JP3378996B2 - Waveform evaluation device using neural network - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a waveform evaluation device for evaluating a waveform measured by a synchroscope or other measuring device, and in particular, a plurality of neural network modules are independently provided for each determination target. Individually formed, these nerve weights relate to a waveform evaluation device using a neural network that is determined by learning by an ideal teacher of a signal and an opposite teacher corresponding to an element to be determined. . The present invention is, for example, an optimum waveform evaluation device for an inspection device for inspecting a good or defective PC board or the like.

[0002]

2. Description of the Related Art Conventionally, in order to evaluate a waveform measured by a synchroscope or the like, there has been only a method of creating a program or the like according to the waveform shape or making a judgment by an observer visually. .

[0003]

However, in order to evaluate a waveform, since the waveform has various shapes, it is necessary to create a program for evaluating each waveform when creating a program or the like. However, there was a problem that it was extremely complicated. Furthermore, these programs have the problem that they are not versatile and inefficient. The evaluation of the analog waveform is difficult to describe by a program, and the visual evaluation has a problem that it is not possible to carry out a quantitative evaluation due to individual differences among observers.

The electric signal waveform obtained from the object to be judged is composed of many elements, some of which are the objects of waveform evaluation and some of which are unnecessary.

Therefore, for these reasons, it is difficult to mechanize and automate, and without writing a program,
There has been a strong demand for the appearance of a waveform evaluation device that can evaluate all waveforms.

[0006]

SUMMARY OF THE INVENTION The present invention has been devised in view of the above problems, and is composed of at least three layers of an input layer, a hidden layer and an output layer to which a signal to be judged is inputted as an input signal, A plurality of neural network modules having a nerve weight ratio are independently formed for each determination target including at least a phase of a waveform and a cycle, and the nerve weight ratio of each neural network module is equal to that of an ideal signal. A model teacher and a determination unit for determining pass / fail of a waveform based on a plurality of identification outputs of the neural network module, which is determined by learning with a face teacher corresponding to each element to be determined Is characterized by. Further, according to the present invention, a waveform pre-processing unit for performing at least waveform cutting, enlarging and reducing waveform pre-processing on the signal to be judged is provided, and the output of the waveform pre-processing unit is input to the neural network module. It can also be configured. Further, the determination target of the present invention can be configured such that parameters such as waveform loss, amplitude value, effective value, maximum value, minimum value, and frequency are selected. Then, the present invention provides a plurality of neural network modules independent for each waveform data of each part of the measurement object, and based on the identification result thereof, logical processing, fuzzy inference processing, etc. It is also possible to provide an integrated reasoning determination unit for reasoningly determining the inconvenient part of the object and the overall performance.

[0007]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention configured as described above is
A determination target including a neural network module having at least three layers of an input layer, a hidden layer, and an output layer to which a determination target signal is input as an input signal and including at least a waveform phase and a cycle A plurality of independent neural networks are formed for each neural network module, and the neural weight of each neural network module is determined by learning with an ideal signal model teacher and a counter teacher corresponding to each element to be judged, and the judgment unit However, the quality of the waveform can be determined based on the plurality of identification outputs of the neural network module. Further, the waveform pre-processing unit of the present invention, at least the waveform cut-out for the determination target signal,
It is also possible to perform waveform pre-processing of expansion and contraction and input the output of the waveform pre-processing unit to the neural network module. Further, parameters such as waveform loss, amplitude value, effective value, maximum value, minimum value, and frequency can be selected as the determination target of the present invention. Then, the present invention provides a plurality of neural network modules which are independent for each waveform data of each part of the measurement object, and performs logical processing, fuzzy inference processing, or the like based on the identification results thereof, and an integrated inference determination unit. However, it is also possible to infer and determine the inconvenient part of the measurement object, the overall performance, and the like.

[0008]

【Example】

An embodiment of the present invention will be described with reference to the drawings.

A case where the present invention is applied to the waveform evaluation apparatus 1 for judging the quality of a waveform will be described. As shown in FIG. 1, the waveform evaluation apparatus 1 according to the present embodiment includes a learning execution system A and a waveform evaluation system B. The learning execution system A outputs a signal from the determination target object 10 to a waveform statistics collection unit. 21, the teacher waveform determination unit 22, the waveform identification unit 23, and the determination unit 24. In the waveform evaluation system B, the signal from the determination target 10 is the waveform sampling unit 25, the waveform preprocessing unit 26, and the waveform identification unit. 23 and the determination unit 24. Note that the determination target object 10
Is a device under test that outputs a signal whose waveform is to be evaluated.

The waveform statistical sampling unit 21 is for sampling the waveform data from the object 10 to be judged one after another and statistically processing these to form an ideal ideal waveform. That is, the waveform statistic sampling unit 21 has a statistical processing function and a function of sampling only the waveform to be sampled. The waveform slicing function, the waveform enlarging function, the waveform reducing function, the waveform enlarging function, the waveform portion reducing function, and the waveform shape. The change function is provided. These functions facilitate the formation of an appropriate teacher waveform for facilitating waveform feature extraction from the acquired waveform data.

The teacher waveform determining unit 22 corresponds to the characteristic elements of the waveform identified by the waveform identifying unit 23, and the model teacher, the first reversal teacher, and the second reversal teacher based on the ideal waveform for each characteristic element. Is determined and output to the input layer formed in each neural network module.

The waveform identifying section 23 is provided with an independent neural network module having the number of characteristic elements of the waveform to be identified. For example, a chip, a phase, an amplitude value, a period, etc. correspond to the characteristic element of this waveform. Then, the waveform identifying unit 23 determines the nerve weight ratio during learning, and determines the waveform during determination.

The determination unit 24 is for making a comprehensive determination from the identification result of the waveform identification unit 23, and is configured to send the determination result to the control calculation unit 27 and the display unit 28.

The waveform sampling unit 25 is for sampling a waveform from the measurement object 10, and the waveform preprocessing unit 26 is
This is for performing waveform preprocessing such as cutting out, enlarging, and reducing the collected waveform.

The control calculation unit 27 is for switching the input waveform signal between the learning execution system A and the waveform evaluation system B. The display unit 28 is for displaying the determination result of the determination unit 24.

Next, the structure of the multilayer neural network that constitutes the neural network used in this embodiment will be described.

A neural network is composed of a plurality of nerve cells (neurons), and one neuron is composed of a cell body, dendrites (signal input portion), and axons (signal output portion). . Axon (signal output part)
Are synaptically linked to the dendrites of other neurons, forming a network.

The learning method applied to this neural network is called a backpropagation method, and the structure of the neural network has a multilayer structure of an input layer, an intermediate layer, and an output layer.

The backpropagation method used for learning the multilayer neural network will be briefly described below.

A back propagation execution subroutine will be described. First, an initial value of nerve weight ratio is set by a random number. Then, iterative calculation is performed for the number of learning data. The output of each cell is calculated, and the evaluation function is further calculated. Next, the nerve weight correction amount is calculated in order to minimize this evaluation function value. Then, the nerve weight is corrected based on the nerve weight correction amount, and the calculation is repeated. As a result, when the sum of the evaluation functions is less than or equal to the predetermined value, the final nerve weight is output.

Next, the waveform identification section 23 of this embodiment will be described in detail. As shown in FIG. 2, the waveform identifying unit 23 uses a relatively simple n neural network module 231 (a), 231.
(B) ... 231 (n). As shown in FIG. 3, each neural network module 231 includes at least three layers of an input layer 2311, a hidden layer 2312, and an output layer 2313, and has a nerve weight.

Each neural network module 231,
The input data of 231, ..., For example, as shown in FIG. 4, waveform data is sampled at fixed time intervals,
The time interval is assigned to one neuron. The input value of each neuron may be, for example, a value normalized to 1 in the amplitude range determined from the maximum voltage value and the minimum voltage value of the waveform. The number of input neurons is set according to the shape of the waveform that the observer wants to observe. Also, each neural network module 231, 231.
The waveforms input to .. may be the same or different, for example, the waveforms whose characteristics are well represented may be used. That is, the number of input neurons may be different or the same for each module. Further, the respective network configurations may have different numbers of neurons in the hidden layer and the output layer, or may have the same number.

Each neural network module 231, 23
1 ... Is individually configured for each feature element to be identified, and is configured, for example, for each phase, amplitude, shape, and other feature elements to be particularly extracted.

Further, the teacher signal is formed by the teacher waveform determining unit 22, and each neural network module 231, 231 ...
.. is selected as a model teacher, and the waveform desired to be identified as no is set to the representative waveform as a teacher. The model teacher, for example, causes one output neuron to learn so that the output becomes 1, while the opposite teacher learns so that the response of the output neuron becomes 0. Or
Another output neuron may be prepared, and the neuron may be configured to react with 1 or 0. Further, it is not limited to 1 or 0, and any one can be used as long as it can be distinguished by a threshold value.

Next, the operation of this embodiment will be described. The waveform evaluation apparatus 1 of the present embodiment is configured to perform pre-learning for evaluation when performing waveform evaluation. Therefore, the learning procedure of the waveform evaluation apparatus 1 will be described with reference to FIG.

First, learning is started in step 1 (hereinafter abbreviated as S1), and the control calculation unit 27 switches so that the input waveform is guided to the learning execution system A. Next, in S2, the waveform data to be sampled from the measurement object 10 is input to the waveform statistical sampling unit 21, and the waveform statistical sampling unit 21 performs the waveform statistical processing to form an ideal waveform. And S
In 3, the waveform statistical sampling unit 21 determines the waveform to be sampled,
Collect.

Next, in S4, the neural network module 231 (a), 231 (b), ... To be learned is selected.
The neural network module 231 selected here is shown in FIG.
Neural network module 231 for “chip detection” shown in FIG.
In the case of (a), the waveform of FIG. 6 (a) is used as a model teacher in S5, and the waveform of FIG. 6 (b) is used as a first side teacher, FIG. 6 (c).
Is determined as the second opposite teacher. Note that FIG.
The ideal waveform is adopted as it is in the model teacher of the waveform of (a).

Then, learning is executed based on the determined teacher waveform. First, in S6, the nerve weight is initialized. Next, in S7, the evaluation function of the output layer in the state where the model teacher waveform is input is calculated. And in S8, S
It is determined whether or not the evaluation function value calculated in 7 is within the predetermined range, and if it is within the predetermined range, the process proceeds to S9 to determine whether or not all teacher learning is finished. In S8, S
If the evaluation function value calculated in 7 is outside the predetermined range, S1
The process proceeds to 0 to calculate the nerve severity correction amount, and in S11, S1
The nerve weight is corrected based on the nerve weight correction amount of 0. Then, the process returns to S7 to perform the evaluation function calculation again, and this routine is repeated until the nerve weight ratio in which the evaluation function value falls within the predetermined range is determined.

If it is determined in S9 that learning has been completed for all the teacher waveforms of the selected neural network module, the process proceeds to S12 and all neural network modules 231 (a), 231 ( b) It is judged whether or not learning has been completed. If it is determined in S12 that the learning has been completed for all the neural network modules 231 (a), 231 (b), ... Then, the procedure proceeds to S13 and all learning is completed.

If it is determined in S9 that learning has not been completed for all the teacher waveforms of the selected neural network module 231 (a), 231 (b) ,.
In S14, the learning waveform that has not ended is used as an input signal, and the process returns to S6 to perform learning.

Further, in S12, when it is judged that the learning is not completed in all the neural network modules 231 (a), 231 (b) ..., the processing proceeds to S15, in which the neural network is not completed. Change the learning target to module 231,
Returning to S5, the changed neural network module 231
Learning is performed by the teacher waveform corresponding to. If the teacher signal is unsuitable and the nerve weight cannot be determined, or if proper judgment cannot be made, the teacher signal is changed and re-learning is performed.

The waveform evaluation apparatus 1 of this embodiment for performing waveform evaluation (judgment) using the pre-learning configured as described above.
The operation of will be described.

As shown in FIG. 7, step 1 (hereinafter referred to as S1
Waveform evaluation (judgment) is started, and the control calculation unit 2
7 switches so that the input waveform is guided to the waveform evaluation system B. Next, in S2, the waveform sampling unit 25 samples a waveform from the measurement object 10. Then, in S3, the waveform pre-processing unit 26 performs waveform pre-processing such as waveform cutting, enlargement, reduction, etc. on the waveform sampled in S2. And in S4,
The waveform identifying unit 23 identifies the waveform pre-processed in S3, and in S5, the determining unit 24 determines the waveform of the identified waveform. Further, in S6, the display unit 28 displays the determination result determined in S5 by a display method such as Good / NoGood. In S7,
It is determined whether or not the measurement is completed. If the measurement is completed, the process proceeds to S8 to end the measurement. If it is determined in S7 that the measurement has not ended, the process returns to S2. The display unit 28 includes a waveform statistics collection unit 21, a teacher waveform determination unit 22,
It is also possible to display the data of the waveform sampling unit 25 and the waveform preprocessing unit 26.

Next, the case where the waveform evaluation apparatus 1 of this embodiment evaluates a specific waveform will be described.

[Detection of chip] will be described first. Using the analog waveform shown in FIG. 6A as a model teacher,
When it is desired to discriminate the waveform as shown in FIG. 6B, the waveform shown in FIG. 6B is set as the first representative teacher as a representative waveform. Further, the second kind of DC data shown in FIG.
However, it can be set as a teacher.

With the settings as described above, the response of one output neuron becomes 1 by using the waveform as shown in FIG. 6A as a model teacher, and the waveform as shown in FIG. On the other hand, if learning is performed so as to react with 0 as a teacher, at least a waveform similar to FIG. 6A becomes close to 1, and a waveform similar to FIG. 6B shows a reaction close to 0. Therefore, these waveforms can be easily identified by setting a threshold value at a specific place.

The neural network module 231 that has performed such learning does not impose any restrictions on the phase when focusing on the phase, and thus the phases are slightly shifted in the shapes of FIGS. 6A and 6B. The waveform that is present can be identified.
Moreover, since the generalization ability is high, the unfavorable waveform shown in FIG. 6D can be identified. In addition, the second opposite teacher shown in FIG. 6C is the model teacher shown in FIG.
It is possible to perform the discrimination in the same manner by learning any output neuron together with the first side teacher shown in (b), and it is possible to improve the discrimination ability.

Next, the case of [phase shift determination] will be described.
Neural network module 231 (b) for detecting phase shift
In FIG. 8, a waveform having a typical phase as shown in FIG. 8A is selected as a model teacher, and the first phase is shown in FIG.
The waveforms with a 45 ° phase delay as shown in (b) are selected, and the waveforms with a 45 ° phase advance as shown in FIG. 8 (c) are selected as the second opposite teacher to learn these waveforms. You can do it. That is, the first reversal teacher and the second reversal teacher are typical examples of unfavorable phases. The neural network module for detecting this phase shift is independent of the amplitude and the shape of the waveform, and has at least the ability to identify whether the phase is good or bad.

When it is desired to simultaneously identify two characteristic elements such as a waveform as shown in FIG. 6B and a phase shift waveform, the neural network module 231 (a) and the neural network module 231 ( The two neural network modules 231 in b) may be identified. Therefore, when it is desired to identify a plurality of characteristic elements, the plurality of neural network modules 231 (a) and 231 which focus on each characteristic are emphasized.
(B) ... should be created.

As described above, at least the waveform to be identified can be identified only by learning the amplitude, shape, and phase, especially the place to be observed, etc., by giving a representative model teacher and a face teacher to each network. .

Further, as shown in FIG. 6 (e), when the place to be observed changes, a neural network module 231 for detecting the defect waveform of FIG. 6 (e) may be prepared. When it is desired to detect a defect waveform as shown in FIG. 6F, a neural network module 231 for detecting the defect waveform as shown in FIG. 6F may be additionally created. A plurality of neural network modules 231 (a),
As another learning example without creating 231 (b) ..., the defect shown in FIG. 6 (b), FIG. 6 (e), and FIG. 6 (f) for one neural network module 231. It is also possible to have the ability to detect. In this case, the representative teachers of FIGS. 6 (b), 6 (e), and 6 (f) may be the opposite teachers, and FIG. 6 (a) may be the model teacher or the like. Even with such a configuration, since the respective teachers detect the defects of the corresponding waveforms, the teachers shown in FIGS.
(E), the characteristic element of the defect shown in FIG. 6 (f) can be detected.

6 (a) to 6 (d) are analog waveforms, the same applies to rectangular waveforms shown in FIGS. 6 (g) to 6 (i). Has discriminating ability. That is, by setting the representative ones shown in FIGS. 6G and 6H as the model teacher and the opposite teacher,
The unfavorable waveform shown in (h) or FIG. 6 (i) can be detected. Thus, if the number of characteristic elements to be identified is one to several for each neural network module 231, 231, ..., Generalization ability is increased, and at least the characteristic element to be identified can be detected. 231 can be created.

In the waveform evaluation apparatus 1 of the present embodiment configured as described above, the determination unit 24 outputs from the neural network module 231 (a), 231 (b), ... Of the waveform identification unit 23. Good / N of the waveform based on the plurality of identification results
It is designed to judge oGood. However, the waveforms to be determined are those of the neural network module 231 (a), 2
When it is obviously inconvenient without checking the output result of 31 (b) ..., NoGood is directly set. In such a case, for example, the amplitude of the input waveform is extremely large or small, and it is set to NoGood without checking the result of the waveform identifying unit 23.

And a plurality of neural network modules 231
(A), 231 (b) ... When the determination unit 24 makes a determination, all the neural network module modules 231 (a), 231 (b) ... The waveform is determined to be Good when the response is equal to or more than the set threshold value, and the waveform is determined to be NoGood when no response occurs.

When the number of characteristic elements to be discriminated increases,
Neural network module 2 for discriminating its characteristic element
It is only necessary to add 31 for learning, and it is not necessary to re-learn other neural network module 231 that has already been learned. Further, the determination of the determination unit 24 may be added to the quality determination based on the output result of the model teacher identification neuron of the added network module. In this way, each neural network module 231 (a), 231 (b).
.. can be executed at high speed, the number of neurons does not need to fall to a local minimum, and a small memory capacity can be executed.
In addition, the additional learning is performed by the corresponding neural network module 23.
This is possible only by increasing 1, and there is an effect that it can be easily realized.

Next, a first modified example of the present invention will be described.
In the first modification, as shown in FIG. 9, for the input neurons of the neural network module 231 (a), 231 (b), ... Parameters such as evaluation values related to are added. That is, parameters such as the effective value, the average value, the maximum value, the minimum value, the frequency, or some other evaluation value (which may be the output value of the neural network module 231) of the waveform are set to 1 as an analog value. By normalizing and adding,
More neural network modules 231 (a), 231 (b)
The identification ability of ... Can be improved.

A second modified example of the present invention will be described with reference to FIG. In this second modification, an integrated inference determination unit 30 for performing an inference determination function is added, and the integrated inference determination unit 30 includes a plurality of neural network module 231 (a),
231 (b) ... are further connected in parallel. That is, a plurality of neural network modules 231
(A), 231 (b) ... Are made independent for each waveform data of each part of the measurement object 10 (waveform 1, waveform 2, ... Waveform n), and the output results thereof are provided to the integrated inference determination unit 30. It is connected. As shown in FIG. 11, this second modified example has neural network module 231 (a), 23 of each part.
From the identification result of 1 (b) ..., the measured value of the waveform, and the like, it is possible to infer and determine the inconvenient portion of the measurement object 10, the overall performance, and the like. The integrated inference determination unit 30 is composed of logic processing, another neural network, fuzzy inference processing, or the like. By providing such an additional function, it is possible to provide a sophisticated and intelligent waveform evaluation apparatus 1.

[0049]

[Effect] The present invention configured as described above is composed of at least three layers of an input layer, a hidden layer, and an output layer, which are input with a determination target signal as an input signal, and has a nerve weight ratio. A plurality of circuit modules are independently formed for each determination target including at least the phase and period of the waveform, and the neural weight of each neural network module is an ideal signal model teacher and each determination target element. On the other hand, it is determined by learning with a teacher corresponding to the above, and based on a plurality of identification outputs of the neural network module, a determination unit for determining the quality of the waveform is provided. There is an effect that it is not necessary to create the waveform, and the waveform can be evaluated only by performing simple learning. That is, since a plurality of neural network modules are configured for each characteristic element and a teacher and a teacher are set for each characteristic element to be extracted, it is possible to evaluate any waveform by simply performing simple learning. There is. In particular, it is configured such that the input waveform is first decomposed or extracted by the signal processing into the characteristic elements to be the object of the quality judgment, the teacher and the opposite teacher are created based on this, and the judgment is made after learning the neural network. Therefore, there is an outstanding effect that a good judgment can be made.

Further, even if the number of characteristic elements to be identified increases, there is an excellent effect that the neural network module corresponding to the new characteristic element only needs to be added without changing the already prepared neural network module. Further, the present invention has an effect that it is possible to provide a waveform evaluation system that is highly versatile, has no individual difference in determination, and is easy to mechanize and automate.

[0051]

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of an exemplary embodiment of the present invention.

FIG. 2 is a diagram illustrating a neural network module of this embodiment.

FIG. 3 is a diagram illustrating a neural network module of this embodiment.

FIG. 4 is a diagram for explaining input data of the neural network module of this embodiment.

FIG. 5 is a diagram illustrating a learning procedure of the waveform evaluation apparatus 1 according to the present embodiment.

FIG. 6 is a diagram illustrating a case where the waveform evaluation apparatus 1 according to the present embodiment performs “chip detection”.

FIG. 7: Waveform evaluation (judgment) of the waveform evaluation apparatus 1 of the present embodiment.
It is a figure explaining.

FIG. 8 is a diagram illustrating a case where the waveform evaluation apparatus 1 of the present embodiment performs “phase detection”.

FIG. 9 is a diagram illustrating a first modified example of the present invention.

FIG. 10 is a diagram illustrating a second modified example of the present invention.

FIG. 11 is a diagram illustrating a second modified example of the present invention.

[Explanation of symbols]

1 Waveform evaluation device 10 Object to be judged 21 Waveform Statistics Collection Unit 22 Teacher waveform determination unit 23 Waveform identification section 24 Judgment section 25 Waveform sampling section 26 Waveform preprocessing unit 27 Control calculation unit 28 Display

Claims (4)

(57) [Claims] 1. A neural network module having at least three layers, an input layer, a hidden layer, and an output layer, which are inputted with a signal to be judged as an input signal, and which has a nerve weight ratio.
A plurality of the neural network modules are independently formed for each determination target including at least the phase and period of the waveform, and the neural weight of each neural network module corresponds to each of the ideal signal model teacher and the determination target element. It is determined by learning by a teacher and is based on a plurality of identification outputs of the neural network module.
A waveform evaluation device using a neural network, characterized by comprising a determination unit for determining the quality of the waveform. 2. At least a waveform of the signal to be judged
Waveform preprocessing unit that performs waveform preprocessing such as cutout, enlargement, and reduction
The output of this waveform preprocessing section is
The neural circuit according to claim 1, wherein the neural circuit is configured to be input to a tool.
Waveform evaluation device using a net. 3. The judgment target is a waveform lack, an amplitude value, and an effective value.
Parameters such as value, maximum value, minimum value and frequency are selected.
A waveform evaluation device using the neural network according to claim 1.
Place 4. The waveform data of each part of the measuring object is independently set.
Provide multiple neural network modules that are
Logical processing or fuzzy inference processing based on the identification result
By performing the
A contract provided with an integrated inference determination unit for inferring and determining performance, etc.
A waveform evaluation apparatus using the neural network according to claim 1.