes a measurement unit 11 that performs voltage drop measurement on the first light emitting element 20 and acquires a first voltage measurement value _Vm1 , and a storage unit 11 such as a memory that stores the classifier α. and processing included in a CPU (Central Processing Unit) or the like for predicting the thermal resistance R _th from the first voltage measurement value V _m1 acquired from the measurement unit 11 by referring to the classifier α stored in the storage unit 11 and an output unit 14 such as a monitor that acquires the predicted thermal resistance _Rth from the processing unit 12 and outputs it to the outside.

According to the characteristic tester 10, in step S1 of the above-described thermal resistance prediction method, for example, the measurement unit 11 causes the first light emitting element 20 to generate the current I (for example, I _M or I _H described later) is passed to obtain a first voltage measurement value V _m1 .

In step S2, the processing unit 13 acquires the first voltage measurement value V _m1 from the measurement unit 11, refers to the classifier α stored in the storage unit 12, and converts the first voltage measurement value V _m1 into a thermal Predict the resistance R _th .

In step S3, the processing unit 13 transmits the predicted thermal resistance _Rth to the output unit 14 and outputs it to the outside.

(Creating a classifier)
FIG. 3 is a flow chart showing the flow of creating a classifier according to the embodiment of the present invention. Generating a classifier will be described below with reference to FIG.

First, a characteristic tester is used to measure voltage drop under a plurality of conditions for a plurality of LEDs included in an LED package as a sample for creating a classifier (step S11). Hereinafter, the LED included in the LED package as a sample for creating this classifier is referred to as a second light emitting element, and the voltage drop measurement value obtained by this measurement is referred to as a second voltage measurement value _Vm2 . The second light emitting element and its package are of the same standard as the first light emitting element and its package.

The second voltage measurement V _m2 includes _ΔVF , which is the difference between the forward voltage _VF1 and the forward voltage _VF2 in the _ΔVF measurement of the _ΔVF method. In addition, the forward voltage V _F1 , the forward voltage V _F2 , the forward voltage in the LED circuit short test, the forward voltage in the open test, the forward voltage and reverse voltage in the leak test (when the reverse current is applied) voltage drop), etc., can be included in the second voltage measurement _Vm2 . V _F1 , V _F2 , and ΔV _F in the ΔV _F measurement will be described later. Table 1 below shows an example of a second voltage measurement V _m2 .

As shown in Table 1, V _F1 , V _F2 , and ΔV _F are obtained by performing multiple ΔV _F measurements while changing the magnitude of the current I _H for heat generation, and acquiring each multiple times. As the magnitude of the current _IH , it is preferable to use, for example, the maximum rated current value of the LED (eg, 2.2 A) or a commonly used current value (eg, 1.8 A).

Also, each of the second voltage measurements V _m2 (VF1-1, VF1-2, dVF1, . The plurality of second light emitting elements are light emitting elements of the same standard. By using a plurality of second light emitting elements, variation in characteristics of individual elements can also be learned in machine learning, so prediction accuracy of the classifier is improved.

Table 2 below shows an example of numerical values when each of the second voltage measurements V _m2 is taken for the 16 second light emitting elements (A to P). Note that "R _m " in Table 2 is the thermal resistance measurement value of the second light emitting elements (A to P) obtained in the next step S12.

Next, a thermal resistance measuring instrument is used to measure the thermal resistance of the second light emitting element from which the second voltage measurement value _Vm2 has been obtained (step S12). The measured value of thermal resistance obtained by this measurement is hereinafter referred to as the measured thermal resistance value _Rm . Note that the order in which steps S11 and S12 are performed is not limited.

Next, a causal relationship between the second voltage measurement value _Vm2 and the thermal resistance measurement value _Rm is derived by machine learning using the second voltage measurement value _Vm2 as an explanatory variable and the thermal resistance measurement value _Rm as an objective variable. , create a classifier for predicting the thermal resistance from the voltage drop measurements (step S13). A specific example of step S13 will be described later.

Machine learning is performed using a computer, and can be performed using various algorithms such as support vector machines (SVM), random forests, deep learning, etc., depending on the content of the second voltage measurement value _Vm2 , etc. can. Also, machine learning by a plurality of algorithms may be combined.

The classifier obtained by machine learning is stored in the LED characteristic tester, the server on the computer network, etc., as described above, and used in step S2 of the thermal resistance prediction method according to the present embodiment.

(Overview of _ΔVF measurement)
A _ΔVF measurement to obtain V _F1 , V _F2 , and _ΔVF used as the second voltage measurement V _m2 will now be described.

4(a)-(c) are graphs relating to _ΔVF measurements. 1A, 1B, and 1C, the forward voltage of the second light emitting element (voltage drop when forward current is applied), the junction temperature of the second light emitting element, and the 2 shows the forward current flowing through the light emitting element No. 2, and the horizontal axis indicates time.

First, a measurement current Im having a _magnitude (for example, 1 mA) that does not affect the junction temperature is passed through the second light emitting element (step S21). The forward voltage of the second light emitting element in step S21 is _VF1 .

Next, a current _IH having a predetermined magnitude, which is larger than the measured current _Im , is passed through the second light emitting element to raise the junction temperature from the room temperature _TR (step S22). Step S22 is preferably continued until the junction temperature approaches saturation.

Next, the current _IH flowing through the second light emitting element is switched to the measurement current _Im (step S23). The forward voltage of the second light emitting element immediately after the measurement current _Im starts to flow in step S23 is _VF2 , and the forward voltage _VF2 is smaller than the forward voltage _VF1 in step S21. The difference between the forward voltage _VF1 and the forward voltage _VF2 is _ΔVF .

In the _ΔVF method conventionally used to measure the thermal resistance of LEDs, the thermal resistance is calculated from ΔVF and _ΔTj , which is the difference between the junction temperature at step _S21 and the junction temperature at the end of step S22.

Here, the temperature-voltage characteristic of the LED measured in advance is used to obtain ΔT _j . Since the junction temperature in step S21 and the junction temperature at the end of step S22 are obtained from the temperature-voltage characteristics, the forward voltage V _F1 and the forward voltage V _F2 , _ΔTj is calculated. However, as described above, many man-hours are required to measure the temperature-voltage characteristics of LEDs.

(Flow of machine learning)
FIG. 5 is a flowchart showing an example of the flow of machine learning in step S13 described above. A specific example of machine learning will be described below with reference to FIG.

First, as preprocessing for machine learning, null values and measurement abnormal values are deleted as missing data from the second voltage measurement value _Vm2 (step S131). Null values are values that cannot be calculated in machine learning, and measurement abnormal values such as open values cause over-learning and abnormal classification. This deletion of missing data may be automatically performed by a computer program.

Next, as preprocessing for machine learning, the second voltage measurement value _Vm2 , which is an explanatory variable, is normalized (step S22). By normalizing explanatory variables, it is possible to prevent the degree of dependence on values with large classification results from increasing and to make the weights of all explanatory variables the same.

6A and 6B show the predicted value of the thermal resistance _Rth and the measured value of the thermal resistance _Rth , that is, the thermal resistance measured value R 10 is a graph showing an example of the relationship with _m . The straight line in FIG. 6(b) was obtained by fitting with linear regression.

In the example of the second voltage measurements V _m2 shown in Table 2, each of the second voltage measurements V _m2 (VF1-1, VF1-2 , dVF1,...) are normalized so that their standard deviation σ is 1.

Next, the importance of the explanatory variables is calculated by machine learning using random forest as an algorithm, and the explanatory variables are narrowed down based on the importance (step S23). Since explanatory variables with low importance are variables with low effects and variables that cause variability, by removing these and leaving only explanatory variables with relatively high importance, that is, by narrowing down explanatory variables, Over-learning can be suppressed.

Here, the present inventor confirmed that the value of _ΔVF (dVF1, dVF2 in the example of Table 1) tends to be more important among the second voltage measurement values _Vm2 , which are explanatory variables. It is For this reason, the explanatory variables that are finally used in machine learning often include at least part of the explanatory variables that are the values of _ΔVF .

FIG. 7 shows an example of decision trees formed in a random forest. In a random forest, explanatory variables are randomly selected to form a decision tree as shown in FIG. The decision tree shown in FIG. 7 includes dVF4, dVF3, VF3, VF2 of the second voltage measurement _Vm2 as explanatory variables.

To obtain the importance of explanatory variables, the following processing is performed. First, for a certain explanatory variable included in the decision tree, the order of the data is randomly shuffled, and whether or not the accuracy of the decision tree changes before and after shuffling is compared. Then, if the accuracy changes significantly, the explanatory variable is important, and if it does not change, it is judged to be unimportant. Furthermore, similar processing is repeated for a plurality of decision trees. The results obtained from a plurality of decision trees are voted by majority to determine the importance of explanatory variables.

Next, machine learning using SVM as an algorithm is performed to create a classifier (step S24). By performing machine learning by SVM after narrowing down the explanatory variables by random forest, the generalizability of the classifier (prediction performance for unknown data not used for learning) can be improved.

FIG. 8 is a conceptual diagram for explaining SVM. In the example shown in FIG. 8, two types of data with thermal resistance measurement values _Rm of 5.4°C/W or more and less than 5.4°C/W are linearly separated into two-dimensional data related to two explanatory variables dVF4 and dVF3. classified into two classes. A straight line between the data group in which the measured thermal resistance value _Rm is 5.4°C/W or more and the data group in which the measured thermal resistance value _Rm is less than 5.4°C/W is the boundary line between the two classes. , the method for drawing this boundary line is SVM.

In order to draw the boundary line, not all data are used, but several points of data near the boundary (data surrounded by dotted lines in FIG. 8) are used. Boundaries are drawn so that the distance (margin) between the few points near these boundaries and the boundary is maximized.

A classifier is obtained by defining the boundaries and classifying the classes. In this example, by inputting data with two explanatory variables dVF4 and dVF3 into this classifier, it is possible to determine to which of the two classes _the or less than 5.4°C/W.

If the SVM cannot classify classes (for example, if classes cannot be classified by straight lines in a two-dimensional space as shown in FIG. 8, classes can be classified by planes in a three-dimensional space). is not possible, or the class cannot be classified by a hyperplane in a space of four or more dimensions), classification can be made possible by introducing a kernel method using a kernel function. When using the kernel method, the space in which data exists (input space) is transformed into a space (feature space) in which classes can be classified by linear separation, and then the classes are classified by SVM.

Kernel functions used for SVM mainly include Linear, rbf, poly, and sigmoid. Among these kernel functions, the one with the highest coefficient of determination R2 is selected. The _coefficient of determination R2 indicates how well the predicted value matches the measured _value . It indicates how well it agrees with a given thermal resistance measurement _Rm .

When the measured value of the voltage drop of the LED included in the LED package is the explanatory variable and the thermal resistance is the objective variable, using Linear as the SVM kernel function often increases the coefficient of determination R2. confirmed by the inventor.

9A and 9B respectively show the predicted value of the thermal resistance _Rth when Linear and rbf are used as the kernel function of the SVM, and the thermal resistance measured value _Rm , which is the measured value of the thermal resistance _Rth . is a graph showing an example of the relationship between Further, FIG. 9C is a graph showing an example of a case where machine learning by random forest is performed as a comparative example. For the graphs shown in FIGS. 9A and 9B, the hyperparameters described below are also optimized. The straight lines in FIGS. 9(a) to 9(c) were obtained by linear regression fitting.

In addition, in machine learning using SVM, optimization of a regularization parameter C called a hyperparameter and an allowable value ε of residuals is performed. The regularization parameter C is a value that indicates the extent to which erroneous determinations are allowed. The residual tolerance ε is a value that indicates how much deviation from the regression curve a data point (explanatory variable) is penalized. Create a (classifier). This optimization is performed by a method called grid search.

FIGS. 10A and 10B show the predicted value of the thermal resistance R _th and the measured thermal resistance R _th before and after the hyperparameter optimization, respectively. 10 is a graph showing an example of the relationship with _m . The straight lines in FIGS. 10(a) and 10(b) were obtained by linear regression fitting. The values of the regularization parameter C and residual tolerance ε in this example are C=10 and ε=0.1 before optimization and C=1000 and ε=0.01 before optimization.

Cross-validation is preferably performed on the created classifier to evaluate its generalizability. As a result, if necessary, machine learning can be performed again to improve the prediction accuracy of the classifier.

As described above, the measurement of the first voltage measurement V _m1 in step S1 is performed under the same conditions as the measurement of the second voltage measurement V _m2 that was ultimately used as an explanatory variable in the machine learning described above. be done. For example, in the machine learning in step S13, when "dVF1", "OP", ... in Table 1 are finally used as explanatory variables, " _IH is 2.2 A ΔV _F ” _, “forward voltage in open test”, .

(Effect of Embodiment)
According to the thermal resistance prediction method according to the embodiment of the present invention, a classifier created by machine learning can be used to predict the thermal resistance from the measured value of the voltage drop of the LED, so the thermal resistance can be directly measured. The measurement can be performed using a characteristic inspection machine that does not have a function to perform the measurement, and the thermal resistance of the light-emitting element can be accurately obtained in a small number of steps.

Also, in creating the classifier, the measured value of the voltage drop of the LED is used as an explanatory variable, and the temperature-voltage characteristics of the LED and the like are not required. Therefore, a classifier can be created in a small number of steps using a characteristic inspection machine that does not have a function of directly measuring thermal resistance.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and various modifications can be made without departing from the scope of the invention. For example, the thermal resistance prediction method according to the above embodiments may be applied to other light-emitting elements such as laser diodes (LD) instead of LEDs.

Moreover, the above embodiments do not limit the invention according to the scope of claims. Also, it should be noted that not all combinations of features described in the embodiments are essential to the means for solving the problems of the invention.

10 Characteristic Inspection Machine 11 Measurement Unit 12 Storage Unit 13 Processing Unit 14 Output Unit

Claims (7)

A thermal resistance prediction method for predicting the thermal resistance of a first light emitting element using a classifier created using a plurality of second light emitting elements,
For the first light emitting element constituting the package, voltage drop measurement is performed under a plurality of conditions in which at least one of the magnitude and direction of the current flowing through the first light emitting element is different, obtaining a voltage measurement of
predicting a thermal resistance of the first light emitting element from the first voltage measurement using a classifier;
including
The classifier is
For the plurality of second light emitting elements constituting the package, voltage drop measurement is performed under a plurality of conditions in which at least one of the magnitude and direction of the current flowing through the second light emitting element is different, obtaining a second voltage measurement;
measuring the thermal resistance of the second light emitting element using a thermal resistance measuring instrument to obtain a thermal resistance measurement;
performing machine learning with the second voltage measurement value as an explanatory variable and the thermal resistance measurement value as an objective variable to derive a causal relationship between the second voltage measurement value and the thermal resistance measurement value;
made by a method comprising
a classifier for classifying and predicting the thermal resistance of the first light emitting element from the first voltage measurement based on the causal relationship;
The plurality of conditions for the voltage drop measurement for obtaining the first voltage measurement value are the second conditions finally used as the explanatory variables as a result of narrowing down the explanatory variables in the machine learning. is the same as the plurality of conditions of the voltage drop measurement for obtaining a voltage measurement;
Thermal resistance prediction method. The second voltage measurement is greater than the measured current and the forward voltage VF1 of the second light emitting element under a measured current that does not affect the junction temperature of the second light emitting element. ΔVF, which is the difference between the forward voltage VF2 of the second light emitting element immediately after the measurement current is applied after the junction temperature is raised by applying a current of a predetermined magnitude,
The thermal resistance prediction method according to claim 1. prior to the machine learning, removing null values and measurement outliers from the second voltage measurements;
The thermal resistance prediction method according to claim 1 or 2. normalizing the second voltage measurement prior to the machine learning;
The thermal resistance prediction method according to any one of claims 1 to 3. In the machine learning, the importance of the explanatory variables is calculated by first machine learning using a random forest as an algorithm, and the explanatory variables are narrowed down based on the importance.
The thermal resistance prediction method according to any one of claims 1 to 4. In the machine learning, after the first machine learning, a second machine learning using a support vector machine as an algorithm,
The thermal resistance prediction method according to claim 5. using Linear as the kernel function of the support vector machine;
The thermal resistance prediction method according to claim 6 .

Images