Internal process estimation system and internal process estimation method

Abstract

The sequence of unknown internal processes is estimated and visualized. [Solution] One aspect of the present invention is an internal process estimation system comprising a search unit, an actual measurement value input unit, and an output unit, wherein the search unit applies multiple types of element models to multiple element regions of a virtual model, which is divided into multiple element regions in a continuous unidirectional manner, in multiple sequences to create multiple virtual models that mimic the process of state transition from a first state to a second state of an actual device; the actual measurement value input unit receives actual measurement values ​​in the second state of the actual device; the search unit evaluates the virtual models based on the difference between predicted data output by inputting values ​​based on the first state into the multiple virtual models and the actual measurement values; and the output unit outputs the results evaluated by the search unit.

Description

[Technical Field]

[0001] The present invention relates to a technique for estimating internal processes of manufacturing equipment and the like. [Background Art]

[0002] Along with advances in mass production technology, inspection technology, and process informatics, machine learning-related technologies have come to be used to propose process treatment conditions for manufacturing facilities (such as rotary kilns, twin-screw extruders, and the like) that have low transparency of the internal state of the equipment and tend to depend on the experience of skilled workers. By using an AI model trained on knowledge of process treatment conditions accumulated in the past, it has become possible to reduce the number of trial experimental productions before mass production when developing processes for new materials and the like.

[0003] For example, Patent Document 1 discloses a process estimation device comprising: a regression model creation processing unit that performs machine learning on relationships between process data and creates a regression model representing correlation between these data; and a process estimation processing unit that estimates the process data to be estimated using the regression model created by the regression model creation processing unit.

[0004] Patent Document 2 discloses a method of optimizing a sequential combination process using a surrogate model. [Prior Art Documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Unexamined Patent Application Publication No. 2024-066339 [Patent Document 2] Japanese National Publication of International Patent Application No. 2007-534038 [Summary of the Invention] [Problem to be Solved by the Invention]

[0006] Process conditions proposed using machine learning models can be difficult for even experienced engineers to understand—why those conditions worked, or what might have been happening internally. It is almost always difficult to understand the true flow of the physical internal process that reproduces the results of a small amount of real-world experimental data, and even identifying possible internal process flows is challenging for those who are not experienced engineers.

[0007] If a simulator is available to simulate the process on a computer, the user deductively constructs a possible physical model in the assumed sequence (e.g., solid melts → liquid is heated → chemical reaction) and then performs the calculation. The accuracy of this calculation is often indirectly confirmed by the small difference between the calculation result and already obtained experimental data from the actual machine.

[0008] If there is a large difference between the calculation results and the actual experimental data, the calculation conditions, such as the physical model settings, are reviewed again, and the process of checking the difference between the calculation results and the actual experimental data under the next calculation conditions is repeated. However, in this process, the user needs to have a certain level of knowledge about the process in question and assume a reasonable candidate sequence of physical processes. Proposing a reasonable candidate sequence of physical processes is not easy for an unskilled engineer. Therefore, a technology that estimates and visualizes the sequence of unknown internal processes is desired. [Means for solving the problem]

[0009] One aspect of the present invention is an internal process estimation system comprising a search unit, an actual measurement value input unit, and an output unit, wherein the search unit applies multiple types of element models to multiple element regions of a virtual model, which is divided into multiple element regions in a continuous unidirectional manner, in multiple sequences to create multiple virtual models that mimic the process of state transition from a first state to a second state of an actual device; the actual measurement value input unit receives actual measurement values ​​in the second state of the actual device; the search unit evaluates the virtual models based on the difference between predicted data output by inputting values ​​based on the first state into the multiple virtual models and the actual measurement values; and the output unit outputs the results evaluated by the search unit.

[0010] Another aspect of the present invention is an internal process estimation method performed on an information processing apparatus comprising a search unit, an actual measurement value input unit, and an output unit, wherein the search unit applies multiple types of element models to multiple element regions of a virtual model, which is divided into multiple element regions in a continuous unidirectional manner, in multiple possible orders to create multiple virtual models that mimic the process of state transition from a first state to a second state of an actual device; the actual measurement value input unit inputs actual measurement values ​​of the second state of the actual device; the search unit evaluates the virtual models based on the difference between predicted data output by inputting values ​​based on the first state into the multiple virtual models and the actual measurement values; and the output unit outputs the results evaluated by the search unit. [Effects of the Invention]

[0011] It is possible to estimate and visualize the sequence of unknown internal processes. [Brief explanation of the drawing]

[0012] [Figure 1] This is a schematic cross-sectional view of a twin-screw extruder. [Figure 2] This is a conceptual diagram illustrating the surrogate model database in the example. [Figure 3A] This is a schematic diagram illustrating the input / output format of the surrogate model used in the example. [Figure 3B]It is a schematic diagram for explaining other input / output formats of the surrogate model used in the embodiment. [Figure 4] It is a block diagram for explaining the system configuration of the present invention. [Figure 5] It is an explanatory diagram for briefly outlining the embodiment. [Figure 6] It is an explanatory diagram for specifically outlining the embodiment. [Figure 7] It is a graph for explaining an example of the visualization method of the embodiment. [Figure 8] It is a flowchart showing the procedure from the start to outputting the estimated internal process permutation. [Figure 9] It is a block diagram of a hardware configuration example of an information processing apparatus. [Figure 10] It is a block diagram of programs and data stored in a storage apparatus. [Figure 11] It is a block diagram explaining the operation of the internal process estimation program. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present invention will be described with reference to the drawings. However, the present invention should not be construed as being limited to the description of the embodiments shown below. Those skilled in the art will readily understand that specific configurations of the present invention can be changed without departing from the spirit or scope of the present invention. In addition, the positions, sizes, shapes and the like of respective components shown in the drawings and the like in the present specification may not represent actual positions, sizes, shapes and the like, in order to facilitate understanding of the present invention. Therefore, the present invention is not limited to the positions, sizes, shapes and the like disclosed in the drawings and the like.

[0014] In the configuration of the embodiments described below, the same reference numerals are commonly used for the same parts or parts having similar functions across different drawings, and repeated description may be omitted.

[0015] When there are a plurality of elements having identical or similar functions, they may be described by adding different subscripts to the same reference numeral. However, when there is no need to distinguish between the plurality of elements, the description may be given with the subscripts omitted.

[0016] In the present specification and the like, notations such as "first", "second", and "third" are used to identify constituent elements, and do not necessarily limit the number, order, or content thereof. Further, numbers for identifying constituent elements are used for each context, and a number used in one context does not necessarily indicate the same configuration in another context. Furthermore, this does not preclude a constituent element identified by a certain number from also serving the function of a constituent element identified by another number.

[0017] The technology described in the embodiment can contribute to solving the problem that it is difficult to estimate the internal process sequence of a manufacturing apparatus. An example of the configuration of the embodiment is a permutation optimization calculation system that uses a virtual model in which a section from a raw material input port to a discharge port is continuously divided into a plurality of element regions in one direction, and an outline of the constituent elements is as follows. (1) A surrogate model database including a plurality of surrogate regression models (element models) assignable to each of the plurality of element regions. (2) An input unit that inputs measurement values of a substance at the discharge port. (3) A search unit that searches for a permutation of the regression models so as to minimize an evaluation function based on a difference between predicted data output by an integrated model obtained by serially connecting two or more of the regression models selected from the regression models included in the surrogate model database and the measured value of the substance. (4) An output unit that outputs a permutation optimization calculation result of the regression models calculated by the search unit.

[0018] This configuration allows for the display of plausible process sequence candidates for limited real-world experimental data by utilizing various surrogate models trained on simulation data as elemental models in processing processes where simulation costs are relatively lower than the cost of acquiring actual experimental data. Specifically, it comprehensively examines permutations of surrogate models and outputs a ranking of the combinations of models that best explain the limited experimental data. This allows users to estimate the order of the processing unit's internal processes from the trends of the top-ranked model permutations. [Examples]

[0019] In Example 1, a twin-screw extruder is used as a manufacturing apparatus for plastic materials. We consider a scenario in which we can present candidate physical processes that led to the extrusion of the resin inside the apparatus under operating conditions used to optimize the quality of the extruded plastic resin.

[0020] Figure 1 is a schematic diagram of a cross-section of a twin-screw extruder, viewed from the side. In a twin-screw extruder 100, it is difficult to accurately measure the state of the resin being processed inside. One of the main reasons for this is that the inside of the twin-screw extruder 100 is sealed, which limits visibility. Furthermore, because the work is carried out under high temperature and high pressure conditions, and because it involves long-term continuous operation as a mass production facility, it is difficult to place measuring devices inside, and even if measuring devices are installed, there are known concerns about the accuracy of the measurement data. For this reason, a thermometer 101 and a pressure gauge 102 are usually installed near the discharge port of the twin-screw extruder 100, but measuring devices for the purpose of directly observing the internal process of the device are often not installed.

[0021] In a twin-screw extruder 100, it is necessary to understand process parameters that are important for the quality of the manufactured product, such as material heating, discharge rate per unit time (throughput), and screw configuration. However, because there are limited means of directly measuring the internal state, it is difficult to obtain this information and to obtain guidelines for quality improvement. Therefore, the development of technologies for optimizing operating condition parameters using machine learning techniques is progressing.

[0022] Furthermore, since twin-screw extruders are generally large-scale mass production equipment, even small ones consume several kilograms of raw material per hour, and medium to large ones consume several hundred kilograms per hour. Therefore, it is not easy to carefully investigate what the important process parameters are using single-factor testing methods.

[0023] Operating parameter settings optimized using machine learning models are often difficult for experienced twin-screw extruder technicians to understand. This is because, during the training of machine learning models, they can discover complex patterns and correlations from the training data that are difficult for humans to grasp intuitively. Therefore, when problems occur with the manufacturing equipment due to the operating parameter settings proposed by the machine learning model, or when users judge that good results have been obtained by chance, it is difficult to obtain guidance on what modifications to the operating parameters should be made next.

[0024] Let's consider the specific physical processes involved in the process of feeding raw materials into a twin-screw extruder 100 and performing resin mixing, specifically how the extruded material is discharged through the internal machinery. First, the raw material 103 is fed into the twin-screw extruder 100 in solid form from the upstream inlet. This solid raw material is transported by the conveyor screw 104 to the molten zone 105 where the cylinder is heated, and there it melts into a fluid molten resin.

[0025] The molten resin is continuously heated and transported in the cylinder, eventually reaching the mixing zone 106. In the mixing zone 106, the molten resin is subjected to strong shear stress, causing chemical changes and an increase in its viscosity. After that, it is again passed under the degassing port 107 by the transport screw 104, removing gaseous components from the molten resin as it reaches the die 108.

[0026] Because the inside of die 108 is a narrow channel, the molten resin is discharged from the discharge port 109 under strong pressure from the resin accumulated just before die 108. During discharge, the temperature and pressure of the discharged resin are constantly measured by thermometer 101 and pressure gauge 102.

[0027] However, unlike the temperature and pressure at the discharge port, the physical processes occurring inside the extruder, as described above, cannot always be directly observed. Instead, they are expected to some extent by the user, a skilled technician, based on the extruder's screw configuration and cylinder temperature settings. Because many items cannot be actually observed, it is not easy to answer questions such as, "Has the solid raw material not fully melted before reaching the mixing zone due to insufficient heating or some other reason?" or "Has degassing already begun in the resin portion within the molten zone due to a low internal filling rate?" This embodiment provides a reasonably reliable answer to these questions. We consider applying this embodiment to understand the internal processes of the twin-screw extruder 100.

[0028] Figure 2 illustrates the management of surrogate models included in the surrogate model database 1 applied to the embodiment. The surrogate model database 1 stores surrogate models that have been trained on simulation data sets obtained by computational methods such as the 3D finite element method. A surrogate model is one of the methods that replaces physical simulations with machine learning, and its general creation method is known.

[0029] For example, twin-screw extruders handle fluids and fine powders, and numerical calculation of the equations of motion can be used to simulate the behavior of fluids and fine powders within a twin-screw extruder. Numerical calculations to accurately reproduce the movement of fluids and fine powders take time, but by training a surrogate model using the simulation results as training data, it is possible to generate a surrogate model as an estimation model that can quickly replace the simulation.

[0030] While neural network models are generally assumed as surrogate models, other models that can describe the relationships between values ​​analytically, such as regression models (e.g., multiple regression models) or models that can be expressed using known analytical physical formulas, are also acceptable. Furthermore, although surrogate models were used as an example in this embodiment, other types of models that can rapidly simulate physical phenomena can be used. In addition, while simulation data is advantageous and commonly used as training data for surrogate models from a cost perspective, experimental data or other methods can also be used.

[0031] The surrogate model database 1 manages various surrogate models 7 from two perspectives: physical process element name and input / output format. In this embodiment, four physical processes are candidates: liquid heating process, solid heating and melting process, degassing process, and chemical change process, and these four types were managed. Figure 2 assumes two types of input formats, input format A and input format B. Both the number of physical processes and input formats may be greater or less than the number in this example.

[0032] As is well known, surrogate models are models that replace various physical simulations. Physical simulations offer a high degree of flexibility, as the underlying conditions and parameters can be freely set, but they are computationally intensive and difficult to speed up. Therefore, machine learning, using the inputs and outputs of physical simulations as training data, generates inference models that replace physical simulations.

[0033] The surrogate model 7 may be a generalized model applicable to multiple types of instruments. Alternatively, a model specific to a particular instrument may be created to enable high-precision analysis (a so-called custom-made model).

[0034] For example, considering a twin-screw extruder as shown in Figure 1, there are various screw shapes. Specifically, there are shapes such as full-flight screws, kneading screws, and mixing screws, and surrogate models are prepared based on separate simulations. Kneading screws are further distinguished into three types (forward kneading FK, neutral kneading NK, and back kneading BK), so it is conceivable to prepare five types of surrogate models. Furthermore, the extent to which differences in shape are pursued and differentiated, such as by screw thread pitch (e.g., 5mm and 10mm), becomes a trade-off between accuracy and cost. In the example in Figure 2, each physical model section for each input / output type contains two models corresponding to full-flight screws and kneading screws.

[0035] In the example in Figure 2, four physical processes are modeled, but the types of models are not limited to these. The physical simulations on which the models are based offer a high degree of flexibility, so they can be used as surrogate models as needed according to the user's purpose and constraints. For example, there can be various classifications within the physical process of "chemical change." These include molecular weight changes due to polymerization of a single chemical species, chemical reactions between two chemical species, and degradation and decomposition of resins due to high temperatures. The appropriate physical process model should be chosen according to the equipment and materials assumed by the user. The classification can also be divided into categories such as "heating," "melting," "degassing," "polymerization," and "decomposition." Furthermore, different models can be prepared depending on the accuracy of the surrogate model and the training data.

[0036] In a custom-made surrogate model, for example, in the case of the twin-screw extruder 100 shown in Figure 1, if the flow path length of the device is L and the number of element domains is N, then each surrogate model should be configured to perform a physical simulation of the material moving over a distance of L / N. An element domain is the area in which one surrogate model substitutes for a physical phenomenon. In this case, the number of N is determined by the model design, so for example, if you want to halve N, you can use two identical models connected in series as one unit and apply it to the element domain.

[0037] Furthermore, in the case of a twin-screw extruder, the equipment specifications are largely influenced by the cylinder diameter of the flow path and the screw shape. The screw shape is as described above. The size of the cylinder diameter is limited and is generally the outer diameter of the screw plus a predetermined value, so the screw shape (and diameter) is important when considering variations in the surrogate model.

[0038] In the case of a twin-screw extruder, the screw length is almost always expressed as L / D. L is the screw length, and D is the screw diameter. When using this system, it is preferable that N corresponds to the number of screw pieces, or a multiple thereof. For example, if all screw pieces have a length of L / D=1, then the number of screw pieces that can be put into an L / D=40 machine will be 40. In that case, when operating this system, it is conceivable to set N=40 or N=80.

[0039] Figures 3A and 3B schematically illustrate two types of input formats. There are broadly two types of input / output formats. One is the pattern where the output and input formats of the surrogate model 7 are the same, as shown in Figure 3A. That is, the input material state variable 11 and the output material state variable 12 are the same.

[0040] Another pattern, as shown in Figure 3B, is an input format that includes all the output formats of the surrogate model 7, and in addition, accepts operating conditions D(x) that depend on the position x in which the surrogate model is placed as part of the input. In other words, the input material state variable 11 is the output material state variable 12 plus the operating parameters 13.

[0041] Examples of input and output material state quantities 11 and 12 include physical quantities such as temperature, pressure, viscosity, flow velocity, temperature distribution (dispersion), or the particle shape of the second phase (the phase with the smaller area) when the material is separated into multiple phases, or characteristic quantities based on these physical quantities. These are just examples, and other physical quantities may also be used.

[0042] In the twin-screw extruder 100, examples of operating conditions D that depend on position x include the heater set temperature at that cylinder position and the screw configuration. When using operating conditions D as input, it is necessary to prepare prior simulation data corresponding to possible condition levels for screw type and heater set temperature, and to have a surrogate model database trained using this data as training data.

[0043] In both Figure 3A and Figure 3B, all values ​​on the output side also exist on the input side. Specifically, the features that are the output of each type of element model include all the features that are the inputs of each type of element model. By imposing this constraint on the surrogate models, the positions of the surrogate models arranged in series become commutative, enabling permutation optimization calculations by the search unit.

[0044] Figure 4 shows a functional block diagram of the internal process estimation system in the embodiment. The internal process estimation system 40 can be configured with a general information processing device, as will be described later.

[0045] Based on the results of actual machine experiments using the twin-screw extruder 100, the true values ​​(actual measured values) obtained by the thermometer 101 and pressure gauge 102 are input to the actual measurement value input unit 2. While it is acceptable to use only one of these two values, it is preferable to have both, and additional values ​​obtained by other measurement means may also be added.

[0046] The search unit 3 uses input from the actual measurement input unit 2 to search for the order of surrogate models 7 within the virtual model 5 in order to estimate the processes occurring inside the twin-screw extruder 100. For this purpose, it searches for permutations of surrogate models 7 that minimize an evaluation function based on the difference with actual measurement values.

[0047] The search constraint input unit 4 is an optional function that can reduce the amount of search performed by the search unit 3. In the virtual model 5, element regions 500, where each surrogate model is placed, are arranged in series. If the number N of element regions 500 in the virtual model 5 is large, and there are many types of physical process element names, the search unit 3 will incur a heavy computational load on its exhaustive search. Therefore, by defining user-defined constraints in the search constraint input unit 4, the calculation of unnecessary model combination permutations can be reduced.

[0048] In this embodiment, one example of a user-defined constraint is inputting the knowledge that "no chemical change occurs before the melting of a solid" into the search constraint input unit 4. This skips the calculation of virtual models for surrogate model combination permutations where the chemical change process is upstream of the solid's heating and melting process. As a result, if the search algorithm performed by the search unit 3 is a brute-force search, the computational complexity can be reduced by half.

[0049] User-defined constraints can be freely set by the user empirically or by referring to literature, and it is preferable that they be input from outside the internal process estimation system 40 via the search constraint input unit 4 as appropriate.

[0050] The search unit 3 performs calculations to construct a virtual model by connecting surrogate models 7 in series. The calculations of this virtual model predict the pressure and temperature when the raw material fed into the twin-screw extruder 100 from the upstream side is discharged from the outlet at room temperature.

[0051] Figure 5, for the sake of simplicity, shows the predicted pressure and temperature values ​​for each of the two types of virtual models, with the number of element domains N=4. In reality, in addition to the two combinations shown in (1) and (2) in Figure 5, calculations are performed on virtual models that include all physical process element surrogates in all four element domains, as long as they do not violate user-defined constraints. In other words, in the example in Figure 5, a user-defined constraint is imposed that "no chemical reactions occur before melting," which reduces the computational complexity to half of the 4^4 possible virtual models when there is no constraint, resulting in 4^4 / 2 = 128 possible virtual models. Here, as a combination of surrogate models, two virtual models are shown in which the order of combination of chemical change and degassing is reversed.

[0052] We compare the predicted values ​​of the two virtual models with the measured values. In virtual model (1), where the chemical change is coupled upstream, both the predicted pressure and temperature values ​​are closer to the measured pressure and temperature values ​​input from the measured value input unit 2.

[0053] It is believed that setting a larger value for the number of element regions N increases the process resolution. Furthermore, appropriately setting user-defined constraints accordingly reduces the computational complexity of the search unit 3. The search unit 3 sorts all calculation results in order of proximity between predicted and actual values. If there are multiple predicted values, the values ​​can be standardized across all calculation results to align the scale, or they can be multiplied by a user-defined coefficient, summed, and then sorted.

[0054] Figure 6 is a schematic diagram showing the results of sorting the predicted values ​​calculated by the search unit 3 in the case of N=24, starting with those closest to the measured values. When N is a large value, many virtual models return predicted values ​​that are almost identical to the measured values. This is because the degrees of freedom of the combination permutations of surrogate models are very high, increasing the likelihood of permutations appearing whose predicted results happen to be close to the measured values.

[0055] Figure 7 shows the search and sorting results from the search unit 3, visualized as a histogram by the output unit 6, showing the frequency of surrogate occurrences at each position according to a threshold (e.g., top 1%) provided by the user. From this histogram, the user can infer the following process.

[0056] When creating a histogram, in addition to the user specifying the percentiles to be accumulated as described above, other options include weighting higher-ranking data more heavily, specifying a tolerance for deviation from measured values, and accumulating data within a specified range. These options can be selected according to the device configuration and purpose.

[0057] In the example shown in Figure 7, the input raw material is first heated and melted while still in solid form, and the heating continues even after it becomes liquid. Subsequently, a chemical reaction occurs, leading to degassing. During this process, there is a certain period in which solid and liquid states coexist, and from around this point, degassing actually begins continuously in parallel with all other processes, and degassing is constantly occurring from the start to the end of the chemical reaction. Furthermore, the system remains in a liquid state, and even after the chemical reaction, the heating continues slightly in the liquid state.

[0058] In the examples shown in Figures 6 and 7, the constraints set in the search constraint input unit 4 are used to impose a constraint that prevents the chemical surrogate model from appearing before the solid melting surrogate model, thereby reducing the computational complexity.

[0059] In this embodiment, it is possible to estimate the phenomena occurring within the device and visualize the process. For example, by checking the difference in process sequence between when the desired result is obtained and when the desired result is not obtained in the actual device, it is possible to estimate the reason why the desired result is not obtained.

[0060] Figure 8 is a flowchart showing the steps from the start to outputting the estimated internal process permutations. Below, the operating principle of the system that outputs the estimated internal process permutations will be explained following the flowchart in Figure 8.

[0061] First, the number of element regions N of the virtual model is determined (A1). This can be a value specified by the user, or a value registered in one of the storage devices as an initial value. The value specified by the user is entered via the input device 44, which will be described later.

[0062] Subsequently, the target actual measured value is determined from the actual measured value input unit 2 (A2). The actual measured value is, for example, in the case of the twin-screw extruder shown in Figure 1, the value measured by the thermometer 101 and pressure gauge 102 when a predetermined raw material 103 is fed into the apparatus under predetermined conditions and the raw material 103 becomes molten resin and is discharged from the discharge port 109. The actual measured value can be measured using a known method with an actual machine. There can be any number of types of actual measured values, but they must be physical quantities that can be output by the surrogate model 7.

[0063] Subsequently, the search unit 3 refers to the surrogate model database 1 to confirm the type of physical process surrogate to be used in the permutation search (A3). This requires selecting a physical process that the user can understand.

[0064] Subsequently, user-defined search constraints are determined from the search constraint input unit 4 (A4). This is to reduce computation time, and it is acceptable to leave it blank. At this time, the type of surrogate model used in the search unit 3 may be limited.

[0065] Subsequently, the search unit 3 retrieves surrogate models 7 from the surrogate model database 1 and performs permutation and combination calculations (A5). The search unit 3 creates multiple virtual models by fitting multiple types of surrogate models 7 to multiple element domains in multiple possible orders according to the constraints. Then, for the multiple virtual models, in the case of the twin-screw extruder in Figure 1, the input conditions of the raw material 103 when actual measured values ​​were obtained are used as input, and an output (estimated value) is obtained. The input conditions are, for example, temperature and pressure.

[0066] The search unit 3 compares the actual measured value with multiple estimated values ​​and performs an evaluation as shown in Figure 6. If the number of types of surrogate models corresponding to the physical process or N is large, a search algorithm other than exhaustive search may be used, such as Bayesian optimization using user settings or the results of the previous processing as initial values, without trying all combinations.

[0067] Subsequently, the output unit 6 (output device 45, described later) outputs the search results and visualization results (A6). For example, a histogram image like the one shown in Figure 7 is displayed on the screen of the output device 45. Note that the order of A1 to A4 is arbitrary as long as it is before A5.

[0068] Figure 9 is a block diagram showing the hardware configuration of the internal process estimation system 40. It is implemented by an information processing device that primarily includes a processor (CPU) 41, memory 42, storage device 43, input device 44, output device 45, communication device 46, and bus 47, as shown in Figure 9.

[0069] The processor 41 functions as a functional unit (functional block) that provides predetermined functions by executing processing according to a program loaded into memory 42. The storage device 43 stores the program that enables the functional unit to function, as well as data used by the functional unit. The storage device 43 uses a non-volatile storage medium such as an HDD (Hard Disk Drive) or SSD (Solid State Drive). The input device 44 is a keyboard, pointing device, etc., and the output device 45 is a display, etc. The communication device 46 enables communication with other information processing devices via a network. These are connected to each other via a bus 47.

[0070] In this embodiment, functions equivalent to those configured in software can also be implemented using hardware such as FPGAs (Field Programmable Gate Arrays), ASICs (Application Specific Integrated Circuits), and GPUs (Graphics Processing Units). Such embodiments are also included within the scope of this embodiment.

[0071] Furthermore, the internal process estimation system 40 does not need to be implemented on a single information processing device; it may be implemented on multiple information processing devices. Also, some or all of the functions of the internal process estimation system may be implemented as a cloud-based application.

[0072] Figure 10 shows the programs and data stored in the storage device 43. The internal process estimation program 51 is loaded into memory 42 and executed by the processor 41, thereby enabling the processor 41 to function as the search unit 3. The internal process estimation program 51 includes a database (DB) search program 52, a search / optimization program 53, and a virtual model calculation program 54 as subprograms. These subprograms are also loaded into memory 42 and executed by the processor 41, enabling the processor 41 to function as the DB search unit, the search / optimization unit, and the virtual model calculation unit. The surrogate model database 1 used by the internal process estimation system is also stored in the storage device 43.

[0073] Figure 11 illustrates the operation of a subprogram of the internal process estimation program 51. The DB search program 52 is a program that specifies the surrogate model to be used by referring to the surrogate model database 1. Specifically, it has the function of displaying the contents of the surrogate model database 1 on the output device 45 and allowing the user to specify the surrogate model to be used from the input device 44. At this time, the conditions and definition constraints of the surrogate model to be applied may also be specified from the input device 44.

[0074] The search / optimization program 53 is a program that enumerates candidate combinations of surrogate models (i.e., virtual models that do not violate user-defined constraints) and aggregates and visualizes the calculation results. The search / optimization program 53 can be equipped with the function to search for the virtual model that best represents the internal processes of the actual device from among multiple virtual models and to optimize the virtual model.

[0075] The virtual model calculation program 54 is a program that causes the virtual model created by the search / optimization program 53 to perform inference calculations. The inference calculations can be performed independently and in parallel with the calculations of the DB search program 52 and the search / optimization program 53.

[0076] Surrogate models are primarily intended to be constructed using deep learning modeling such as neural networks. However, even performing inference (i.e., obtaining an output for a given input) with just one type of model can take several seconds to tens of seconds depending on the size of the model. For this reason, as shown in Figure 10, it is preferable to alleviate the execution time bottleneck of the overall system (internal process estimation program 51) by simultaneously executing multiple virtual model calculation programs 54 in parallel.

[0077] Furthermore, in this embodiment, increasing the number of element regions N improves the process resolution, but increases the processing time for search / optimization. One method to shorten the processing time is to apply constraints as described above. Another method is to randomly perform virtual model calculations, display a ranking as shown in Figure 6 on the output device 45 in real time or at predetermined intervals, and when the ranking stabilizes to a certain extent, the user can decide that optimization has been achieved and instruct the input device 44 to stop the calculation. Alternatively, the system could be configured to automatically stop under predetermined conditions on behalf of the user.

[0078] Alternatively, instead of randomly executing virtual model calculations, it is possible to use the search / optimization program 53 to create a list of calculation candidates in advance using experimental design (DoE) or similar methods, ensuring that the virtual models are unique and not exhaustive, and only include a limited number of possibilities, before instructing the virtual model calculation program 54 to perform the calculations.

[0079] Another concrete example involves determining a list of candidate virtual models in DoE as the initial value, and then checking whether the accuracy of the virtual model improves by slightly (for example, by one) swapping surrogate models from the initial virtual model permutation. If it improves, the initial value is swapped again, and this process is repeated until the virtual model permutation analysis progresses to the point where swapping by one does not improve the accuracy of the virtual model. At that point, it can be said that a local optimum has been found from the initial value. The search / optimization program 53 may be configured to automatically execute this process. [Examples]

[0080] Although the above explanation used a twin-screw extruder as an example, the embodiments are not limited to twin-screw extruders and can be applied to other devices. Other types of extruders, such as single-screw extruders, or even completely different devices, are also acceptable. For example, in sake production, saccharification by koji mold and fermentation by yeast occur simultaneously. Since it is difficult to measure the physical phenomena inside the brewing tank from the outside, the simulation of the physical phenomena in the target device can be replaced with a surrogate model, and the sequence of the internal processes can be estimated using the same method as in the embodiments described above.

[0081] As demonstrated above, it is possible to estimate the processes occurring inside the device without conducting experiments with an actual machine. This results in lower energy consumption, reduced carbon emissions, prevention of global warming, and contribution to the realization of a sustainable society. [Explanation of Symbols]

[0082] 1: Surrogate model database, 2: Actual measurement input unit, 3: Search unit, 4: Search constraint input unit, 5: Virtual model, 6: Output unit, 7: Surrogate model, 11: Input material state quantities, 12: Output material state quantities, 13: Operating condition D, 40: Internal process estimation system, 41: Processor (CPU), 42: Memory, 43: Storage device, 44: Input device, 45: Output device, 46: Communication device, 47: Bus, 51: Internal process estimation program, 52: Database search program, 53: Search / optimization program, 54: Virtual model calculation program, 101: Thermometer, 102: Pressure gauge, 103: Raw material, 104: Conveyor screw, 105: Melting zone, 106: Mixing zone, 107: Degassing port, 108: Die, 109: Discharge port

Claims

1. It comprises a search unit, a measured value input unit, and an output unit. The search unit, Multiple types of element models are applied in multiple sequences to multiple element regions of a virtual model, which is divided into multiple element regions in a continuous unidirectional manner, thereby creating multiple virtual models that mimic the state transition process from the first state to the second state of an actual device. The aforementioned actual measured value input unit is, The actual measured values ​​in the second state of the actual device are input, The search unit, Based on the difference between the predicted data output by inputting values ​​based on the first state into the plurality of virtual models and the actual measured values, the virtual models are evaluated. The output section is, The search unit outputs the results of its evaluation. Internal process estimation system.

2. An internal process estimation system according to claim 1, The aforementioned element model is an estimation model that replaces the simulation of a physical process, and is an internal process estimation system.

3. An internal process estimation system according to claim 1, An internal process estimation system characterized in that the output of each type of element model of the aforementioned multiple types of element models includes all of the inputs of each type of element model.

4. An internal process estimation system according to claim 1, The search unit is characterized by ranking the plurality of virtual models in order of the smallest difference between the predicted data and the actual measured values.

5. An internal process estimation system according to claim 1, It has a search constraint input section that accepts user-defined constraints, The search unit is an internal process estimation system characterized by fitting the multiple types of element models to the multiple element regions under the constraints of the user-defined constraints.

6. An internal process estimation system according to claim 1, The internal process estimation system is characterized in that the element model is at least one selected from a surrogate model constructed by machine learning and a regression model capable of analytically describing the relationships between values.

7. An internal process estimation system according to claim 1, The aforementioned element model is an internal process estimation system characterized by simulating at least one physical process selected from four physical processes: a liquid heating process, a solid heating and melting process, a degassing process, and a chemical change process.

8. An internal process estimation system according to claim 1, The aforementioned element model is a surrogate model that mimics the process of state transition of at least one substance selected from fluids and fine powders. An internal process estimation system characterized in that the training data used by the surrogate model is the result of numerical calculation of the equations of motion.

9. An internal process estimation system according to claim 1, The output unit is characterized by outputting, in histogram format, the order of element models fitted to multiple element regions of a plurality of virtual models, for a plurality of virtual models whose difference from the predicted data and the actual measured value satisfies predetermined conditions.

10. An internal process estimation system according to claim 1, The actual apparatus is an extruder, the first state is the state of the substance at the inlet of the extruder, the second state is the state of the substance at the discharge port of the extruder, and the actual measured value is at least one value of the pressure and temperature of the substance, characterized in that it is an internal process estimation system.

11. An internal process estimation system according to claim 1, The search unit includes a plurality of virtual model calculation units for inputting and outputting values ​​based on the first state to the plurality of virtual models, and the internal process estimation system is characterized by operating the plurality of virtual model calculation units in parallel.

12. This is executed in an information processing device that includes a search unit, an actual measurement value input unit, and an output unit. The search unit, Multiple types of element models are applied in multiple sequences to multiple element regions of a virtual model, which is divided into multiple element regions in a continuous unidirectional manner, thereby creating multiple virtual models that mimic the state transition process from the first state to the second state of an actual device. The aforementioned actual measurement value input unit, The actual measured values ​​in the second state of the actual device are input, The search unit, Based on the difference between the predicted data output by inputting values ​​based on the first state into the plurality of virtual models and the actual measured values, the virtual models are evaluated. The output unit, The search unit outputs the results of its evaluation. Internal process estimation method.

13. An internal process estimation method according to claim 12, The internal process estimation method is characterized in that the element model is an estimation model that replaces the simulation of a physical process.

14. An internal process estimation method according to claim 12, An internal process estimation method characterized in that the output of each type of element model of the aforementioned multiple types of element models includes all of the inputs of each type of element model.

15. An internal process estimation method according to claim 12, The internal process estimation method is characterized in that the element model is at least one selected from a surrogate model constructed by machine learning and a regression model capable of analytically describing the relationships between values.

Images