JP2020177287A - Specimen evaluation system and method for constructing the same - Google Patents

Abstract

To make it possible to accurately extract the feature quantity of a specimen after testing and accurately determine the quality of the specimen based on the feature quantity.SOLUTION: A feature quantity extraction unit of a specimen evaluation system of the present disclosure corresponds to an encoder that constitutes an autoencoder together with a decoder, and is adapted by repeating learning for the autoencoder to restore input image data, with the decoder, from a feature quantity extracted by the encoder, and a threshold used by a determination unit of the specimen evaluation system is determined by clustering the plurality of feature quantities extracted, by the adapted feature extraction unit, from a plurality of normal specimens.SELECTED DRAWING: Figure 1

Description

The present disclosure relates to a test product evaluation system for evaluating a test product after a test and a method for constructing the same.

Conventionally, an auto encoder (self-encoder) that dimensionally compresses input data, converts the abstract features of the input data into feature vectors (feature quantities), and reproduces the input data from the obtained feature vectors has been known. (See, for example, Non-Patent Document 1). The self-encoder of Non-Patent Document 1 includes an encoder and a decoder which are multi-layer neural networks, and is configured by initializing the parameters of the neural network by unsupervised learning and then re-learning by supervised learning. ..

"Reducing the dimensinality of data with neural networks" by Geoffrey Hinton, Ruslan Salakhutdinov, Science, 313, pp. 504-507 (2006)

By extracting the characteristics of the test product after the test with the self-encoder as described above, it will be possible to judge the quality of the test product based on the extracted characteristics. However, in order to improve the accuracy of the quality judgment of the test product after the test, it is necessary to improve the extraction accuracy of the characteristics of the test product and appropriately set the threshold value used for the quality judgment of the test product. ..

Therefore, the main purpose of the present disclosure is to provide a test product evaluation system and a method for constructing the same, which can accurately extract the feature quantity of the test product after the test and accurately judge the quality of the test product based on the feature quantity. To do.

The test product evaluation system of the present disclosure is a test product evaluation system that evaluates a test product after a test, and is a feature extraction that extracts a feature amount of the test product from image data acquired for the test product after the test. The feature extraction unit includes a unit and a determination unit for determining the quality of the test product based on the feature amount extracted by the feature extraction unit and a predetermined threshold value, and the feature extraction unit is a self-encoder together with a decoder. It corresponds to the encoder constituting the system, and is adapted by repeating the learning of the self-encoder that restores the input image data by the decoder from the feature amount extracted by the encoder. A plurality of feature quantities extracted from a plurality of normal test products by the matched feature extraction unit are clustered and determined.

In the test product evaluation system of the present disclosure, the feature extraction unit that extracts the feature quantity of the test product corresponds to an encoder that constitutes a self-encoder together with the decoder, and is input by the decoder from the feature quantity extracted by the encoder. It is adapted by repeating the training of the self-encoder so that the image data is restored. This makes it possible to more accurately extract the feature amount of the test product from the image data acquired for the test product after the test in which the quality distribution does not generally correspond to the specific distribution. Further, the threshold value used for determining the quality of the test product is determined by clustering a plurality of feature quantities extracted from a plurality of normal test products by the matched feature extraction unit. As a result, even if the number of samples of the test product in which an abnormality is found is significantly smaller than the number of samples of the normal test product and the quality of the test product must be determined artificially, the test is performed. The threshold value used for determining the quality of the product can be set more appropriately. As a result, according to the test product evaluation system of the present disclosure, it is possible to accurately extract the feature quantity of the test product after the test and accurately judge the quality of the test product based on the feature quantity.

The method for constructing the test product evaluation system of the present disclosure includes a feature extraction unit that extracts a feature amount of the test product from the image data acquired for the test product after the test, and the feature amount extracted by the feature extraction unit. A method for constructing a test product evaluation system including a determination unit for determining the quality of the test product based on a predetermined threshold, and a self-encoder including an encoder and a decoder corresponding to the feature extraction unit. The feature extraction unit is constructed by repeating the learning of the self-encoder that prepares and restores the input image data by the decoder from the feature amount extracted by the encoder and matching the encoder. A plurality of feature quantities extracted from a plurality of normal test products by the constructed feature extraction unit are clustered to determine the threshold value.

According to the test product evaluation system constructed by such a method, it is possible to accurately extract the feature quantity of the test product after the test and accurately judge the quality of the test product based on the feature quantity.

It is a block diagram which shows the test article evaluation system of this disclosure. It is a schematic diagram for demonstrating the image data set used for abnormality discrimination by the test product evaluation system of this disclosure. It is a block diagram which shows the self-encoder used for constructing the feature extraction part of the test article evaluation system of this disclosure. It is a flowchart for demonstrating the construction procedure of the test article evaluation system of this disclosure. It is a flowchart for demonstrating the construction procedure of the test article evaluation system of this disclosure. It is explanatory drawing for demonstrating the construction procedure of the test article evaluation system of this disclosure.

Next, a mode for carrying out the invention of the present disclosure will be described with reference to the drawings.

FIG. 1 is a block diagram showing the test product evaluation system 1 of the present disclosure. The test product evaluation system 1 shown in the figure is a time-varying test for verifying the time-dependent change of a test product in which external factors are changed, such as a thermal test, a vibration test, an impact test, a temperature / humidity test, an acoustic test, and an electrostatic test. The test product is evaluated based on the image data obtained from the test product, and is mounted on a computer (hardware) (not shown) including a CPU, ROM, RAM, input / output interface, storage device, etc. To. The test product evaluation system 1 has a feature extraction unit 2 that extracts a feature vector (feature amount) from image data acquired for the test product after the test, a feature vector extracted by the feature extraction unit 2, and a predetermined threshold value. It includes a determination unit 3 for determining the quality of the test product based on the above.

The feature extraction unit 2 is a convolutional neural network including a plurality of convolutional layers, a pooling layer (partial sampling layer), and a fully connected layer, respectively. In the present embodiment, the feature extraction unit 2 captures a plurality of continuous image data (grayscale image data, hereinafter, obtained by imaging a test product such as an industrial product at predetermined intervals along a predetermined direction. , Such a plurality of image data is appropriately referred to as an "image data set"), and after convolving with a filter having a number larger than the number of image data (number of channels), further convolution and pooling (dimension compression) are performed. This is done to output a feature vector of a predetermined latent order (for example, a secondary order in this embodiment). As a plurality of image data (image data sets) continuously acquired from the test product, for example, as shown in FIG. 2, the solder C for joining the electronic component B to the circuit board A is in the height direction by X-rays or the like. A plurality of continuous image data obtained by imaging at predetermined intervals in (upper and lower directions in the figure) can be mentioned. The latent order is not limited to the secondary order.

Further, in the present embodiment, the feature extraction unit 2 corresponds to the encoder 12 of the self-encoder 11 as shown in FIG. The self-encoder 11 includes an encoder 12 which is a convolutional neural network and a decoder 13. The decoder 13 of the self-encoder 11 includes a layer corresponding to the inverted fully connected layer of the feature extraction unit 2, a plurality of deconvolutional layers and an amplifiering layer, respectively, and is extracted from the image data by the encoder 12. It is a convolutional neural network that restores the input image data by expanding the dimensions of the feature vector. The self-encoder 11 is adapted by repeating learning so that the input image data is restored by the decoder 13 from the feature vector extracted by the encoder 12. Then, the test product evaluation system 1 of the present embodiment uses the encoder 12 adapted by the learning of the self-encoder 11 as the feature extraction unit 2.

Next, the procedure for constructing the test product evaluation system 1 will be specifically described with reference to FIGS. 4 to 6.

FIG. 4 is a flowchart showing a learning routine for fitting the self-encoder 11 including the encoder 12 which is the feature extraction unit 2 of the test product evaluation system 1. At the start of the learning routine of FIG. 3, the CPU of the computer on which the self-encoder 11 is mounted sets the learning epoch number epoch to zero (step S100), and whether or not the epoch number epoch is equal to or less than the specified number of times NB_EPOCH. (Step S110). When it is determined that the epoch number epoch is equal to or less than the specified number of times NB_EPOCH (step S110: YES), the CPU sets the variable itter to zero (step S120), and then the variable itter is a parameter of each network predetermined. It is determined whether or not the number of updates is less than or equal to Steps_EPOCH (step S130). When it is determined in step S130 that the variable iter is less than or equal to the update count Steps_EPOCH (step S130: YES), the CPU executes the above-mentioned learning of the self-encoder (step S140).

When learning the self-encoder in step S140, as shown in FIG. 5, a predetermined number of image data sets selected from a large number (plural) image data sets are given to the encoder 12 as teacher data (step). S141), the encoder 12 executes convolution of a plurality of image data and extracts (encodes) the feature vector of each image data set (step S143). In the present embodiment, a large number of image data sets serving as teacher data are a plurality of image data continuously acquired from a normal test product before the test and a test product (good product) having a good (normal) test result. A small number (for example, about 5 to 20% of the total) including a large number of non-defective image data sets including and a plurality of image data continuously acquired from a test product (defective product) in which an abnormality was found after the test. Includes non-defective image datasets.

The feature vector of each image data set extracted by the encoder 12 is given to the decoder 13, and the decoder 13 restores (decodes) each image data set by performing deconvolution of the feature vector and the like (step S145). ). Then, according to the loss function Loss _AE shown in the following equation (1), the image data set based on the feature vector extracted by the encoder 12 is more accurately represented, and the image data set restored by the decoder 13 is used as the base. The parameters of the encoder 12 and the decoder 13 (network) are updated to be closer to the image data set (step S147). However, in the equation (1), "MSE" indicates a square error, "En" indicates a coding operation, "De" indicates a decoding operation, and "x" is input in step S141. Shows the image dataset. Further, the square error MSE is expressed by the following equation (2), “y” in the equation (2) indicates a set (batch) of the restored image data images, and “t” is an image data set (batch). It indicates a collection (batch) of teacher data), and "BS" indicates the total number of batches.

When the learning of the self-encoder in step S140 is executed once, the CPU increments the variable iter as shown in FIG. 4 (step S150), and the variable iter is equal to or less than the update number Steps_EPOCH. Whether or not it is determined (step S130). Then, while it is determined in step S130 that the variable iter is less than or equal to the update count Steps_EPOCH (step S130: YES), learning of the self-encoder using the same teacher data (step S140) is repeatedly executed.

Further, when it is determined in step S130 that the variable iter exceeds the update count Steps_EPOCH (step S130: NO), the CPU increments the epoch number epoch (step S160), and then the epoch number epoch is the specified number of times NB_EPOCH. It is determined whether or not it is as follows (step S110). When it is determined that the epoch number epoch is equal to or less than the specified number of times NB_EPOCH (step S110: YES), the CPU sets the variable itter to zero again (step S120), and then the variable itter is a parameter of each network predetermined. It is determined whether or not the number of updates of Steps_EPOCH is less than or equal to (step S130).

When the variable iter is determined in step S130 to be less than or equal to the update count Steps_EPOCH (step S130: YES), the CPU determines that the variable iter exceeds the update count Steps_EPOCH in step S130. The self-encoder learning (step S140) using other teacher data selected from the data set is repeatedly executed. Then, when the processes of steps S120-S160 are repeatedly executed and it is determined in step S110 that the epoch number epoch exceeds the specified number of times NB_EPOCH (step S110: NO), the learning routine of FIG. 4 ends. Become.

Here, in addition to the learning of the self-encoder 11 as described above, the learning of the discriminator (Discriminator) and the learning of the encoder 12 using a discriminator (not shown), that is, so-called hostile learning (for example, "Adversarial"). By executing "autoencoders" by Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, Brendan Frey, ArXiv preprint, arXiv: 1511.05644 (2015)), the feature extraction unit 2 extracts, for example, a feature vector that follows a normal distribution. It will be possible to obtain the feature vector of the industrial product whose various variations generally follow the normal distribution with higher accuracy. However, the quality distribution of the test product after the test generally does not apply to a specific distribution. Based on this, the feature extraction unit 2 of the test product evaluation system 1 of the present embodiment is a self-encoder 11 that allows the decoder 13 to restore the input image data from the feature amount extracted by the encoder 12. It is adapted by repeating only the learning (step S140). As a result, a feature extraction unit 2 capable of more accurately extracting the feature vector of the test product from a plurality of image data continuously acquired for the test product after the test whose quality distribution does not generally correspond to a specific distribution is constructed. can do.

After the feature extraction unit 2 is obtained, the threshold value used by the determination unit 3 of the test product evaluation system 1 is set using the feature extraction unit 2. When setting the threshold, as shown in FIG. 6, a large number (plurality) of good product image data sets acquired for a normal test product before the test and a test product (good product) with good (normal) test results have been adapted. It is given to the feature extraction unit 2 of the above, and the feature extraction unit 2 extracts the feature vector of each non-defective image data set. Further, the identification boundary is determined by clustering with a support vector machine, that is, one class-SVM, which uses only the feature vectors of a plurality of non-defective image data sets extracted by the feature extraction unit 2 as training data. Then, the threshold value used for determining the quality of the test product is determined based on the identification boundary. As a result, even if the number of samples of the test product in which an abnormality is found is significantly smaller than the number of samples of the normal test product and the quality of the test product must be determined artificially, the test is performed. The threshold value used for determining the quality of the product can be set more appropriately. As a result, according to the test product evaluation system 1, the feature vector of the test product after the test is accurately extracted, and it is determined whether or not the feature vector is included in the threshold range based on the identification boundary. It is possible to accurately judge the quality of the test product after the test.

In the above embodiment, the test product evaluation system 1 has been described as determining the quality of the test product based on a plurality of image data continuously acquired for the test product after the test. It is not limited. That is, the test product evaluation system 1 may be configured to determine the quality of the test product based on a single grayscale image data or multi-channel color image data acquired for the test product after the test. .. Further, the threshold value used for the quality judgment of the test product by the judgment unit 3 is a semi-supervised mixed Gauss model or self-organization of a plurality of features extracted from a plurality of normal subjects by the matched feature extraction unit 2. It may be defined by clustering with a map. Further, the test product may be one used in a test other than the aging test. Further, the application target of the test product evaluation system 1 is not limited to industrial products. Further, the test product evaluation system 1 can be applied to the determination of the presence or absence of abnormality of the subject in the medical field, the determination of abnormal noise from the voice waveform, and the like.

As described above, the test product evaluation system of the present disclosure is a test product evaluation system (1) that evaluates the test product after the test, and the test is performed from the image data acquired for the test product after the test. A feature extraction unit (2) for extracting the feature amount of the product, and a determination unit (determining unit) for determining the quality of the test product based on the feature amount extracted by the feature extraction unit (2) and a predetermined threshold value. The feature extraction unit (2) including 3) corresponds to the encoder (12) constituting the self-encoder (11) together with the decoder (13), and the features extracted by the encoder (12). The feature extraction unit (2), which is adapted by repeating the learning of the self-encoder (11) so that the input image data is restored by the decoder (13) from the quantity, and the threshold is adapted. It is determined by clustering a plurality of feature quantities extracted from a plurality of normal test products.

In the test product evaluation system of the present disclosure, the feature extraction unit that extracts the feature quantity of the test product corresponds to an encoder that constitutes a self-encoder together with the decoder, and is input by the decoder from the feature quantity extracted by the encoder. It is adapted by repeating the training of the self-encoder so that the image data is restored. This makes it possible to more accurately extract the feature amount of the test product from the image data acquired for the test product after the test in which the quality distribution does not generally correspond to the specific distribution. Further, the threshold value used for determining the quality of the test product is determined by clustering a plurality of feature quantities extracted from a plurality of normal test products by the matched feature extraction unit. As a result, even if the number of samples of the test product in which an abnormality is found is significantly smaller than the number of samples of the normal test product and the quality of the test product must be determined artificially, the test is performed. The threshold value used for determining the quality of the product can be set more appropriately. As a result, according to the test product evaluation system of the present disclosure, it is possible to accurately extract the feature quantity of the test product after the test and accurately judge the quality of the test product based on the feature quantity.

Further, the threshold value may be determined by clustering the plurality of feature quantities extracted from the plurality of normal test products by the matched feature extraction unit (2) by a support vector machine. This makes it possible to set the threshold value more appropriately and further improve the accuracy of the quality determination of the test product.

Further, the test may be a time-varying test for verifying the time-dependent change of the test product when an external factor is changed.

The method for constructing the test product evaluation system of the present disclosure is extracted by a feature extraction unit (2) for extracting the feature amount of the test product from the image data acquired for the test product after the test and the feature extraction unit (2). A method for constructing a test product evaluation system (1) including a determination unit (3) for determining the quality of the test product based on the feature amount and a predetermined threshold value, wherein the feature extraction unit ( A self-encoder (11) including an encoder (12) and a decoder (13) corresponding to 2) is prepared, and the input image data is restored by the decoder (13) from the feature amount extracted by the encoder (12). The feature extraction unit (2) is constructed by repeating the learning of the self-encoder (11) to fit the encoder (12), and a plurality of the feature extraction units (2) constructed. The threshold value is determined by clustering a plurality of feature quantities extracted from the normal test product.

According to the test product evaluation system constructed by such a method, it is possible to accurately extract the feature quantity of the test product after the test and accurately judge the quality of the test product based on the feature quantity.

It goes without saying that the invention of the present disclosure is not limited to the above-described embodiment, and various changes can be made within the extension of the present disclosure. Furthermore, the above-described embodiment is merely a specific embodiment of the invention described in the column of the outline of the invention, and does not limit the elements of the invention described in the column of the outline of the invention.

The invention of the present disclosure can be used in the manufacturing industry of a test product evaluation system for evaluating a test product after a test.

1 Test product evaluation system, 2 Feature extraction unit, 3 Judgment unit, 11 Self-encoder, 12 Encoder, 13 Decoder, A circuit board, B electronic component, C solder.

Claims (4)

A test product evaluation system that evaluates test products after testing.
A feature extraction unit that extracts the feature amount of the test product from the image data acquired for the test product after the test, and a feature extraction unit.
A determination unit for determining the quality of the test product based on the feature amount extracted by the feature extraction unit and a predetermined threshold value is provided.
The feature extraction unit corresponds to an encoder that constitutes a self-encoder together with a decoder, and of the self-encoder that allows the decoder to restore input image data from the feature amount extracted by the encoder. Adapted by repeating learning,
The threshold value is a test product evaluation system defined by clustering a plurality of feature quantities extracted from a plurality of normal test products by the matched feature extraction unit. In the test product evaluation system according to claim 1,
The threshold value is a test product evaluation system in which the plurality of feature quantities extracted from the plurality of normal test products by the matched feature extraction unit are clustered by a support vector machine. In the test product evaluation system according to claim 1 or 2.
The test is a test product evaluation system that is a time-varying test for verifying the time-dependent change of the test product when an external factor is changed. The test product after the test The test product is based on a feature extraction unit that extracts the feature amount of the test product from the acquired image data, the feature amount extracted by the feature extraction unit, and a predetermined threshold value. It is a method of constructing a test product evaluation system including a judgment unit for judging the quality of the product.
A self-encoder including an encoder and a decoder corresponding to the feature extraction unit is prepared.
The feature extraction unit is constructed by repeating the learning of the self-encoder that restores the input image data by the decoder from the feature amount extracted by the encoder and matching the encoder.
A method for constructing a test product evaluation system in which a plurality of feature quantities extracted from a plurality of normal test products by the constructed feature extraction unit are clustered to determine the threshold value.

Images