JP6789460B1 - Fracture surface analysis device, trained model generator, fracture surface analysis method, fracture surface analysis program, and trained model - Google Patents

Abstract

From the fracture surface image of the resin molded body, the cause of destruction of the resin molded body can be easily identified. Machine learning in which the fracture surface analysis device 1 uses the acquisition unit 11 for acquiring the target fracture surface image of the fracture surface of the resin molded body to be analyzed as teacher data, and the fracture surface image and the label that is the cause of the fracture. It has an analysis unit 13 for inputting the target fracture surface image into the trained model generated by the above and identifying the cause of destruction of the resin molded body.

Description

The present invention relates to a technique for analyzing a fracture surface of a resin molded product.

Fracture surface analysis is performed by observing the fracture surface of the fractured part and searching for the cause of the fracture from the state. Fracture surface analysis is a method widely used to estimate the fracture cause or fracture mechanism from the characteristics of fracture surface patterns, mainly in the field of metals. Similar to the metal material, the resin molded body also shows various fracture surface patterns depending on the cause of fracture, so that the same fracture surface analysis method is adopted.

In the case of a resin molded product, for example, the presence or absence of bleed-out products, foreign substances, or deterioration products using a scanning electron microscope (SEM) to observe the surface microstructure of the fracture surface at high magnification and chemical composition analysis. The cause of destruction is estimated by confirming (Patent Document 1).

Japanese Patent Application No. 2016-33494

Although the condition of the fracture surface differs depending on the cause of fracture, it is a difficult technique to identify the cause of fracture by observing the surface microstructure of the fracture surface unless it is a skilled engineer. For example, finding small features that are key points in a fracture surface and distinguishing between similar fracture surfaces requires experience and knowledge (a kind of sense) of observing many fracture surfaces. ..

The present invention has been made in view of the above problems, and an object of the present invention is to easily identify the cause of destruction from a fracture surface image.

In order to achieve the above object, the fracture surface analysis apparatus of the present invention has an acquisition unit that acquires an image of the fracture surface of the resin molded product to be analyzed, a fracture surface image, and a label that is the cause of destruction. It has an analysis unit for inputting the target fracture surface image into the trained model generated by machine learning as training data and identifying the cause of destruction of the resin molded product.

The learned model generator of the present invention includes an acquisition unit that acquires teacher data in which a fracture surface image of a fracture surface of a resin molded product is imaged and a label that is a cause of destruction of the resin molded product, and a teacher. It has a learning unit that generates a learned model for identifying the cause of destruction of the resin molded product by machine learning using data.

In the fracture surface analysis method performed by the computer of the present invention, the acquisition step of acquiring the target fracture surface image obtained by imaging the fracture surface of the resin molded product to be analyzed, and the fracture surface image and the label causing the fracture are used as teacher data. An analysis step of inputting the target fracture surface image into the trained model generated by machine learning to identify the cause of destruction of the resin molded product is performed.

The fracture surface analysis program of the present invention causes a computer to function as the fracture surface analysis apparatus.

The trained model of the present invention is a trained model for making a computer function so as to identify the cause of destruction of the resin molded body based on a fracture surface image obtained by imaging the fracture surface of the resin molded body to be analyzed. Then, when the pixel value of the fracture surface image is input to the input layer by machine learning using the teacher data in which the fracture surface image is associated with the label that is the cause of destruction, the output layer determines the probability of the cause of destruction. The weighted value is learned so as to be output, and the pixel value of the fracture surface image input to the input layer is calculated based on the learned weighted value, and the cause of destruction is caused by the output layer. Make the computer function to output the probability.

According to the present invention, the cause of destruction can be easily identified from the fracture surface image.

It is a figure which shows the whole structure of the fracture surface analysis system which concerns on embodiment of this invention. This is an example of a fracture surface image of fatigue fracture. This is an example of a fracture surface image of ductile fracture. This is an example of a fracture surface image of brittle fracture. It is a flowchart of a trained model generation process. It is an example of the trained model of the first embodiment. It is a flowchart of a fracture surface analysis process. This is an example of the fracture surface analysis screen of the first embodiment. It is a figure which shows the example of the teacher data. It is a figure which shows the resin database example. It is an example of the trained model of the second embodiment. This is an example of the fracture surface analysis screen of the second embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

<First Embodiment>
FIG. 1 is an overall configuration diagram of a fracture surface analysis system according to the first embodiment of the present invention. The fracture surface analysis system of the present embodiment identifies the cause of fracture from the fracture surface image of the resin molded product (resin molded product). The fracture surface analysis system shown in the figure includes a fracture surface analysis device 1, a model generation device 2, and a user terminal 3. These devices are connected to a network and can communicate with each other via the network. The network is, for example, a LAN, the Internet, or the like.

The user terminal 3 is a terminal device used by a user who requests fracture surface analysis. The user terminal 3 receives the user's instruction and transmits an analysis request including a fracture surface image of the resin molded product acquired by the scanning electron microscope 4 (hereinafter referred to as “SEM”) to the fracture surface analysis device 1. To do. Further, the user terminal 3 receives the cause of destruction from the fracture surface analysis device 1 as an analysis result. The user stores the fracture surface image of the resin molded product imaged and output by the SEM4 in a recording medium such as a USB memory, inputs the fracture surface image to the user terminal 3 using the recording medium, and inputs the fracture surface analysis device 1 to the user terminal 3. Send to. When the SEM4 has a communication function, the SEM4 may be used as the user terminal 3. Here, the SEM4 is not limited to a scanning electron microscope as long as the fracture surface of the resin molded body can be observed with a sufficient magnification and resolution for analysis, and is not limited to a scanning electron microscope, an optical microscope such as a stereomicroscope or a metallurgical microscope, or a high-magnification digital microscope. You can also use a camera.

The user terminal 3 may be a terminal installed on the same site as the fracture surface analysis device 1 (for example, a factory, a research institute, etc.) and used by an engineer on the site. Alternatively, the user terminal 3 is located at a location away from the fracture surface analysis device 1, and a processor that manufactures the resin molded product, a customer who purchased the resin molded product, a sales staff member who sells the resin material or the resin molded product, or the like It may be the terminal to be used.

The fracture surface analysis device 1 analyzes the fracture surface image of the resin molded product in response to the request from the user terminal 3 and identifies the cause of destruction of the resin molded product. The fracture surface analysis device 1 shown in the figure includes an acquisition unit 11, a change unit 12, an analysis unit 13, an output unit 14, and a model storage unit 15.

The acquisition unit 11 acquires a fracture surface image (target fracture surface image) of the fracture surface of the resin molded product to be analyzed, which is transmitted from the user terminal 3. The changing unit 12 resizes (resizes) the acquired fracture surface image to an image having a predetermined number of pixels. The analysis unit 13 inputs the fracture surface image into the trained model generated by machine learning using the fracture surface image and the label that is the cause of the fracture as teacher data, and identifies the cause of the fracture of the resin molded body. The output unit 14 outputs the cause of destruction specified by the analysis unit 13. Specifically, the output unit 14 transmits the identified cause of destruction to the user terminal 3.

The model storage unit 15 stores the trained model generated by the model generation device 2. The trained model is a model trained to output the cause of destruction when a fracture surface image is input. In this embodiment, the trained model outputs each probability of a plurality of fracture causes such as fatigue fracture, brittle fracture, and ductile fracture.

The model generation device 2 includes a data acquisition unit 21, a learning unit 22, a data storage unit 23, and a model storage unit 24. The data acquisition unit 21 acquires from the data storage unit 23 the teacher data in which the fracture surface image of the fracture surface of the resin molded product is associated with the label that is the cause of the destruction of the resin molded product. The learning unit 22 generates a learned model for identifying the cause of destruction of the resin molded product by machine learning using a plurality of teacher data.

Next, the cause of destruction of the resin molded body will be described. In this embodiment, three types of fracture causes are assumed: "fatigue fracture", "ductile fracture", and "brittle fracture". However, the present invention is not limited to this, and other causes of destruction may be specified. As shown in FIGS. 2A to 2C, the condition of the fracture surface differs depending on the cause of fracture.

FIG. 2A is an example of a fracture surface image due to fatigue fracture. Fatigue fracture is fracture caused by repeated loading. A striped pattern (stration pattern) appears in the fracture surface image due to fatigue fracture. The image on the right of FIG. 2A is an enlarged view of the area surrounded by the dotted line of the fracture surface image on the left.

FIG. 2B is an example of a fracture surface image due to ductile fracture. Ductile fracture is fracture caused by a static load. When a load is applied at a low speed, the resin is plastically deformed with necking and whitening, resulting in fracture. Therefore, a portion where the resin is stretched appears in the fracture surface image. The image on the right of FIG. 2B is an enlarged view of the area surrounded by the dotted line of the fracture surface image on the left.

FIG. 2C is an example of a fracture surface image due to brittle fracture. Brittle fracture is fracture caused by an impact force. In the fracture surface image, patterns appear radially from the starting point of fracture, and the cross section shows a scaly pattern. The image on the right of FIG. 2C is an enlarged view of the area surrounded by the dotted line of the fracture surface image on the left.

Next, the trained model generation process performed by the model generation device 2 will be described.

FIG. 3 is a flowchart of a trained model generation process (machine learning process). First, the data acquisition unit 21 acquires, as teacher data, teacher data in which the fracture surface image and the fracture cause (label) of the fracture image are associated with each other from the data storage unit 23 (S11). It is assumed that a plurality of teacher data are stored in the data storage unit 23 in advance. The learning unit 22 performs machine learning using the acquired teacher data, and generates and updates a learned model for identifying the cause of destruction of the resin molded product (S12).

FIG. 4 shows an example of the trained model of the present embodiment. The trained model includes a neural network and parameters (weighted values, biases). Machine learning is the process of optimizing the parameters of a neural network. Before learning, initial values are set for the parameters. In this embodiment, machine learning is performed by a neural network composed of a combination of neurons (perceptrons, nodes). Specifically, the pair of the fracture surface image data included in the teacher data and the label is given to the neural network, and the weight value and the bias of each neuron are set so that the output of the neural network becomes the same as the label. Repeat learning while changing. In this way, the fracture surface image of the teacher data is trained, and a trained model for estimating the cause of destruction from the fracture surface image is generated.

The illustrated neural network includes an input layer, an intermediate layer, and an output layer, and each layer is composed of neurons. In the illustrated example, the intermediate layer is one layer, but it may be composed of a plurality of layers in order to perform advanced image recognition.

The pixel value of each pixel of the fracture surface image is input to each neuron in the input layer. In this embodiment, a fracture surface image of 150 × 150 pixels (22500 pixels) is used. In this case, the number of neurons in the input layer is 22500. Further, in the present embodiment, it is assumed that the fracture cause is any one of "fatigue fracture", "ductile fracture", and "brittle fracture". Therefore, the illustrated output layer is composed of three neurons. The number of neurons in the output layer is set according to the number of possible causes of destruction.

In a neural network, the output of each neuron is input to all neurons in the next layer. The output value y of each neuron is expressed by the following equation.

y = (x1 x w1) + (x2 x w2) + ... + (xn x wn) + b
“Xn” is an input value from the nth neuron in the previous layer, “wn” is a weighted value for xn, and “b” is a bias.

In S12, the learning unit 22 updates the weighted value wn and the bias b of each layer by machine learning by an error reverse propagation method or the like. Specifically, the learning unit 22 inputs each pixel value of the fracture surface image into the input layer of the neural network and performs an calculation, and obtains the output value of each neuron in the output layer and the label corresponding to the fracture surface image. Is output, and the weighting value wn and the bias b of each layer are updated so that the error becomes small (approaches 0).

Then, the learning unit 22 determines whether or not to end the machine learning (S13). As a condition for ending machine learning, for example, when machine learning is repeated with a preset number of teacher data, the learning unit 22 may end machine learning. For example, a predetermined number of teacher data (fracture surface image and label) are prepared for each cause of destruction, and machine learning is repeated using the teacher data. Further, as an end condition, the learning unit 22 may end the machine learning when the error between the label and the output of the output layer becomes equal to or less than a predetermined value.

When it is determined that the machine learning is not completed (S13: NO), the model generation device 2 returns to S11 and repeats the subsequent processing. When it is determined that the machine learning is finished (S13: YES), the learning unit 22 outputs the generated learned model to the model storage unit 24 and stores it. Further, the learning unit 22 transmits the generated learned model to the fracture surface analysis device 1 via the network (S14). As a result, the trained model generated by the model generation device 2 is stored in the model storage unit 15 of the fracture surface analysis device 1.

As described above, the trained model of the present embodiment is for causing the computer to function so as to identify the cause of the destruction of the resin molded product based on the fracture surface image obtained by capturing the fracture surface of the resin molded product to be analyzed. When the pixel value of the fracture surface image is input to the input layer by machine learning using the teacher data that associates the fracture surface image with the label that is the cause of destruction, the output layer is destroyed. The weighted value is learned so as to output the probability of the cause. Further, the trained model performs a calculation based on the trained weighted value on the pixel value of the fracture surface image input to the input layer, and performs fracture surface analysis so as to output the probability of the cause of destruction from the output layer. Make device 1 work.

Next, the fracture surface analysis process performed by the fracture surface analysis device 1 will be described.

FIG. 5 is a flowchart of the fracture surface analysis process. The acquisition unit 11 acquires a fracture surface image of the fracture surface of the resin molded product transmitted from the user terminal 3 (S21).

Then, the changing unit 12 resizes the acquired fracture surface image (S22). Specifically, the changing unit 12 changes the size (size) of the acquired fracture surface image to be the same size (predetermined size) as the fracture surface image used in the teacher data of the trained model. The trained model neural network (input layer) stored in the model storage unit 15 includes the same number of neurons as the number of pixels of the fracture surface image of the teacher data. Therefore, if a fracture surface image having a different number of pixels is input, it cannot be handled. Therefore, as a preprocessing, the changing unit 12 uses the acquired fracture surface image as the same number of pixels as the fracture surface image of the teacher data (for example, 150 × 150 pixels). ) To resize.

For example, when the acquired fracture surface image is larger than the fracture surface image of the teacher data, the changing unit 12 uses the average of the pixel values of adjacent pixels, and when the acquired fracture surface image is smaller than the fracture surface image of the teacher data. Interpolates a new pixel between adjacent pixels and resizes. Also, even if the aspect ratio of the acquired fracture surface image is different from the aspect ratio of the fracture surface image of the teacher data, the acquired fracture surface image is resized so that the aspect ratio is the same as that of the fracture surface image of the teacher data. ..

As a result, even when the user acquires and transmits fracture surface images of various image sizes according to the specifications or settings of SEM4, the fracture surface analysis device 1 performs fracture surface analysis based on the fracture surface image. It can be performed. If the acquired fracture surface image and the fracture surface image of the teacher data have the same size, the resizing process is not performed.

Then, the analysis unit 13 identifies (determines) the cause of destruction by using the fracture surface image after resizing (S23). Specifically, the analysis unit 13 reads out the trained model (neural network and parameters) stored in the model storage unit 15, and inputs the n pixel values included in the resized fracture surface image to the neurons in the input layer, respectively. Input and calculate the trained model. That is, each pixel input to the input layer is output to the intermediate layer and the output layer while being calculated using the parameters of the trained model, and finally the identification result of the cause of destruction is output from each neuron in the output layer. To. In the example shown in FIG. 4, the probability of fatigue fracture is m1%, the probability of brittle fracture is m2%, and the probability of ductile fracture is m3%. The total of m1%, m2%, and m3% is 100%. The analysis unit 13 identifies the cause of destruction based on the probability of each cause of destruction output by the trained model. For example, the analysis unit 13 identifies the cause of destruction having the highest probability. Then, the output unit 14 transmits the cause of destruction identified by the analysis unit 13 to the user terminal 3 (S24).

FIG. 6 shows an example of a screen for fracture surface analysis displayed on the display of the user terminal 3. In the illustrated example, one screen is used as an input screen and an output screen. The user selects a desired fracture surface image (image file) from the data stored in the storage unit of the user terminal 3 using the reference button 81 on the screen, and selects by clicking the upload button 82. The fracture surface image is determined as the analysis target, and the analysis button 83 is clicked. As a result, the user terminal 3 transmits a fracture surface analysis request including the fracture surface image selected by the user to the fracture surface analysis device 1.

Then, when the fracture surface analysis process shown in FIG. 5 is completed, the specified fracture cause (“brittle fracture” in the illustrated example) is displayed in 84 in the display area of the analysis result on the screen. In the illustrated screen example, one cause of destruction with the maximum probability is displayed as the analysis result, but a plurality of causes of destruction output by the trained model and their probabilities may be displayed as the analysis result. Good. In addition, the analysis results can be sent to a separately specified address by e-mail etc. to check from other terminals, or uploaded to a specified URL or SNS site etc. so that they can be viewed with a web browser or application. You may.

The fracture surface analysis device 1 of the present embodiment described above teaches the acquisition unit 11 that acquires the target fracture surface image of the fracture surface of the resin molded product to be analyzed, and the fracture surface image and the label that is the cause of the fracture. It has an analysis unit 13 for inputting the target fracture surface image into the trained model generated by machine learning as data and identifying the cause of destruction of the resin molded product. As a result, in the present embodiment, even if the engineer is not a skilled engineer with experience and knowledge, the cause of the fracture can be easily identified in a short time with high accuracy simply by acquiring the fracture surface image of the fractured resin molded product. Can be (estimated). Thereby, the convenience of the user can be improved.

<Second embodiment>
The fracture surface analysis system of this embodiment includes a fracture surface analysis device 1, a model generation device 2, and a user terminal 3 as in the first embodiment shown in FIG. In the first embodiment, the cause of fracture is identified from the fracture surface image, but in the second embodiment, in addition to the fracture surface image, peripheral information that causes the resin molded body to be destroyed is included in the teacher data. The teacher data used by the model generator 2 of the present embodiment will be described below.

FIG. 7 is an example of teacher data of this embodiment. The illustrated teacher data includes a label that is the cause of destruction, image data of a fracture surface image, and peripheral information. The peripheral information is information about the resin molded product, and is information that causes the resin molded product to be destroyed. Specifically, the peripheral information includes, for example, composition information (material), molding conditions, usage environment, shape, and the like. For composition information, resin grade is used here.

FIG. 8 is an example of a resin grade DB (database). In the resin grade DB, the type and ratio of the base resin and the type and ratio of at least one additive are defined for each resin grade, and represent the composition information of the resin molded product. The data storage unit 23 of the model generation device 2 of the present embodiment stores the teacher data of FIG. 7 and the resin grade DB of FIG.

Next, with reference to FIG. 3, the trained model generation process performed by the model generation device 2 of the present embodiment will be described. First, the data acquisition unit 21 acquires teacher data from the data storage unit 23 as teacher data (S11). In the present embodiment, the data acquisition unit 21 converts the resin grade of the teacher data into the corresponding composition information by using the resin grade DB. Then, the learning unit 22 performs machine learning using the teacher data obtained by converting the resin grade into composition information, and generates and updates a learned model for identifying the cause of destruction of the resin molded product (S12).

FIG. 9 shows an example of the trained model of the present embodiment. Here, the differences from the trained model (FIG. 4) of the first embodiment will be mainly described, and duplicate description will be omitted. In the illustrated example, the intermediate layer of the neural network is one layer, but it may be composed of a plurality of layers in order to perform advanced image recognition.

In the neural network of the present embodiment, an input layer and an intermediate layer are formed for each data group of the teacher data, the output of the intermediate layer for each data group is input to the output layer, and the output layer outputs the probability of the cause of destruction. The parameters (weighted value, bias) have been learned. A data group is a data set for each category. The data group includes a fracture surface image data group and a plurality of types of data groups included in the peripheral information (for example, composition information data group, molding condition data group, usage environment data group, shape data group, etc.). There is.

In the illustrated example, the data group A is a data group of a fracture surface image, each pixel value of the fracture surface image is input to each neuron of the input layer, and a calculation result (output) using the parameters of each neuron is obtained. It is input to each neuron in the middle layer of the data group A. The input layer and the intermediate layer of the data group A are the same as the trained model of the first embodiment. The data group B is a data group of composition information, and each data (ratio of each composition) of the composition information is input to each neuron of the input layer, and the calculation result using the parameters of each neuron is the data group B. It is input to each neuron in the middle layer of. Then, the outputs of all the neurons in the intermediate layer for each data group are input to each neuron in the output layer through the calculation using the parameters, and each neuron in the output layer outputs the probability of the corresponding destruction cause.

By forming the input layer and the intermediate layer for each data group in this way, it is possible to generate a trained model according to the properties of the data, and it is possible to realize high analysis accuracy. That is, since the properties of data (quantitative data, categories) are different for each data group, high analysis accuracy can be realized by constructing and using an intermediate layer according to the properties of each data. Specifically, for the image data of the fracture surface image, an intermediate layer of a network structure suitable for image analysis is constructed, and for various data groups of peripheral information, an intermediate layer suitable for the properties of each data group is constructed. To do. The number of layers constituting the intermediate layer may be different for each data group.

Similar to the first embodiment, the learning unit 22 updates the parameters of each layer by machine learning by an error reverse propagation method or the like. Then, the learning unit 22 determines whether or not to end the machine learning (S13), and if it determines that the machine learning is not completed (S13: NO), returns to S11 and repeats the subsequent processing. When it is determined to end (S13: YES), the learning unit 22 stores the generated learned model in the model storage unit 24, and transmits the generated learned model to the fracture surface analysis device 1 (S14). As a result, the trained model is stored in the model storage unit 15 of the fracture surface analysis device 1.

As described above, the trained model of the present embodiment is for causing the computer to function so as to identify the cause of the destruction of the resin molded product based on the fracture surface image obtained by capturing the fracture surface of the resin molded product to be analyzed. This is a trained model of the above, and is generated by machine learning using teacher data in which a plurality of data groups including a fracture surface image, peripheral information that causes destruction, and a label that is the cause of destruction are associated with each other. Then, in the neural network of the trained model, an input layer and an intermediate layer are formed for each data group, the output of the intermediate layer for each data group is input to the output layer, and the output layer outputs the probability of the cause of destruction. The weighted value is learned. Further, in the trained model, the fracture surface analysis device 1 is used so as to perform an operation based on the trained weighted values on a plurality of data groups input to the input layer and output the probability of the cause of destruction from the output layer. Make it work.

Next, with reference to FIG. 5, the fracture surface analysis process performed by the fracture surface analysis device 1 of the present embodiment will be described. Here, the difference from the fracture surface analysis process of the first embodiment will be mainly described. The acquisition unit 11 acquires a fracture surface image and peripheral information of the fracture surface of the resin molded product transmitted from the user terminal 3 (S21).

FIG. 10 shows an example of a screen for fracture surface analysis displayed on the display of the user terminal 3. The screen of the present embodiment is different from the screen of the first embodiment in that an input field (input area) 85 for peripheral information is set. The user selects a desired fracture surface image (image file) from the data stored in the storage unit of the user terminal 3 using the reference button 81 on the screen, and selects by clicking the upload button 82. The fracture surface image is determined as the analysis target. Then, the user inputs the peripheral information of the resin molded product to be analyzed in each input field 85 of the peripheral information. In the illustrated example, the resin grade is input for the composition information. As a result, the input load and input error of the user can be reduced. The user clicks the analysis button 83 after selecting the fracture surface image and inputting the peripheral information. As a result, the user terminal 3 transmits a fracture surface analysis request including the fracture surface image and peripheral information selected by the user to the fracture surface analysis device 1, and the acquisition unit 11 of the fracture surface analysis device 1 transmits the fracture surface. Receive images and peripheral information.

The acquisition unit 11 converts the received peripheral information resin grade into composition information using the resin grade DB. It is assumed that the fracture surface analysis device 1 of the present embodiment stores the resin grade DB in a storage unit (not shown). Then, the changing unit 12 resizes the acquired fracture surface image (S22), and the analysis unit 13 identifies the cause of destruction by using the fracture surface image after resizing and the peripheral information (S23).

Specifically, the analysis unit 13 reads out the trained model (neural network and parameters) stored in the model storage unit 15, inputs each data group to the neurons in the input layer, and performs the calculation of the trained model. .. In the present embodiment, the input layer and the intermediate layer are calculated for each data group, the intermediate layer for each data group is output to the output layer, and finally the identification result of the corresponding destruction cause is obtained from each neuron in the output layer. It is output.

The analysis unit 13 identifies the cause of destruction based on the probability of each cause of destruction output by the trained model. For example, the analysis unit 13 identifies the cause of destruction having the highest probability. Then, the output unit 14 transmits the cause of destruction identified by the analysis unit 13 to the user terminal 3 (S24).

As a result, the identified fracture cause (“brittle fracture” in the illustrated example) is displayed in 84 in the display area of the analysis result of the screen of the user terminal 3. In the screen example of FIG. 10, one destruction cause with the maximum probability is displayed as the analysis result, but a plurality of destruction causes output by the trained model and their probabilities are also displayed as the analysis result. May be good.

In the fracture surface analysis device 1 of the present embodiment described above, an acquisition unit that acquires an image of the fracture surface of the resin molded product to be analyzed and peripheral information that causes the fracture, and the fracture surface image and its surroundings. It has an analysis unit that identifies the cause of destruction of the resin molded body by inputting the target fracture surface image and peripheral information into the trained model generated by machine learning using the information and the label that is the cause of destruction as teacher data. .. As a result, in the present embodiment, the cause of fracture can be easily and quickly identified (estimated) with high accuracy in consideration of the peripheral information of the resin molded product. That is, even if it is difficult to determine the cause of fracture from the fracture surface image alone, the cause of fracture can be comprehensively analyzed by adding peripheral information such as the composition of the fractured resin molded body, and the cause of fracture can be determined with high accuracy. Can be identified.

The fracture surface analysis device 1 and the model generation device 2 described above include, for example, a CPU (Central Processing Unit, processor), a memory, a storage (HDD: Hard Disk Drive, SSD: Solid State Drive), and a communication device. A general-purpose computer system including an input device and an output device can be used. In this computer system, each function of each device is realized by executing a predetermined program loaded on the memory by the CPU. For example, each function of the fracture surface analysis device 1 and the model generation device 2 is such that the CPU of the fracture surface analysis device 1 in the case of the program for the fracture surface analysis device 1 and the model generation device in the case of the program for the model generation device 2. It is realized by executing each of the two CPUs. As the user terminal 3, a general-purpose computer system (for example, a PC or the like) including a CPU, a memory, a communication device, and the like can be used.

The program for the fracture surface analysis device 1 and the program for the model generation device 2 should be stored in a computer-readable recording medium such as an HDD, SSD, USB memory, CD-ROM, DVD-ROM, or MO. It can also be delivered over the network. Further, since the model generation device 2 has a large amount of machine learning calculation amount, it may be provided with a GPU (Graphics Processing Unit) to perform high-speed processing.

Further, the present invention is not limited to the above embodiment, and many modifications can be made within the scope of the gist thereof. For example, there are the following modified examples.

<Modification example 1>
In the first embodiment, the trained model is used to identify the cause of destruction from the fracture surface image, but template matching may be used to identify the cause of destruction from the fracture surface image. In this case, the fracture surface analysis device stores, instead of the trained model, a partial image showing a characteristic portion in the fracture surface image in the storage unit as a template image for each cause of destruction. As the template image, for example, a partial image of a region surrounded by a dotted line of each fracture surface image shown in FIGS. 2A to 2C can be used.

The fracture surface analysis device 1 performs template matching with each of the plurality of template images stored in the storage unit for the fracture surface image transmitted from the user terminal 3, and calculates the degree of similarity. In template matching, the similarity between a part of the input image and the template image is calculated, the location with the highest similarity is searched, and the similarity (maximum similarity) at the location is output. Methods for calculating the similarity of template matching include SSD (Sum of Squared Difference), SAD (Sum of Absolute Difference), and NCC (Normalized Cross Correlation).

Then, the fracture surface analysis device 1 identifies the cause of destruction according to the similarity output for each template. That is, the fracture surface analysis device 1 identifies the cause of destruction corresponding to the template image having the highest degree of similarity as the cause of destruction of the received fracture surface image.

As described above, the fracture surface analysis device of the modified example 1 has a storage unit that stores a partial image showing a characteristic portion in the fracture surface image as a template image for each cause of destruction, and a resin molded product to be analyzed. The degree of similarity for each template image is obtained by performing template matching with the acquisition unit that acquires the target fracture surface image obtained by imaging the fracture surface, the target fracture surface image, and each of the plurality of template images stored in the storage unit. Is included, and an analysis unit for identifying the cause of destruction of the resin molded product based on the similarity is provided. That is, the fracture surface analysis device of the modified example 2 stores a template image for each cause of destruction instead of storing the learned model in the model storage unit 15 of the fracture surface analysis device 1 of FIG.

<Modification 2>
The fracture surface analysis system of the first and second embodiments receives the fracture surface image transmitted from the user terminal 3 via the network, and transmits the analysis result of the destruction cause to the user terminal 3. And said. However, in the present invention, the user may directly input the fracture surface image to the fracture surface analysis device 1 using a recording medium such as a USB memory. Further, the analysis result may be transmitted to the user terminal 3 via the network, or may be output to an output device (display, printer, etc.) included in the fracture surface analysis device 1.

<Modification example 3>
In the second modification, the user inputs a fracture surface image to the fracture surface analysis device 1 using a recording medium such as a USB memory. However, in the present invention, the fracture surface analysis device 1 and the SEM4 may be integrally configured, and the fracture surface image of the resin molded product imaged and output by the SEM4 may be directly input to the fracture surface analysis device 1. Further, the analysis result may be transmitted to the user terminal 3 via the network as in the modification 2, or may be output to an output device (display, printer, etc.) included in the fracture surface analysis device 1. .. In the latter case, the movement of data via the recording medium and the transmission of data via the user terminal 3 can be omitted.

<Modification example 4>
In the first embodiment, the second embodiment, and the first to third embodiments, the user terminal 3 is connected to the fracture surface analysis device 1 via a network as a configuration different from that of the fracture surface analysis device 1. However, in the present invention, the user terminal 3 may be integrally incorporated in the same housing as the fracture surface analysis device 1 and / or SEM4.

<Modification 5>
In the second embodiment, the acquisition unit 11 converts the received peripheral information resin grade into composition information using the resin grade DB. However, in the present invention, composition information such as the base resin and additives may be input individually. For this purpose, for example, in the resin grade field of the peripheral information input field 85 shown in FIG. 10, "composition information input" is set as one of the options (drop-down list). Then, when the user selects "composition information input", the composition information input field is displayed so that the composition information input by the user is transmitted to the fracture surface analysis device 1 as peripheral information. You should keep it. By inputting the composition information individually by the user in this way, it is possible to analyze the resin-formed body which is not registered in the resin grade DB by adding the composition information as peripheral information.

<Modification 6>
In the above modification 5, the composition information already grasped in advance is input individually. However, in the present invention, the composition analyzer may be further provided integrally with or separately from any one or more of the SEM 4, the user terminal 3, and the fracture surface analysis apparatus 1. The composition analyzer performs composition analysis of the resin molded product to be analyzed at the same time as imaging of the fracture surface and / or before and / or after imaging, and the composition information obtained there is used as peripheral information for fracture surface analysis. It may be input to the device 1. In this case, the acquisition unit 11 of the fracture surface analysis device 1 acquires composition information using the analysis result of the composition analysis device.

Examples of the composition analyzer used here include known composition analyzers such as an infrared spectroscopic analyzer, a nuclear magnetic resonance apparatus, a mass spectrometer, an X-ray diffractometer, a fluorescent X-ray analyzer, an atomic emission analyzer, and various types of chromatography. One type or a combination of two or more types of the above can be used. In this case, by inputting the teacher data including the data group of the composition information according to the composition analysis method of the apparatus to be used into the model generation apparatus 2, it is possible to generate the trained model according to the composition analysis method to be used. .. As a result, even in a resin molded product whose composition information is unclear, more accurate analysis becomes possible.

<Modification 7>
In the first and second embodiments and the modified examples 1 to 6, it is decided that the user selects a specific part of the fracture surface and captures the image data as the fracture surface image. However, in the present invention, any one or more of the SEM4, the user terminal 3, and the fracture surface analysis device 1 is provided with a feature portion specifying function for specifying a characteristic portion of the fracture surface, and the fracture surface image of the portion is provided. It may be automatically acquired and entered. The fracture surface image of the characteristic portion may be obtained by selecting and enlarging a specific range from the image obtained by photographing the entire fracture surface of the resin molded body, or the fracture surface of the resin molded body is enlarged for each part. A specific image may be selected and acquired from a plurality of images taken in the above. Here, the selection of the characteristic part in the characteristic part identification function may be specified by using a model generated by machine learning in the same manner as the identification of the cause of destruction, or the characteristic part stored in advance may be specified. It may be specified from the similarity with the template. As the fracture surface image to be input, the image selected by the feature part identification function may be used as it is, or an appropriate image may be selected and used by the user from the images narrowed down by the feature part identification function. , The image selected by the feature part identification function from the image narrowed down in the range and the number of sheets by the user may be used.

1: Fracture surface analysis device 11: Acquisition unit 12: Change unit 13: Analysis unit 14: Output unit 15: Model storage unit 2: Model generation device 21: Data acquisition unit 22: Learning unit 23: Data storage unit 24: Model storage Part 3: User terminal 4: SEM (scanning electron microscope)

Claims (17)

It is a fracture surface analysis device,
An acquisition unit that acquires an image of the fracture surface of the resin molded product to be analyzed,
An analysis unit that inputs the target fracture surface image to the trained model generated by machine learning using the fracture surface image and the label that is the cause of destruction as teacher data to identify the cause of destruction of the resin molded body. A fracture surface analysis device characterized by having. The fracture surface analysis apparatus according to claim 1.
The teacher data includes peripheral information that causes destruction.
The acquisition unit acquires the peripheral information of the resin molded product, and obtains the peripheral information.
The analysis unit is a fracture surface analysis apparatus characterized in that the peripheral information is input to the trained model together with the target fracture surface image to identify the cause of destruction of the resin molded product. The fracture surface analysis apparatus according to claim 2.
The fracture surface analysis apparatus, characterized in that the peripheral information includes at least one of the composition information, molding conditions, usage environment and shape of the resin molded product. The fracture surface analysis apparatus according to claim 3.
The acquisition unit is a fracture surface analysis apparatus, characterized in that the composition information is acquired by using the analysis result of the resin grade or the composition analyzer. The fracture surface analysis apparatus according to claim 3.
The acquisition unit is a fracture surface analysis device characterized by acquiring the composition information individually input by the user. The fracture surface analysis apparatus according to any one of claims 1 to 5.
A fracture surface analysis apparatus characterized by having a changing portion for resizing the acquired target fracture surface image to an image having a predetermined number of pixels. The fracture surface analysis apparatus according to any one of claims 1 to 6.
The acquisition unit acquires the target fracture surface image transmitted from the user terminal, and obtains the target fracture surface image.
A fracture surface analysis device characterized by having an output unit that transmits the cause of destruction identified by the analysis unit to the user terminal. The fracture surface analysis apparatus according to any one of claims 1 to 7.
The acquisition unit is the target fracture surface image stored in a recording medium, or the target fracture surface image input from one selected from the group consisting of a scanning electron microscope, a stereomicroscope, a metallurgical microscope, and a digital camera. To get and
A fracture surface analysis apparatus having an output unit that outputs the cause of destruction identified by the analysis unit to the fracture surface analysis apparatus or transmits it to a user terminal. The fracture surface analysis apparatus according to any one of claims 1 to 8.
The trained model outputs each probability of multiple causes of destruction.
The analysis unit is a fracture surface analysis apparatus characterized in that the cause of destruction of the resin molded product is identified based on the probability. The fracture surface analysis apparatus according to any one of claims 1 to 9.
A fracture surface analysis device characterized by being an device integrated with a user terminal and / or a scanning electron microscope. The fracture surface analysis apparatus according to any one of claims 1 to 10.
A fracture surface analysis apparatus characterized by having a characteristic portion specifying portion for specifying a characteristic portion of the resin molded product. It is a fracture surface analysis device,
A storage unit that stores a partial image showing a characteristic part in the fracture surface image as a template image for each cause of destruction,
An acquisition unit that acquires an image of the fracture surface of the resin molded product to be analyzed,
The target fracture surface image and each of the plurality of template images stored in the storage unit are subjected to template matching to calculate the similarity for each template image, and the resin molded product is destroyed based on the similarity. A fracture surface analysis device characterized by having an analysis unit for identifying the cause. A trained model generator
An acquisition unit that acquires teacher data in which a fracture surface image obtained by imaging a fracture surface of a resin molded product and a label that is the cause of destruction of the resin molded product are associated with each other.
A trained model generation device comprising a learning unit that generates a trained model for identifying a cause of destruction of the resin molded product by machine learning using the teacher data. It is a fracture surface analysis method performed by a computer.
The acquisition step of acquiring the target fracture surface image of the fracture surface of the resin molded body to be analyzed, and
An analysis step of inputting the target fracture surface image into a trained model generated by machine learning using the fracture surface image and the label as the cause of destruction as teacher data to identify the cause of destruction of the resin molded product. A fracture surface analysis method characterized by performing. A fracture surface analysis program according to any one of claims 1 to 11, wherein a computer functions as the fracture surface analysis apparatus. It is a trained model for making a computer function so as to identify the cause of destruction of the resin molded product based on a fracture surface image obtained by imaging the fracture surface of the resin molded product to be analyzed.
When the pixel value of the fracture surface image is input to the input layer by machine learning using the teacher data that associates the fracture surface image with the label that is the cause of destruction, the output layer outputs the probability of the cause of destruction. The weighted value is learned as
Learned to make the computer function so that the pixel value of the fracture surface image input to the input layer is calculated based on the learned weighted value and the probability of the cause of destruction is output from the output layer. model. The trained model according to claim 16.
The teacher data is data in which a plurality of data groups including a fracture surface image, peripheral information causing destruction, and a label causing destruction are associated with each other.
In the neural network of the trained model, an input layer and an intermediate layer are formed for each data group, the output of the intermediate layer for each data group is input to the output layer, and the output layer determines the probability of the cause of destruction. The weighted value is trained to be output,
A trained model for making a computer function so as to perform an operation based on the learned weighting value on the plurality of data groups input to the input layer and output the probability of the cause of destruction from the output layer.