Optimization support device, method, and program
By employing a two-step conversion process to distinguish between operating condition and state parameters, the optimization support device enhances prediction accuracy in product production processes, addressing overtraining issues and ensuring products meet quality standards.
Patent Information
- Application Number
- JP2021529910
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-07-03
- Filing Date
- 2020-05-19
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2040-05-19
AI Technical Summary
Existing prediction models, particularly neural networks, struggle with accurate predictions when faced with a large number of process condition parameters in product production processes due to overtraining, leading to inaccurate quality parameter predictions for inputs outside the training data and complex function or table structures.
The optimization support device employs a two-step conversion process using learning models to convert operating condition parameters into state parameters and then quality parameters, or vice versa, reducing the number of input parameters and preventing overfitting by distinguishing between operating condition parameters and state parameters.
This approach allows for accurate prediction of various conditions in product production processes, ensuring high accuracy in deriving target operating condition parameters from target quality parameters, thereby producing products with desired quality.
Smart Images

Figure 0007577657000001 
Figure 0007577657000002 
Figure 0007577657000003
Abstract
Description
[Technical field]
[0001] The present disclosure relates to an optimization support device, method, and program that support the optimization of various parameters in a production process of a product. [Background technology]
[0002] In a process for producing a product, a method has been proposed in which various calculations using, for example, a neural network are used to associate process condition parameters that specify the operating conditions of the process with parameters of the quality of the product, and to predict the quality parameters from the process condition parameters as a forward problem, or to predict the optimal solution of the process condition parameters from the quality parameters as an inverse problem.
[0003] For example, Patent Document 1 proposes a method of dividing a product's production process to define multiple prediction models, and deriving an optimal solution for the production plan in the reverse order of the flow of producing the product from the raw materials in order to optimize the production plan for the product. The method described in Patent Document 1 corresponds to a method of predicting process condition parameters from the quality parameters of the product by inverse problem by learning the correlation between the process condition parameters and the quality parameters. Also, Patent Document 2 proposes a method of modeling a production process by operation variables, state variables, and quality variables, and deriving optimal operation variables to improve the quality of the product. The method described in Patent Document 2 corresponds to a method of predicting quality parameters from process condition parameters as a forward problem. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] JP 2003-345416 A [Patent Document 2] JP 2006-323523 A Summary of the Invention [Problem to be solved by the invention]
[0005] In a process for producing a product, there are a considerable number of process condition parameters that specify the operating conditions of the process. When such a large number of condition parameters are input into a prediction model to predict the quality parameters of the product, there is a possibility that accurate prediction will not be possible. In particular, when the prediction model is a neural network, the neural network will be overtrained if there are a large number of process condition parameters. When a neural network is overtrained in this way, it will be difficult to predict accurate quality parameters when process condition parameters other than the training data are input. in It is also possible to obtain quality parameters from process condition parameters by using a function or a table instead of a prediction model. However, even in such a case, if the number of process condition parameters is large, the structure of the function or table becomes complicated, making it difficult to create an appropriate function or table, and as a result, there is a possibility that accurate prediction cannot be performed.
[0006] The present invention has been made in consideration of the above circumstances, and an object of the present invention is to make it possible to accurately predict various conditions in a production process of a product. [Means for solving the problem]
[0007] A first optimization support device according to the present disclosure includes a first conversion unit that converts operating condition parameters representing operating conditions of a process for producing a product into state parameters representing states of the process; and a second conversion unit that converts the state parameter into a quality parameter that represents the quality of the product.
[0008] In the first optimization support device according to the present disclosure, the first conversion unit has a first learning model that has been trained to output a state parameter in response to an input of an operating condition parameter, The second conversion unit may include a second learning model that has been trained to output a quality parameter in response to an input of a state parameter.
[0009] In addition, the first optimization support device according to the present disclosure includes a third conversion unit that converts the quality parameters into state parameters; The system may further include a fourth conversion unit that converts the state parameters into operating condition parameters.
[0010] In addition, in the first optimization support device according to the present disclosure, the third conversion unit converts the quality parameters into state parameters based on the control parameters of the second learning model, The fourth conversion unit may convert the state parameters into the operating condition parameters based on the control parameters of the first learning model.
[0011] A second optimization support device according to the present disclosure includes a third conversion unit that converts a quality parameter representing a quality of a product generated in a process for producing a product into a state parameter representing a state of the process; and a fourth conversion unit that converts the state parameters into operating condition parameters that represent operating conditions of the process.
[0012] In addition, in the second optimization support device according to the present disclosure, the third conversion unit may convert the quality parameter into a state parameter based on a control parameter of a learning model that has been trained to output a quality parameter in response to an input of a state parameter.
[0013] In this case, the system may further include a learning model that has been trained to output a quality parameter in response to an input of a state parameter, and a second conversion unit that converts the state parameter into a quality parameter that represents the quality of the product.
[0014] In addition, in the second optimization support device according to the present disclosure, the fourth conversion unit may convert the state parameters into operating condition parameters based on control parameters of a learning model that has been trained to output state parameters in response to input of operating condition parameters.
[0015] In this case, the system may further include a learning model that has been trained to output a state parameter in response to an input of an operating condition parameter, and a first conversion unit that converts the operating condition parameter into the state parameter.
[0016] Furthermore, in the first and second optimization support devices according to the present disclosure, the process may be a flow synthesis process, a cell culture process, a vacuum film formation process, or a coating process.
[0017] A first optimization support method according to the present disclosure includes converting an operating condition parameter representing an operating condition of a process for producing a product into a state parameter representing a state of the process; The state parameters are converted into quality parameters that represent the quality of the product.
[0018] A second optimization support method according to the present disclosure includes converting a quality parameter representing a quality of a product generated in a process for producing a product into a state parameter representing a state of the process; The state parameters are converted into operating condition parameters that represent the operating conditions of the process.
[0019] The first and second optimization support methods according to the present disclosure may be provided as a program executed by a computer.
[0020] A third optimization support device according to the present disclosure includes a memory that stores instructions to be executed by a computer; and a processor configured to execute the stored instructions, the processor comprising: converting operating condition parameters representative of operating conditions of a process for producing a product into state parameters representative of a state of the process; A process is carried out to convert the state parameters into quality parameters representing the quality of the product.
[0021] A fourth optimization support device according to the present disclosure includes a memory that stores instructions to be executed by a computer; and a processor configured to execute the stored instructions, the processor comprising: converting a quality parameter representing a quality of a product generated in a process for producing a product into a state parameter representing a state of the process; A process of converting the state parameters into operating condition parameters representing the operating conditions of the process is executed. Effect of the Invention
[0022] According to the present disclosure, various conditions in the production process of a product can be predicted with high accuracy. [Brief description of the drawings]
[0023] [Figure 1] A schematic block diagram showing the configuration of a production facility to which an optimization support device according to the present embodiment is applied. [Diagram 2] A schematic block diagram showing the configuration of a production facility including a flow reactor. [Diagram 3] A schematic block diagram showing the configuration of an optimization support device according to the present embodiment. [Figure 4] Conceptual diagram of the layered structure of a neural network [Diagram 5] Conceptual diagram of the layered structure of a neural network [Figure 6] Diagram of the training data set [Figure 7] Conceptual diagram of processing performed in this embodiment [Figure 8] A flowchart showing a process performed when generating a first learning model of the first conversion unit and a second learning model of the second conversion unit in this embodiment. [Figure 9]1 is a flowchart showing a process for deriving an operating condition parameter from a target quality parameter. [Figure 10] FIG. 13 is a schematic block diagram showing the configuration of an optimization support device according to another embodiment. [Figure 11] FIG. 13 is a schematic block diagram showing the configuration of an optimization support device according to still another embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0024] Hereinafter, an embodiment of the present disclosure will be described with reference to the drawings. Fig. 1 is a schematic block diagram showing the configuration of a production facility to which an optimization support device according to an embodiment of the present disclosure is applied. As shown in Fig. 1, the production facility 1 according to this embodiment includes a production device 2 and an optimization support device 3 according to this embodiment.
[0025] In this embodiment, the production apparatus 2 includes a flow reactor. The flow reactor is an apparatus for obtaining a product by performing a flow reaction process in which raw materials are continuously reacted while flowing. FIG. 2 is a schematic block diagram showing the configuration of a production facility 1 including a flow reactor. As shown in FIG. 2, the production apparatus 2 includes a flow reactor 11 and a controller 12. The flow reactor 11 includes a first supply unit 21, a second supply unit 22, a reaction section 23, and a recovery section 26. The operation of each part of the flow reactor 11 is controlled by the controller 12. The controller 12 is connected to the optimization support device 3.
[0026] In the flow reactor 11, the first supply section 21 and the second supply section 22 are each connected to the upstream end of the reaction section 23 by piping, and the recovery section 26 is connected to the downstream end of the reaction section 23 by piping.
[0027] The flow reaction performed in the flow reactor 11 may be, for example, a synthesis reaction for synthesizing a compound that is a monomer, a polymerization reaction for generating a polymer by reacting monomers with each other, or an elementary reaction such as an initiation reaction and a termination reaction of an anionic polymerization reaction. Therefore, the reactant to be subjected to the flow reaction may be, for example, a compound in the growth stage to be subjected to a termination reaction. In this embodiment, a termination reaction for terminating the growth of polystyryllithium with methanol is performed by the flow reaction.
[0028] The first supply unit 21 is for supplying a first raw material for the flow reaction to the reaction section 23. The first raw material in this embodiment is, for example, a first liquid in which polystyryllithium is dissolved in a solvent, and polystyryllithium is an example of a reactant for the flow reaction process. The first supply unit 21 is equipped with a pump (not shown), and the flow rate of the first raw material to the reaction section 23 is adjusted by adjusting the rotation speed of the pump.
[0029] The second supply unit 22 is for supplying the second raw material of the flow reaction to the reaction section 23. The second raw material in this embodiment is a mixture of methanol and water, i.e., an aqueous methanol solution, and methanol is used as a terminator for the termination reaction. Like the first supply unit 21, the second supply unit 22 is also provided with a pump (not shown), and the flow rate of the second raw material to the reaction section 23 is adjusted by adjusting the rotation speed of the pump.
[0030] The reaction section 23 is for performing a termination reaction as a flow reaction, and includes a junction 31, a reaction section 32, a temperature control section 33, an irradiation section 34, and a first detection section 35. The junction 31 is a pipe branched into a T shape, that is, a T-shaped pipe. A cross pipe may be used instead of the T-shaped pipe. The first pipe section 31a of the junction 31 is connected to the first supply section 21, the second pipe section 31b is connected to the second supply section 22, and the third pipe section 31c is connected to the reaction section 32. As a result, the first raw material and the second raw material guided to the reaction section 23 are joined and sent to the reaction section 32 in a mixed state. The reaction section 32 has a predetermined reaction path length and reaction path diameter. The reaction path length and reaction path diameter can be changed by changing the tubular member constituting the reaction section 32.
[0031] The inside of the reaction section 32 is a flow path for a mixture of the first raw material and the second raw material (hereinafter referred to as the mixed raw material), and the hollow space inside the tube is defined as a reaction site. As the mixed raw material passes through the reaction section 32, a termination reaction of the anionic polymerization is promoted, and polystyrene is produced.
[0032] The temperature control unit 33 is composed of, for example, a heater and is for controlling the temperature of the flow reaction (hereinafter referred to as the reaction temperature). The temperature control unit 33 controls the temperature (reaction temperature) of the mixed raw material flowing through the confluence unit 31 and the reaction unit 32.
[0033] The irradiation section 34 has a light source that emits light such as ultraviolet light, and is used to irradiate the reaction section 32 with light such as ultraviolet light when a photoreaction is carried out as a flow reaction.
[0034] The first detection unit 35 detects the state of the mixed raw material in the reaction section 23 and outputs the result to the optimization support device 3. The parameters representing the state of the mixed raw material (hereinafter referred to as state parameters) are parameters representing the physical properties of the mixed raw material obtained when the mixed raw material is reacted according to the input operating condition parameters and the environment in the reaction section. Examples of the state parameters include at least one of the reaction temperature, color, pH, amount of dissolved oxygen of the mixed raw material, pressure in the reaction section 23, and the shape of a spectrum representing the physical properties of the product (infrared absorption spectrum, Raman spectroscopic waveform, and nuclear magnetic resonance waveform). For this purpose, the first detection unit 35 includes a temperature sensor, an imaging unit, a pH sensor, a dissolved oxygen sensor, a spectrometer, and the like, all of which are not shown.
[0035] The recovery section 26 is for recovering polystyrene, which is a product of the flow reaction. In the recovery section 26, polystyrene is precipitated from the polystyrene solution guided from the reaction section, the precipitated polystyrene is collected from the mixed liquid, and the collected polystyrene is dried, thereby obtaining polystyrene.
[0036] Moreover, the recovery section 26 includes a second detection unit 36. The second detection unit 36 detects the quality of the product, which is the processing result of the flow reaction, and outputs it to the optimization support device 3. A parameter representing the quality of the product (hereinafter, referred to as a quality parameter) is a parameter that serves as a measure for judging whether or not the product obtained as a result of the reaction has appropriate quality. Specifically, the quality parameter may be at least one of the product concentration and the impurity concentration, but in addition to these, at least one of the purity, molecular weight, molecular weight dispersity, and yield of the product may also be used. Furthermore, when the product is obtained in the recovery section 26 in a solution state, for example, dissolved in a solvent, the concentration of the product in the solution (molar concentration, etc.) may also be detected as a quality parameter.
[0037] The reaction section and the recovery section are not limited to the above examples, and may be appropriately changed depending on at least one of the type of flow reaction and the type of product. For example, a container may be provided instead of the recovery section 26, and the polystyrene solution guided from the reaction section 23 may be temporarily stored in this container. In this case, the stored polystyrene solution may be guided to the recovery section 26, where polystyrene may be precipitated, collected, and dried to obtain the polystyrene.
[0038] The controller 12 performs overall control of the flow reactor 11. The controller 12 is connected to the pumps of the first supply unit 21 and the second supply unit 22, the temperature adjustment unit 33, the irradiation unit 34, the first detection unit 35, and the second detection unit 36. The controller 12 adjusts the flow rates of the first and second raw materials by adjusting the rotation speeds of the pumps of the first supply unit 21 and the second supply unit 22. The controller 12 also controls the temperature of the mixed raw materials by adjusting the temperature adjustment unit 33. The controller 12 also controls the irradiation of light such as ultraviolet light to the reaction section 23 by instructing the irradiation unit 34. The controller 12 also detects state parameters and quality parameters by instructing the first detection unit 35 and the second detection unit 36.
[0039] The controller 12 also sets the operating conditions of the flow reactor 11. The parameters representing the operating conditions (hereinafter referred to as operating condition parameters) are used to set reaction conditions, which are processing conditions for the flow reaction process, and are parameters for driving each part of the flow reactor 11 to produce a product of appropriate quality. Examples of the operating condition parameters include at least one of the flow rate of the first raw material, the flow rate of the second raw material, the reaction time, the reaction temperature, the mixing ratio, the UV illuminance, the flow channel depth, and the reagent equivalent when a reagent is used. The controller 12 has an operation unit (not shown) and sets the operating condition parameters by inputting an operation signal from the operation unit, thereby controlling the flow reactor 11 to the set operating conditions. For example, the operating condition parameters are set by clicking or selecting with the mouse of the operation unit and / or inputting characters from the keyboard.
[0040] In addition, the controller 12 is connected to the optimization support device 3, and in addition to or instead of the operation signal from the above-mentioned operation unit, sets the operating conditions to the target operating conditions output by the optimization support device 3, thereby controlling the flow reactor 11 to the predetermined operating conditions.
[0041] The optimization support device 3 provides support for accurately determining target operating condition parameters for the flow reaction process performed by the flow reactor 11. The optimization support device 3 is, for example, a computer in which the optimization support program of this embodiment is installed. The optimization support program is stored in a storage device of a server computer connected to a network or in a network storage in a state accessible from the outside, and is downloaded to the computer and installed in response to a request from an operator. Alternatively, the program is recorded on a recording medium such as a DVD (Digital Versatile Disc) or a CD-ROM (Compact Disc Read Only Memory) and distributed, and is installed in the computer from the recording medium.
[0042] Fig. 3 is a schematic block diagram showing the configuration of an optimization support device realized by installing an optimization support program in a computer. As shown in Fig. 2, the optimization support device 3 includes a CPU (Central Processing Unit) 41, a memory 42, and a storage 43 as a standard computer configuration. In addition, a display unit 44 such as a liquid crystal display, and an input unit 45 such as a keyboard and a mouse are connected to the optimization support device 3.
[0043] The storage 43 is made up of a hard disk drive or the like, and stores various information including information necessary for the optimization support processing.
[0044] An optimization support program is also stored in the memory 42. The optimization support program prescribes, as processes to be executed by the CPU 41, a first conversion process for converting operating condition parameters representing operating conditions of a process for producing a product into state parameters representing the state of the process, a second conversion process for converting the state parameters into quality parameters representing the quality of the product, a third conversion process for converting the quality parameters into state parameters, a fourth conversion process for converting the state parameters into operating condition parameters, and a learning process for learning a neural network, which will be described later.
[0045] The CPU 41 executes these processes according to the program, so that the computer functions as a first conversion unit 51, a second conversion unit 52, a third conversion unit 53, a fourth conversion unit 54, and a learning unit 55.
[0046] The first conversion unit 51 derives state parameters by converting operating condition parameters representing operating conditions of a process (flow reaction in this embodiment) for producing a product into state parameters representing the state of the process. For this purpose, the first conversion unit 51 has a first learning model M1 that has been trained to output state parameters in response to input of operating condition parameters. The first learning model M1 is constructed by the learning unit 55 learning a neural network or the like as described below.
[0047] The second conversion unit 52 derives the quality parameters by converting the state parameters into quality parameters representing the quality of the product (polystyrene in this embodiment). For this purpose, the second conversion unit 52 has a second learning model M2 that has been trained to output the quality parameters in response to input of the state parameters. The second learning model M2 is constructed by the learning unit 55 learning a neural network or the like as described below.
[0048] The third conversion unit 53 derives the state parameters by converting the quality parameters representing the quality of the product into state parameters representing the state of the process. In this embodiment, the third conversion unit 53 converts the quality parameters into the state parameters based on the control parameters of the second learning model M2 in the second conversion unit 52.
[0049] The fourth conversion unit 54 derives the operating condition parameters by converting the state parameters representing the state of the process into operating condition parameters representing the operating conditions of the process. In this embodiment, the fourth conversion unit 54 converts the state parameters into the operating condition parameters based on the control parameters of the first learning model M1 in the first conversion unit 51.
[0050] Here, the model used for the first learning model M1 is a prediction model that predicts state parameters by inputting operating condition parameters. The model used for the second learning model M2 is also a prediction model that predicts quality parameters by inputting state parameters. As the prediction model, a machine learning model can be used. Examples of the machine learning model include linear regression, Gaussian process regression, support vector regression, decision tree, ensemble method, bagging method, boosting method, and gradient boosting method. In addition, an example of the machine learning model is a neural network model. Examples of the neural network model include a simple perceptron, a multilayer perceptron, a deep neural network, a convolutional neural network, a deep belief network, a recurrent neural network, and a probabilistic neural network.
[0051] Another example of an ensemble method for machine learning models is the Random Forest. The Random Forest is a learning model that uses randomly sampled training data and randomly selected explanatory variables to create multiple groups of decision trees with low correlation, and then integrates and averages their prediction results to improve prediction accuracy. Note that the control parameters of the Random Forest model include the number of explanatory variables and the number of branches in the decision trees.
[0052] Further, the neural network model may be a deep neural network. Compared with machine learning models other than neural networks, deep neural networks have a large number of model control parameters and can be flexibly combined, so that they can exhibit high performance for various data configurations. Examples of control parameters of deep neural networks include the number of network layers, the number of nodes, the type of activation function, the dropout rate, the mini-batch size, the number of epochs, and the learning rate. These models have a plurality of execution frameworks, from which an appropriate one can be selected. For example, the execution framework can be selected from Tensorflow, CNTK, Theano, Caffe, mxnet, Keras, PyTorch, Chainer, Scikit-learn, Caret, Matlab (registered trademark), and the like. In this embodiment, a neural network is used as a prediction model.
[0053] The learning unit 55 learns a neural network as a prediction model to construct a first learning model M1 and a second learning model M2. First, the first learning model M1 of the first conversion unit 51 will be described. The learning unit 55 learns a neural network with the driving condition parameters as explanatory variables and the state parameters as objective variables to derive a function that represents the association between the driving condition parameters and the state parameters, thereby constructing the first learning model M1. In this embodiment, the learning unit 55 generates the following functions (1A), (1B), and (1C) that represent the association between the driving condition parameters and the state parameters.
[0054] y1=w u1y1 / [1+exp{-(w x1u1 ×x1+w x2u1 ×x2+…+w x5u1 ×x5)}] +w u2y1 / [1+exp{-(w x1u2 ×x1+w x2u2 ×x2+…+w x5u2 ×x5)}] +w u3y1 / [1+exp{-(w x1u3 ×x1+w x2u3 ×x2+…+w x5u3 ×x5)}] …(1A) y2=w u1y2 / [1+exp{-(w x1u1 ×x1+w x2u1 ×x2+…+w x5u1 ×x5)}] +w u2y2 / [1+exp{-(w x1u2 ×x1+w x2u2 ×x2+…+w x5u2 ×x5)}] +w u3y2 / [1+exp{-(w x1u3 ×x1+w x2u3 ×x2+…+w x5u3 ×x5)}] …(1B) y3=w u1y3 / [1+exp{-(w x1u1 ×x1+w x2u1 ×x2+…+w x5u1 ×x5)}] +w u2y3 / [1+exp{-(w x1u2 ×x1+w x2u2 ×x2+…+w x5u2 ×x5)}] +w u3y3 / [1+exp{-(w x1u3 ×x1+w x2u3 ×x2+…+w x5u3 ×x5)}] …(1C)
[0055] In the above functions (1A) to (1C), xi (i is a natural number) is the value of the operating condition parameter, and the maximum value of i is the number of operating condition parameters. In this embodiment, if five parameters, for example, the flow rate of the first raw material, the flow rate of the second raw material, the reaction time, the reaction temperature, and the mixing ratio, are used as the operating condition parameters, then i=5. ym (m is a natural number) is the value of the state parameter, and the maximum value of m is the number of state parameters. In this embodiment, if, for example, the color, pressure, and pH of the mixed raw materials are used as the state parameters, then m=3. ul (l is a natural number) is a node of the hidden layer L2 in the neural network described later, and the maximum value of l is the number of nodes. In this embodiment, l=3. w xiul , w ulym is the weighting coefficient that represents the connection weight of the neural network. Specifically, w xiul is the weighting coefficient between xi and ul, and w ulym is the weighting factor between ul and ym.
[0056] FIG. 4 is a diagram for explaining the layer structure of a neural network for constructing a first learning model M1 in this embodiment. As shown in FIG. 4, the neural network 60 has a three-layer structure of an input layer L1, a hidden layer L2, and an output layer L3. The input layer L1 is composed of values x1 to x5 of operating condition parameters, which are explanatory variables. The hidden layer L2 is composed of three nodes u1 to u3, and is one layer in this embodiment. Each of the nodes u1 to u3 is assigned a weighting coefficient w xiulThe output layer L3 is composed of state parameter values y1 to y3, which are the objective variables. The state parameter values y1 to y3 are calculated by using nodes u1 to u3 and the weighting coefficients w ulym The black circles in Figure 4 indicate the weighting coefficient w xiul ,w ulym The weighting coefficient w xiul ,w ulym is derived by learning the neural network 60 using the teacher data. Note that the layer structure of the neural network 60 is not limited to that shown in FIG.
[0057] Next, the second learning model M2 of the second conversion unit 52 will be described. The learning unit 55 performs learning of a neural network using the state parameters as explanatory variables and the quality parameters as objective variables to derive a function that represents the association between the state parameters and the quality parameters, thereby constructing the second learning model M2. In this embodiment, the learning unit 55 generates the following functions (2A) and (2B) that represent the association between the operating condition parameters and the state parameters.
[0058] z1=w u1z1 / [1+exp{-(w y1u11 ×y1+w y2u11 ×y2+w y3u11 ×y3)}] +w u2z1 / [1+exp{-(w y1u12 ×y1+w y2u12 ×y2+w y3u12 ×y3)}] +w u3z1 / [1+exp{-(w y1u13 ×y1+w y2u13 ×y2+w y3u13 ×y3)}] …(2A) z2=w u1z2 / [1+exp{-(w y1u11 ×y1+w y2u11 ×y2+w y3u11 ×y3)}] +w u2z2 / [1+exp{-(w y1u12×y1+w y2u12 ×y2+w y3u12 ×y3)}] +w u3z2 / [1+exp{-(w y1u13 ×y1+w y2u13 ×y2+w y3u13 ×y3)}] …(2B)
[0059] In the above functions (2A) and (2B), ym (m is a natural number) is the value of the state parameter. In this embodiment, if the reaction temperature, color, and pH of the mixed raw material are used as the state parameters as described above, m=3. zk (k is a natural number) is the value of the quality parameter, and the maximum value of k is the number of quality parameters. In this embodiment, if the product concentration and impurity concentration are used as the quality parameters, k=2. u1l (l is a natural number) is a node of the hidden layer L12 in the neural network described later, and the maximum value of 1l is the number of nodes. In this embodiment, 1l=3. w ymu1l , w u1lzk is the weighting coefficient that represents the connection weight of the neural network. Specifically, w ymu1l is the weighting factor between ym and u1l, and w u1lzk is the weighting factor between u1l and zk.
[0060] FIG. 5 is a diagram for explaining the layer structure of a neural network for constructing a second learning model M2 in this embodiment. As shown in FIG. 5, the neural network 70 has a three-layer structure of an input layer L11, a hidden layer L12, and an output layer L13. The input layer L11 is composed of state parameter values y1 to y3, which are explanatory variables. The hidden layer L12 is composed of three nodes u11 to u13, and is one layer in this embodiment. Each of the nodes u11 to u13 is assigned a weighting coefficient w ymu1l The output layer L13 is composed of quality parameter values z1 and z2, which are the objective variables. The quality parameter values z1 and z2 are each weighted by the weighting coefficient w u1lzkThe black circles in Figure 5 indicate the weighting coefficient w ymu1l ,w u1lzk The weighting coefficient w ymu1l ,w u1lzkym is derived by learning the neural network 70 using the teacher data. The layer structure of the neural network 70 is shown in FIG. 5 is not limited to those shown in
[0061] The learning unit 55 uses a plurality of pre-generated teacher data to train the neural networks 60, 70 to construct the first learning model M1 and the second learning model M2. Here, the teacher data includes the operating condition parameters, state parameters, and quality parameters when a product of desirable quality is obtained. A plurality of teacher data are prepared and stored in the storage 43. FIG. 6 is a diagram showing an example of the teacher data. As shown in FIG. 6, the teacher data includes the operating condition parameters, state parameters, and quality parameters. The operating condition parameters include five parameters, namely, the flow rate of the first raw material, the flow rate of the second raw material, the reaction time, the reaction temperature, and the mixing ratio. The state parameters include three parameters, namely, the color, pressure, and pH of the mixed raw materials. The quality parameters include two parameters, namely, the product concentration and the impurity concentration.
[0062] During learning, the learning unit 55 uses the teacher data stored in the storage 43 to train the neural networks 60, 70, for example, according to the backpropagation method. Specifically, for the neural network 60, the learning unit 55 inputs the operating condition parameters included in one of the teacher data sets to the neural network 60, and causes the neural network 60 to output state parameters. Then, the learning unit 55 determines the weighting coefficient w so as to minimize the difference between the state parameters output by the neural network 60 and the state parameters included in the teacher data. xiul ,w ulym The neural network 60 is trained by deriving
[0063] For the neural network 70, the learning unit 55 inputs state parameters included in one of the teacher data sets to the neural network 70 and causes the neural network 70 to output quality parameters. The learning unit 55 then determines the weighting coefficient w so as to minimize the difference between the quality parameters output by the neural network 70 and the quality parameters included in the teacher data. ymu1l ,w u1lzk The neural network 70 is trained by deriving
[0064] When the learning is completed and the first learning model M1 and the second learning model M2 are constructed, the learning unit 55 stores in the storage 43 the functions represented by the first learning model M1 and the second learning model M2.
[0065] In this embodiment, the third conversion unit 53 and the fourth conversion unit 54 derive operating condition parameters for obtaining a target quality parameter from an unknown target quality parameter (hereinafter referred to as a target quality parameter). For this purpose, the third conversion unit 53 converts the target quality parameter into a state parameter based on the control parameter of the second learning model M2 possessed by the second conversion unit 52. The state parameter obtained by converting the target quality parameter is referred to as a target state parameter. The control parameter is a weighting coefficient w ymu1l ,w u1lzk In this embodiment, the third conversion unit 53 derives the following functions G21, G21 to generate the target state parameters.
[0066] G21=zt1-[w u1z1 / [1+exp{-(w y1u11 ×y1+w y2u11 ×y2+w y3u11 ×y3)}] +w u2z1 / [1+exp{-(w y1u12 ×y1+w y2u12 ×y2+w y3u12 ×y3)}] +w u3z1 / [1+exp{-(w y1u13×y1+w y2u13 ×y2+w y3u13 ×y3)}]] …(3A) G22=zt2-[w u1z2 / [1+exp{-(w y1u11 ×y1+w y2u11 ×y2+w y3u11 ×y3)}] +w u2z2 / [1+exp{-(w y1u12 ×y1+w y2u12 ×y2+w y3u12 ×y3)}] +w u3z2 / [1+exp{-(w y1u13 ×y1+w y2u13 ×y2+w y3u13 ×y3)}]] …(3B)
[0067] In the functions G21 and G22, zt1 and zt2 are target quality parameters, respectively. The third conversion unit 53 derives a state parameter ym for minimizing the absolute values of the functions G21 and G22 as a target state parameter ytm. Specifically, the target quality parameters zt1 and zt2 are used as explanatory variables, and the target state parameter ytm is used as an objective variable, and the target state parameter ytm is derived by performing multivariate analysis such as multiple regression analysis and principal component analysis, or multi-objective optimization such as genetic algorithm, multi-objective particle swarm optimization, and Bayesian optimization.
[0068] The fourth conversion unit 54 converts the target state parameters into driving condition parameters based on the control parameters of the first learning model M1 held by the first conversion unit 51. The driving condition parameters obtained by converting the target state parameters are regarded as target driving condition parameters. The control parameters are expressed by the weighting coefficient w xiul , w ulym In this embodiment, the fourth conversion unit 54 generates the following functions G11, G12, and G13 to generate the target operating condition parameters.
[0069] G11=yt1-[w u1y1 / [1+exp{-(w x1u1 ×x1+w x2u1×x2+…+w x5u1 ×x5)}] +w u2y1 / [1+exp{-(w x1u2 ×x1+w x2u2 ×x2+…+w x5u2 ×x5)}] +w u3y1 / [1+exp{-(w x1u3 ×x1+w x2u3 ×x2+…+w x5u3 ×x5)}]] …(4A) G12=yt2-[w u1y2 / [1+exp{-(w x1u1 ×x1+w x2u1 ×x2+…+w x5u1 ×x5)}] +w u2y2 / [1+exp{-(w x1u2 ×x1+w x2u2 ×x2+…+w x5u2 ×x5)}] +w u3y2 / [1+exp{-(w x1u3 ×x1+w x2u3 ×x2+…+w x5u3 ×x5)}]] …(4B) G13=yt3-[w u1y3 / [1+exp{-( wx1u1 ×x1+w x2u1 ×x2+…+w x5u1 ×x5)}] +w u2y3 / [1+exp{-(w x1u2 ×x1+w x2u2 ×x2+…+w x5u2 ×x5)}] +w u3y3 / [1+exp{-(w x1u3 ×x1+w x2u3 ×x2+…+w x5u3 ×x5)}]] …(4C)
[0070] In the functions G11 to G13, yt1, yt2, and yt3 are target state parameters, respectively. The fourth conversion unit 54 derives the operating condition parameters xi for minimizing the absolute values of the functions G11 to G13 as the target operating condition parameters xti. Specifically, the target operating condition parameters xti are derived by performing multivariate analysis such as multiple regression analysis and principal component analysis, or multi-objective optimization such as genetic algorithm, multi-objective particle swarm optimization, and Bayesian optimization, with the target state parameters yt1, yt2, and yt3 as explanatory variables and the target operating condition parameters xti as objective variables.
[0071] The processes performed by the above-mentioned first conversion unit 51, second conversion unit 52, third conversion unit 53, and fourth conversion unit 54 will be conceptually explained with reference to Fig. 7. The first conversion unit 51 predicts state parameters from operating condition parameters as a forward problem. The second conversion unit 52 predicts quality parameters from state parameters as a forward problem. The third conversion unit 53 predicts state parameters from quality parameters as an inverse problem. The fourth conversion unit 54 predicts operating condition parameters from state parameters as an inverse problem. In Fig. 7, the open arrows represent forward problems, and the filled-in-black arrows represent inverse problems.
[0072] 7, in this embodiment, the quality parameters are derived by a two-stage process from the operating condition parameters through the state parameters, and the operating condition parameters are derived by a two-stage process from the quality parameters through the state parameters.
[0073] The optimization support device 3 outputs the target operating condition parameters xti to the controller 12. The controller 12 controls the operation of the flow reactor 11 according to the target operating condition parameters xti. This results in the production of a product having a target quality.
[0074] Next, the process performed in this embodiment will be described. FIG. 8 is a flowchart showing the process performed when constructing the first learning model M1 of the first conversion unit 51 and the second learning model M2 of the second conversion unit 52 in this embodiment. The learning unit 55 reads one piece of teacher data from the multiple pieces of teacher data stored in the storage 43 (step ST1), and causes the neural network of the first conversion unit 51 to learn the association between the operating condition parameter and the state parameter (first learning; step ST2). Next, the learning unit 55 causes the neural network of the second conversion unit 52 to learn the association between the state parameter and the quality parameter (second learning; step ST3). Then, the process returns to step ST1, reads the next teacher data, and repeats the processes of steps ST2 and ST3. As a result, the first learning model M1 and the second learning model M2 are constructed.
[0075] The learning unit 55 repeats the learning until the difference between the state parameter generated by the first conversion unit 51 and the state parameter of the teacher data, and the difference between the quality parameter generated by the second conversion unit 52 and the quality parameter of the teacher data become equal to or smaller than a predetermined threshold value. The number of repetitions may be a predetermined number.
[0076] Next, a process for deriving the operating condition parameters from the unknown target quality parameters in this embodiment will be described. Fig. 9 is a flowchart showing a process for deriving the operating condition parameters from the target quality parameters. Input section 4 When the target quality parameters are input from the optimization support device 3 (step ST11), the third conversion unit 53 converts the target quality parameters into target state parameters based on the control parameters of the second learning model M2 (step ST12). conditions The data is converted into parameters (step ST13), and the process ends.
[0077] In this manner, in the present embodiment, quality parameters are derived from process condition parameters in the past, but the process condition parameters are divided into operating condition parameters and state parameters, the first conversion unit 51 converts the operating condition parameters into state parameters, and the second conversion unit 52 converts the state parameters into quality parameters, thereby deriving the quality parameters through a two-step process. Therefore, compared to the conventional process in which the operating condition parameters and the state parameters are not distinguished from each other and are used as process condition parameters to obtain the quality parameters, the number of parameters input in the processes performed by the first conversion unit 51 and the second conversion unit 52 can be reduced. This makes it possible to prevent overlearning, particularly when a learning model constructed by learning such as a neural network is used for conversion. Therefore, according to the present embodiment, various conditions in the production process of the product can be predicted with high accuracy.
[0078] In addition, in this embodiment, the process of deriving the target operating condition parameters from the target quality parameters is divided into two processes by being performed by the third conversion unit 53 and the fourth conversion unit 54, respectively. Therefore, compared to the conventional process of determining the process condition parameters from the quality parameters without distinguishing between the operating condition parameters and the state parameters, it is possible to reduce the number of parameters input in the processes performed by the third conversion unit 53 and the fourth conversion unit 54. As a result, according to this embodiment, the target operating condition parameters can be accurately predicted from the target quality parameters.
[0079] In addition, in this embodiment, the target quality parameters are converted into target state parameters using the control parameters of the second learning model M2, and the target state parameters are converted into target operating condition parameters using the control parameters of the first learning model M1. Here, the control parameters of the first learning model M1 and the second learning model M2 are capable of accurately predicting various conditions in the production process of the product. Therefore, according to this embodiment, the target operating condition parameters can be derived with high accuracy from the target quality parameters. Therefore, by producing a product using the target operating condition parameters, a product having a target quality can be obtained.
[0080] Furthermore, when predicting process condition parameters from quality parameters by learning the correlation between the process condition parameters and quality parameters as in the method described in the above Patent Document 1, it is as if no correlation between the operating condition parameters and the state parameters has been learned as in this embodiment. Therefore, even if a production process is carried out using the predicted process condition parameters, a product having the target quality may not be produced. In other words, when predicting process condition parameters from quality parameters by learning the correlation between the process condition parameters and quality parameters, it is assumed that the state parameters can be realized by the operating condition parameters. However, this assumption does not hold in most cases. For this reason, the conventional method of predicting process condition parameters from quality parameters does not predict the target quality parameters. Ta However, it is not possible to accurately predict the process condition parameters that can achieve this.
[0081] In this embodiment, the process condition parameters are divided into operating condition parameters and state parameters, and the state parameters are predicted from the quality parameters, the operating condition parameters are predicted from the state parameters, or the state parameters are predicted from the operating condition parameters, and the quality parameters are predicted from the state parameters. Therefore, the operating condition parameters, state parameters, and quality parameters can be predicted under the assumption that the state parameters can be realized by the operating condition parameters included in the process condition parameters. Therefore, according to this embodiment, various conditions in the production process of the product can be predicted with high accuracy.
[0082] In the above embodiment, the first conversion unit 51 converts the driving condition parameters into state parameters by using the first learning model M1 constructed by learning the neural network 60, but the present invention is not limited to this. For example, the driving condition parameters may be converted into state parameters by a formula for converting the driving condition parameters into state parameters, or a table in which the driving condition parameters correspond to the state parameters.
[0083] In this case, the fourth conversion unit 54 may convert the state parameters into the operating condition parameters using a formula for converting the state parameters into the operating condition parameters, or a table that associates the state parameters with the operating condition parameters, or the like.
[0084] In the above embodiment, the second conversion unit 52 converts the state parameters into quality parameters by using the second learning model M2 constructed by learning the neural network 70, but the present invention is not limited to this. For example, the state parameters may be converted into quality parameters by a formula for converting the state parameters into quality parameters, or a table in which the state parameters and the quality parameters are associated with each other.
[0085] In this case, the third conversion unit 53 may convert the quality parameters into the state parameters using a formula for converting the quality parameters into the state parameters, or a table in which the quality parameters correspond to the state parameters, or the like.
[0086] In the above embodiment, the optimization support device 3 includes the first conversion unit 51, the second conversion unit 52, the third conversion unit 53, and the fourth conversion unit 54, but is not limited thereto. As shown in FIG. 10, the optimization support device 3 may include only the first conversion unit 51 and the second conversion unit 52. In this case, the first conversion unit 51 may convert the operating condition parameters into state parameters by the first learning model M1, or may convert the operating condition parameters into state parameters by a formula, a table, or the like. The second conversion unit 52 may convert the state parameters into quality parameters by the second learning model M2, or may convert the state parameters into quality parameters by a formula, a table, or the like.
[0087] 11, the optimization support device 3 may have only the third conversion unit 53 and the fourth conversion unit 54. In this case, the third conversion unit 53 may convert the quality parameters into state parameters using the control parameters of the second learning model M2 constructed by a device separate from the optimization support device 3, or may convert the quality parameters into state parameters by a formula, a table, or the like. The fourth conversion unit 54 may convert the state parameters into operating condition parameters using the control parameters of the first learning model M1 constructed by a device separate from the optimization support device 3, or may convert the state parameters into operating condition parameters by a formula, a table, or the like.
[0088] In the above embodiment, a flow synthesis process is used as a process for producing a product, but the present invention is not limited thereto. A cell culture process may be used as a process for producing a product. In a cell culture process, the product is a cell and an antibody produced by the cell, and the cell culture conditions are used as operation condition parameters, and the quality of the cell and the antibody produced by the cell is used as a quality parameter. The production apparatus 2 shown in FIG. 1 is a cell culture apparatus. For example, when the cell culture process is an antibody cell culture process, the operation condition parameters may be, for example, a perfusion ratio, a stationary phase cell count, a stirring speed of the liquid in the culture tank, a bottom air amount, a bottom oxygen amount, a bottom nitrogen amount, a top carbon dioxide amount, a top air amount, an antifoaming agent amount, and a surfactant amount. Examples of the state parameters that can be used include pH, pO2 (oxygen partial pressure), pCO2 (carbon dioxide partial pressure), Gln (glutamine concentration), Glu (glutamic acid concentration), Gluc (glucose concentration), Lac (lactic acid concentration), NH4+ (ammonia ion concentration), Na+ (sodium ion concentration), K+ (potassium ion concentration), Osmol (osmotic pressure), and Kla (oxygen transfer coefficient), etc. Examples of the quality parameters that can be used include antibody concentration, by-product concentration, and viable cell count (number of living cells), etc.
[0089] A vacuum film-forming process may be used as a process for producing a product. In the vacuum film-forming process, the product is a film formed on the surface of a substrate such as glass, resin, or metal. The production apparatus 2 shown in FIG. 1 is a vacuum film-forming apparatus. The vacuum film-forming conditions are controlled by an operating temperature. conditions The parameters are used as the conditions, and the quality of the film formation is used as the quality parameter. Examples of the operating condition parameters that can be used include vacuum pressure, applied voltage, bias voltage, substrate temperature, film formation time, gas flow rate, gas concentration, and line speed. The state parameters are the state of the plasma, and examples of the parameters that can be used include plasma stability, plasma color development, ICP-OES / ICP-AES (plasma optical emission spectroscopy) waveform spectrum, plasma density, anti-adhesion plate deposit thickness, and device potential difference. Examples of the quality parameters that can be used include film quality and barrier performance.
[0090] A roll-to-roll coating process may also be used as a process for producing the product. In the coating process, the product is a liquid crystal retardation film and an anti-glare film in which a coating film is formed on the surface of a resin substrate such as TAC (triacetyl cellulose), PET (polyethylene terephthalate), and COP (cycloolefin polymer). The production device 2 shown in FIG. 1 is a coating device. The coating conditions are used as operating condition parameters, and the quality of the coating film is used as a quality parameter. Examples of operating condition parameters that can be used include the coating flow rate, liquid temperature, drying temperature, line speed, and transport roll temperature. Examples of state parameters that can be used include the pulsation of the liquid, the fluctuation of the drying wind speed, the amount of volatilization during drying, and the amount of liquid seeping into the substrate. Examples of quality parameters that can be used include the film quality, the film thickness distribution of the coating film, and the adhesion between the coating film and the substrate.
[0091] In the above embodiment, the following various processors can be used as the hardware structure of the processing unit that executes various processes, such as the first conversion unit 51, the second conversion unit 52, the third conversion unit 53, the fourth conversion unit 54, and the learning unit 55. As described above, the above various processors include a CPU, which is a general-purpose processor that executes software (programs) and functions as various processing units, as well as a programmable logic device (PLD), which is a processor whose circuit configuration can be changed after manufacture, such as an FPGA (Field Programmable Gate Array), and a dedicated electric circuit, which is a processor having a circuit configuration designed specifically for executing specific processes, such as an ASIC (Application Specific Integrated Circuit).
[0092] A single processing unit may be configured with one of these various processors, or may be configured with a combination of two or more processors of the same or different types (e.g., a combination of multiple FPGAs or a combination of a CPU and an FPGA). Also, multiple processing units may be configured with a single processor.
[0093] As an example of configuring multiple processing units with one processor, first, there is a form in which one processor is configured with a combination of one or more CPUs and software, as typified by computers such as client and server, and this processor functions as multiple processing units. Second, there is a form in which a processor is used that realizes the functions of the entire system including multiple processing units with one IC (Integrated Circuit) chip, as typified by System On Chip (SoC), etc. In this way, the various processing units are configured using one or more of the above various processors as a hardware structure.
[0094] Furthermore, more specifically, the hardware structure of these various processors can be an electric circuit that combines circuit elements such as semiconductor elements. [Explanation of symbols]
[0095] 1. Production facilities 2. Production Equipment 3. Optimization Support Device 11 Flow Reactor 12 Controller 21 1st supply section 22 2nd supply section 23 Reaction Section 26 Collection Section 31 Junction 31a~31c 1st pipe section ~ 3rd pipe section 32 Reaction section 33 Temperature control section 34 Irradiation unit 35 First detection section 36 Second detection section 41 CPU 42 Memory 43 Storage 44 Display section 45 Input section 51 First conversion unit 52 Second conversion section 53 Third conversion section 54 4th conversion section 55 Learning Department 60,70 Neural Networks L1,L11 Input layer L2, L12 Hidden layer L3, L13 output layer xi Operating condition parameter values ul, u1l node values ym State parameter value zk Quality parameter value
Claims
1. a first conversion unit that converts operating condition parameters representing operating conditions of a process for producing a product into state parameters representing a state of the process, the first conversion unit having a first learning model that has been trained to output the state parameters in response to input of the operating condition parameters, with the operating condition parameters being explanatory variables and the state parameters being objective variables; and a second conversion unit that converts the state parameters into quality parameters that represent the quality of the product.
2. The optimization support device as described in Claim 1, wherein the second conversion unit has a second learning model that has been trained to output the quality parameter in response to the input of the state parameter.
3. a third conversion unit for converting the quality parameters into the state parameters; 2. The optimization support device according to claim 1, further comprising a fourth conversion unit that converts the state parameters into operating condition parameters.
4. a third conversion unit for converting the quality parameters into the state parameters; 3. The optimization support device according to claim 2, further comprising a fourth conversion unit that converts the state parameters into operating condition parameters.
5. The third conversion unit converts the quality parameters into the state parameters based on a control parameter of the second learning model; The optimization support device according to claim 4 , wherein the fourth conversion unit converts the state parameters into the operating condition parameters based on control parameters of the first learning model.
6. a third conversion unit that converts a quality parameter representing a quality of a product generated in a process for producing a product into a state parameter representing a state of the process; and a fourth conversion unit that converts the state parameters into operating condition parameters that represent operating conditions of the process, the fourth conversion unit converting the state parameters into the operating condition parameters based on control parameters of a learning model that has been trained to output the state parameters in response to input of the operating condition parameters, with the operating condition parameters as explanatory variables and the state parameters as objective variables.
7. 7. The optimization support device according to claim 6, wherein the third conversion unit converts the quality parameter into the state parameter based on a control parameter of a learning model that has been trained to output the quality parameter in response to an input of the state parameter.
8. 8. The optimization support device according to claim 7, further comprising a second conversion unit that has a learning model that has been trained to output the quality parameter in response to the input of the state parameter, and that converts the state parameter into a quality parameter that represents the quality of the product.
9. 9. The optimization support device according to claim 8, further comprising a learning model that has been trained to output the state parameter in response to an input of the operating condition parameter, and a first conversion unit that converts the operating condition parameter into the state parameter.
10. The optimization support device according to claim 1 , wherein the process is a flow synthesis process.
11. The optimization support device according to claim 1 , wherein the process is a cell culture process.
12. The optimization support device according to claim 1 , wherein the process is a vacuum film formation process.
13. The optimization support device according to claim 1 , wherein the process is a coating process.
14. converting the operating condition parameters into the state parameters by a first learning model that has been trained to output the state parameters in response to input of the operating condition parameters, the operating condition parameters being explanatory variables and state parameters being objective variables that represent the state of the process; An optimization support method for converting the state parameters into quality parameters representative of the quality of the product.
15. converting quality parameters representative of the quality of a product produced in a process for producing a product into state parameters representative of a state of the process; An optimization support method for converting an operating condition parameter representing an operating condition of the process into an operating condition parameter based on a control parameter of a learning model that has been trained to output the state parameter in response to an input of the operating condition parameter, the state parameter being an explanatory variable and the state parameter being a target variable.
16. a step of converting an operating condition parameter representing an operating condition of a process for producing a product into a state parameter by a first learning model that has been trained to output the state parameter in response to an input of the operating condition parameter, the state parameter representing the state of the process being a response variable; and a procedure for converting the state parameters into quality parameters that represent the quality of the product.
17. A step of converting quality parameters representing the quality of a product produced in a process for producing a product into state parameters representing the state of the process; and a procedure for converting the state parameters into the operating condition parameters based on control parameters of a learning model that has been trained to output the state parameters in response to input of the operating condition parameters, the operating condition parameters being explanatory variables and the state parameters being objective variables.
Citation Information
Patent Citations
Coke strength estimation device
JP1997304376A
Optimization method of production planning and optimization program of production planning
JP2003345416A
Operation variable selection device, operation variable selection method, operation variable selection program, and computer readable recording medium recording the same
JP2006323523A
Optimal operation condition setting system
JP2007241954A
Method for manufacturing functional film
JP2011194728A
