JP2021096148A - Inspection system, inspection device, learning device and program - Google Patents

Abstract

To provide an inspection system which can quickly execute additional learning of a learning model for use in inspection.SOLUTION: An inspection device 100 comprises: an electromagnetic wave irradiation unit 111; a transmitted quantity detection unit 113 detecting the data indicating distribution of quantity of electromagnetic wave transmitted through an inspection object W; a learning model storage unit 115 storing a learning model that outputs information indicating the possibility of presence of prescribed abnormality in the inspection object; inspection result output means 116 inputting the transmitted quantity distribution data of the inspection object to the learning model and outputting an inspection result; and learned model receive means 118 receiving a learning model having undergone additional learning from a learning device 200 and storing the received learning model in the learning model storage unit. The learning device comprises a learning data storage unit 212 storing, as learning data, the transmitted quantity distribution data of the inspection object for both cases when prescribed abnormality exists and when it does not exist; and learning means 213 causing the learning model received from the inspection device to be additionally learnt using the learning data.SELECTED DRAWING: Figure 1

Description

The present invention relates to an inspection system, an inspection device, a learning device, and a program for inspecting an abnormality inside an inspection object based on distribution data of a transmission amount obtained by irradiating the inspection object with an electromagnetic wave.

When the inspection target is irradiated with electromagnetic waves such as X-rays, ultraviolet rays, visible rays, infrared rays, or microwaves and the transmitted amount is detected, the strength of the transmitted amount due to the presence or absence of foreign matter in the inspection target and the material of the foreign matter is determined. A distribution is obtained. By generating a two-dimensional image in which this intensity distribution is expressed by the color tone of each pixel (tone of color tone, shade, intensity, etc.), the situation inside the inspection object that cannot be seen from the outside is visualized. be able to.

As described above, the strength of the permeation amount depends on the presence or absence of foreign matter in the inspection object, the material of the foreign matter, etc., but depending on the combination of the foreign matter and the non-foreign matter, the difference in the permeation amount between the two is very small and is generated. It may be difficult to distinguish between the two from the two-dimensional image.

As a solution to such a problem, a method using a machine-learned learning model has recently been proposed. For example, a large number of distribution data of the permeation amount of the inspection object having a predetermined abnormality inside are collected, the feature amount is extracted using these as training data, and the extracted feature amount is used inside the inspection object. Generate a learning model to detect a given anomaly. Then, by inputting data showing the distribution of the permeation amount of the inspection object into this learning model, it is possible to accurately detect the presence or absence of an abnormality.

For example, by executing machine learning using an X-ray inspection image of a product in which a plurality of articles are overlapped as a learning image and X-ray inspection of the product using the feature amount obtained thereby, the products overlap each other. Patent Document 1 discloses an inspection device capable of suppressing a decrease in inspection accuracy of a product containing a plurality of articles. In the inspection device described in Patent Document 1, a CPU (Central Processing Unit) in the inspection device executes machine learning processing.

The learning model can be used as it is once generated, but the detection accuracy can be further improved by additional learning using new learning data.

Japanese Patent No. 6537008

When the machine learning process is executed by the CPU (Central Processing Unit) in the inspection device, the processing capacity of the CPU provided in the inspection device is generally insufficient, and learning takes a long time. Further, this may cause the CPU to continue to be in a high heat state for a long time, resulting in damage to the CPU, surrounding circuits, and the like. Further, since the CPU is responsible for not only the machine learning process but also the collection of learning data and the execution control of the inspection, other processes cannot be performed or the processing speed is significantly reduced during the execution of the machine learning process.

An object of the present invention is an inspection system, an inspection device, and a learning device that can quickly execute additional learning of a learning model and have a small load on the inspection device when inspecting an inspection object using the learning model. And to provide the program.

The inspection system of the present invention includes an inspection device and a learning device. The inspection device includes an electromagnetic wave irradiation unit that generates electromagnetic waves and irradiates the inspection target object, a transmission amount detection unit that detects the transmission amount distribution data of the electromagnetic waves transmitted through the inspection target object, and a transmission amount distribution data of the inspection target object. A learning model storage unit that stores in advance a learning model that outputs information indicating the possibility of existence of a predetermined abnormality in the inspection object by inputting, and a distribution of the transmission amount of the inspection object detected by the transmission amount detection unit. It includes a test result output means for inputting data into a learning model and outputting a test result based on the output information, and a learning model transmitting means for transmitting the learning model. The learning device stores the learning model receiving means for receiving the learning model transmitted from the inspection device and the distribution data of the permeation amount of the inspection object both when there is a predetermined abnormality and when there is no abnormality as the learning data. A learning data storage unit for learning, a learning means for additionally learning a learning model received by a learning model receiving means using learning data, a learning model storage unit for storing an additional learning learning model, and additional learning for learning. A trained model transmission means for reading a model from a learning model storage unit and transmitting the model is provided. The inspection device further includes a trained model receiving means that receives the additionally trained learning model from the learning device and stores it in the learning model storage unit.

The inspection device further includes a learning data transmission means for transmitting the learning data collected by the transmission amount detection unit, and the learning device receives the learning data transmitted from the inspection device and stores the learning data in the learning data storage unit. Further receiving means may be provided.

The learning data transmitting means may transmit the learning data with a label indicating the presence or absence of a predetermined abnormality, and the learning means may perform additional learning with reference to the label.

The inspection device of the present invention includes an electromagnetic wave irradiation unit that generates electromagnetic waves and irradiates the inspection object, a transmission amount detection unit that detects distribution data of the transmission amount of the electromagnetic waves that have passed through the inspection object, and a transmission amount of the inspection object. A learning model storage unit that stores in advance a learning model that outputs information indicating the possibility of the existence of a predetermined abnormality in the inspection object by inputting the distribution data of A test result output means that inputs quantity distribution data into a training model and outputs a test result based on the output information, a learning model transmission means that sends a learning model, and a learning model that has been additionally trained are received and trained. A trained model receiving means for storing in a model storage unit is provided.

The learning device of the present invention is a learning model receiving means for receiving a learning model, and learning data that stores distribution data of the permeation amount of an inspection object both when there is a predetermined abnormality and when there is no abnormality as learning data. The storage unit, the learning means for additionally learning the learning model received by the learning model receiving means using the learning data, the learning model storage unit for storing the additionally learned learning model, and the learning model for which the additional learning has been completed are learned. A trained model transmission means for reading and transmitting from the model storage unit is provided.

The functions of each means of the inspection device and the learning device of the present invention may be realized by describing them in a program and causing a computer to execute them.

According to the present invention, there is provided an inspection system, an inspection apparatus, a learning apparatus and a program capable of promptly executing additional learning of the learning model and having a small load on the inspection apparatus when performing an inspection using the learning model. can do.

It is a figure which shows the functional structure example of the inspection system 10. It is a figure which shows the configuration example of the inspection system 10 which is realized by describing the function of each means in a program and executing it by a computer.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description and drawings, the same functional parts are designated by the same reference numerals, and the functional parts once described will be omitted or described to the extent necessary.

FIG. 1 is a diagram showing a functional configuration example of the inspection system 10 of the present invention. The inspection system 10 includes an inspection device 100 and a learning device 200. The inspection device 100 and the learning device 200 are communicably connected by an arbitrary wired communication method or wireless communication method via an arbitrary network NW such as peer-to-peer, LAN, and WAN. Further, one learning device 200 may be shared by a plurality of inspection devices 100.

The inspection device 100 includes an electromagnetic wave irradiation unit 111, a transport unit 112, a transmission amount detection unit 113, a detection data storage unit 114, a learning model storage unit 115, an inspection result output means 116, a learning model transmission means 117, and a learned model receiving means 118. To be equipped.

The learning device 200 includes a learning model receiving means 211, a learning data storage unit 212, a learning means 213, and a learned model transmitting means 215.

The electromagnetic wave irradiation unit 111 irradiates an inspection object W mounted on the transport unit 112 and transported in the Y-axis direction in FIG. 1 with electromagnetic waves such as X-rays, ultraviolet rays, visible rays, and infrared rays. It is the source.

The transport unit 112 is an arbitrary transport mechanism that transports the inspection object W placed on the transport surface forming the XY plane in the three-dimensional Cartesian coordinate system at a predetermined speed in the Y-axis direction. The specific configuration method of the transport unit 112 is arbitrary. For example, as shown in FIG. 1, it may be configured as two belt conveyors arranged with a gap so that the electromagnetic wave transmitted through the inspection object W directly reaches the transmission amount detection unit 113, or the electromagnetic wave transmission property. It may be composed of one high belt conveyor.

The transmission amount detection unit 113 detects the distribution data of the transmission amount of the electromagnetic wave transmitted through the inspection object W. The method of configuring the transmission amount detection unit 113 is arbitrary, and may be configured as, for example, a line sensor composed of a plurality of detection elements arranged in the X-axis direction. In this case, the amount of electromagnetic waves transmitted from the inspection target object W is repeatedly detected at a cycle corresponding to the transport speed by the transport unit 112 until the entire inspection target object W passes over the line sensor. As a result, distribution data of the amount of transmission in each region obtained by linearly dividing the inspection object W in each cycle is obtained, and in each cycle from the start of the inspection target W passing on the line sensor to the end of the passage. By arranging the data in the order of acquisition, it is possible to obtain distribution data of the permeation amount for the entire inspection object W.

The detection data storage unit 114 stores the distribution data of the permeation amount of the inspection object W detected by the permeation amount detection unit 113.

The learning model storage unit 115 stores in advance a learning model that outputs information indicating the possibility of existence of a predetermined abnormality in the inspection object W by inputting distribution data of the permeation amount of the inspection object W. The learning model stored in advance may be one generated in advance in the inspection device 100 or may be generated by another device.

Predetermined abnormalities include, for example, when the inspection target W is a package, the presence of foreign matter inside, an abnormality in the shape of the contents, biting of the contents into the seal portion, and the like, as well as the inspection target W. Abnormal shape of itself, cracks, chips, etc. can be mentioned.

The learning model is known based on the difference in the color tone of each pixel and the difference in the mode of change in the color tone between the pixels, using the distribution data of the transmission amount of the inspection object W having a predetermined abnormality as the learning data. It is generated by performing machine learning by any method. For example, in supervised machine learning, transparent image data with a label indicating the presence or absence of a predetermined abnormality is input to the learning model, and the parameters are adjusted so that the learning model outputs the result according to the label. And so on. At this time, in order to generate a learning model with higher detection accuracy, a deep learning method using a multi-layer neural network may be applied to the learning algorithm.

As the information indicating the possibility of existence of a predetermined abnormality output from the learning model, for example, an inference value indicating the possibility of a predetermined abnormality can be mentioned.

The inspection result output means 116 inputs the distribution data of the permeation amount of the inspection object W read from the detection data storage unit 114 into the learning model read from the learning model storage unit 115, and determines a predetermined value in the output inspection object W. The inspection result is output based on the information indicating the possibility of the existence of the abnormality.

When the information indicating the possibility of existence of a predetermined abnormality output from the learning model is an inferred value indicating the possibility of a predetermined abnormality, the existence or nonexistence of the predetermined abnormality is obtained by comparing this inferred value with a predetermined threshold value. Can be specified and output as an inspection result.

Specifically, for example, the inference value is defined in the range of 0 to 1 in which the case where the possibility of abnormality is extremely low is 0 and the case where the possibility of abnormality is extremely high is 1, and the output inference value is constant. When it is equal to or more than the value of (for example, 0.6), it can be determined that there is an abnormality.

The learning model transmitting means 117 reads a learning model from the learning model storage unit 115 and transmits it to the learning device 200 (learning model receiving means 211).

The learning model receiving means 211 receives the learning model from the inspection device 100 (learning model transmitting means 117).

The learning data storage unit 212 stores the distribution data of the permeation amount of the inspection target W in both the case where there is a predetermined abnormality and the case where there is no abnormality as the learning data. The more learning data to be stored, the more desirable it is. When supervised learning is performed by the learning means 213, if the stored learning data is not labeled with a predetermined abnormality, it is attached based on input by the user or the like.

As the distribution data of the permeation amount of the inspection target W in both the case where there is a predetermined abnormality and the case where there is no abnormality, for example, the distribution data collected in advance outside the inspection system 10 may be stored.

Further, the inspection device 100 collects the distribution data of the permeation amount of the inspection target W in each of the cases where there is a predetermined abnormality and the case where there is no abnormality, which is the learning data, and stores this in the learning data storage unit 212. You may. In this case, for example, the learning data transmitting means 119 is provided in the inspection device 100, and the learning data receiving means 216 is provided in the learning device 200.

The learning data transmitting means 119 is the learning data detected by the transmission amount detection unit 113 and stored in the detection data storage unit 114. The distribution data of the transmission amount is read from the detection data storage unit 114 and transmitted to the learning data receiving means 216. When transmitting the learning data, it may be transmitted with a label indicating the presence or absence of a predetermined abnormality.

The learning data receiving means 216 receives the learning data transmitted from the learning data transmitting means 119 and stores it in the learning data storage unit 212.

As a result, in the inspection device 100, the inspection object W having a predetermined abnormality or having no abnormality is irradiated with an electromagnetic wave from the electromagnetic wave irradiation unit 111, and the transmission amount detection unit 113 detects the transmission amount distribution data, and the detection data. It can be stored in the storage unit 114, read by the learning data transmitting means 119 and transmitted, and further received by the learning data receiving means 216 and stored in the learning data storage unit 212.

A plurality of inspection devices 100 are connected to one learning device 200, and distribution data of the permeation amount is collected from one of the inspection devices 100 for the inspection target W when there is a predetermined abnormality and when there is no abnormality. The inspection system 10 may be configured to collect and store the data in the learning data storage unit 212.

The learning means 213 reads the learning model received by the learning model receiving means 211 from the learning data storage unit 212, and distributes the amount of permeation for the inspection object W both when there is a predetermined abnormality and when there is no abnormality. Additional learning is performed using the data, and the additional learning training model is output. The additional learning may be performed by any known machine learning method based on the difference in the color tone of each pixel in the transmission amount distribution data, the difference in the mode of the change in the color tone between the pixels, and the like. For example, in supervised machine learning, transparent image data with a label indicating the presence or absence of a predetermined abnormality is input to the learning model, and the parameters are adjusted so that the learning model outputs the result according to the label. And so on. At this time, in order to generate a learning model with higher detection accuracy, a deep learning method using a multi-layer neural network may be applied to the learning algorithm.

The learned model storage unit 214 stores the additionally learned learning model output by the learning means 213. As the learning model that has been additionally learned, in addition to the learning model that has been additionally learned most recently, the learning model that has been additionally learned in the past may be further stored.

The learned model transmitting means 215 reads the additionally learned learning model from the learned model storage unit 214 and transmits it to the inspection device 100 (learned model receiving means 118).

The learned model receiving means 118 receives the additionally learned learning model from the learning device 200 (learned model transmitting means 215) and stores it in the learning model storage unit 115.

After that, the inspection result output means 116 executes the inspection using the learning model that has been additionally trained.

The functions of the means of the inspection device 100 and the learning device 200 described above may be realized by describing them in a program and causing a computer to execute them.

FIG. 2 shows a configuration example of an inspection system 10 (inspection apparatus 100 and learning apparatus 200) realized by describing the functions of each means in a program and executing them on a computer.

The inspection device 100 includes, in addition to the electromagnetic wave irradiation unit 111, the transport unit 112, and the transmission amount detection unit 113, for example, an input unit 121, a storage unit 122, a CPU 123, an output unit 125, and a communication unit 126.

The input unit 121 inputs an execution instruction for inspection / learning, selects and inputs detection data to be training data, adds information indicating the presence or absence of a predetermined abnormality to the training data, and selects and inputs a learning model to be used for inspection / learning. , To input instructions for sending and receiving data between the inspection device 100 and the learning device 200, and to input information necessary for the inspection device 100 to function, such as a learning model generated outside the inspection system 10. Any interface of. The interface for inputting instructions and selecting input may be configured as a physical switch, or a display may be provided as an output unit 125 to display an input screen, and input may be performed from a pointing device, keyboard, or the like. It may be configured.

The storage unit 122 takes on the functions of the detection data storage unit 114 and the learning model storage unit 115, and stores an arbitrary program in which the functions of each means, the control information of each unit, the remote control information of the learning device 200, and the like are described. It is a means of memory. For the storage unit 122, for example, in addition to storage means such as magnetic disks and optical disks, storage means such as semiconductor memories such as non-volatile memory and volatile memory can be adopted.

The CPU 123 executes a program in which the functions of each means stored in the storage unit 122, the control information of each unit, the remote control information of the learning device 200, and the like are described, and the functions of each means, the control of each unit, and the learning device 200 Realize remote control, etc.

The output unit 125 is an output means for displaying an operation screen used for executing inspection and learning, displaying the inspection result output by the inspection result output means 116, printing, and the like. The specific configuration depends on the output form. For example, when displaying an operation screen or an inspection result, it is configured as various displays, and when printing, it is configured as various printing machines.

The communication unit 126 is a learning device 200 (communication unit 226) when the functions of the learning model transmitting means 117, the learned model receiving means 118, and the learning data transmitting means 119 are realized, or when the learning device 200 is remotely controlled by the inspection device 100. It is a communication interface according to the form and communication method of the network NW to be adopted, which sends and receives information to and from.

The learning device 200 includes, for example, an input unit 221, a storage unit 222, a CPU 223, and a communication unit 226.

The input unit 221 is an arbitrary interface into which information necessary for operating the learning device 200, such as distribution data of the permeation amount of the inspection object W collected outside the inspection system 10, is input.

The storage unit 222 is an arbitrary storage means that bears the function of the learning data storage unit 212 and stores a program in which the functions and the like of each means are described. For the storage unit 222, for example, in addition to storage means such as magnetic disks and optical disks, storage means such as semiconductor memories such as non-volatile memory and volatile memory can be adopted.

The CPU 223 executes a program in which the functions and the like of each means stored in the storage unit 222 are described, and realizes the functions and the like of each means.

The communication unit 226 communicates with the inspection device 100 (communication unit 126) when the functions of the learning model receiving means 211, the learned model transmitting means 215, and the learning data receiving means 216 are realized or when remote control is performed from the inspection device 100. It is a communication interface that sends and receives information according to the form and communication method of the network NW to be adopted.

In order to realize the function of the learning means 213 in the learning device 200, it generally takes time to execute the machine learning process only by the CPU. Therefore, in addition to the CPU 223, a GPU (Graphics Processing Unit) 224, which has an architecture specialized for parallel computing and can execute machine learning processing at high speed and with a low load, is provided, and the functions of the learning means 213 are performed by the CPU 223 and the GPU 224. It may be realized by cooperation.

According to the inspection system 10 of the present invention, when the inspection is performed using the learning model, the additional learning of the learning model is executed by the learning device 200 different from the inspection device 100, so that the speed of the machine learning process can be increased. At the same time, it is possible to prevent the inspection device 100 from interfering with the processing and causing damage to the CPU and surrounding circuits. Further, by providing the GPU in the learning device 200, the machine learning process can be further speeded up.

The present invention is not limited to the above embodiments. The above-described embodiment is an example, and any object having substantially the same configuration as the technical idea described in the claims of the present invention and exhibiting the same effect and effect is the present invention. It is included in the technical scope of the invention. That is, it can be appropriately changed within the scope of the technical idea expressed in the present invention, and the form in which such a change or improvement is added is also included in the technical scope of the present invention.

10 ... Inspection system 100 ... Inspection device 111 ... Electromagnetic wave irradiation unit 112 ... Transport unit 113 ... Transmission amount detection unit 114 ... Detection data storage unit 115 ... Learning model storage unit 116 ... Inspection result output means 117 ... Learning model transmission means 118 ... Learning Finished model receiving means 119 ... Learning data transmitting means 121, 221 ... Input unit 122, 222 ... Storage unit 123, 223 ... CPU
125 ... Output unit 126 ... Communication unit 211 ... Learning model receiving means 212 ... Learning data storage unit 213 ... Learning means 214 ... Learned model storage unit 215 ... Learned model transmitting means 216 ... Learning data receiving means 224 ... GPU
NW ... Network W ... Inspection target

Claims (6)

An electromagnetic wave irradiation unit that generates electromagnetic waves and irradiates the inspection object,
A transmission amount detection unit that detects distribution data of the transmission amount of the electromagnetic wave transmitted through the inspection object, and a transmission amount detection unit.
A learning model storage unit that stores in advance a learning model that outputs information indicating the possibility of existence of a predetermined abnormality in the inspection object by inputting distribution data of the permeation amount of the inspection object.
An inspection result output means that inputs the distribution data of the permeation amount of the inspection object detected by the permeation amount detection unit into the learning model and outputs the inspection result based on the output information.
A learning model transmitting means for transmitting the learning model and
With an inspection device equipped with
A learning model receiving means for receiving the learning model transmitted from the inspection device, and
As learning data, a learning data storage unit that stores distribution data of the permeation amount of the inspection object both when there is a predetermined abnormality and when there is no abnormality, and
A learning means for additionally learning the learning model received by the learning model receiving means using the learning data, and
A learning model storage unit that stores the additionally learned learning model,
A trained model transmission means for reading and transmitting the additionally trained learning model from the learning model storage unit,
With a learning device equipped with
With
The inspection system further includes a learned model receiving means that receives the additionally learned learning model from the learning device and stores it in the learning model storage unit. The inspection device further includes a learning data transmitting means for transmitting the learning data collected by the permeation amount detecting unit.
The inspection system according to claim 1, wherein the learning device further includes a learning data receiving means that receives the learning data transmitted from the inspection device and stores the learning data in the learning data storage unit. The learning data transmitting means transmits the learning data with a label indicating the presence or absence of the predetermined abnormality.
The inspection system according to claim 2, wherein the learning means executes additional learning with reference to the label. An electromagnetic wave irradiation unit that generates electromagnetic waves and irradiates the inspection object,
A transmission amount detection unit that detects distribution data of the transmission amount of the electromagnetic wave transmitted through the inspection object, and a transmission amount detection unit.
A learning model storage unit that stores in advance a learning model that outputs information indicating the possibility of existence of a predetermined abnormality in the inspection object by inputting distribution data of the permeation amount of the inspection object.
An inspection result output means that inputs the distribution data of the permeation amount of the inspection object detected by the permeation amount detection unit into the learning model and outputs the inspection result based on the output information.
A learning model transmitting means for transmitting the learning model and
A trained model receiving means that receives an additionally trained learning model and stores it in the learning model storage unit,
Inspection device equipped with. A learning model receiving means for receiving a learning model,
As training data, a learning data storage unit that stores distribution data of the permeation amount of the inspection object both when there is a predetermined abnormality and when there is no abnormality, and
A learning means for additionally learning the learning model received by the learning model receiving means using the learning data, and
A learning model storage unit that stores the additionally learned learning model,
A trained model transmission means for reading and transmitting the additionally trained learning model from the learning model storage unit,
A learning device equipped with. A program for operating a computer as each means according to any one of claims 1 to 5.

Images