Prediction device, processing system, method, and program

Abstract

To provide a prediction device that solves at least one of the problems of conventional technology. [Solution] A predictive device for predicting the mechanical properties of a processed product obtained as a result of applying stress to a polymer material that has been plasticized by heating for a predetermined residence time, comprising a trained model that generates predictive information of the mechanical properties based on input information generated from the change over time of a parameter representing the fluidity of the polymer material within a time period determined based on the residence time.

Description

Technical Field

[0001] The present disclosure relates to a prediction device, a processing system, a method, and a program.

Background Art

[0002] In order to realize a resource - recycling economy, it is required to promote material recycling with minimized energy consumption. Promoting the recycling of plastic waste discharged from households (hereinafter also referred to as "household - discharged plastics") always occupies an important position in the overall plan.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] Generally, plastic waste contains different types of polymer components and is likely to have non - uniform material properties. Therefore, it is difficult to reproduce the original physical properties and grades by material recycling, and it is often difficult to stably use it as a raw material for recycling.

[0005] In particular, in the case of household - discharged plastics, although separate collection is being promoted, the level of separation varies greatly depending on various conditions such as local government rules and individual awareness. Therefore, compared with plastic waste separately collected from factories, etc., when considered as a raw material for material recycling, the purity of household - discharged plastics often lacks more stability.

[0006] Conventionally, methods for separating various polymer components from plastic waste have been considered as one way to achieve stable material recycling (Patent Document 1). However, even after various separation processes, it remains difficult to separate household plastic waste into multiple types of pure polymer components. Consequently, the performance of recycled materials obtained from household plastic waste is often still unstable.

[0007] However, the fact that the performance of recycled materials is "unstable" does not mean that "they are not all of low quality." Even recycled materials made from household plastic waste can sometimes have high performance. Interestingly, obtaining high-performance recycled materials is not always synonymous with the raw material being a polymer material consisting of only one type of polymer component.

[0008] In other words, high-performance recycled materials can sometimes be obtained from household plastic waste without strict sorting, or to put it another way, without investing a lot of resources and energy into strict sorting. This can happen when the ratio of polymer components in the raw materials happens to be appropriate for a particular purpose. For example, high-performance recycled materials in terms of impact resistance can sometimes be obtained from polymer materials that contain elastic polymer components.

[0009] However, it is practically difficult to selectively use only high-quality recycled materials. While conventional recycled materials do contain a certain amount of high-quality material, the guaranteed quality remains at a minimum standard. This is because 100% inspection is difficult.

[0010] In the actual operation of material recycling, the types and composition of polymer components contained in the supplied raw materials can constantly fluctuate. This tendency is particularly pronounced when using household waste plastics as raw materials. Therefore, in order to extract high-quality recycled materials, it is essential to perform component analysis of the raw materials and / or performance testing of the recycled materials at least for each lot, and in some cases, continuously. Such component analysis and performance testing will require a greater investment of resources and energy compared to sorting. This disclosure solves at least one of the problems of the prior art described above. [Means for solving the problem]

[0011] The first prediction device of this disclosure is a prediction device for predicting the mechanical properties of a processed product in a processing apparatus that applies stress to a polymer material plasticized by heating for a predetermined residence time and obtains a processed product as a result, and comprises a trained model that generates prediction information of the mechanical properties based on input information generated from the change over time of a parameter representing the fluidity of the polymer material within a time determined based on the residence time.

[0012] The first prediction method of this disclosure is a prediction method for predicting the mechanical properties of a processed product in an apparatus that applies stress to a polymer material plasticized by heating for a predetermined residence time, and as a result obtains a processed product, the method comprising generating prediction information of the mechanical properties by a trained model based on input information generated from the change over time of a parameter representing the fluidity of the polymer material within a time determined based on the residence time.

[0013] The first program of the present disclosure is a program for causing a processing apparatus that applies stress to a polymer material plasticized by heating for a predetermined residence time to obtain a processed product, to execute a prediction method for predicting the mechanical properties of the processed product, the program including the step of generating prediction information on the mechanical properties by a learned model based on input information generated from the change over time of a parameter representing the fluidity of the polymer material within a time determined based on the residence time.

Advantages of the Invention

[0014] According to the present disclosure, at least one of the problems of the above-described conventional techniques can be solved.

Brief Description of the Drawings

[0015] [Figure 1] It is a hardware configuration diagram of a processing system. [Figure 2] It is an explanatory diagram (partial cross-sectional view) of a processing apparatus. [Figure 3] It is a functional block diagram of a processing system. [Figure 4] It is a flow chart showing a series of procedures including data acquisition to model training. [Figure 5] It is a flow chart of generating prediction information using a prediction apparatus. [Figure 6] It is a flow chart showing a series of procedures including data acquisition to model training in the case of continuous processing. [Figure 7] It is a flow chart showing a series of procedures including data acquisition to transfer learning of a learned model. [Figure 8] It is a flow chart of generating prediction information using a prediction apparatus provided with a second model. [Figure 9] It is a dendrogram showing the result of hierarchical clustering. [Figure 10] It is a superposition of melt viscosity curves by class. [Figure 11] It is a superposition of stress-strain curves by class. [Figure 12] It is a superposition of stress-strain curves by class. [Figure 13] It is a diagram showing the structure of the model used. [Figure 14] It is a diagram showing a confusion matrix representing the verification results of the accuracy and sensitivity of the model. [Figure 15] It is a confusion matrix of prediction information obtained based on Raw data. [Figure 16] It is a confusion matrix of prediction information obtained based on two-channel data consisting of Raw data and centered (Offset) data. [Figure 17] It is a confusion matrix of prediction information obtained based on two-channel data consisting of Raw data and standardized data. [Figure 18] It is a confusion matrix of prediction information obtained based on two-channel data consisting of centered data and standardized data. [Figure 19] The melt viscosity curves of each channel of Raw data, centered data (Mean Zero Offset), and standardized data are classified by class. [Figure 20] It is a confusion matrix of prediction information obtained when the machine learning of the model is newly performed using the second training dataset. [Figure 21] It is a confusion matrix of prediction information obtained when the feature extraction method is pre-learned and transferred to the Charpy impact prediction model.

Mode for Carrying Out the Invention

[0016] The first prediction device of this disclosure is a prediction device for predicting the first mechanical properties of a processed product in a processing apparatus that applies stress to a polymer material plasticized by heating for a predetermined residence time and obtains a processed product as a result, and comprises a trained model that generates prediction information of the first mechanical properties based on input information generated from the change over time of a parameter representing the fluidity of the polymer material within a time determined based on the residence time.

[0017] Polymer materials that have been plasticized by heating possess a certain degree of fluidity, although to varying degrees. In this state, stress is applied to the polymer material within the processing apparatus for a predetermined time. The parameters representing fluidity may reflect the properties of the polymer material itself, i.e., its composition and state. Furthermore, the parameters representing fluidity may also reflect the magnitude of the applied stress, the strain rate, and the temperature. Moreover, the changes in the parameters representing fluidity over time may reflect the history of the above changes.

[0018] The first prediction device generates prediction information for the first mechanical properties of the processed material based on input information generated from the time-dependent changes in parameters representing fluidity. The time-dependent changes in the parameters representing fluidity may reflect the properties of the raw material, the thermal history due to processing, and the stress history, etc. Based on this information of "time-dependent changes in parameters representing fluidity" which can be obtained as a result of processing the raw material, the first prediction device makes it possible to predict at least the first mechanical properties of the processed material without having to perform a full inspection. If the first prediction device is applied to the production of recycled materials (i.e., processed materials) by material recycling of polymer materials, high-quality processed materials can be extracted without performing a full inspection.

[0019] The second prediction device of this disclosure is a prediction device in which, in the first prediction device, the input information is multi-channel data obtained by applying a transformation that preserves the reference axis to the change in the parameters over time using at least one transformation method.

[0020] Parameters representing fluidity change under the influence of various factors. For example, variations in the amount of raw material input to the processing device can affect the fluctuation of the absolute value of the fluidity parameter. To mitigate the effects of such scale-related or non-reproducible errors, the input information in the second prediction device is augmented with data obtained by transforming the time-dependent changes in the fluidity parameter using at least one transformation method. Furthermore, since this transformation method maintains the reference axis before and after the transformation, the original data (time-dependent changes in the fluidity parameter) and the transformed data can be multi-channel data sharing the reference axis. The data in channels other than the original data emphasizes and retains one or more of the various features of the original data, depending on the transformation method. By treating these as a single input, the effects of scale-related or non-reproducible errors are reduced, resulting in more accurate prediction results.

[0021] The third prediction device of this disclosure is a prediction device in which, in the first prediction device, the trained model is a model whose basic structure is a recurrent neural network.

[0022] The recurrent neural network, the basic structure of the third prediction device, can propagate information to the next step while retaining information from the previous step. Therefore, it is suitable for analyzing time series data. The input information generated from the time-dependent changes in parameters representing liquidity also has the characteristics of time series data, and by adopting an RNN-based model, more accurate prediction information can be obtained.

[0023] The fourth prediction device of this disclosure is a prediction device in which, in the first prediction device, at least one selected from the group consisting of the composition and state of the polymer material is unknown.

[0024] One of the key features of the fourth prediction device is that, based on information on the "time-dependent changes in parameters representing fluidity" that can be obtained as a result of processing the raw materials, it is possible to predict at least the first mechanical properties of the processed material without having to perform a full inspection. This remains true even if the composition and state of the polymer material (raw material) are unknown. In other words, predictive information on the first mechanical properties of the processed material can be obtained without performing a component analysis of the raw material. Recycling raw materials such as household waste plastics often have unknown compositions, and their state, including the state of degradation, often varies. If the fourth prediction device is applied to the production of recycled materials, high-quality recycled materials can be easily extracted even in such cases.

[0025] A fifth prediction device of this disclosure is a prediction device in which, in the first prediction device, the prediction information includes at least one type of information selected from the group consisting of classification information of the results of a first mechanical properties test of the processed object based on the distance of the pattern, and predicted values ​​of the results of the first mechanical properties test.

[0026] Classification information obtained through clustering is more suitable as predictive information for trained models that have been machine-learned using training datasets generated from a small number of samples. In this case, more accurate results are likely to be obtained than with predictive information that is the result (predicted value) of the first mechanical property test itself, and it is easier for operators with less skill and experience to understand and use. If identification information such as a name is assigned to each class in advance, it becomes easier to understand the appropriate application of the processed object from the predictive information. Furthermore, by changing the clustering criteria, it is easy to generate trained models that output different predictive information, making it easier to expand or change the application of the prediction device. This only requires changing the annotations of the same training dataset, so the cost can be kept lower.

[0027] The predicted values ​​of the results of the first mechanical properties test are more suitable as predictive information for a trained model that has been machine-trained using a training dataset generated from a sufficiently large number of samples. In this case, even the predicted values ​​of the first mechanical properties test itself are likely to yield more accurate results. The type of predictive information can be appropriately selected depending on the size (number of samples) and purpose of the training dataset. It may also be configured to output both types of information.

[0028] The sixth prediction device of this disclosure is a prediction device in which, in the fifth prediction device, the class is set in advance based on at least one feature quantity extracted from the result.

[0029] Depending on the intended use of the processed material, the required mechanical properties may be known in advance. For example, if one or more of the following properties, obtained from mechanical property tests, are required for the processed material—such as tensile strength, modulus of elasticity, elongation at break, yield stress, flexural modulus, bending resistance, impact resistance, compressive strength, hardness, wear resistance, and coefficient of thermal expansion—then using these characteristics obtained from the results of the first mechanical property test as clustering criteria makes it easier to obtain processed materials of the appropriate quality for the intended use.

[0030] The seventh prediction device of this disclosure is a prediction device that, in the first prediction device, includes a rotating mechanism driven by a motor, and is a processing device that applies shear stress as the stress to the polymer material contained in a container by the rotating mechanism.

[0031] Processing equipment equipped with a motor-driven rotation mechanism that applies shear stress to polymer materials contained in a container is widely used for the mixing and transfer of polymer materials. To stabilize the performance of the processed material, attention has been paid to the thermal and stress history of the raw materials during use. However, when handling various raw materials, many factors are intertwined, and even stabilizing the performance of the processed material, let alone extracting high-quality processed material, has required a great deal of trial and error and subjective skill. The seventh prediction device is suitable for application to these processing equipment that have been used in general, and can generate highly accurate prediction information regardless of the composition of the polymer material.

[0032] The eighth prediction device of this disclosure is a prediction device in which, in the seventh prediction device, the change over time of the parameter representing fluidity is calculated based on the change over time of an index representing the drive parameter of the motor.

[0033] The time-dependent changes in an index representing the motor's drive parameters can be easily obtained as a result of processing the raw materials. Furthermore, since the motor is an essential component for any processing device equipped with a rotating mechanism, the eighth prediction device can acquire predetermined data and generate prediction information more easily and without the need for special mechanisms or sensors.

[0034] The ninth prediction device of this disclosure is a prediction device in which, in the eighth prediction device, the drive parameters include the torque of the motor and the rotational speed of the motor.

[0035] Motor torque and rotational speed, or their respective indicator values ​​such as current and power values, are generally monitored during the operation of the processing device. The ninth prediction device, configured to calculate the time-dependent changes in parameters representing fluidity using these values, can acquire predetermined data and generate prediction information more easily and without the need for special mechanisms or sensors.

[0036] The tenth prediction device of this disclosure is a prediction device in which, in the third prediction device, the trained model is a model whose basic structure is a bidirectional recurrent neural network.

[0037] The pre-trained model of the 10th prediction device is based on a bidirectional recurrent neural network (BiRNN). BiRNN is an RNN that enables the propagation of information in the reverse direction. The input information generated from the time-dependent changes in parameters representing fluidity has characteristics such as changes after the raw material is introduced into the processing device, and the state just before the end of the residence time, which has a significant impact on the next process such as pelletizing. Forward analysis of time-series data alone is not always sufficient. Furthermore, it is necessary to consider long-range dependence as time-series data. The basic structure of the pre-trained model in the 10th prediction device employs a BiRNN optimized for the input information, thus enabling more accurate prediction information.

[0038] Furthermore, if the basic structure of the trained model is a Gated BiRNN (Bidirectional Recurrent Neural Network) with a gate mechanism, long-term dependencies are more easily maintained compared to a standard BiRNN. In particular, when the input data is long time-series data, adopting a Gated BiRNN improves training efficiency and accuracy. As a result, more accurate predictive information can be obtained.

[0039] The eleventh prediction device of this disclosure is a prediction device in which, in the second prediction device, the transformation method includes centralization and standardization, and the multi-channel data includes at least three of the data before transformation, the centralized data and the standardized data.

[0040] The unconverted data, centered data, and standardized data can each characteristically reflect the absolute value, the trend of value change, and the pattern of change over time for the parameters representing liquidity. By using multi-channel data containing these elements as input information, more accurate prediction results can be obtained.

[0041] The twelfth prediction device of this disclosure is a prediction device that, in the first prediction device, includes a learning unit, the learning unit performs transfer learning on the trained model to generate a second model, and the second model generates prediction information of a second mechanical property based on the input information.

[0042] When using processing equipment to recycle polymer materials and produce recycled materials as processed products, attention is focused not only on the performance of the processed products themselves, but also on the performance of the processed products when they are molded and processed into new products. In most cases, the processed materials are supplied as pellets, and new products are manufactured by molding these pellets.

[0043] The twelfth prediction device includes a learning unit. This learning unit performs transfer learning on a trained model to generate a second model. The second model generates prediction information for the second mechanical properties based on the input information used when predicting the first mechanical properties. The second mechanical properties may be the same as the first mechanical properties, but it is preferable that they be different.

[0044] According to the 12th prediction device equipped with a second model, for example, in the production of recycled pellets (processed material) using waste plastic as a raw material, and in the production of products processed from recycled pellets, it is possible to predict the second mechanical properties of the "product" based on input information obtained incidentally during the recycling of waste plastic (polymer material, raw material). In this case, by employing transfer learning that utilizes a feature extraction method from the time-dependent change of parameters representing the fluidity of polymer materials, highly accurate predictive information of the product's mechanical properties (second mechanical properties) can be generated even if the processing process from processed material to product goes through a different thermal process than the manufacturing process of recycled pellets from which the time-dependent change data of the parameters representing fluidity was obtained.

[0045] The first processing system of this disclosure is a processing system that includes one of the first to twelfth prediction devices and the processing device.

[0046] Because this processing system is equipped with the aforementioned prediction device, it possesses the same characteristics as the prediction device, facilitates collaboration with the processing unit, and can be easily applied to feedback control and other applications that utilize prediction information.

[0047] Embodiments will be described in detail below with reference to the attached drawings. The following examples illustrate apparatuses and methods for realizing the technical concept of the disclosure, and the technical concept of this disclosure is not limited to those described below. The technical concept of this disclosure can be modified in various ways within the scope of the claims. In particular, it should be noted that the drawings are schematic and may differ from actual ones.

[0048] Figure 1 is a hardware configuration diagram of the processing system (Example 1). Figure 2 is an explanatory diagram (partial cross-sectional view) of the processing unit. Figure 3 is a functional block diagram of the processing system.

[0049] The processing system 1 comprises a prediction device 10 and a processing device 20. The prediction device 10 is a computer including a processor 11, memory 12, communication I / F 13, processing device I / F 14, and input / output I / F 15. The processing device 20 comprises a controller 21, a motor 22, a rotating mechanism 23, and a heater 24. The processing device 20 drives the rotating mechanism 23 with the motor 22 under the control of the controller 21, which is an MCU (Microcontroller Unit). It also heats and plasticizes the polymer material with the heater 24. The processing device applies shear stress to the polymer material that has been plasticized by heating over a residence time.

[0050] The processor 11 is the central component of the prediction device 10 and has the function of executing the program stored in the memory 12 and processing data. Specifically, it has the function of executing instructions included in the program, arithmetic functions, the function of reading data from the memory 12 etc., processing it and writing it back to the memory 12 etc. as needed, and the function of controlling other things (clock management, power management, etc.). Examples of processors 11 include CPU (Central Processing Unit), GPU (Graphics Processing Unit), DSP (Digital Signal Processor), and FPGA (Field Programmable Gate Array).

[0051] Memory 12 consists of volatile memory and non-volatile memory. Volatile memory is mainly used for storing temporary data or data being processed. Hardware that has this function includes RAM (Random Access Memory), cache memory, and SRAM (Static RAM). On the other hand, non-volatile memory is mainly used for long-term data storage. Hardware that has this function includes flash memory, SSD (Solid State Drive), ROM (Read Only Memory), MRAM (Magnetoresistive RAM), and PRAM (Phase-change RAM).

[0052] The communication interface 13 provides functions such as data transmission and reception between the prediction device 10 and other devices and networks, consistency verification and error checking of transmission and reception, network connection management, and security assurance. Hardware that provides such functions includes wired communication interfaces such as optical fiber interfaces and Ethernet® interfaces, and wireless communication interfaces such as Wi-Fi® modules and Bluetooth® modules.

[0053] The processing unit interface (I / F) 14 connects the prediction device 10 and the processing unit 20, and performs functions such as sending and receiving drive parameters and control signals. Hardware that provides such functions includes serial interfaces, Ethernet® interfaces, USB (Universal Serial Bus) interfaces, and fieldbuses. If the processing unit 20 is equipped with a wireless communication module, the communication interface (I / F) 13 may also serve as the processing unit interface (I / F) 14. In this case, the prediction device 10 does not need to be equipped with a processing unit interface (I / F) 14.

[0054] The I / O interface (I / F5) transmits and receives data to and from external input / output devices (such as keyboards, mice, printers, and monitors). It may also supply power to connected external hardware. Examples of hardware with such functionality include USB (Universal Serial Bus), HDMI (registered trademark), DisplayPort, Bluetooth (registered trademark), and serial ports.

[0055] The processing apparatus 20 (Figure 2) is a twin-screw extruder comprising a controller 21, a motor 22, a reducer 23A, a screw 23B, a heater 24, a barrel 25, a die 26, and a feeder 27. The motor 22 is driven by power supplied from a power source (not shown) under the control of the controller 21. The rotating mechanism 23, consisting of the reducer 23A and the screw 23B, rotates by the driving force from the motor 22. Multiple heaters 24 are arranged around the outer circumference of the barrel 25, which is a cylindrical metal container. When the barrel 25 is heated by the heaters 24, the polymer material contained within the barrel 25 is heated and plasticized. Shear stress is applied to the plasticized polymer material by the two screws 23B. At this time, the torque T and rotational speed N of the motor 22 are stored as time-series data.

[0056] The polymer material is fed from the feeder 27 into the barrel 25. The feeder 27 comprises an inverted cone-shaped hopper 27A, a feed motor 27B, and a feed screw 27C. The hopper 27A holds a predetermined amount (e.g., one lot) of raw material (polymer material). The feed screw 27C is rotated at a constant speed, typically by the feed motor 27B. The raw material contained in the hopper 27A is sequentially pushed out by the feed screw 27C, which is located adjacent to the outlet, and fed into the barrel 25. The feeder 27 may also be equipped with a feed heater for heating the raw material.

[0057] The raw material fed from the feeder 27 into the barrel 25 is heated by the heater 24 and sequentially plasticized. Simultaneously with the progress of plasticization, shear stress is applied by the rotation of the two screws 23B. The raw material moves toward the tip of the screws 23B while being subjected to shear stress. In this example, the screws 23B have a tapered shape, becoming narrower from the feeder 27 side (i.e., the motor 22 side, the base side) toward the tip. At the base side, the clearance between the barrel 25 and the screws 23B is set narrowly, resulting in high shear stress and shear rate applied to the polymer material, which facilitates sufficient mixing. On the other hand, at the tip side, the clearance between the barrel 25 and the screws 23B is set wide, which suppresses pressure loss and ensures a stable extrusion amount from the die 26. The shear stress and shear rate applied to the polymer material (raw material) can constantly change during the residence time. Furthermore, since heating proceeds sequentially, the fluidity of the polymer material can constantly change during the residence time. Note that the above is just one example, and the diameter and shape of the screw 23B, the clearance with the barrel 25, etc., can be appropriately changed depending on the application.

[0058] Here, residence time refers to the time during which the polymer material remains in the container and stress is applied. In this example, it refers to the time during which the polymer material remains in the barrel 25 and is subjected to shear stress by the rotating mechanism 23. Residence time begins when the polymer material is fed from the feeder 27 into the barrel 25 and ends when the fed polymer material is discharged from the die 26. The feeding of polymer material from the feeder 27 into the barrel 25 can be detected by monitoring the current value, power value, etc., of the feed motor 27B. In this example, the polymer material feeding process proceeds in batch form. That is, one lot of polymer material contained in the hopper 27A is sequentially fed into the barrel 25 and discharged from the die 26. During this time, no new polymer material is fed in, and the residence time of one lot of polymer material is clearly defined. The processing device 20 may also be equipped with a circulation path to return the polymer material from the die 26 back into the barrel 25. In this case, the residence time ends when the circulation is completed and the processed material is discharged from the barrel 25.

[0059] The residence time can vary depending on the method of inputting the raw materials and the method of discharging the processed material. Classification by input and discharge method is as follows: First, input and discharge are performed simultaneously. This is a batch system in the strict sense. In this case, the residence time is from the input of the raw materials to the discharge of the processed material. Second, input is performed sequentially and discharge is performed simultaneously. In this case, the time spent in barrel 25 differs between the first and last inputted raw materials in the strict sense, but the "residence time" in this case is defined as from the start of raw material input to the completion of discharge of the processed material. Third, input is performed simultaneously and discharge is performed sequentially. In this case, the time spent in barrel 25 differs between the first and last discharged processed material in the strict sense, but the "residence time" is defined as from the input of the raw materials to the completion of discharge of all processed material. Fourth, both input and discharge are performed sequentially. In this case, assuming that no new raw materials are added until the discharge of all processed materials is complete, the "residence time" is defined as the time from the start of raw material input until the complete discharge of all processed materials.

[0060] While the definitions of these residence times do not directly affect the accuracy of the prediction information, it is preferable that they be the same when acquiring data for the training dataset and when generating prediction information for new polymer materials. This is because if the system is trained using a training dataset acquired by one of the first to fourth methods, the accuracy will be highest when generating prediction information for polymer materials processed in the same way. In this embodiment, the following explanation is based on the premise that no new raw materials are introduced until all of the introduced raw materials (e.g., one lot) have been discharged (the fourth pattern). However, as will be described later, the prediction device 10 can be applied not only to the four batch processing patterns described above, but also to continuous input of raw materials and continuous discharge of processed materials (continuous processing). The handling of continuous processing will be described later.

[0061] Furthermore, although the processing apparatus 20 in this example is a twin-screw extruder, the processing apparatus to which the prediction device 10 can be applied and the processing apparatus that can be incorporated into the processing system 1 are not limited to those described above. Any known processing apparatus can be used as long as it applies stress to a polymer material that has been plasticized by heating over a residence period of time, and as a result obtains a processed product. Examples of such processing apparatuses include those that apply shear stress, such as single-screw extruders, kneaders, roll mills, and Banbury mixers. Alternatively, those that apply compressive / tensile stress, such as press molding machines, compression molding machines, and stretchers, may also be used. Among these, those that apply shear stress are preferred in that they provide more accurate predictive information, and devices that apply shear force to a polymer material contained in a container by a rotating mechanism (screw, blades, and rollers, etc.) (e.g., extruders, kneaders, and Banbury mixers, etc.) are more preferred.

[0062] Next, the functions of each part of the processing system 1 will be explained with reference to Figure 3. The prediction device 10 comprises a control unit 101, a storage unit 102, a data acquisition unit 103, a processing unit 104, a prediction unit 105, a learning unit 106, an output unit 107, and a communication unit 108.

[0063] The control unit 101 controls the entire prediction device 10. The functions of the control unit 101 are realized by the processor 11, etc. The storage unit 102 stores programs and necessary data. The functions of the storage unit 102 are realized by the memory 12, etc.

[0064] The data acquisition unit 103 transmits control commands to the processing unit 20 and receives time-dependent changes in the drive parameters of the motor 22 from the processing unit 20. The functions of the data acquisition unit 103 are realized by the execution of a program stored in the processing unit I / F 14 (or communication I / F 13) and memory 12 by the processor 11.

[0065] The drive parameters refer to parameters that can typically be obtained incidentally when driving the motor 22, and may include torque, rotational speed, acceleration, temperature, and vibration. In particular, from the viewpoint of easy conversion to parameters representing the fluidity of polymer materials, the drive parameters preferably include torque and / or rotational speed, and more preferably include torque and rotational speed. The data acquisition unit 103 may be configured to acquire the change over time of an index representing the drive parameters. In addition to the above-mentioned drive parameters, the index representing the drive parameters may include values ​​that can be easily converted to drive parameters. Examples of such values ​​include the current value, voltage value, and power value of the motor 22. That is, the data acquisition unit 103 may be configured to acquire the power value, voltage value, and / or power value of the motor 22 as an index representing the drive parameters, in addition to or instead of the drive parameters themselves. While expensive sensors may be required to directly measure the drive parameters themselves, such as torque, the power value, voltage value, and / or power value of the motor 22 can be measured more easily.

[0066] The processing unit 104 generates input information from the time-dependent changes in the drive parameters (represented by index values) acquired by the data acquisition unit 103. The function of the processing unit 104 is realized when a program stored in memory 12 is executed by the processor 11.

[0067] The input information is generated from the time-dependent changes in parameters representing the fluidity of the polymer material within the residence time. The processing unit 104 converts the time-dependent changes in index values ​​representing the drive parameters into time-dependent changes in parameters representing fluidity. Parameters representing the fluidity of the polymer material in a plasticized state due to heating (hereinafter also referred to as Rh parameters) include elastic modulus, viscosity, relaxation time, melt viscosity, and melt flow rate (MFR). In this example, the processing unit 20 is a twin-screw extruder, and the Rh parameter is preferably melt viscosity. The type of Rh parameter can be appropriately selected depending on the type and structure of the processing unit. Furthermore, the Rh parameter does not need to be a so-called true value. This value is used for model learning and prediction, and the processing unit 20 used in this process is a specific device. Therefore, the Rh parameter may be a relative value calculated using a method specific to the processing unit. Consequently, even if the Rh parameter is "melt viscosity," the actual Rh parameter may be an index value related to melt viscosity calculated by a predetermined calculation formula, so to speak, an "apparent melt viscosity."

[0068] If the Rh parameter is melt viscosity, the processing unit 104 calculates the shear stress and shear rate from the motor torque and rotational speed included in the drive parameters, as well as structural parameters related to the processing unit 20, such as the diameter of the screw 23B and the clearance with the barrel 25. From the shear stress and shear rate, the processing unit 104 calculates the melt viscosity (apparent melt viscosity) and its change over time, taking into account predetermined corrections as necessary. The change over time of the Rh parameter refers to the collection of points plotted in chronological order based on the measured (calculated) values ​​of the Rh parameter at each point in time.

[0069] Depending on the operating method of the processing unit 20, the time-dependent fluctuation pattern of the index value representing the drive parameter and the time-dependent fluctuation pattern of the Rh parameter may coincide or be similar. For example, when the processing unit 20 is operated with the motor 22 rotation speed constant, the fluctuation pattern of torque (or its index value, such as the current value of the motor 22), which is one of the index values ​​representing the drive parameter, and the fluctuation pattern of melt viscosity, which is an Rh parameter, may coincide in almost the same way. In such cases, the processing unit 104 may treat the time-dependent change of the index representing the drive parameter as the time-dependent change of the Rh parameter.

[0070] The processing unit 104 may further perform predetermined data adjustments on the time-dependent changes in index values ​​representing the driving parameters and / or the time-dependent changes in the Rh parameters (hereinafter also referred to as "time-dependent changes in Rh parameters, etc."). Examples of data adjustments include dimensionality reduction and augmentation. Data adjustments may also include processing of the entire dataset. Data adjustments are optional processes.

[0071] The method of dimensionality reduction is not particularly limited, but examples include methods that preserve the characteristics of time-series data, such as interval averaging and downsampling, and methods that do not preserve the characteristics of time-series data, such as PCA (Principal Component Analysis). Among these, dimensionality reduction methods that preserve the characteristics of time-series data are preferred from the viewpoint of obtaining more accurate predictive information. Interval averaging is excellent as a dimensionality reduction method for time-series changes of Rh parameters, etc., because it can reduce data redundancy while retaining the effects of variability and short-term fluctuations. The width of the interval is not particularly limited, but it is preferably between 1% and 20% of the dwell time. In addition to interval averaging, moving averages can also be used. In this case, by adjusting the overlap amount of the sliding window, it is possible to achieve both dimensionality reduction and capture the characteristics of time-series data fluctuations.

[0072] The augmentation method is not particularly limited, but a method that maintains the reference axis as time series data (time axis, or a reference axis after dimensionality reduction derived from the time axis) is preferred. Examples of such methods include noise addition, scaling, and translation, with noise addition being preferred. Examples of noise addition include Gaussian noise addition, frequency noise addition, spike noise addition, and drift noise addition. Among these, Gaussian noise addition is preferred.

[0073] Processing the entire dataset may include, for example, normalization (overall standardization) based on the mean and variance of the entire dataset. While data augmentation using the transformation methods described later is performed on a sample-by-sample basis, i.e., on a single entry of the change in the Rh parameter over time, data adjustment is a process performed on the entire dataset, which includes multiple changes in the Rh parameter over time.

[0074] The processing unit 104 further expands the data on the time-dependent changes in the Rh parameter by applying a transformation that maintains the reference axis using a predetermined transformation method, thereby creating multi-channel data. Examples of such transformation methods include standardization, normalization, centering, logarithmic transformation, translation, and clipping. Standardization is a transformation method that sets the mean of the data to 0 and the variance to 1. Normalization is a transformation method that sets the minimum value of the data to 0 and the maximum value to 1 (or another specified range). Centering is a transformation method that subtracts the mean of the entire data from each value. Logarithmic transformation is a transformation method that takes the logarithm of the data values. Translation is a transformation method that adds a certain offset to the entire data. Clipping is a transformation method that limits the data values ​​to a specified upper limit or subtraction.

[0075] All of these conversion methods can preserve a reference axis, that is, the time axis, or a reference axis derived from the time axis. Therefore, the data obtained after conversion shares the same reference axis as the data before conversion. Data that has changed over time and shares such a reference axis is called multi-channel data.

[0076] The Rh parameter changes under the influence of various factors. For example, variations in the amount of raw material input into the processing unit can affect the fluctuation of the absolute value of the parameter representing fluidity. Raw data, such as the change in the Rh parameter over time, contains valuable data that reflects the characteristics of the sample. On the other hand, placing too much emphasis on the difference in its absolute value can lead to leaving room for improvement in the accuracy of predictive information.

[0077] The conversion method is selected appropriately depending on the purpose. Purposes of conversion include, for example, emphasizing various features of the data before conversion or facilitating comparisons between data sets. From the perspective of obtaining more accurate predictive information, it is preferable to select one or more of the following conversion methods, each with a different purpose, and to multi-channel the data before conversion.

[0078] The first type of transformation method is one that adjusts the scale. In this specification, standardization and normalization are classified as such. Scaling functions to adjust for the spread of data, such as the variation in absolute values ​​between samples.

[0079] The second transformation method is a transformation method that adjusts the distribution. In this specification, logarithmic transformations and the like are classified as such. This transformation method compresses the range of possible values ​​for the data and can reduce the impact of outliers.

[0080] A third transformation method is a transformation method that adjusts the position. In this specification, centering and translation are classified as such. This transformation method allows for changing the position of data while maintaining its scale, making it easier to compare data points.

[0081] The fourth transformation method is one that restricts the range of data. In this specification, clipping and similar methods fall into this category. This transformation method can reduce the impact of outliers and extreme values.

[0082] From the viewpoint of obtaining more accurate predictive information, it is preferable that the multi-channel data is a composite of data transformed by multiple different transformation methods and the data before transformation. The number of transformation methods to be selected is not particularly limited, but it is preferably 1 to 10, more preferably 1 to 5, and even more preferably 2 to 4. In particular, from the viewpoint of obtaining more accurate predictive information, it is preferable to have composite data of 3 to 5 channels, including one or more data transformed by a first transformation method, one or more data transformed by a third transformation method, and the data before transformation.

[0083] By performing the conversion while maintaining the reference axis using a predetermined conversion method, multi-channel conversion becomes possible. With multi-channel data in which multiple different features are enhanced for each channel and the magnitude of the influence of each feature is appropriately adjusted, the effects of scaling or non-reproducible errors are reduced, resulting in more accurate prediction results.

[0084] Multi-channeling is particularly effective when the temporal changes in the Rh parameter contain a greater amount of scale-dependent or non-reproducible errors. When the input information is multi-channel data, it is preferable that each entry in the first and second training datasets, described later, is also multi-channel data. This allows the prediction device to generate more accurate prediction information.

[0085] On the other hand, if the time-dependent changes in the Rh parameter do not contain many scale-dependent or non-reproducible errors, in other words, if the data is of higher quality, it may be possible to generate highly accurate predictive information without multi-channelization. In this case, the input information can be the time-dependent changes in the Rh parameter as is, or data that has undergone the aforementioned predetermined data adjustments (excluding multi-channelization), such as dimensionality reduction (in this specification, such data may be referred to as "Raw data" below). Similarly, it is preferable to use "Raw data" for the training dataset. Using "Raw data" can make computation and learning easier, and if the dataset is of higher quality, more accurate predictive information can be obtained.

[0086] Returning to Figure 3, the prediction unit 105 generates prediction information for the first mechanical properties using a trained model based on the input information generated by the processing unit 104. The function of the prediction unit 105 is realized by the execution of a program stored in the memory 12 by the processor 11.

[0087] The prediction unit 105 includes a pre-trained model. Preferably, the pre-trained model is based on deep learning. Examples of such models include CNNs (Convolutional Neural Networks), RNNs (Recurrent Neural Networks), and Transformers.

[0088] The input information provided to the trained model is generated from the time-dependent changes in parameters representing liquidity. Therefore, the input information has the characteristics of time-series data. When extracting features from time-series data, a model that can efficiently capture long-term dependencies is preferable. From this perspective, the prediction unit 105, which includes a trained model with an RNN and a Transformer as its basic structure, can generate more accurate prediction information. In particular, RNN-based models are more suitable for analyzing time-series data because they have the characteristic of being able to propagate information to the next step while retaining information from the previous step.

[0089] Bidirectional recurrent neural networks (BiRNNs) are an extension of RNNs that enables the propagation of information in the reverse direction. BiRNNs are well-suited architectures for learning both past and future dependencies in time-series data. According to BiRNNs, it is possible to efficiently capture features such as the rapid decrease in melt viscosity immediately after the start of processing, and the final state just before the processed material is moved to the next process (e.g., injection molding).

[0090] This trend becomes even more pronounced with bidirectional RNNs that incorporate a gate mechanism. Examples of RNNs with a gate mechanism include LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit). These are extensions of the RNN model.

[0091] A gate mechanism refers to a mechanism for controlling the selection (retention, forgetting) of information. For example, a GRU includes a Reset Gate and an Update Gate. Models with gate mechanisms exhibit superior performance in terms of learning long-term existing relationships and mitigating gradient vanishing. As mentioned above, the input information to the prediction unit 105 (trained model) has the characteristics of time-series data. Therefore, when extracting features from time-series data, using a bidirectional RNN-based model with a gate mechanism can yield even more accurate prediction information.

[0092] The prediction information generated by the prediction unit 105 includes the predicted results of the mechanical properties of the processed material. Mechanical properties are, in principle, obtained as a result of various mechanical property tests. Based on the input information, the prediction unit 105 outputs the predicted results of the mechanical properties of the processed material without performing mechanical property tests.

[0093] The first mechanical property is preferably at least one selected from the group consisting of tensile strength, elastic modulus, elongation at break, yield stress, flexural modulus, bending resistance, impact resistance, compressive strength, hardness, wear resistance, thermal expansion coefficient, impact resistance, toughness, and brittleness, preferably including at least one selected from the group consisting of tensile strength, elastic modulus, elongation at break, yield stress, impact resistance, toughness, and brittleness, and more preferably including two or more selected from the group consisting of tensile strength, elastic modulus, elongation at break, yield stress, impact resistance, toughness, and brittleness. The first mechanical property may be one of these or a combination of several. The form of the first mechanical property may be the physical property values ​​themselves, mechanical property curves related thereto (e.g., stress-strain curve, dynamic viscoelastic test curve, and thermal expansion curve, etc.), one feature quantity obtained from the mechanical property curve, and / or a combination of two or more feature quantities obtained from the mechanical property curve. Note that the above feature quantities may also include the physical property values ​​themselves. For example, the characteristic quantities obtained from a stress-strain curve may include tensile strength, elastic modulus, elongation at break, and yield stress.

[0094] The prediction information may include the results of the first mechanical properties test itself, which is used to obtain the first mechanical properties. For example, the prediction information may include predicted values ​​from the results of the mechanical properties test. The prediction information may also include the results of a judgment regarding the similarity between the mechanical properties curve obtained as a result of the first mechanical properties test and a reference curve. Furthermore, the prediction information may be classification information for the class of the results of the first mechanical properties test. These can be arbitrarily adjusted through annotation of the training dataset and model design, as described later. For example, if predicted values ​​are to be output as prediction information, a regression model should be selected, and for similarity judgment and class classification, a classification model should be selected. In particular, when the training dataset is large, a regression model can be used to generate more accurate prediction information. Also, when the training dataset is relatively small, a classification model may generate more accurate prediction information.

[0095] The mechanical properties of a processed material change under the influence of numerous factors, including the composition and state (definition of which will be described later) of the polymer material, stress history, and thermal history. Because these factors are intricately intertwined, it is difficult to predict the mechanical properties of a processed material with high accuracy using simulations or engineering models. However, if we limit ourselves to the mechanical properties of processed materials manufactured by heating, plasticization, and applied stress, these factors can be summarized in the Rh parameter, that is, the fluidity response of the polymer material to externally applied forces. Furthermore, the change in the Rh parameter over time reflects the history of its fluidity response. A data-driven approach using input information generated from the change in the Rh parameter over time can accurately predict the mechanical properties of a processed material.

[0096] The predictive information is not particularly limited, but preferably includes at least one piece of information selected from the group consisting of classification information of the results of a first mechanical property test of the processed object based on pattern distance (mathematical distance), and predicted values ​​of the results of the first mechanical property test. Of these, the classification information of the results of the mechanical property test is preferred from the viewpoint of making the application of the output data easier. In one form, it is preferable that the classification information is of patterns with high similarity in the results of the mechanical property test.

[0097] Classification of the results of the first mechanical property test based on pattern distance can be performed using various algorithms. Examples of such algorithms include k-nearest neighbors, support vector machines, neural networks (such as Siamese), and clustering algorithms. Among these, clustering algorithms are preferred because they allow clustering through unsupervised learning without labels, thus reducing the cost required for classification. Clustering algorithms may include k-means, hierarchical clustering, and DBSCAN (Density-based Spatial Clustering of Applications with Noise).

[0098] Among these methods, hierarchical clustering is preferable because it does not require the number of clusters to be determined in advance, and the grouping can be flexibly adjusted based on the hierarchical structure visualized in a dendrogram. Classification is performed in advance based on predetermined criteria. In this case, if hierarchical clustering is adopted as the classification algorithm, flexible adjustment becomes possible as described above.

[0099] Clustering may be performed based on user-specified features and / or automatically generated features. Such features may be obtained from the results of a first mechanical property test. User-specified features may include, for example, material properties related to at least one mechanical property selected from the group consisting of tensile strength, modulus of elasticity, elongation at break, yield stress, flexural modulus, bending resistance, impact resistance, compressive strength, hardness, wear resistance, and coefficient of thermal expansion. Among these, at least one value selected from the group consisting of tensile strength, modulus of elasticity, elongation at break, and yield stress is preferred in that it yields more accurate prediction results, and combinations of these are even more preferred. Automatically generated features may include various indices generated from the variation patterns of the mechanical property curve.

[0100] On the other hand, algorithms such as linear regression, ridge regression, lasso regression, support vector regression, decision tree regression, random forest regression, gradient boosting regression, kernel ridge regression, and neural network regression can be used to generate predicted values ​​for the results of the first mechanical property test. These can be selected based on the amount of data and linearity.

[0101] The learning unit 106 executes the machine learning process of the model of the prediction unit 105 based on the training dataset (first training dataset). The function of the learning unit 106 is realized by the execution of a program stored in memory 12 by the processor 11. The structure of the first training dataset will be described later, but typically it is a collection of data records in which classification information of the results of a first mechanical property test of the processed object is attached as a label to input information generated from the change over time of a parameter representing the fluidity of a polymer material. The learning unit 106 uses the first training dataset to train the model by appropriately combining supervised learning, semi-supervised learning, and transfer learning.

[0102] Furthermore, it is preferable that the learning unit 106 includes a function to perform transfer learning on the trained model and generate a second model. The second model generates prediction information for a second machine property based on the input information.

[0103] The second machine property may be the same as or different from the first machine property described above. Generally, the relationship between the second machine property and the input information is presumed to be different from the relationship between the first machine property and the input information. Therefore, the learning unit 106 performs transfer learning on the trained model using the second training dataset, which will be described later.

[0104] The method of transfer learning is not particularly limited, but one method is to fix the intermediate layers of the trained model and adjust the last few layers or the output layer, that is, to use the trained model as a feature extractor. The structure of the second training dataset for transfer learning may be the same as that of the first training dataset, except that the input information is labeled based on the second test results.

[0105] Transfer learning by the learning unit 106 can reduce the cost of retraining, and can generate highly accurate predictive information (predictive information for the second machine property) even when there is little new data.

[0106] The output unit 107 outputs output information from each unit and generated operation screens to the outside. The functions of the output unit 107 are realized when a program stored in memory 12 is executed by the processor 11 and the input / output interface 15 is controlled. The output information and operation screens are output to an external output device connected via the input / output interface 15.

[0107] The communication unit 108 is responsible for sending and receiving data between the prediction device 10 and the outside, and sending and receiving various commands and data between the prediction device 10 and the processing unit 20. The functions of the communication unit 108 are realized when a program stored in the memory 12 is executed by the processor 11, and the communication interface 13 and processing unit interface 14 are controlled.

[0108] Next, we will describe the data acquisition procedure for generating the first training dataset using processing system 1, and the model training procedure. Figure 4 is a flowchart showing the series of steps from data acquisition to model training. The procedure consists of steps S10 to S16, of which steps S10 to S15 can be considered the procedure for generating the first training dataset.

[0109] First, in step S10, stress is applied to the polymer material that has been plasticized by heating, and processing is started. Specifically, in one embodiment, the barrel 25 is heated by the heater 24, the motor 22 is driven, the screw 23B is started to rotate, and a predetermined amount of polymer material is sequentially fed into the barrel 25 from the feeder 27. At this time, the type and composition of polymer components contained in the polymer material do not need to be known. One data record that constitutes the training dataset is acquired over the processing process of one batch of a predetermined amount of polymer material fed into the processing device 20. The amount of polymer material provided as one batch is not particularly limited, and it does not need to be strictly the same between batches.

[0110] Next, in step S11, data on the time-dependent change of an index representing the motor drive parameters during the residence time is acquired. The residence time is measured cumulatively from the time the polymer material is introduced into the barrel 25 until the polymer material is discharged from the barrel 25 (corresponding to the fourth method described above). The time-dependent change of the index representing the motor drive parameters is acquired by the data acquisition unit 103. The values ​​acquired at this time may be the motor drive parameters themselves, such as torque and rotational speed, or they may be current values, voltage values, and power values, etc., which are indicators of these parameters. The acquired time-dependent change of the index representing the motor drive parameters may be stored in the storage unit 102 until it is used by the processing unit 104. Typically, data processing by the processing unit 104 is performed after all the data for the residence time has been acquired. Therefore, the time-dependent change of the index representing the motor drive parameters can be sequentially stored in the storage unit 102 until data for the entire residence time has been acquired.

[0111] Next, in step S12, the change over time of the parameters representing fluidity is calculated and predetermined data adjustments are performed. The processing unit 104 calculates the change over time of the parameters representing fluidity from the change over time of the index representing the motor drive parameters during the residence time, which is acquired by the data acquisition unit 103. The parameters representing fluidity and the calculation method are as described above. The parameters representing fluidity and the calculation method can be selected as appropriate and determined in advance, but the same calculation method is used for all data records that make up the first training dataset. In addition, data adjustments may include dimensionality reduction and augmentation. Data adjustments are performed by the processing unit 104, and the method is as described above.

[0112] Next, in step S13, the time-dependent changes in the parameters representing fluidity are converted into multi-channel data to generate input information. The processing unit 104 generates input information by multi-channeling the time-dependent changes in the Rh parameters adjusted in step S12. The method of multi-channeling is as described above.

[0113] Next, in step S14, a first mechanical properties test is performed on the obtained processed material, and the test results are clustered based on predetermined feature quantities. The clustering criteria may also be the shape or movement of the change in the Rh parameter over time. In summary, this can be described as clustering of the results of the first mechanical properties test based on pattern distance. This step is performed in order to assign labels in a later stage (step S15). Therefore, if labels are assigned based on the feature quantities extracted in a later stage, clustering in this step is unnecessary. In this case, one or more feature quantities are extracted manually or automatically from the results of the mechanical properties test, and these are used as labels.

[0114] Next, in step S15, labels (class labels) obtained from the clustering results are assigned to the corresponding input information to generate a first training dataset. The generation of the first training dataset may be performed by the user, or it may be performed by the processing unit 104 in conjunction with step S14 based on the input results of the first mechanical properties test.

[0115] The first training dataset consists of multiple data records. Each data record is labeled input information, i.e., the time-dependent changes in multi-channel Rh parameters, which are assigned class labels. The generated first training dataset is stored in the storage unit 102.

[0116] Next, in step S16, the machine learning process of the model is executed using the first training dataset. The training process is carried out by the training unit 106. After this step, a trained model is generated. The prediction unit 105 includes this trained model.

[0117] The procedure for generating prediction information using the prediction device 10, which includes the trained model generated as described above, will now be explained. Figure 5 is a flowchart of the prediction information generation using the prediction device 10.

[0118] Steps S10 to S13 are the same as steps S10 to S13 in Figure 4. The polymer material in question may be a pure polymer consisting of one type of polymer component, or it may contain two or more types of polymer components. Each of the polymer components may be a copolymer. The composition and state of the polymer material may be unknown. In this specification, the "composition" of the polymer material is defined by the types, number, and relative ratios of the polymer components contained. Secondary components such as additives, fillers, and other impurities are also considered part of the composition. Secondary components may include metal layers formed on the film-like polymer material. Metal layers are often placed to improve gas barrier properties or impart conductivity. Metal layers can be formed by lamination, spray coating, vapor deposition, and chemical deposition methods. Here, the "types" of polymer components refer to variations in the primary structure, molecular weight, molecular weight distribution, stereoregularity, and chemical modifications of one or more polymer components contained in the polymer material. In other words, polymer components that differ in any of these are defined as belonging to different "types." Furthermore, "states" include degradation states due to thermal history, stress history, oxidation, and hydrolysis, phase separation states including microphase separation structures, crystalline states, and molecular orientation states.

[0119] Polymer materials (raw materials) discharged and recovered from factories are generally sorted by "composition." However, their condition may differ. For example, even polymer materials with the same apparent "composition" may have different thermal histories, stress histories, etc., resulting in various variations in their degradation state, and these can be mixed together. For instance, even polymer materials recovered as "extruded polypropylene" can often end up in an unknown "state" different from that of new materials due to the mixing of polymer materials in different "states."

[0120] On the other hand, with household plastic waste, the accuracy of sorting can vary depending on the collection day and location. For example, even if collected as "polypropylene," the composition may be unknown due to the presence of different types of polymer components, such as polyethylene and polystyrene. Furthermore, in all cases, considering the inclusion of impurities, when using recycled plastic waste as raw material, both the composition and state are generally unknown. The prediction device of Example 1 is particularly useful for such polymer materials.

[0121] Examples of impurities include metal components, non-metallic inorganic components, and organic compounds other than polymers. Polymer materials may contain the above-mentioned secondary components as long as polymer components constitute the main component (preferably 50% or more by mass, more preferably 70% or more, and even more preferably 90% or more). In this specification, a polymer means a molecular structure composed of repeating monomer units of at least one type and having a molecular weight of 1000 or more.

[0122] The polymer components are not particularly limited, but examples include polyolefins, polyesters, polycarbonates, polystyrenes, polyurethanes, fluoropolymers, acrylic polymers, polyamides, polyimides, polyepoxides, polysiloxanes, vinyl polymers, polysulfones, polyacetals, and polyphenylene ethers.

[0123] Examples of polyolefins include polyethylene, polypropylene, and polybutene. Examples of polyesters include polyethylene terephthalate and polybutylene terephthalate. Examples of polystyrene include polystyrene.

[0124] Polyolefins, polystyrenes, and polyesters constitute a significant portion of household plastic waste. Polyolefins, in particular, are used in a wide variety of applications. Therefore, the performance and quality of the polyolefins themselves in household plastic waste vary considerably. Furthermore, when not completely separated from other polymer components, the performance and quality of the polymer materials as recycled raw materials exhibit a wide range of variations.

[0125] The general procedure for material recycling these polymer materials involves plasticizing the raw material (polymer material) by heating, applying stress over a residence time to knead, homogenize, and pelletize it. The prediction device 10 can process the data acquired as a result of this kneading and homogenization process into input information and use it. Therefore, when the polymer material contains at least one polymer component selected from the group consisting of polyolefins, polystyrene, and polyester, and furthermore, when this polymer material is a raw material for recycling, the prediction device 10 is excellent in that it can generate highly accurate prediction information. In particular, the prediction device 10 is more suitable for application to polymer materials mainly composed of polyolefins.

[0126] Returning to Figure 5, after the input information is generated in step S13, the prediction unit 105 provides the input information to the trained model and generates prediction information (step S20). The generated prediction information includes classification information for the results of the first mechanical properties test into classes, and / or predicted values ​​for the results of the first mechanical properties test.

[0127] Classification information obtained through clustering is easier for operators with less skill and experience to understand and use than predictive information, which is the result of the first mechanical property test itself (physical property values ​​and mechanical property curves). If names are assigned to each class in advance, it becomes easier to understand the suitable application of the processed material from the predictive information. This will be explained in detail in the demonstration test described later, but let's explain it with a concrete example. First, the modulus of elasticity, yield stress, and elongation at break from the tensile test results are selected as features, and clustering is performed based on these. Then, the three classes obtained from the hierarchical clustering results are named, for example, "Soft and Ductile," "Hard and Ductile," and "Hard and Brittle." When the predictive information includes the above-mentioned identification information set for each class, operators can more easily understand the general performance of the processed material. When the predictive information includes identification information set for each class (for example, naming for each class), it becomes easier to use the predictive information. Note that the same procedure can be used to generate predictive information that includes the predicted values ​​of the first mechanical property test.

[0128] As described above, the prediction device 10 and the processing system 1 equipped therewith make it possible to predict at least the first mechanical properties of the processed material without having to perform a full inspection, based on information on "time-dependent changes in parameters representing fluidity" that can be obtained as a result of processing the raw materials, and to extract high-quality processed material. In other words, the prediction information generated by the prediction device 10 can be used for sorting the processed material. One possible application is to change the collection destination of the processed material for each lot based on the prediction information. This sorting may be performed by an operator or by a known sorting device.

[0129] Next, we will describe an application example in which the residence time is not clearly defined, as the input of polymer material to the processing apparatus 20 is continuous and the discharge of processed material is continuous. Depending on the operation method of the processing system 1, the input of polymer material to the barrel 25 and the removal of processed material from the barrel 25 may be performed continuously. In this specification, such operation is referred to as continuous processing.

[0130] In continuous processing, it is preferable to estimate the true residence time of the polymer material through prior testing or simulation, and to set the "estimated residence time" as a predetermined time range based on the true residence time. Furthermore, it is preferable to use this estimated residence time in place of the "residence time" in Example 1. The estimated residence time is a time determined based on the (true) residence time. In this specification, the time determined based on the residence time may include the residence time and the estimated residence time.

[0131] Figure 6 is a flowchart illustrating the sequence of steps, from data acquisition to model training, in the case of continuous processing.

[0132] First, in step S30, the true residence time in the polymer material barrel 25 is estimated, and the estimated residence time is determined. The estimated residence time is preferably an approximation of the true residence time, but it may be a different value. In principle, the accuracy of the prediction information is improved if the estimated residence time matches the true residence time. On the other hand, a discrepancy between the estimated residence time and the true residence time does not hinder the generation of highly accurate prediction information. Even if there is a discrepancy between the estimated residence time and the true residence time, prediction information with sufficient accuracy for practical purposes can be generated. The method for determining the estimated residence time is not particularly limited, but preliminary tests and / or simulations can be used.

[0133] Next, in step S10, processing by the processing unit is initiated. Then, in step S31, data on the time-dependent change of an index representing the motor drive parameters within the estimated dwell time is acquired. This step is the same as in the flow of Figure 4, except that the dwell time in step S11 is changed to the estimated dwell time. The following steps S12 and S13 are also the same as in the flow of Figure 4.

[0134] Next, in step S32, a mechanical properties test is performed on the processed material discharged within the estimated residence time, and the test results are clustered based on predetermined features. In the next step S15, labels obtained from the clustering results are assigned to the input information to generate the first training dataset.

[0135] In the batch processing case of Example 1, when one batch of processed material corresponding to one batch of polymer material is removed from the processing device 20, a first mechanical properties test is performed on the processed material. This is then linked to input information generated from the change in Rh parameters over time. The same principle applies to continuous processing. Specifically, first, a first mechanical properties test is performed on the processed material discharged within a time corresponding to the estimated residence time (for convenience, referred to as "one unit"). Then, input information is generated from the change in Rh parameters over time for a time period preceding the estimated residence time from the end of discharge of the one unit of processed material. This input information is then linked to a label obtained from the clustering results, becoming one of the data records in the training dataset. In continuous processing, the calculation of Rh parameters can be performed continuously, which is different from batch processing in that input information must be generated from data preceding the estimated residence time from the end of the estimated residence time.

[0136] In the above example, both the time interval for acquiring time-series data and the time interval for the frequency of the first mechanical properties test are based on the estimated dwell time. However, the time interval for the frequency of the first mechanical properties test can be determined independently of the estimated dwell time. For example, the time interval for the frequency of the first mechanical properties test may be shorter than the estimated dwell time. For instance, if the estimated dwell time is 600 seconds, the time interval for the frequency of the first mechanical properties test may be 30 seconds. In other words, the first mechanical properties test may be performed on a sample taken every 30 seconds. In this case, input information can be generated every 30 seconds from the time-dependent change in the Rh parameter going back 600 seconds from that point, and linked to the class label of the first mechanical properties test every 30 seconds. A prediction device equipped with a model trained on the training dataset generated in this way can generate highly accurate prediction information in real time during continuous processing.

[0137] Next, we will describe the procedure for generating prediction information for a second mechanical properties test, which is different from the first mechanical properties test, using the prediction device 10. One method for this is to generate a new model (second model) using a new training dataset (second training dataset), and then use this model to generate prediction information. Below, we will describe the procedure for acquiring data to generate the second training dataset using the processing system 1, and the method for performing transfer learning of the trained model.

[0138] Figure 7 is a flowchart illustrating the series of steps, from data acquisition to transfer learning of the trained model. This flowchart assumes that machine learning using the first training dataset of the trained model has been completed using the procedure shown in Figure 4 or Figure 6.

[0139] This flowchart differs from those in Figures 4 and 6 in that it describes the procedure for obtaining a product from a processed material, and that the product undergoes a second mechanical properties test.

[0140] First, in step S40, a product is manufactured from the processed material. The manufacturing method is not particularly limited, but a method of applying stress to the processed material that has been plasticized by heating is preferred. In this case, a second processing apparatus can be used. The second processing apparatus may be a different apparatus from the first processing apparatus, or it may be the same apparatus. If the same processing apparatus is used, it may be operated under different conditions. The second processing apparatus may be the same apparatus as the first processing apparatus. The preferred form is also the same. The second processing apparatus may be configured to acquire the change over time of parameters representing fluidity from the processed material to the product. However, according to the prediction apparatus of this example, highly accurate prediction information can be generated without using the change over time of parameters representing fluidity from the processed material to the product.

[0141] Next, in step S41, a second mechanical properties test is performed on the obtained product, and the test results are clustered based on predetermined feature quantities. The second mechanical properties test may be different from or the same as the first mechanical properties test. The same test as the first mechanical properties test may be used for the second mechanical properties test. The same applies to preferred forms.

[0142] Next, in step S42, labels obtained from the clustering results are assigned to the input information to generate a second training dataset. This step is the same as step S15 in Figure 4, and the preferred form is also the same. The characteristic of this step is that labels based on the results of the second mechanical property test of the "product" are assigned to the input information generated from the time-dependent changes in parameters representing the fluidity of the "polymer material". Information regarding the manufacturing of the product from the processed material is not included in the input information.

[0143] Then, in step S43, a transfer learning process of the trained model is performed using a second training dataset. The transfer learning process can be performed by the learning unit 106. Typically, the intermediate layers are frozen, and the last few layers (or output layers) of the trained model are adjusted using the second training dataset. By repurposing the trained model as a feature extractor and applying the output layers to new data, the cost of retraining is reduced, and overfitting is suppressed even if the second training dataset is small, resulting in a transfer-trained model capable of generating highly accurate predictive information. In other words, this flow provides a prediction device including a prediction unit equipped with a transfer-trained model. Although the above flow is a batch process similar to that in Example 1, in this example as well, it can be applied to continuous processing using a flow similar to that in Figure 6.

[0144] Next, we will describe the procedure for generating prediction information using a prediction device equipped with a transfer-learned model. Figure 8 is a flowchart of the generation of prediction information using a prediction device equipped with a second model. In step S50, when input information is given to the second model, the second model generates prediction information for the second machine characteristic. The prediction device may also be equipped with a trained model and a second model. That is, after generating a trained model based on the relationship between the input information and the first machine characteristic, a model capable of generating prediction values ​​for other machine characteristics can be generated by transfer learning, thereby allowing it to be used as a prediction device for various types of machine characteristics.

[0145] The following describes the results of the demonstration experiment on the generation of predictive information using the prediction device described above.

[0146] (Experimental Example 1) • Sample preparation The samples used to generate the training dataset were plastic waste collected in Toyama Prefecture. This plastic waste originated from households. The plastic waste consisted of 23 different samples collected on different days between April and June 2023. These samples were pre-sorted by Toyama Environmental Services Co., Ltd. using density, magnetic, and optical methods. As a result, all samples had a polypropylene (PP) purity of 80%, with approximately 20% being other polymer components (polyethylene, polystyrene, etc.).

[0147] • Test method A small twin-screw extruder (Xplore MC 15HT; Xplore Instruments) was used as the processing unit. The barrel was heated to 220°C, and each sample was circulated in the barrel for 600 seconds while shear stress was applied to it by a screw rotating at a constant speed, before being removed. During the circulation in the barrel, the melt viscosity value at 600 seconds was obtained for each sample at 1-second intervals.

[0148] Furthermore, two dogbone-shaped test specimens for tensile testing were created from the extracted processed material by injection molding. These dogbone-shaped specimens were subjected to testing using a tensile testing machine (AG-10KNX; Shimadzu Corporation). From the resulting stress-strain curves, the mechanical properties (elastic modulus, yield stress, and fracture strain) of each sample were calculated. These results were correlated with the change in melt viscosity over time for each sample.

[0149] Clustering Hierarchical clustering was performed on 23 types of treated materials based on their extracted mechanical properties. Figure 9 shows a dendrogram representing the results of the hierarchical clustering. Based on this dendrogram, the 23 types of treated materials were cut off to classify them into three classes, and class labels were assigned to the corresponding changes in melt viscosity over time (melt viscosity curves).

[0150] Figure 10 shows the superimposed melt viscosity curves for each class. Figures 11 and 12 show the superimposed stress-strain curves for each class. All samples classified as Class 0 had low modulus of elasticity and high fracture strain. Based on these characteristics, the Class 0 samples were named "Soft and Ductile" as they were soft and ductile. All samples classified as Class 1 had high modulus of elasticity and high fracture strain. Based on these characteristics, the Class 1 samples were named "Hard and Ductile" as they were hard and ductile. Class 1 was defined as "high quality" as the performance generally required for recycled PP. Needless to say, the definition of "high quality" varies depending on the application of the processed material. In this example, Class 1 was defined as high quality, but depending on the application, Class 0 or Class 2 may also be considered "high quality".

[0151] All samples classified as Class 2 exhibited high elastic modulus and low fracture strain. Based on these characteristics, Class 2 samples were described as "Hard and Brittle," due to their hardness and low ductility.

[0152] • Generation of training datasets The melt viscosity curve consists of melt viscosity measurements at 600 time points. First, to simplify the model, this data was averaged every 20 seconds and compressed into 30-dimensional data. Next, this compressed data was standardized to have a mean of zero and a variance of 1. Similarly, the overall mean calculated for each sample was subtracted from this compressed data to generate zero-centered data.

[0153] From the compressed, standardized, and centered data obtained in this way, a 3-channel dataset with 30 time points, serving as a common reference axis, was acquired. Furthermore, to accommodate the extremely small dataset of 23 samples, the dataset was expanded tenfold by randomly adding Gaussian noise with a standard deviation of 0.1 to the melt viscosity curve. The training dataset includes multiple data records consisting of the 3-channel data and corresponding class labels based on tensile test results.

[0154] • Model training Figure 13 shows the structure of the model used. The input layer 200 consists of 3-channel data with the class labels described above. The recurrent layer 201 is a 16-channel bidirectional Gated Recurrent Unit (GRU) network implemented using the PyTorch package. The final bidirectional state is then input to the GRU output layer 202 and the fully connected (FC) layer 203, where the probability of belonging to each class is calculated using the softmax function, and the class with the highest probability is adopted. The learning process aimed to minimize the cross-entropy error, and the learning rate, a parameter that adjusts the progress of learning, was set to 0.01. The entire data was divided into small groups (minibatches) of 50 data points each for learning, and the entire dataset was trained 300 times.

[0155] • Model accuracy verification Leave-one-out cross-validation (LOOCV) was used to validate the model's accuracy. Of the 23 samples, one original melt viscosity curve and all the augmented data generated from it were excluded from the validation, treating it as an unknown class. Specifically, a predictive model was constructed by assigning class labels to 22 of the 23 melt viscosity curves, and the unknown melt viscosity curve was input to predict its class. This prediction was repeated 23 times to obtain the confusion matrix shown in Figure 14. The vertical axis represents the predicted label, and the vertical axis represents the true label. Class 0 and Class 1 achieved approximately 83% accuracy, Class 2 approximately 82%, and the overall classification accuracy was approximately 83%. Further accuracy improvements are expected with larger datasets.

[0156] (Experimental Example 2) • Verification of the effectiveness of multi-channeling The confusion matrix in Figure 14 showed the accuracy of the prediction information obtained based on 3-channel data consisting of compressed data, standardized data, and centered data. Therefore, under the same conditions as in Experimental Example 1, the accuracy and sensitivity of the prediction information were compared by changing the number of channels in the input information, and the effect of multi-channelization was verified. Note that the following tests were the same as in Experimental Example 1 except for changing the number of channels in the input information.

[0157] Figure 15 shows the confusion matrix of prediction information obtained based on raw data. Figure 16 shows the confusion matrix of prediction information obtained based on two-channel data consisting of raw data and centered (offset) data. Figure 17 shows the confusion matrix of prediction information obtained based on two-channel data consisting of raw data and standardized data. Figure 18 shows the confusion matrix of prediction information obtained based on two-channel data consisting of centered data and standardized data. Note that the raw data has been processed to have a mean of zero and a variance of 1 for the entire dataset as a data adjustment. Centering is performed by setting the time series mean of each sample to zero, and standardization is performed by setting the time series mean of zero and the variance of 1 for each sample.

[0158] Figure 19 shows the melt viscosity curves for each channel—Raw data, Mean Zero Offset data, and Standardized data—classified by category.

[0159] As is clear from Figure 19, the raw data retains absolute information about the melt viscosity curves between samples. The trend in melt viscosity can be described as Class 0 > Class 1 > Class 2. However, there is significant variation within each class, and considerable overlap between classes.

[0160] Looking at the corresponding confusion matrix (Figure 15), the prediction accuracy (precision; the ratio of samples predicted as Class 2 that were actually Class 2) and sensitivity (sensitivity; the ratio of samples that were actually Class 2 that were correctly predicted as Class 2) for Class 2 were 70%, which is within the practical range. On the other hand, the prediction accuracy for Class 0 and Class 1 was not necessarily high. This was thought to be due to the somewhat unusual shape of the melt viscosity curve for Class 2, as shown in the melt viscosity curve in Figure 19.

[0161] Centered (ZeroOffset) data loses absolute value information but retains the time-series difference. It is clear that the difference is Class 0 > Class 1 > Class 2. In the confusion matrix (Figure 16) when centered data is converted to 2 channels, there is absolutely no confusion between Class 0 and Class 1.

[0162] Standardized data is obtained by further scaling zero-offset data to have a variance of 1. In standardized data, the differences between samples in time series deviations are lost, but the characteristics of time series changes (curve shape) are emphasized. In particular, class 2 differs significantly from class 0 and 1. In the confusion matrix (Figure 17) when standardized data is converted to two channels, both the accuracy and sensitivity of class 2 are high.

[0163] In the confusion matrix (Figure 18) obtained when using a 2-channel setup with centered and standardized data without using raw data, relatively good accuracy and sensitivity can be obtained even without using raw data. However, it does not match the 3-channel input using raw data, indicating that the absolute value of melt viscosity contained in the raw data is also important information.

[0164] Based on the results above, it was found that multi-channel data with two or more channels is preferable to raw data, and in that case, it is even preferable to include at least standardized data. Furthermore, it was found that better prediction results can be obtained using 3-channel data consisting of raw data, centered data, and standardized data.

[0165] (Experimental Example 3) • Prediction of Charpy impact energy Next, we adopted "Charpy impact energy" as the second mechanical property and performed predictions based on it. The melt viscosity curve used in this prediction was obtained for the "treated material" in Experimental Example 1. The impact values ​​for each sample are measurement data from "test pieces" (products) molded using the treated material. This data was provided by Toyama Environmental Services Co., Ltd. Three classes, Class 0, 1, and 2, were assigned in descending order of impact value, and unknown labels were predicted using Leave-one-out cross-validation (LOOCV).

[0166] In this class labeling process, since the original data consists of scalar values ​​(tensile properties are 3D data), there is a clear hierarchy between classes (low, medium, and high impact values). Therefore, instead of one-hot coding the correct labels (for 3 classes, class 0 is [1,0,0], class 1 is [0,1,0], and class 2 is [0,0,1]), soft labeling was performed considering the distance between classes. By providing the correct labels as probabilities with a distribution (class 0: [0.75,0.2,0.05], class 1: [0.2,0.6,0.2], class 2: [0.05,0.2,0.75]), the distance between classes can be considered. The loss function used for training the model was defined by the 3-class probabilities output by the model and the KL divergence between the soft labels. The model structure itself is the same as in Experimental Example 1, only the way the correct labels are provided is different. Table 1 shows the results of the tensile test (Experimental Example 1), the results of the Charpy impact test (Experimental Example 3), the class classified based on the tensile test results (Tensile class: Experimental Example 1), and the class of impact value classified based on the Charpy impact test (Impact class: Experimental Example 3) for each sample.

[0167] [Table 1]

[0168] The impact test was performed five times. The standard deviation was 0.07 to 0.1 kJ / m 2 It exists to some extent. <3.9kJ / m 2 Class 0, 3.9~4.1 kJ / m 2 Class 1, >4.1kJ / m 2 Clustering that classifies the above as Class 2 is also an option. In Experimental Example 3, if this clustering is adopted, there will only be 3 samples each for Class 0 and Class 2, resulting in <3.95 kJ / m³. 2 Class 0, 3.95~4.0 kJ / m 2 Class 1, >4.0 kJ / m 2 This is labeled as Class 2. In this case, since there are samples that fall into different classes with similar values, particularly between Class 1 and Class 2, class prediction is difficult due to the nature of the data. That is, while a relative comparison between Figure 20 and Figure 21, which will be discussed later, is meaningful, it is not appropriate to judge practicality based solely on the results of Figure 20. In other words, the form of Figure 20 is also considered to have sufficient practicality.

[0169] In Experimental Example 3, the input information is the melt viscosity curve obtained when processing the raw material acquired in Experimental Example 1, and does not reflect the thermal history information experienced by the test piece (product) during its molding. Therefore, when attempting to directly predict the Charpy impact value class, there was room for improvement in prediction accuracy and sensitivity. Figure 20 shows the confusion matrix of predicted information obtained when machine learning of a new model was performed using a second training dataset in which class labels based on impact test results were added to the input information generated in Experimental Example 1. As shown in Figure 20, there was a characteristic of frequent misidentification between good products (Class 2) and poor products (Class 0).

[0170] On the other hand, Figure 21 shows the confusion matrix of predicted information obtained when, while using the same second training dataset, the feature extraction method was pre-trained from the trained model obtained in Experiment Example 1 and then transferred to the second model. In this case, Class 0 (fragile, poor-quality products) could be detected with a sensitivity of 7 / 10 (70%). Furthermore, the prediction accuracy for good products (Class 2) exhibiting high impact values ​​was 4 / 6 (66%), but this included only one poor-quality product, and sufficient accuracy was obtained for the purpose of avoiding downward leveling. [Explanation of Symbols]

[0171] 1 Processing system, 10 Prediction device, 20 Processing unit, 22 Motor, 23 Rotation mechanism, 23B Screw, 25 Barrel, 101 Control unit, 102 Storage unit, 103 Data acquisition unit, 104 Processing unit, 105 Prediction unit, 106 Learning unit, 107 Output unit, 108 Communication unit

Claims

1. A predictive device for predicting the first mechanical properties of a processed product in a processing apparatus that applies stress over a residence time to a polymer material plasticized by heating, thereby obtaining a processed product as a result, A prediction device comprising a trained model that generates prediction information for the first mechanical properties based on input information generated from the change over time of a parameter representing the fluidity of the polymer material within a time period determined based on the residence time.

2. The prediction device according to claim 1, wherein the input information is multi-channel data obtained by expanding the change in the parameter over time by applying a transformation that maintains the reference axis using at least one transformation method.

3. The prediction device according to claim 1, wherein the pre-trained model is a model whose basic structure is a recurrent neural network.

4. The prediction device according to claim 1, wherein at least one selected from the group consisting of the composition and state of the polymer material is unknown.

5. The prediction device according to claim 1, wherein the prediction information includes at least one type of information selected from the group consisting of classification information of the results of a first mechanical properties test of the processed object based on the distance of the pattern, and predicted values ​​of the results of the first mechanical properties test.

6. The prediction device according to claim 5, wherein the class is set in advance based on at least one feature obtained from the results of the first mechanical properties test.

7. The predictive apparatus according to claim 1, wherein the apparatus comprises a rotating mechanism driven by a motor, and applies shear stress as the stress to the polymer material contained in a container by the rotating mechanism.

8. The prediction device according to claim 7, wherein the change over time of the parameter representing fluidity is calculated based on the change over time of an index representing the drive parameter of the motor.

9. The prediction device according to claim 8, wherein the drive parameters include the torque of the motor and the rotational speed of the motor.

10. The prediction device according to claim 3, wherein the trained model is a model whose basic structure is a bidirectional recurrent neural network.

11. The aforementioned transformation method includes centralization and standardization. The prediction device according to claim 2, wherein the multi-channel data includes at least three: data before conversion, centered data, and standardized data.

12. Equipped with a learning department, The learning unit performs transfer learning on the trained model to generate a second model. The prediction device according to claim 1, wherein the second model generates prediction information of a second mechanical characteristic based on the input information.

13. A processing system comprising a prediction device according to any one of claims 1 to 12 and the processing device.

14. A prediction method for predicting the mechanical properties of a processed material obtained as a result of applying stress to a polymer material that has been plasticized by heating for a predetermined residence time, in an apparatus for which a processed material is obtained as a result, A prediction method comprising generating predictive information of the mechanical properties using a trained model based on input information generated from the time-dependent change in a parameter representing the fluidity of the polymer material within a time period determined based on the residence time.

15. On the computer, A program for executing a prediction method to predict the mechanical properties of a processed material in a processing apparatus that applies stress to a polymer material plasticized by heating for a predetermined residence time, thereby obtaining a processed material as a result, A program comprising the step of generating predictive information for the mechanical properties by a trained model based on input information generated from the time-dependent change in a parameter representing the fluidity of the polymer material within a time period determined based on the residence time.

Images