JPH0816207A - Method for adjusting electronic circuit package and automatic adjusting device - Google Patents

Abstract

PURPOSE:To provide a method capable of being applied to the adjustment of various electronic circuit pacakges, being used by everybody independently of skillness, shortening manhour and time and facilitating automation. CONSTITUTION:A normal signal is supplied to an electronic circuit package to be adjusted, the output value Xs of (n) outputs at the time of fixing (m) adjusting trimmers on an initializing position (setting of a beta sort) is applied to a GA operation part 1 for executing genetic algorithm and the rotational variable of each adjusting trimmer for adjusting the output of the package to a reference output value Xo is found out and used as a teacher signal Zs to allow a neural network 2 to learn it. The learning operation is executed for the initializing position of the beta sort. At the time of adjustment, the output value X of the package to be practically adjusted is inputted to the learned neural network 2 and the trimmer is rotated in accordance with the outputted rotational variable Y. If the output of the package is not adjusted to the reference output value Xo by one rotation adjustment, the output value after rotation is inputted and similar operation is repeated.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an electronic circuit package adjusting method and an automatic adjusting apparatus.

[0002]

2. Description of the Related Art Adjustment of an electronic circuit package having an adjusting trimmer such as a variable resistor or a variable capacitor is often necessary as a final step for giving a desired output characteristic to a product when manufacturing the electronic circuit package. .
Conventionally, such adjustment of an electronic circuit package is performed by an operator 10 using an adjustment trimmer 7 using a bit 12 while observing an output value from the electronic circuit package 6 with a measuring device 11, as shown in FIG. It is done by rotating.

On the other hand, as a conventional technique for automating the adjustment work of such an electronic circuit package without depending on the skill of the operator, for example, an electronic circuit adjustment device described in Japanese Patent Laid-Open No. 2-92004. There is. This electric circuit adjusting device uses two variable resistors for gain and offset adjustment of an electronic circuit by fuzzy reasoning from the AC component and the DC component of the output of the electronic circuit when a sine wave input having a specified amplitude is applied. The adjustment amount (trimmer rotation amount) is obtained, and the operation of rotating the adjustment trimmer of each variable resistor by that amount by the resistance value adjuster of the motor drive is repeated until each adjustment amount converges to "0". .

[0004]

In the above-mentioned conventional method for adjusting an electronic circuit package by an operator, the operator rotates the adjusting trimmer while performing observation while observing with a measuring instrument, and therefore the time required for the adjustment is reduced. It depends largely on the degree of skill of the operator. Especially, when the scale of the electronic circuit package becomes large and the number of trimmers for adjustment increases and mutual interference occurs, not only the adjustment time becomes longer and the degree of dependence on the skill increases, but also the degree of skill is insufficient. There is a problem that the operator may not be able to make adjustments.

On the other hand, the adjusting device using the above-mentioned fuzzy inference is not affected by the skill level of the operator, but since it is based on the fuzzy inference, it is generally necessary to make some repeated adjustments (rotating operation of the adjusting trimmer). In addition to increasing the time, there is the problem that the membership function, which is the rule for performing fuzzy inference, must be determined and set in advance. In other words, an engineer who is fully familiar with the configuration and operation of the electronic circuit package that is the target of adjustment,
In addition to the fact that the membership function must be set in advance, the setting of this membership function becomes complicated and difficult as the number of trimmers for adjustment increases.

An object of the present invention is to use a neural network having a learning function, which has been attracting attention in recent years, and a genetic algorithm method which is an efficient and practical search method for a huge search space. It is an object of the present invention to provide a universally usable electronic circuit package adjusting method and an automatic adjusting device in which the drawbacks are eliminated.

[0007]

According to another aspect of the present invention, there is provided an electronic circuit package adjusting method, wherein an output value when a specified input signal is applied is within a predetermined allowable range with respect to a reference output value. The amount of rotation of the trimming trimmer of is determined by a method of a genetic algorithm for a plurality of initial setting states of the trimming trimmer, and after learning these in a hierarchical neural network having an intermediate layer as a teacher signal, The rotation amount of the adjustment trimmer is obtained from the neural network based on the output value when the specified input signal is applied to each electronic circuit package to be adjusted, and the adjustment trimmer is set according to the obtained rotation amount. It is characterized by rotating so that the output value is adjusted to fall within a predetermined allowable range with respect to the reference output value.

According to another aspect of the present invention, there is provided an electronic circuit package automatic adjusting apparatus, wherein a measuring section for applying a specified input signal to an electronic circuit package to be adjusted to measure an output value, and an adjusting trimmer for the electronic circuit package according to control information. The rotation amount of the adjusting trimmer is obtained by the learned neural network by inputting the output value of the electronic circuit package from the measuring unit including a hierarchical neural network having an intermediate layer And a control section having an adjusting means for outputting the control information to the mechanical section.

An electronic circuit package automatic adjusting apparatus according to a third aspect is the electronic circuit package automatic adjusting apparatus according to the second aspect, wherein the control section outputs a prescribed input signal to a plurality of initial setting states of the adjusting trimmer. The rotation amount of the trimmer for adjustment of the electronic circuit package for keeping the output value when given within the predetermined allowable range with respect to the reference output value is obtained based on the method of the genetic algorithm, and these are used as teacher signals for the neural network. It is characterized by including a learning means for learning the network.

[0010]

The basic technical idea of the present invention will be described. As described above, the object of the present invention is to provide an adjusting method and an automatic adjusting device which can be universally applied to various electronic circuit packages without depending on the skill of an operator, and achieve this object. In order to do so, a neural network having a learning function and a method of a genetic algorithm are used.
More specifically, the basic technical idea of the present invention is as follows: learning of a neural network by the method of the genetic algorithm shown in FIG. 1 (a) and adjustment by a learned neural network shown in FIG. 1 (b). It is divided into two stages:

The neural network is an information processing system formed by modeling the human brain nerve cell structure, and is a large-scale parallel network composed of a large number of nonlinear processing units interconnected by links having a variable weighting function. . In a hierarchical network having a hidden unit forming an intermediate layer between an input layer composed of a processing unit to which an input pattern is supplied and an output layer composed of a processing unit generating an output pattern, a known input / output relationship is established. By learning as a teacher signal, the variable weighting of links is modified and the internal state of the network changes, so that an appropriate output pattern can be output even for similar input patterns other than those learned. It is known to have a function.

The electronic circuit package targeted by the present invention generally has k inputs, n outputs, and m adjusting trimmers (k, n, and m are all positive integers). It is an electronic circuit, and in a state where a specified input signal is applied to k inputs, m adjustment trimmers are set so that n output values are within a predetermined allowable range with respect to each reference output value. Shall be designed to be adjustable using. The example of the gain and offset adjustment of the electronic circuit by the fuzzy inference described in the above-mentioned Japanese Patent Application Laid-Open No. 2-92004 corresponds to the case where the target electronic circuit is k = 1, m = n = 2.

At the time of learning, one or a few electronic circuit packages equivalent to the circuit to be adjusted are used. A specified input signal is supplied to each of the k inputs of the electronic circuit package, and the m adjustment trimmers are fixed to one of the set initial states (combination of the initial positions of the adjustment trimmers). The output values Xs of n outputs are input to the GA operation unit 1 and the neural network 2 which execute the method of the genetic algorithm, as shown in FIG. First,
Output value Xs of the electronic circuit package by the GA operation unit 1
The amount of rotation of the m trimming trimmers necessary for adjusting to the reference output value Xo is obtained, and the neural network 2 is made to learn this amount as the teacher signal Zs. This learning operation is
This is executed inside the neural network 2 by back propagation that returns the error between the rotation amount Ys output from the neural network 2 and the teacher signal Zs to the input layer side. Here, the output value Xs, the reference output value Xo,
The rotation amount Ys and the teacher signal Zs are, as shown in FIG. 1A, a matrix of one column consisting of n and m elements, respectively, and the relation of input and output of the neural network 2 is Ys = F (Xs ), F is a complex function of m × n order.

By repeating the above operation β times by changing the initial state of the electronic circuit package, the neural network 2 changes the internal state and completes learning, and also for similar input patterns other than those learned. A proper output pattern can be obtained. This learning is a specific one
Although it may be performed with only one electronic circuit package, some electronic circuit packages may be replaced.

The method of the genetic algorithm is developed as a method for efficiently solving the search problem of a huge area, inspired by the process of evolution of living organisms, and it is a group of genes expressing multi-bit information. Assuming that, the generation alternation simulation is repeated by adding a series of operations such as selection, proliferation, genotype crossover, and mutation to this population, and an optimal solution that matches the conditions is obtained. Specifically, the rotation amount of the m trimming trimmers of the electronic circuit package is coded into one gene, and the output value and the reference that is the regulation target when the trimming trimmer is rotated based on the gene information as the fitness. The optimum solution (teaching signal) is obtained using the error from the output value. Starting from a group of genes with randomly selected gene information, it is possible to obtain the evolved gene of interest by repeatedly executing simple basic operations. Anyone can easily create a teacher signal without knowing the technical contents.

Next, at the time of adjustment, as shown in FIG. 1 (b), the neural network 2 after learning is used,
The output value X of the electronic circuit package to be actually adjusted is input, the rotation amount Y of the adjustment trimmer corresponding to the input is calculated, and the adjustment trimmer is rotated according to the calculated rotation amount Y to execute the adjustment. To do. If the desired reference output value is not reached with one adjustment, repeat the same operation.

For neural networks, especially for hierarchical networks and backpropagation,
For example, "PDP Model-Cognitive Science and Search for Neuron Networks- (First Edition)" (DE Lamelhardt et al., PDP Research Group, translated by Shunichi Amari, Sangyo Tosho, 1989).
Detailed description is given in Chapter 8 (published February 27, 2012) (P322 to P333). Furthermore, regarding genetic algorithms, for example, "Genetic Algorithm (First Edition)"
(Takeshi Yasuiin and Tomoharu Nagao, Shokodo, September 1993 2
It is generally described in Chapter 1 (P1-P15) of the 2nd issue).

[0018]

Embodiments of the present invention will now be described with reference to the drawings.

FIG. 2 is a block diagram showing the configuration of an embodiment of the electronic circuit package automatic adjusting apparatus according to the present invention.

The electronic circuit package automatic adjusting apparatus of this embodiment, as shown in FIG. 2, applies a specified input signal from the measuring instrument to the electronic circuit package 6 to be adjusted, measures the output value, and outputs it as digital information. The measuring unit 3 and the mechanical unit 4 that rotates the adjusting trimmer 7 of the electronic circuit package 6 according to the control information by the robot mechanism 8 including the adjusting head 9, and the electronic unit that controls the measuring unit 3 and the mechanical unit 4. It is composed of a control unit 5 for adjusting the package 6. The control unit 5 is composed of, for example, an information processing device such as a workstation or a personal computer, has a function of realizing a neural network having a learning ability by software simulation, and a function of updating information based on a genetic algorithm method. A function of controlling the measuring unit 3 and the mechanism unit 4 to create a teacher signal by the method of the genetic algorithm using the electronic circuit package 6 and allowing the neural network to learn, and an electronic circuit package to be adjusted using the learned neural network. 6 and the function of executing the adjustment.

In the case of learning, the control unit 5 controls the measuring unit 3 to receive the output value from the electronic circuit package 6 as digital information, and while controlling the mechanism unit 4, creates a teacher signal by the method of the genetic algorithm. Then, the operation of making the neural network learn is repeated according to the prescribed procedure. When the adjustment is performed, the digital information of the output value of the electronic circuit package received from the measuring unit 3 is input to the neural network, the mechanism unit 4 is controlled by the obtained result, and the adjusting trimmer 7 of the electronic circuit package 6 is controlled. To rotate. As a result, if the digital information from the measuring unit 3 matches the reference output value, the adjustment ends. If they do not match, readjustment is performed in the same manner.

FIG. 3 is a flow chart showing the detailed operation at the time of learning. Hereinafter, the operation at the time of learning will be described with reference to FIGS. 2 and 1A. First, the number of teacher signals to be used for learning (β) is determined, and only the number of different initial setting conditions ( A combination of the initial positions of m trimming trimmers 7 is set (step S1). continue,
A plurality (α that stores the rotation amount information of the m adjustment trimmers 7)
Individual genes) are set, and all genes are randomly initialized (step S2). When the amount of rotation of the adjusting trimmer 7 is represented by 8-bit digital information, one gene has an information amount of 8 m bits. Next, it is determined which of the initial setting conditions previously set in step S1 is used (step S3).

The control section 5 controls the mechanical section 4 to adjust the position of the adjusting trimmer 7 of the electronic circuit package 6 to the designated initial setting condition (step S4), and then the measuring section 3
Supplies a prescribed input signal to the electronic circuit package 6 (step S5), measures the output value, and stores it as the output value Xs for the initial setting condition (step S6). Then, one of the α genes set in step S2 is selected (step S7), and each adjusting trimmer 7 is rotated based on the rotation amount information coded in the gene (step S8). The output value from the electronic circuit package 6 is measured again and stored (step S9). After the processing from step S4 to step S9 is executed for all the set α genes (step S1
0), the output value stored in step S9 is compared with the reference output value Xo of the adjustment target to calculate the fitness (step S1).
1). For the fitness, the error with respect to the reference value corresponding to each of the n outputs is evaluated, and based on them, the comprehensive quantitative evaluation of the fitness is also performed. Based on this evaluation result, it is determined whether or not there is a gene whose fitness for all outputs satisfies a predetermined condition (step S12). If there is no gene that satisfies the condition, gene manipulation is performed by the method of the genetic algorithm (step S13), the process returns to step S4 and the above-described processing is continued. On the other hand, if there is a gene that satisfies the condition, the gene with the best overall evaluation is selected, and the teacher signal Z is made to correspond to the output value Xs previously stored in step S6.
It is registered as s (step S14). Next, the number of registered teacher signals is determined (step S15), and if β has not been set in step S1, step S15 is performed.
Returning to step 2, the unprocessed initial setting condition is selected and the same process is repeated. When the processing for all the initial setting conditions is completed and β teacher signals are registered, the learning of the neural network 2 is started.

The neural network 2 is a hierarchical network for inputting the output values of the n outputs of the electronic circuit package 6 and outputting the rotation amount of the m adjustment trimmers 7, and the output values and the rotations. When each quantity is represented by 8-bit digital information, at least 8n binary input terminals are provided in the input layer and 8m binary output terminals are provided in the output layer. The output value Xs is applied to the input terminal of the neural network 2 and the teacher signal Zs is applied to the output terminal to sequentially perform the learning operation (step S16). When learning for all registered teacher signals is completed (step S1).
7), all the operations of the learning process are completed.

FIG. 4 is a flowchart showing another operation example during learning. In FIG. 4, step S21
The operation of each step from step S33 to step S33 is the same as the operation of each step from step S1 to step S13 shown in FIG. In the case of FIG. 4, step S33.
Repeating the gene operation in step S32, if there is a gene that satisfies the condition as a result of the fitness determination in step S32, the gene with the best overall evaluation is selected, and it is used as the teacher signal Zs to immediately learn the neural network 2 in step S34. To do. After learning, it is determined whether or not all learning of the learning conditions (the number of teacher signals and the initial setting condition) set in step S21 is completed (step S35), and if not completed, the process returns to step S22 and the same processing is repeated. , The learning process is completed when the set number of learnings are all performed.

FIG. 5 is a flow chart showing the operation at the time of adjustment in the embodiment of FIG. First, an input signal is supplied from the measuring unit 3 to the electronic circuit package 6 to be adjusted (step S41), and the output value X of n outputs is measured (step S42). The control unit 5 receives this as digital information and inputs it to the learned neural network 2 to determine the corresponding rotation amount Y (step S43).
Then, the mechanical section 4 is controlled based on the determined rotation amount Y, and the m adjustment trimmers 7 of the electronic circuit package 6 are rotated (step S44). After the rotation, the output value of the electronic circuit package 6 is measured again (step S45), and the output value after this rotation (denoted as Xa) is compared with the target reference output value Xo. It is determined whether or not the adjustment target condition is satisfied (step S4
6) If the condition is not satisfied, the process returns to step S43, the output value Xa after rotation is input to the neural network 2 and the same process is continued, and if the condition is satisfied, the adjustment ends.

In the above-described embodiment, the control unit has both a learning function of creating a teacher signal by the method of the genetic algorithm to learn the neural network and an adjusting function of actually adjusting the learned neural network. I was prepared. However, as is clear from the above description, the learning of the neural network and the execution of the adjustment can be executed independently, so the automatic adjustment device does not include the function of creating the teacher signal by the genetic algorithm. Good. Further, although the neural network is realized by software simulation by a computer, it may be realized together with hardware or by hardware alone.

[0028]

As described above, the electronic circuit package adjusting method and the automatic adjusting apparatus of the present invention use the genetic algorithm to determine the rotation amount of the adjusting trimmer corresponding to the output value from the electronic circuit package to be adjusted in advance. In order to perform adjustment by determining the required rotation amount of the trimmer for adjustment by using the learned neural network at the time of adjustment,
Even an unskilled operator can easily make adjustments, which has the effect of reducing adjustment man-hours and time.

Further, since the teacher signal necessary for learning the neural network is obtained by the method of the genetic algorithm, if only the mechanical operation procedure is known, the configuration and characteristics of the target electronic circuit package need not be known. The learning of the neural network is possible, and there is an effect that it can be applied to the adjustment of a wide variety of general electronic circuit packages without particularly limiting the users.

[Brief description of drawings]

FIG. 1 is a block diagram for explaining a basic technical idea of the present invention.

FIG. 2 is a block diagram showing a configuration of an exemplary embodiment of the present invention.

FIG. 3 is a flowchart showing an example of an operation at the time of learning in the embodiment of FIG.

FIG. 4 is a flowchart showing another example of the operation at the time of learning in the embodiment of FIG.

5 is a flowchart showing an example of an operation at the time of adjustment in the embodiment of FIG.

FIG. 6 is an explanatory diagram of a conventional general electronic circuit package adjusting method.

[Explanation of symbols]

 1 GA operation part 2 Neural network 3 Measuring part 4 Mechanism part 5 Control part 6 Electronic circuit package 7 Adjustment trimmer 8 Robot 9 Adjustment head 10 Worker 11 Measuring instrument 12 Bit X, Xs Output value Xo Reference output value Y, Ys Rotation amount Zs Teacher signal

Claims (3)

[Claims] 1. A rotation amount of an adjusting trimmer of an electronic circuit package for keeping an output value when a specified input signal is given within a predetermined allowable range with respect to a reference output value. Are obtained by a genetic algorithm method with respect to the initial setting state of, and learned by a hierarchical neural network having an intermediate layer as a teacher signal, and then the specified input signal is applied to each electronic circuit package to be adjusted. The rotation amount of the adjustment trimmer is obtained from the neural network based on the output value when the value is given, and the adjustment trimmer is rotated according to the obtained rotation amount so that the output value is a predetermined value with respect to the reference output value. A method of adjusting an electronic circuit package, which is characterized in that adjustment is performed within an allowable range. 2. A measuring unit for applying a specified input signal to an electronic circuit package to be adjusted to measure an output value, a mechanism unit for rotating an adjusting trimmer of the electronic circuit package according to control information, and an intermediate layer. The output value of the electronic circuit package is input from the measuring unit including the hierarchical neural network having the above, and the rotation amount of the adjustment trimmer is obtained by the learned neural network, and the control information is output to the mechanism unit. An electronic circuit package automatic adjusting device, comprising: 3. The control unit sets an output value when a specified input signal is applied to a plurality of initial setting states of the adjustment trimmer within a predetermined allowable range with respect to a reference output value. 3. The electronic circuit package according to claim 2, further comprising: learning means for obtaining a rotation amount of the trimmer for adjustment of the electronic circuit package based on a method of a genetic algorithm, and using these as a teacher signal to learn the neural network. Automatic adjustment device.