Information processing device, operation method for information processing device, and operation program for information processing device

Abstract

An information processing device comprising a processor, wherein the processor uses a machine learning model outputting a predicted value for the quality of a split section in response to the input of manufacturing parameters for the split section, said split section resulting from splitting a whole product, and derives favorable manufacturing parameters according to which the quality of the entire product will reach a target value, by inputting manufacturing parameters for the split section to the machine learning model, causing predicted values for the split section to be outputted from the machine learning model, and solving an optimization problem seeking manufacturing parameters that minimize the value of an objective function having arguments that include the difference between the predicted value for the split section and a target value for the quality of the split section.

Description

Information processing device, operating method for information processing device, and operating program for information processing device

[0001] The technology of the present disclosure relates to an information processing device, an operating method for an information processing device, and an operating program for an information processing device.

[0002] In the past, in the manufacture of products, suitable manufacturing conditions have been found by trial and error, with many prototypes being manufactured under various manufacturing conditions and the quality of the prototypes being evaluated. For example, in the manufacture of flexible tubes for endoscopes using an extrusion molding machine, as described in JP 2021-166723 A, suitable extrusion rates per unit time of the resin material for the inner and outer layers that coat the flexible tube base material were found by evaluating the elasticity values, etc., of many prototypes in which the extrusion rates per unit time of the resin material for the inner and outer layers were changed and that had different thickness ratios in the axial direction.

[0003] The trial-and-error method of searching for optimal manufacturing conditions, which involves manufacturing a large number of prototypes while varying the manufacturing conditions in various ways, requires a great deal of effort and cost. Therefore, a method can be considered in which a machine learning model that outputs a predicted value of product quality in response to input manufacturing conditions is used to solve an optimization problem and derive optimal manufacturing conditions that achieve a target quality value.

[0004] However, if a machine learning model is specialized for one type of product, it will be unable to be applied to other types of products, which is a problem of versatility. For example, a machine learning model specialized for a large-diameter, long flexible tube for an oral endoscope for the upper gastrointestinal tract will not be able to derive suitable manufacturing conditions for a small-diameter, short flexible tube for a transnasal endoscope for the bronchi.

[0005] One embodiment of the technique of the present disclosure provides an information processing device, an operating method for the information processing device, and an operating program for the information processing device that are capable of deriving generally suitable manufacturing conditions regardless of the type of product.

[0006] The information processing device disclosed herein includes a processor, and the processor uses a machine learning model that outputs a predicted value of the quality of a divided portion into which an entire product is divided in response to input of manufacturing conditions for the divided portion.The processor inputs the manufacturing conditions of the divided portion into the machine learning model, causes the machine learning model to output a predicted value of the divided portion, and solves an optimization problem to find manufacturing conditions that minimize the value of an objective function having a term that includes the difference between the predicted value of the divided portion and the target value of the quality of the divided portion, thereby deriving suitable manufacturing conditions that will result in the quality of the entire product being the target value.

[0007] The objective function preferably has a regularization term for accommodating the manufacturing conditions of the divided parts within the constraints on the manufacturing conditions of the entire product.

[0008] The regularization term is preferably a term that smooths the manufacturing conditions of the divided portions.

[0009] The regularization term is preferably the sum of second-order differential values ​​of the manufacturing conditions of the divided portions.

[0010] The term containing the difference is preferably the sum of the difference divided by the target value of the divided portion.

[0011] It is preferable that at least one of physical property information of the material that constitutes the product and design information of the product is also input to the machine learning model.

[0012] There are multiple types of predicted values, and it is preferable that a machine learning model is prepared for each of the multiple types of predicted values.

[0013] The product is a flexible tube for an endoscope having a flexible tube base material and a resin layer covering the flexible tube base material, the resin layer being composed of an inner layer and an outer layer with different thickness ratios in the axial direction, the divided portions being portions obtained by dividing the flexible tube for an endoscope along the axial direction, the manufacturing conditions preferably including the extrusion amount per unit time of the resin material for the inner layer and the extrusion amount per unit time of the resin material for the outer layer by the extrusion molding machine, and the predicted values ​​preferably including the elasticity value of the divided portions, the thickness of the inner layer of the divided portions, and the thickness of the outer layer of the divided portions.

[0014] The product is a sheet formed by applying a liquid to a long support body that is being transported, the divided portions are portions of the sheet that are divided along the width direction, the manufacturing conditions preferably include a control amount of a tension adjustment roller that is arranged upstream of the liquid application point and adjusts the tension applied to the support body at the application point, and the predicted value preferably includes the thickness of the liquid application in the divided portions.

[0015] Preferably, the product is cells cultured in a culture tank, the divided parts are parts obtained by dividing the culture tank, the manufacturing conditions include cell culture conditions, and the predicted value includes an index value that indicates the suitability of the culture environment for the cells.

[0016] The method of operating the information processing device disclosed herein includes using a machine learning model that outputs a predicted value of the quality of a divided portion obtained by dividing an entire product in response to input of manufacturing conditions for the divided portion; inputting the manufacturing conditions for the divided portion into the machine learning model and outputting a predicted value of the divided portion from the machine learning model; and deriving suitable manufacturing conditions that will result in a target quality for the entire product by solving an optimization problem that finds manufacturing conditions that minimize the value of an objective function having a term that includes the difference between the predicted value of the divided portion and the target value of the quality of the divided portion.

[0017] The operating program of the information processing device disclosed herein causes a computer to execute processes including using a machine learning model that outputs a predicted value of the quality of a divided portion obtained by dividing an entire product in response to input of manufacturing conditions for the divided portion; inputting the manufacturing conditions for the divided portion into the machine learning model and outputting a predicted value of the divided portion from the machine learning model; and deriving suitable manufacturing conditions that will result in a target quality for the entire product by solving an optimization problem that finds manufacturing conditions that minimize the value of an objective function having a term that includes the difference between the predicted value of the divided portion and the target value of the quality of the divided portion.

[0018] According to the technology of the present disclosure, it is possible to provide an information processing device, an operating method for an information processing device, and an operating program for an information processing device that are capable of deriving generally suitable manufacturing conditions regardless of the type of product.

[0019] 1 is a diagram illustrating an information processing device, an extruder, and an endoscope flexible tube. FIG. 2 is a diagram illustrating an extruder. FIG. 3 is a diagram illustrating a divided portion treated as one unit for predicting quality. FIG. 4 is a block diagram illustrating a computer constituting an information processing device. FIG. 5 is a block diagram illustrating a processing unit of a CPU of the information processing device. FIG. 6 is a diagram illustrating input data. FIG. 7 is a diagram illustrating a group of target values. FIG. 8 is a graph illustrating the establishment of elasticity standard values. FIG. 9 is a diagram illustrating a group of prediction models. FIG. 10 is a diagram illustrating processing by a prediction unit for an elasticity value prediction model. FIG. 11 is a diagram illustrating processing by a prediction unit for an inner layer thickness prediction model. FIG. 12 is a diagram illustrating processing by a prediction unit for an outer layer thickness prediction model. FIG. 13 is a diagram illustrating processing by a prediction unit for calculating a total thickness predicted value from a predicted inner layer thickness value and a predicted outer layer thickness value. FIG. 14 is a diagram illustrating a group of predicted values. FIG. 15 is a diagram illustrating learning data. FIG. 16 is a diagram illustrating processing in the learning phase of an elasticity value prediction model. FIG. 17 is a diagram illustrating processing in the learning phase of an inner layer thickness prediction model. FIG. 18 is a diagram illustrating processing in the learning phase of an outer layer thickness prediction model. FIG. 19 is a diagram illustrating an objective function. FIG. 20 is a diagram illustrating processing by a derivation unit and suitable manufacturing conditions. FIG. 21 is a flowchart illustrating the processing procedure of an information processing device. FIG. 22 is a diagram illustrating the derivation of suitable manufacturing conditions for a large-diameter, long endoscope flexible tube. 1 is a diagram showing how suitable manufacturing conditions for a thin-diameter, short flexible tube for an endoscope are derived. FIG. 1 is a diagram showing main parts of a sheet manufacturing apparatus of a second embodiment. FIG. 2 is a diagram showing divided parts of the second embodiment. FIG. 3 is a diagram showing input data of the second embodiment. FIG. 4 is a diagram showing a group of prediction models of the second embodiment. FIG. 5 is a diagram showing processing by a prediction unit for a tension prediction model. FIG. 6 is a diagram showing processing by a prediction unit for a hydraulic pressure prediction model. FIG. 7 is a diagram showing processing by a prediction unit for a coating thickness prediction model. FIG. 8 is a diagram showing a group of predicted values ​​of the second embodiment. FIG. 9 is a diagram showing an objective function of the second embodiment. FIG. 10 is a diagram showing processing by a derivation unit and suitable manufacturing conditions of the second embodiment. FIG. 11 is a diagram showing a culture vessel and a cell-removing filter. FIG. 12 is a diagram showing divided parts of a third embodiment. FIG. 13 is a diagram showing input data of the third embodiment. FIG. 14 is a diagram showing processing by a prediction unit for a culture environment prediction model. FIG. 15 is a diagram showing a group of predicted values ​​of the third embodiment. FIG. 16 is a diagram showing an objective function of the third embodiment. FIG. 17 is a diagram showing processing by a derivation unit and tentative suitable manufacturing conditions of the third embodiment. FIG. 18 is a diagram showing how a more suitable oxygen supply amount is derived from the oxygen concentration of the tentative suitable manufacturing conditions.

[0020] 1 as an example, an information processing device 10 outputs suitable manufacturing conditions 13 for the flexible tube 11 for an endoscope to an extruder 12 that manufactures the flexible tube 11 for an endoscope. The suitable manufacturing conditions 13 are derived using a machine learning model, as will be described later, and are manufacturing conditions under which the overall quality of the flexible tube 11 for an endoscope becomes a target value. Here, "quality becomes the target value" not only refers to a case where the quality exactly matches the target value, but also refers to a case where the quality falls within a range of the target value ±α (α is an allowable error).

[0021] The information processing device 10 is, for example, a desktop personal computer, and has a display 14 that displays various screens, and an input device 15 such as a keyboard, a mouse, a touch panel, or a microphone for voice input. The information processing device 10 is operated, for example, by an operator of the extruder 12. The information processing device 10 and the extruder 12 are connected to each other so as to be able to communicate with each other via a communication network such as a LAN (Local Area Network).

[0022] The endoscopic flexible tube 11 constitutes a part of the endoscope 16. More specifically, the endoscope 16 has an insertion section 17 that is inserted into a body cavity. The insertion section 17 has a tip section 17A, a bending section 17B, and a flexible section 17C. The tip section 17A has a built-in imaging element for capturing images inside the body cavity. The bending section 17B is connected to the tip section 17A and bends up, down, left, and right to change the orientation of the tip section 17A. The flexible section 17C is a soft, elongated tubular section that connects the bending section 17B to a proximal operating section 18 that is provided with an operating knob or the like for operating the bending section 17B, and accounts for the majority of the insertion section 17. The endoscopic flexible tube 11 constitutes this flexible section 17C. The endoscopic flexible tube 11 is an example of a "product" according to the technology of the present disclosure.

[0023] The flexible tube 11 for an endoscope is formed by coating the outer peripheral surface of a flexible tube base material 20 with a resin layer 21. The flexible tube base material 20 is configured by coating a helical tube 22 with a tubular mesh body 23. The helical tube 22 is formed by helically winding a metal strip such as stainless steel. The tubular mesh body 23 is formed by braiding metal wires such as stainless steel fibers. Bases 24 are fitted to both ends of the flexible tube base material 20. The outer diameter of the flexible tube 11 for an endoscope is, for example, 10 mm to 14 mm, and the length is, for example, 50 cm to 150 cm.

[0024] The resin layer 21 is continuously molded on the surface of the flexible tube substrate 20 by the extruder 12. The resin layer 21 has a two-layer structure formed by laminating an inner layer 25 that covers the entire circumferential surface around the axis of the flexible tube substrate 20 and an outer layer 26 that covers the entire circumferential surface around the axis of the inner layer 25. The inner layer 25 is relatively soft, and the outer layer 26 is harder than the inner layer 25. The resin material for the inner layer 25 and the outer layer 26 may be, for example, a resin containing polyurethane elastomer as a main component.

[0025] The combined thickness of the resin layer 21, including the inner layer 25 and the outer layer 26, is substantially the same in the axial direction AD of the flexible tube substrate 20, and is, for example, 0.2 mm to 1.0 mm. However, the thickness ratio between the inner layer 25 and the outer layer 26 varies in the axial direction AD. Specifically, the inner layer 25 is thicker than the outer layer 26 at the distal end 27 of the flexible tube substrate 20 connected to the bending section 17B. Meanwhile, the outer layer 26 is thicker than the inner layer 25 at the proximal end 28 of the flexible tube substrate 20 connected to the proximal operation unit 18. The thickness of the inner layer 25 gradually decreases, and the thickness of the outer layer 26 gradually increases, from the distal end 27 to the proximal end 28. The thickness ratio between the inner layer 25 and the outer layer 26 at the distal end 27 is, for example, inner layer:outer layer = 95:5 to 60:40. On the other hand, the thickness ratio of the inner layer 25 to the outer layer 26 at the proximal end 28 is, for example, inner layer:outer layer=5:95 to 40:60. By making the thickness ratio of the inner layer 25 to the outer layer 26 different in this way, the endoscopic flexible tube 11, and ultimately the flexible section 17C, is soft on the distal end 27 side and hard on the proximal end 28 side. This makes it easier to insert the insertion section 17 into a body cavity.

[0026] The outer surface of the resin layer 21 is further coated with a top coat layer 29. The material of the top coat layer 29 may be any material that is harmless to the human body, is chemical-resistant, and can withstand the high temperatures of steam sterilization, such as fluorine paint. The thickness of the top coat layer 29 is, for example, 10 μm to 200 μm. A scale indicating the length of the insertion portion 17 is printed on the surface of the top coat layer 29.

[0027] As shown in FIG. 2 , the extruder 12 includes extrusion units 35 and 36, a head unit 37, a cooling unit 38, a delivery drum 39, a take-up drum 40, and a control unit 41. The extrusion unit 35 includes a hopper (not shown), a screw 42, and the like. The extrusion unit 35 extrudes the molten resin material of the inner layer 25 toward the head unit 37. Like the extrusion unit 35, the extrusion unit 36 ​​also includes a hopper (not shown), a screw 43, and the like. The extrusion unit 36 ​​extrudes the molten resin material of the outer layer 26 toward the head unit 37. By changing the rotation speeds of the screws 42 and 43, the extrusion amounts per unit time of the resin material of the inner layer 25 and the resin material of the outer layer 26, and therefore the thicknesses of the inner layer 25 and the outer layer 26, can be changed.

[0028] The head portion 37 is composed of a nipple 44, a die 45, and a support 46 that fixedly supports the nipple 44 and the die 45. A circular molding passage 47 is formed in the center of the nipple 44. The entrance of the molding passage 47 is tapered and widened. A flexible tube substrate connected body 20C is inserted into the molding passage 47. The flexible tube substrate connected body 20C is formed by connecting multiple flexible tube substrates 20 with connecting members 48. The flexible tube substrate connected body 20C is introduced into the molding passage 47 with the base ends 28 of the flexible tube substrates 20 facing forward and the tip ends 27 facing backward. The diameter of the molding passage 47 is slightly larger than the outer diameter of the flexible tube substrate 20. Note that the flexible tube substrate connected body 20C may also be introduced into the molding passage 47 with the tip ends 27 of the flexible tube substrates 20 facing forward and the base ends 28 facing backward, inversely to the above.

[0029] The interior of the die 45 is heated to a predetermined molding temperature, for example, about 150°C to 300°C, by a heater (not shown). A resin passage 49 is formed between the nipple 44 and the die 45, surrounding the entire circumference of the molding passage 47. The resin passage 49 is connected to gates 50 and 51 formed in the support body 46. Like the resin passage 49, the gates 50 and 51 are also formed to surround the entire circumference of the molding passage 47. The resin material for the inner layer 25 and the resin material for the outer layer 26 are extruded through the gates 50 and 51 and supplied to the molding passage 47 through the resin passage 49, with the resin material for the inner layer 25 on the bottom and the resin material for the outer layer 26 on top.

[0030] The cooling unit 38 is, for example, a water tank that stores cooling water. The cooling unit 38 cools the resin material of the inner layer 25 and the resin material of the outer layer 26 that have been extruded into the flexible tube base material connected body 20C in the head unit 37.

[0031] The flexible tube substrate connected body 20C before extrusion molding is wound around the feed drum 39. Meanwhile, the flexible tube substrate connected body 20C after extrusion molding is taken up on the take-up drum 40. A motor (not shown) is connected to the feed drum 39 and the take-up drum 40, and the feed drum 39 and the take-up drum 40 rotate in response to the driving of the motor. As the feed drum 39 and the take-up drum 40 rotate, the flexible tube substrate connected body 20C is sent out from the feed drum 39 toward the take-up drum 40 (the head unit 37 and the cooling unit 38).

[0032] The conveying speed of the flexible tube substrate connected body 20C is constant by the delivery drum 39 and the take-up drum 40. By constantly extruding the resin material of the inner layer 25 and the resin material of the outer layer 26 onto the flexible tube substrate connected body 20C conveyed at this constant conveying speed, the resin layer 21 can be continuously formed on multiple flexible tube substrates 20.

[0033] The control unit 41 receives suitable manufacturing conditions 13 from the information processing device 10. The control unit 41 controls the operation of each part of the extruder 12 in accordance with the suitable manufacturing conditions 13. The control unit 41 controls the rotation speeds of the screws 42 and 43 to control the extrusion amount per unit time of the resin material for the inner layer 25 and the resin material for the outer layer 26. The control unit 41 also controls the driving of the motors for the delivery drum 39 and the take-up drum 40 to maintain a constant conveying speed of the flexible tube substrate linked body 20C. In addition, the control unit 41 controls the driving of the heater of the die 45 to control the molding temperature, and controls the temperature of the cooling water in the cooling unit 38 to maintain a constant temperature.

[0034] When molding the resin layer 21 on one flexible tube substrate 20, the control unit 41 controls the outer layer 26 to be thicker than the inner layer 25 at the base end 28, gradually decreasing the thickness of the outer layer 26 and gradually increasing the thickness of the inner layer 25 toward the tip 27, until the inner layer 25 is thicker than the outer layer 26 at the tip 27. After completing the extrusion molding of one flexible tube substrate 20, the control unit 41 switches the extrusion amount per unit time of the resin material for the inner layer 25 and the extrusion amount per unit time of the resin material for the outer layer 26 at the connecting member 48 in preparation for the extrusion molding of the next flexible tube substrate 20.

[0035] The flexible tube substrate connected body 20C wound around the winding drum 40 has the connecting members 48 removed and is separated into individual flexible tube substrates 20. Each flexible tube substrate 20 is then coated with a top coat layer 29. Finally, a mouthpiece 24 is attached to the distal end 27 and proximal end 28 of each flexible tube substrate 20. This completes the flexible tube 11 for an endoscope. The flexible tube 11 for an endoscope is then transferred to the assembly process for the endoscope 16.

[0036] As an example, as shown in FIG. 3 , the information processing device 10 treats divided portions P1, P2, P3, P4, P5, P6, P7, and P8 (arranged in this order from the base end 28) obtained by dividing a single flexible tube substrate 20 along the axial direction AD, and divided portions P9 and P10 (arranged in this order from the base end 28) obtained by dividing a connecting member 48 along the axial direction, as a single unit for predicting quality. Here, divided portion P9 treated as a single unit for predicting quality is the portion of connecting member 48N connecting the current flexible tube substrate 20 to the next flexible tube substrate 20. Furthermore, divided portion P10 treated as a single unit for predicting quality is the portion of connecting member 48P connecting the current flexible tube substrate 20 to the previous flexible tube substrate 20. The divided portions P1 to P10 may have the same length or different lengths. Hereinafter, when there is no need to particularly distinguish between divided portions P1 to P10, they may be referred to as divided portion P.

[0037] 4, the computer constituting the information processing device 10 includes, in addition to the display 14 and input device 15, a storage 55, a memory 56, a CPU (Central Processing Unit) 57, and a communication unit 58. These are interconnected via a bus line 59.

[0038] The storage 55 is a hard disk drive built into a computer constituting the information processing device 10 or connected via a cable or network. Alternatively, the storage 55 is a disk array consisting of multiple hard disk drives. The storage 55 stores control programs such as an operating system, various application programs, and various data associated with these programs. Note that a solid state drive may be used instead of a hard disk drive.

[0039] The memory 56 is a work memory for the CPU 57 to execute processing. The CPU 57 loads programs stored in the storage 55 into the memory 56 and executes processing in accordance with the programs. In this way, the CPU 57 comprehensively controls each part of the computer. The CPU 57 is an example of a "processor" according to the technology of the present disclosure. The memory 56 may be built into the CPU 57. The communication unit 58 controls the transmission of various information to and from external devices such as the extruder 12.

[0040] 5 , an operating program 65 is stored in the storage 55 of the information processing device 10. The operating program 65 is an application program for causing a computer to function as the information processing device 10. In other words, the operating program 65 is an example of an “operating program of an information processing device” according to the technology of the present disclosure. The storage 55 also stores a group of prediction models 66, an objective function 67, and the like.

[0041] When the operating program 65 is started, the CPU 57 of the computer constituting the information processing device 10 works in cooperation with the memory 56 and the like to function as a reception unit 70, a read / write (hereinafter abbreviated as RW (Read Write)) control unit 71, a prediction unit 72, a derivation unit 73, and a distribution control unit 74.

[0042] The reception unit 70 receives various information input by an operator via the input device 15. For example, the reception unit 70 receives input data 75. The input data 75 is data to be input to an elasticity value prediction model 90, an inner layer thickness prediction model 91, and an outer layer thickness prediction model 92 (see FIG. 9 ) that constitute the prediction model group 66. The reception unit 70 outputs the input data 75 to the RW control unit 71.

[0043] The receiving unit 70 also receives a target value group 76. The target value group 76 is a collection of target values ​​for the quality of each divided portion P. The receiving unit 70 outputs the target value group 76 to the RW control unit 71. Although not shown in the figure, the receiving unit 70 also receives an instruction to derive suitable manufacturing conditions 13, an instruction to distribute suitable manufacturing conditions 13, etc.

[0044] The RW control unit 71 controls the storage of various types of information in the storage 55 and the reading of various types of information from the storage 55. For example, the RW control unit 71 stores input data 75 and a target value group 76 from the reception unit 70 in the storage 55. The RW control unit 71 also reads the input data 75 from the storage 55 and outputs the read input data 75 to the prediction unit 72. The RW control unit 71 reads a prediction model group 66 from the storage 55 and outputs the read prediction model group 66 to the prediction unit 72. The RW control unit 71 also reads an objective function 67 and a target value group 76 from the storage 55 and outputs the read objective function 67 and target value group 76 to the derivation unit 73.

[0045] The prediction unit 72 operates in response to an instruction to derive suitable manufacturing conditions 13. The prediction unit 72 inputs input data 75 to the elasticity prediction model 90, the inner layer thickness prediction model 91, and the outer layer thickness prediction model 92 that make up the prediction model group 66. The prediction unit 72 then causes the elasticity prediction model 90, the inner layer thickness prediction model 91, and the outer layer thickness prediction model 92 to output predicted values ​​for the quality of the divided portion P. The prediction unit 72 outputs a prediction value group 77, which is a collection of predicted values ​​output from the elasticity prediction model 90, the inner layer thickness prediction model 91, and the outer layer thickness prediction model 92, to the derivation unit 73.

[0046] Like the prediction unit 72, the derivation unit 73 operates in response to an instruction to derive suitable manufacturing conditions 13. Here, the objective function 67 includes, as variables, target values ​​from the target value group 76 and predicted values ​​from the predicted value group 77. The derivation unit 73 derives suitable manufacturing conditions 13 by solving an optimization problem that finds manufacturing conditions that minimize the value of the objective function 67, for example, by substituting the target values ​​from the target value group 76 and the predicted values ​​from the predicted value group 77 into the objective function 67. The derivation unit 73 outputs the suitable manufacturing conditions 13 to the distribution control unit 74. Well-known black-box optimization methods such as genetic algorithms, evolutionary strategies, and Bayesian optimization can be used to solve the optimization problem.

[0047] The distribution control unit 74 controls the distribution of the suitable manufacturing conditions 13 to the extruder 12 specified in the distribution instruction for the suitable manufacturing conditions 13 .

[0048] As an example, as shown in FIG. 6, the input data 75 includes manufacturing conditions for each of the divided portions P1 to P10, which are treated as a single unit for predicting quality as shown in FIG. 3. More specifically, the input data 75 includes manufacturing conditions 80_P1 for the divided portion P1, manufacturing conditions 80_P2 for the divided portion P2, ..., manufacturing conditions 80_P9 for the divided portion P9, and manufacturing conditions 80_P10 for the divided portion P10. Note that in the figure, the manufacturing conditions 80_P1, etc. are written as "manufacturing conditions (P1)" etc. The same applies to the elasticity prediction value 95, etc., which will appear later. Hereinafter, when there is no need to particularly distinguish between the manufacturing conditions 80_P1 to 80_P10 for the divided portions P1 to P10, they may be written as manufacturing conditions 80.

[0049] The manufacturing conditions 80 include the extrusion amount per unit time of the resin material for the inner layer 25 in each divided portion P (the number of rotations of the screw 42), the extrusion amount per unit time of the resin material for the outer layer 26 in each divided portion P (the number of rotations of the screw 43), and the conveying speed of the flexible tube base material linked body 20C (the number of rotations of the delivery drum 39 and the take-up drum 40). The extrusion amount per unit time of the resin material for the inner layer 25 and the extrusion amount per unit time of the resin material for the outer layer 26 differ for each manufacturing condition 80 depending on the thickness of the inner layer 25 and the outer layer 26 of each divided portion P. In contrast, the conveying speed of the flexible tube base material linked body 20C is common to all manufacturing conditions 80. The operator inputs manufacturing conditions 80 that are considered to result in a target quality for the entire flexible tube for endoscope 11, based on past manufacturing results of flexible tubes 11 for endoscopes and his or her own experience. In addition to the above, the manufacturing conditions 80 may also appropriately include the extrusion pressure and molding temperature of the resin material of the inner layer 25 and the resin material of the outer layer 26. The extrusion pressure and molding temperature are common to all manufacturing conditions 80, as is the conveying speed of the flexible tube substrate connected body 20C.

[0050] The input data 75 includes physical property information 81. The physical property information 81 is information regarding the physical properties of the materials constituting the endoscopic flexible tube 11, in this case the resin material of the inner layer 25 and the resin material of the outer layer 26. The physical property information 81 includes the hardness (e.g., Shore A hardness) of the resin material of the inner layer 25, the tensile tension (e.g., 100% modulus) of the resin material of the inner layer 25, the elasticity (e.g., flexural modulus) of the resin material of the outer layer 26, and the viscosity (e.g., melt viscosity) of the resin material of the outer layer 26. In addition to these, the physical property information 81 may also include the tensile strength of the resin material of the inner layer 25, the elongation of the resin material of the inner layer 25, the viscosity ratio of the resin material of the outer layer 26, and the like, as appropriate.

[0051] The input data 75 also includes design information 82. The design information 82 is information relating to the design values ​​of the endoscopic flexible tube 11. The design information 82 includes the outer diameter of the flexible tube base material 20, the length of each divided portion P, and the like. The length of each divided portion P may be replaced with the distance between two adjacent divided portions P. In addition to the above, the design information 82 may also include, as appropriate, the amount of additives to the resin material of the inner layer 25 and the resin material of the outer layer 26.

[0052] 7, the target value group 76 has standard values ​​for the elasticity of the resin layer 21 of each of the divided portions P1 to P8 (hereinafter referred to as standard elasticity values). More specifically, the target value group 76 has a standard elasticity value 85_P1 for the divided portion P1, a standard elasticity value 85_P2 for the divided portion P2, ..., a standard elasticity value 85_P7 for the divided portion P7, and a standard elasticity value 85_P8 for the divided portion P8. Hereinafter, when there is no need to particularly distinguish between the standard elasticity values ​​85_P1 to 85_P8 for the divided portions P1 to P8, they may be referred to as standard elasticity value 85. The standard elasticity value 85 is an example of a "target value" according to the technology of the present disclosure.

[0053] The target value group 76 also has a set value 86 for the total thickness of the resin layer 21, which is the sum of the thicknesses of the inner layer 25 and the outer layer 26 (hereinafter referred to as the total thickness set value). There is only one total thickness set value 86 because it is common to each divided portion P. Like the elasticity specification value 85, the total thickness set value 86 is also an example of a "target value" according to the technology of the present disclosure.

[0054] 8 , the elasticity specification value 85 is, for example, the median value between the upper and lower limit values ​​of the elasticity of the resin layer 21 of each divided portion P. The upper and lower limit values ​​of the elasticity of the resin layer 21 of each divided portion P, and therefore the elasticity specification value 85 of the resin layer 21 of each divided portion P, decrease from the divided portion P1 on the base end 28 side to the divided portion P8 on the tip end 27 side. This is because the tip end 27 side is relatively soft because the inner layer 25 is thicker than the outer layer 26, and the base end 28 side is relatively hard because the outer layer 26 is thicker than the inner layer 25.

[0055] As an example, as shown in FIG. 9 , the prediction model group 66 includes an elasticity prediction model 90, an inner layer thickness prediction model 91, and an outer layer thickness prediction model 92. The elasticity prediction model 90, the inner layer thickness prediction model 91, and the outer layer thickness prediction model 92 are machine learning models configured, for example, by a neural network. In other words, the elasticity prediction model 90, the inner layer thickness prediction model 91, and the outer layer thickness prediction model 92 are examples of "machine learning models" according to the technology of the present disclosure. Note that, hereinafter, the elasticity prediction model 90, the inner layer thickness prediction model 91, and the outer layer thickness prediction model 92 may be collectively referred to as prediction models 90 to 92.

[0056] The elasticity prediction model 90 outputs a predicted value of the elasticity of the resin layer 21 of the divided portion P (hereinafter referred to as the predicted elasticity value) in response to inputs of the manufacturing conditions 80, the physical property information 81, and the design information 82. The inner layer thickness prediction model 91 outputs a predicted value of the thickness of the inner layer 25 of the divided portion P (hereinafter referred to as the predicted inner layer thickness value) in response to inputs of the manufacturing conditions 80, the physical property information 81, and the design information 82. The outer layer thickness prediction model 92 outputs a predicted value of the thickness of the outer layer 26 of the divided portion P (hereinafter referred to as the predicted outer layer thickness value) in response to inputs of the manufacturing conditions 80, the physical property information 81, and the design information 82. In this way, there are multiple types of predicted values, and prediction models 90 to 92 are prepared for each of the multiple types of predicted values.

[0057] 10 , when predicting the quality of a divided portion P1, the prediction unit 72 inputs a manufacturing condition 80_P1 of the divided portion P1 and manufacturing conditions 80_P10 and 80_P2 of the divided portions P10 and P2 on either side of the divided portion P1 into an elasticity prediction model 90. The prediction unit 72 also inputs physical property information 81 and design information 82 into the elasticity prediction model 90. The prediction unit 72 then outputs a predicted elasticity value 95_P1 of the divided portion P1 from the elasticity prediction model 90. The reason for inputting the manufacturing conditions 80_P10 and 80_P2 in addition to the manufacturing condition 80_P1 into the elasticity prediction model 90 is that the elasticity values ​​of the resin layers 21 of the divided portions P10 and P2 on either side of the divided portion P1 are thought to affect the elasticity value of the resin layer 21 of the divided portion P1.

[0058] As an example, as shown in FIG. 11 , when predicting the quality of a divided portion P1, the prediction unit 72 inputs manufacturing conditions 80A_P1 for the divided portion P1 (excluding information about the outer layer 26), physical property information 81A (excluding information about the outer layer 26), and design information 82 into an inner layer thickness prediction model 91. The prediction unit 72 then outputs a predicted inner layer thickness value 96_P1 for the divided portion P1 from the inner layer thickness prediction model 91. The manufacturing conditions 80A_P1 for the divided portion P1 (excluding information about the outer layer 26) are information excluding, for example, the extrusion amount per unit time of the resin material of the outer layer 26 in the divided portion P1. The physical property information 81A (excluding information about the outer layer 26) is information excluding, for example, the elasticity value and viscosity value of the resin material of the outer layer 26. The information about the outer layer 26 is excluded because the information about the outer layer 26 is considered unnecessary for predicting the predicted inner layer thickness value.

[0059] As an example, as shown in FIG. 12 , when predicting the quality of a divided portion P1, the prediction unit 72 inputs manufacturing conditions 80B_P1 for the divided portion P1 (excluding information about the inner layer 25), physical property information 81B (excluding information about the inner layer 25), and design information 82 into an outer layer thickness prediction model 92. The prediction unit 72 then outputs a predicted outer layer thickness value 97_P1 for the divided portion P1 from the outer layer thickness prediction model 92. The manufacturing conditions 80B_P1 for the divided portion P1 (excluding information about the inner layer 25) are information excluding, for example, the extrusion amount per unit time of the resin material of the inner layer 25 in the divided portion P1. Furthermore, the physical property information 81B (excluding information about the inner layer 25) is information excluding, for example, the hardness and tensile tension of the resin material of the inner layer 25. The reason for excluding the information about the inner layer 25 is that, as with the inner layer thickness prediction model 91, information about the inner layer 25 is considered unnecessary for predicting the predicted outer layer thickness.

[0060] As an example, as shown in Figure 13, the prediction unit 72 adds the inner layer thickness prediction value 96_P1 and the outer layer thickness prediction value 97_P1 of the divided portion P1 to obtain a prediction value 98_P1 of the total thickness of the resin layer 21 of the divided portion P1 (hereinafter referred to as the total thickness prediction value).

[0061] 10 to 13 show an example in which the elasticity prediction value 95_P1, inner layer thickness prediction value 96_P1, and outer layer thickness prediction value 97_P1 for the divided portion P1 are output from each prediction model 90 to 92, and a total thickness prediction value 98_P1 for the divided portion P1 is calculated. However, the prediction unit 72 similarly outputs elasticity prediction values ​​95_P2 to 95_P8 (see FIG. 14), inner layer thickness prediction values ​​96_P2 to 96_P8 (not shown), and outer layer thickness prediction values ​​97_P2 to 97_P8 (not shown) for the remaining divided portions P2 to P8 from each prediction model 90 to 92, and calculates total thickness prediction values ​​98_P2 to 98_P8 (see FIG. 14) for the remaining divided portions P2 to P8. Hereinafter, when there is no need to particularly distinguish between the predicted elasticity values ​​95_P1 to 95_P8, the predicted inner layer thickness values ​​96_P1 to 96_P8, the predicted outer layer thickness values ​​97_P1 to 97_P8, and the predicted total thickness values ​​98_P1 to 98_P8, they may be referred to as the predicted elasticity value 95, the predicted inner layer thickness value 96, the predicted outer layer thickness value 97, and the predicted total thickness value 98, respectively. Note that the predicted elasticity value 95, the predicted inner layer thickness value 96, the predicted outer layer thickness value 97, and the predicted total thickness value 98 are examples of the "predicted value" according to the technology of the present disclosure.

[0062] By calculating the predicted elasticity value 95 and the predicted total thickness value 98 for each divided portion P in this manner, the predicted value group 77 becomes, for example, as shown in Figure 14. That is, the predicted value group 77 has a predicted elasticity value 95_P1 for divided portion P1, a predicted elasticity value 95_P2 for divided portion P2, ..., a predicted elasticity value 95_P7 for divided portion P7, and a predicted elasticity value 95_P8 for divided portion P8. The predicted value group 77 also has a predicted total thickness value 98_P1 for divided portion P1, a predicted total thickness value 98_P2 for divided portion P2, ..., a predicted total thickness value 98_P7 for divided portion P7, and a predicted total thickness value 98_P8 for divided portion P8.

[0063] As an example, as shown in Figure 15, the learning data 100 is data collected from a flexible tube for an endoscope 11 manufactured in the past, for use in learning each of the prediction models 90 to 92. The learning data 100 includes learning input data 75L and correct answer data 101. The learning input data 75L is data corresponding to the input data 75. The learning input data 75L has learning manufacturing conditions 80L, learning physical property information 81L, and learning design information 82L. The learning manufacturing conditions 80L are data corresponding to the manufacturing conditions 80, the learning physical property information 81L is data corresponding to the physical property information 81, and the learning design information 82L is data corresponding to the design information 82.

[0064] As can be seen from the training manufacturing conditions 80L, the training data 100 is a collection of data on divided portions P manufactured using various extrusion amounts per unit time of the resin material for the inner layer 25, various extrusion amounts per unit time of the resin material for the outer layer 26, and various conveying speeds of the flexible tube base material connected body 20C. Furthermore, as can be seen from the training physical property information 81L, the training data 100 is also a collection of data on divided portions P using inner layers 25 and outer layers 26 with various physical properties. Furthermore, as can be seen from the training design information 82L, the training data 100 is also a collection of data on divided portions P with various design values, such as outer diameters of the flexible tube base material 20 of 10 mm, 5 mm, and 8 mm, and lengths of the divided portions P of 100 mm and 50 mm.

[0065] One set of correct answer data 101 is registered corresponding to each piece of learning input data 75L. The correct answer data 101 is data for comparing the outputs from each prediction model 90-92, i.e., the predicted elasticity value 95, the predicted inner layer thickness value 96, and the predicted outer layer thickness value 97. The correct answer data 101 includes an actual measured value of the elasticity of each divided portion P of a previously manufactured flexible tube 11 for an endoscope (hereinafter referred to as an actual elasticity value) 95CA, an actual measured value of the thickness of the inner layer 25 of each divided portion P of a previously manufactured flexible tube 11 for an endoscope (hereinafter referred to as an actual inner layer thickness value) 96CA, and an actual measured value of the thickness of the outer layer 26 of each divided portion P of a previously manufactured flexible tube 11 for an endoscope (hereinafter referred to as an actual outer layer thickness value) 97CA.

[0066] As an example, as shown in FIGS. 16 to 18 , in the learning phase of each prediction model 90-92, learning input data 75L is provided and learning is performed. Specifically, as shown in FIG. 16 , in the learning phase of the elasticity value prediction model 90, learning manufacturing conditions 80L, learning physical property information 81L, and learning design information 82L are input to the elasticity value prediction model 90. The learning manufacturing conditions 80L are the learning manufacturing conditions 80L of the divided portion P that predicts the predicted elasticity value 95, here, the learning manufacturing conditions 80L_P1 of the divided portion P1. The learning manufacturing conditions 80L are also the learning manufacturing conditions 80L of the divided portions P on both sides of the divided portion P that predicts the predicted elasticity value 95, here, the learning manufacturing conditions 80L_P10 of the divided portion P10 and the learning manufacturing conditions 80L_P2 of the divided portion P2. In response to the input of such learning input data 75L, the elasticity prediction model 90 outputs a learning elasticity predicted value 95L, here a learning elasticity predicted value 95L_P1 for the divided portion P1.

[0067] In the learning phase of the elasticity prediction model 90, a loss calculation is performed for the elasticity prediction model 90 using a loss function based on the learning elasticity predicted value 95L, here the learning elasticity predicted value 95L_P1 for the divided portion P1, and the actual elasticity value 95CA corresponding to the learning input data 75L, here the actual elasticity measured value 95CA_P1 for the divided portion P1. Then, depending on the result of the loss calculation, various coefficients of the elasticity prediction model 90 (such as the coefficients of the filter in the convolution layer) are updated and the elasticity prediction model 90 is updated according to the update settings.

[0068] 17 , in the learning phase of the inner layer thickness prediction model 91, learning manufacturing conditions 80LA excluding information on the outer layer 26, learning physical property information 81LA excluding information on the outer layer 26, and learning design information 82L are input to the inner layer thickness prediction model 91. The inner layer thickness prediction model 91 outputs a learning inner layer thickness predicted value 96L in response to the input of such learning input data 75L.

[0069] In the learning phase of the inner layer thickness prediction model 91, a loss calculation is performed for the inner layer thickness prediction model 91 using a loss function based on the learning inner layer thickness predicted value 96L and the inner layer thickness actual measurement value 96CA corresponding to the learning input data 75L. Then, various coefficients of the inner layer thickness prediction model 91 are updated according to the results of the loss calculation, and the inner layer thickness prediction model 91 is updated according to the updated settings.

[0070] 18 , in the learning phase of outer layer thickness prediction model 92, learning manufacturing conditions 80LB excluding information on inner layer 25, learning physical property information 81LB excluding information on inner layer 25, and learning design information 82L are input to outer layer thickness prediction model 92. In response to the input of such learning input data 75L, outer layer thickness prediction model 92 outputs a learning outer layer thickness predicted value 97L.

[0071] In the learning phase of the outer layer thickness prediction model 92, a loss calculation is performed for the outer layer thickness prediction model 92 using a loss function based on the learning outer layer thickness predicted value 97L and the outer layer thickness actual measurement value 97CA corresponding to the learning input data 75L. Then, various coefficients of the outer layer thickness prediction model 92 are updated according to the results of the loss calculation, and the outer layer thickness prediction model 92 is updated according to the updated settings.

[0072] During the learning phase of each prediction model 90-92, the above-mentioned series of processes, including input of learning input data 75L, output of learning elasticity prediction value 95L, learning inner layer thickness prediction value 96L, and learning outer layer thickness prediction value 97L, loss calculation, update setting, and update, are repeatedly performed while the learning input data 75L and the correct answer data 101 are exchanged. The repetition of the above-mentioned series of processes is terminated when the prediction accuracy of the learning elasticity prediction value 95L, learning inner layer thickness prediction value 96L, and learning outer layer thickness prediction value 97L reaches a predetermined set level. Each prediction model 90-92 whose prediction accuracy has reached the set level is stored in the storage 55 of the information processing device 10. Note that learning may be terminated after the above-mentioned series of processes have been repeated a set number of times, regardless of the prediction accuracy of the learning elasticity prediction value 95L, learning inner layer thickness prediction value 96L, and learning outer layer thickness prediction value 97L. Learning may also be continued after storage in the storage 55.

[0073] As an example, as shown in FIG. 19 , the objective function 67 has a first term 105, a second term 106, a third term 107, and a fourth term 108 that are added together. The first term 105 includes the difference between the elastic predicted value 95 of each divided portion P and the elastic normalized value 85 of each divided portion P. More specifically, the first term 105 includes the sum of the squares of the value obtained by dividing the difference between the elastic predicted value 95 and the elastic normalized value 85 by the elastic normalized value 85. The first term 105 is a term obtained by multiplying this sum by a first weighting coefficient W1. The first term 105 is an example of a "term including a difference" according to the technology of the present disclosure.

[0074] The second term 106 includes the difference between the total thickness predicted value 98 of each divided portion P and the total thickness set value 86. More specifically, the second term 106 includes the sum of the squares of the value obtained by dividing the difference between the total thickness predicted value 98 and the total thickness set value 86 by the total thickness set value 86. The second term 106 is a term obtained by multiplying this sum by a second weighting coefficient W2. Like the first term 105, the second term 106 is an example of a "term including a difference" according to the technology of the present disclosure.

[0075] The third term 107 and the fourth term 108 are regularization terms for fitting the manufacturing conditions of each of the divided portions P into the constraints on the manufacturing conditions of the entire flexible tube 11. The constraints on the manufacturing conditions of the entire flexible tube 11 are intended to smoothly connect the manufacturing conditions of each of the divided portions P in order to smoothly change the elasticity value, inner layer thickness, and outer layer thickness throughout the entire flexible tube 11. Therefore, the third term 107 is a term that smooths the extrusion amount R per unit time of the resin material of the inner layer 25 of each divided portion P. Furthermore, the fourth term 108 is a term that smooths the extrusion amount S per unit time of the resin material of the outer layer 26 of each divided portion P.

[0076] More specifically, the third term 107 includes the sum of the second-order differential values ​​of the extrusion amount R per unit time of the resin material of each inner layer 25 of the divided portion P. The third term 107 is a term obtained by multiplying this sum by a third weighting coefficient W3. The fourth term 108 includes the sum of the second-order differential values ​​of the extrusion amount S per unit time of the resin material of each outer layer 26 of the divided portion P. The fourth term 108 is a term obtained by multiplying this sum by a fourth weighting coefficient W4. Here, |R in the third term 107 ｉ－１ +R ｉ＋１ -2R ｉ is an approximation of the second derivative of the extrusion amount R per unit time of the resin material of the inner layer 25 of each divided portion P. ｉ－１ +S ｉ＋１ -2S ｉ | is an approximation of the second derivative of the extrusion amount S of the resin material of each outer layer 26 of the divided portion P per unit time.

[0077] N in the first term 105 to the fourth term 108 is the number of divided portions P. The first weighting coefficient W1 to the fourth weighting coefficient W4 are values ​​that sum to 1 (W1 + W2 + W3 + W4 = 1). For example, W1 to W4 = 0.25. However, the first weighting coefficient W1 to the fourth weighting coefficient W4 do not necessarily have to be the same value. For example, if emphasis is placed on the third term 107, that is, if the extrusion rate R per unit time of the resin material of each inner layer 25 of the divided portions P is to be more smoothed, the third weighting coefficient W3 may be set to a higher value than the others.

[0078] As an example, as shown in FIG. 20 , the preferred manufacturing conditions 13 derived by the derivation unit 73 by solving an optimization problem for determining the manufacturing conditions that minimize the value of the objective function 67 include the preferred extrusion amount per unit time of the resin material of the inner layer 25 in each of the divided portions P1 to P8, the preferred extrusion amount per unit time of the resin material of the outer layer 26 in each of the divided portions P1 to P8, and the preferred conveying speed of the flexible tube substrate linked body 20C in each of the divided portions P1 to P8.

[0079] Here, solving the optimization problem to find the manufacturing conditions that minimize the value of the objective function 67 means searching for suitable manufacturing conditions 13 that minimize the difference between the predicted elasticity value 95 and the specified elasticity value 85 in the first term 105, and the difference between the predicted total thickness value 98 and the set total thickness value 86 in the second term 106. Also, solving the optimization problem to find the manufacturing conditions that minimize the value of the objective function 67 means searching for suitable manufacturing conditions 13 that smooth the extrusion rate R per unit time of the resin material for the inner layer 25 of each divided portion P and the extrusion rate per unit time of the resin material for the outer layer 26 of each divided portion P.

[0080] Next, the operation of the above configuration will be described with reference to the flowchart in Fig. 21. First, when the operating program 65 is started in the information processing device 10, the CPU 57 of the information processing device 10 functions as a reception unit 70, a RW control unit 71, a prediction unit 72, a derivation unit 73, and a distribution control unit 74, as shown in Fig. 5.

[0081] First, an input screen (not shown) for input data 75 and a target value group 76 is displayed on the display 14 of the information processing device 10. The operator inputs the desired input data 75 and target value group 76 on the input screen. The input data 75 and target value group 76 are then received by the reception unit 70 (step ST100). The input data 75 and target value group 76 are output from the reception unit 70 to the RW control unit 71. The RW control unit 71 then stores the input data 75 in the storage 55. As shown in FIG. 6 , the input data 75 includes manufacturing conditions 80, physical property information 81, and design information 82 for each divided portion P. As shown in FIG. 7 , the target value group 76 includes an elasticity specification value 85 and a total thickness setting value 86 for each divided portion P. Note that separate input screens for the input data 75 and the target value group 76 may be used, and the input timings for the input data 75 and the target value group 76 may be different.

[0082] When the receiving unit 70 receives an instruction to derive suitable manufacturing conditions 13, the input data 75 is read from the storage 55 by the RW control unit 71 and output to the prediction unit 72. In addition, the target value group 76 is read from the storage 55 by the RW control unit 71 and output to the derivation unit 73.

[0083] As shown in Fig. 10, in the prediction unit 72, input data 75 is input to an elasticity prediction model 90, and a predicted elasticity value 95 for each divided portion P is output from the elasticity prediction model 90. Also, as shown in Fig. 11, in the prediction unit 72, input data 75 is input to an inner layer thickness prediction model 91, and a predicted inner layer thickness value 96 for each divided portion P is output from the inner layer thickness prediction model 91. Furthermore, as shown in Fig. 12, in the prediction unit 72, input data 75 is input to an outer layer thickness prediction model 92, and a predicted outer layer thickness value 97 for each divided portion P is output from the outer layer thickness prediction model 92 (step ST110). Then, as shown in Fig. 13, in the prediction unit 72, the predicted inner layer thickness value 96 and the predicted outer layer thickness value 97 are added together to determine a predicted total thickness value 98 for each divided portion P. The elasticity predicted value 95 and total thickness predicted value 98 thus obtained are output from the prediction unit 72 to the derivation unit 73 as a predicted value group 77, as shown in FIG.

[0084] 20, the derivation unit 73 solves an optimization problem for determining the manufacturing conditions that minimize the value of the objective function 67 shown in FIG. 19, thereby deriving the suitable manufacturing conditions 13 (step ST120). The suitable manufacturing conditions 13 are output from the derivation unit 73 to the distribution control unit 74.

[0085] When the receiving unit 70 receives an instruction to distribute the suitable manufacturing conditions 13, the distribution control unit 74 distributes the suitable manufacturing conditions 13 to the extruder 12 specified in the distribution instruction (step ST130).

[0086] In the extruder 12, the suitable manufacturing conditions 13 are input from the information processing device 10 to the control unit 41. Then, under the control of the control unit 41, the operation of each part of the extruder 12 is controlled under the suitable manufacturing conditions 13, and the flexible tube 11 for an endoscope is manufactured.

[0087] As described above, the CPU 57 of the information processing device 10 functions as the prediction unit 72 and the derivation unit 73. The prediction unit 72 uses an elasticity prediction model 90, an inner layer thickness prediction model 91, and an outer layer thickness prediction model 92 that output a predicted elasticity value 95, a predicted inner layer thickness value 96, and a predicted outer layer thickness value 97 of a divided portion P in response to input of manufacturing conditions 80 for the divided portions P obtained by dividing the entire endoscopic flexible tube 11. The prediction unit 72 inputs the manufacturing conditions 80 for each of the divided portions P to the elasticity prediction model 90, the inner layer thickness prediction model 91, and the outer layer thickness prediction model 92, and causes the elasticity prediction model 90, the inner layer thickness prediction model 91, and the outer layer thickness prediction model 92 to output the predicted elasticity value 95, the predicted inner layer thickness value 96, and the predicted outer layer thickness value 97 for each of the divided portions P. The derivation unit 73 solves an optimization problem to find manufacturing conditions that minimize the value of an objective function 67 having a first term 105 including the difference between the predicted elasticity value 95 of each divided portion P and the standard elasticity value 85 of each divided portion P, and a second term 106 including the difference between the predicted total thickness value 98 of each divided portion P and the set total thickness value 86 of each divided portion P, thereby deriving suitable manufacturing conditions 13 that will result in the overall quality of the flexible tube 11 for the endoscope becoming the target value.

[0088] 22, for example, for a large-diameter, long flexible tube 11LL for an endoscope used in an oral endoscope for the upper gastrointestinal tract, predicted elastic values ​​95_P1 to 95_P10, etc. of the divided portions P1 to P10 can be output from an elasticity prediction model 90, etc., to derive suitable manufacturing conditions 13. Also, for a small-diameter, short flexible tube 11SS for an endoscope used in a transnasal endoscope for the bronchi, predicted elastic values ​​95_P1 to 95_P6, etc. of the divided portions P1 to P6 can be output from an elasticity prediction model 90, etc., to derive suitable manufacturing conditions 13.

[0089] When predicting the overall quality of the large-diameter, long flexible tube 11LL for an endoscope using a prediction model dedicated to the large-diameter, long flexible tube 11LL, it is necessary to prepare the manufacturing conditions for each of the divided parts P1 to P10 as input data. Therefore, a prediction model dedicated to the large-diameter, long flexible tube 11LL for an endoscope cannot predict the overall quality of the small-diameter, short flexible tube 11SS for an endoscope, since the divided parts P1 to P6 are the only six divided parts, and the number of manufacturing conditions as input data is insufficient. As a result, it is not possible to derive the suitable manufacturing conditions 13 for the small-diameter, short flexible tube 11SS for an endoscope. The same is also true in reverse. In contrast, according to the technology disclosed herein, as described above, the prediction models 90 to 92 can be used to derive the suitable manufacturing conditions 13 for the large-diameter, long flexible tube 11LL for an endoscope, as well as the suitable manufacturing conditions 13 for the small-diameter, short flexible tube 11SS for an endoscope. That is, it is possible to derive suitable manufacturing conditions 13 for general use regardless of the type of the flexible tube 11 for an endoscope.

[0090] As shown in Figure 15, the learning data 100 for each prediction model 90-92 is not data on the entire flexible tube 11 for an endoscope, but data on each divided portion P. This makes it easy to collect a very large amount of data. The greater the amount of learning data 100, the more progress can be made in learning each prediction model 90-92, and the higher the prediction accuracy of each prediction model 90-92 can be. Therefore, it is possible to derive the suitable manufacturing conditions 13 with greater accuracy than when the quality of the entire flexible tube 11 for an endoscope is predicted and the suitable manufacturing conditions 13 are derived based on the prediction results.

[0091] The objective function 67 has a third term 107 and a fourth term 108, which are regularization terms for fitting the manufacturing conditions of each of the divided parts P into the constraints on the manufacturing conditions of the entire flexible tube 11 for an endoscope. Therefore, it is possible to fit the suitable manufacturing conditions 13 into the constraints on the manufacturing conditions of the entire flexible tube 11 for an endoscope.

[0092] The third term 107 and the fourth term 108, which are regularization terms, smooth the extrusion rate R per unit time of the resin material for each inner layer 25 of each divided portion P and the extrusion rate S per unit time of the resin material for each outer layer 26 of each divided portion P. Specifically, the third term 107 is the sum of the second-order differential values ​​of the extrusion rate R per unit time of the resin material for each inner layer 25 of each divided portion P, and the fourth term 108 is the sum of the second-order differential values ​​of the extrusion rate S per unit time of the resin material for each outer layer 26 of each divided portion P. This makes it possible to obtain suitable manufacturing conditions 13 as if they were set by a skilled operator.

[0093] The first term 105 and the second term 106, which are terms including a difference, are the sum of values ​​obtained by dividing the difference by the elasticity specification value 85 and the total thickness setting value 86, which are the target values ​​for the divided portion P. If the terms including a difference were calculated as the sum of the differences without dividing the difference by the target value for the divided portion P, a case in which the target value is 100, the predicted value is 95, and the difference is 5 would be treated as the same as a case in which the target value is 10, the predicted value is 5, and the difference is 5. Therefore, by calculating the terms including a difference as the sum of values ​​obtained by dividing the difference by the target value for the divided portion P, it is possible to distinguish between a case in which the target value is 100, the predicted value is 95, and the difference is 5 (the value obtained by dividing the difference by the target value is 5 / 100 = 0.05) and a case in which the target value is 10, the predicted value is 5, and the difference is 5 (the value obtained by dividing the difference by the target value is 5 / 10 = 0.5). Consequently, it is possible to accurately derive the preferred manufacturing conditions 13.

[0094] Each of the prediction models 90 to 92 also receives input of physical property information 81 of the material that constitutes the flexible tube 11 for an endoscope, and design information 82 of the flexible tube 11 for an endoscope. This allows for higher prediction accuracy of the elasticity prediction value 95, the inner layer thickness prediction value 96, and the outer layer thickness prediction value 97 than when the physical property information 81 and the design information 82 are not input to each of the prediction models 90 to 92.

[0095] Furthermore, in addition to the manufacturing conditions 80 of the division portion P to be predicted, the manufacturing conditions 80 of the division portions P on both sides are input into the elasticity value prediction model 90, which can further improve the prediction accuracy of the predicted elasticity value 95. Furthermore, the manufacturing conditions 80A_P1 and the like excluding information about the outer layer 26 are input into the inner layer thickness prediction model 91, and conversely, the manufacturing conditions 80B_P1 and the like excluding information about the inner layer 25 are input into the outer layer thickness prediction model 92, which eliminates information that is considered unnecessary for prediction, which can further improve the prediction accuracy of the predicted inner layer thickness value 96 and the predicted outer layer thickness value 97.

[0096] It is preferable to input both the physical property information 81 and the design information 82 into the elasticity value prediction model 90, etc., but at least one of the physical property information 81 and the design information 82 may be input into the elasticity value prediction model 90, etc.

[0097] There are multiple types of predicted values, including a predicted elasticity value 95, a predicted inner layer thickness value 96, and a predicted outer layer thickness value 97, and machine learning models are prepared for each of the multiple types of predicted values, including an elasticity value prediction model 90, an inner layer thickness prediction model 91, and an outer layer thickness prediction model 92. Therefore, the predicted elasticity value 95, the predicted inner layer thickness value 96, and the predicted outer layer thickness value 97 can all be predicted with high accuracy.

[0098] The product is a flexible tube 11 for an endoscope, which has a flexible tube base material 20 and a resin layer 21 covering the flexible tube base material 20, the resin layer 21 being composed of an inner layer 25 and an outer layer 26 having different thickness ratios in the axial direction AD. The divided portions P are portions obtained by dividing the flexible tube 11 for an endoscope along the axial direction AD. The manufacturing conditions 80 include the extrusion amount per unit time of the resin material for the inner layer 25 and the extrusion amount per unit time of the resin material for the outer layer 26 by the extruder 12. The predicted values ​​include the elasticity value of the divided portion P (predicted elasticity value 95), the thickness of the inner layer of the divided portion P (predicted inner layer thickness value 96), and the thickness of the outer layer of the divided portion P (predicted outer layer thickness value 97).

[0099] There are a wide variety of types of flexible tubes 11 for endoscopes, such as the large-diameter, long flexible tube 11LL for an oral endoscope for the upper gastrointestinal tract shown in Fig. 22, or the small-diameter, short flexible tube 11SS for a transnasal endoscope for the bronchi shown in Fig. 23. Specifications are also frequently changed, such as by changing the resin materials for the inner layer 25 and outer layer 26 and the outer diameter of the flexible tube base material 20. Therefore, the advantage of being able to derive universally suitable manufacturing conditions 13 regardless of the type of flexible tube 11 for endoscopes can be fully realized.

[0100] The control unit 41 of the extruder 12 may serve as the CPU 57 of the information processing device 10. In other words, the control unit 41 of the extruder 12 may output the predicted values ​​95 to 98 and derive the suitable manufacturing conditions 13. In this case, the extruder 12 itself is an example of the "information processing device" according to the technology of the present disclosure.

[0101] Second Embodiment In the above-described first embodiment, the flexible tube 11 for an endoscope is exemplified as a product, and an example of deriving the suitable manufacturing conditions 13 for the extruder 12 is described, but the technology of the present disclosure is not limited thereto. In a second embodiment, suitable manufacturing conditions 251 (see FIG. 33 ) for a sheet manufacturing apparatus 200 (see FIG. 24 ) are derived.

[0102] As an example, as shown in FIG. 24 , a sheet manufacturing apparatus 200 is an apparatus that applies a liquid to a conveyed long support 201 to manufacture a sheet 202. The sheet 202 is an example of a "product" according to the technology of the present disclosure. The support 201 is, for example, a magnetic tape substrate. In this case, the liquid is a magnetic layer material, and the sheet 202 is a magnetic tape. Alternatively, the support 201 is, for example, an optical film substrate. In this case, the liquid is a photosensitive layer material, and the sheet 202 is an optical film.

[0103] The sheet manufacturing apparatus 200 includes a coating unit 203, conveying rollers 204 and 205, and a tension adjustment roller 206. The coating unit 203 is a giesser or the like for applying a liquid to the support 201. The conveying rollers 204 and 205 are arranged symmetrically on the upstream and downstream sides of the coating unit 203, and are rotated by a motor to convey the support 201 and the sheet 202. The tension adjustment roller 206 is arranged upstream of the coating unit 203 and the conveying roller 204. The tension adjustment roller 206 adjusts the tension applied to the support 201 at the liquid application location in the coating unit 203 by tilting, for example, around center C as the rotation center, relative to the width direction WD of the support 201 and the sheet 202, as shown by the arrow. The application location is a slit 207 having a narrow gap parallel to the width direction WD. The sheet manufacturing apparatus 200 also includes a delivery drum that delivers the roll-shaped support 201 toward the coating unit 203, and a take-up drum that takes up the sheet 202 coated with the liquid from the coating unit 203. The sheet manufacturing apparatus 200 also includes a driven roller that comes into contact with the support 201 or the sheet 202 and rotates in response to the conveyance of the support 201 or the sheet 202, and a dancer roller that oscillates in response to changes in the conveyance speed of the support 201, thereby suppressing fluctuations in the tension applied to the support 201.

[0104] As an example, as shown in FIG. 25, in the second embodiment, the quality of each of divided portions P1, P2, . . . , P7, and P8 obtained by dividing the sheet 202 along the width direction WD is predicted.

[0105] As an example, as shown in FIG. 26 , input data 210 in the second embodiment includes first manufacturing conditions 211 and second manufacturing conditions 212. The first manufacturing conditions 211 include the control amount of the tension adjustment roller 206. In other words, the control amount of the tension adjustment roller 206 is the amount of pressure of the tension adjustment roller 206 in the width direction WD. The second manufacturing conditions 212 include the liquid flow rate from the coating unit 203, the lip clearance, and the taper angle of the pocket portion. The lip clearance is the gap between the slit 207, which is the coating location, and the support 201. The pocket portion is provided within the coating unit 203 and communicates with the slit 207. The pocket portion is a portion for spreading the liquid in the width direction WD. Note that the first manufacturing conditions 211 may also include the amount of oscillation of the dancer roller, etc., as appropriate.

[0106] The input data 210 has first physical property information 213_P1 for the divided portion P1, first physical property information 213_P2 for the divided portion P2, ..., and first physical property information 213_P8 for the divided portion P8. The input data 75 also has second physical property information 214. Hereinafter, when there is no need to particularly distinguish between the first physical property information 213_P1 to 213_P8 for the divided portions P1 to P8, they may be referred to as first physical property information 213.

[0107] The first physical property information 213 and the second physical property information 214 are information about the materials constituting the sheet 202, in this case the support 201 and the physical properties of the liquid. The first physical property information 213 includes the thickness, amount of slack, and amount of waviness of the support 201 in the divided portion P. The thickness, amount of slack, and amount of waviness of the support 201 are values ​​measured for each divided portion P by a sensor (not shown) installed between the coating unit 203 and the conveying roller 204. Therefore, the first physical property information 213 changes from moment to moment as the support 201 is conveyed. Therefore, in the second embodiment, the input data 210 is input each time the support 201 is conveyed a predetermined length. The predetermined length is, for example, 5 cm to 10 cm. The thickness, amount of slack, and amount of waviness of the support 201 are representative values, for example, average values, for the predetermined length. The second physical property information 214 includes the liquid viscosity (for example, extensional viscosity). The first manufacturing conditions 211, the second manufacturing conditions 212, the first physical property information 213, and the second physical property information 214 are not limited to the exemplified contents, and other elements may be added as appropriate.

[0108] 27 as an example, the prediction model group 220 of the second embodiment includes a tension prediction model 221, a hydraulic pressure prediction model 222, and a coating thickness prediction model 223. The tension prediction model 221, the hydraulic pressure prediction model 222, and the coating thickness prediction model 223 are machine learning models configured using, for example, a neural network. In other words, the tension prediction model 221, the hydraulic pressure prediction model 222, and the coating thickness prediction model 223 are examples of "machine learning models" according to the technology of the present disclosure.

[0109] The tension prediction model 221 outputs a predicted value of tension (hereinafter referred to as predicted tension value) applied to the divided portion P at the slit 207, which is the application location, in response to input of the first manufacturing conditions 211 and the first physical property information 213. The hydraulic pressure prediction model 222 outputs a predicted value of pressure of the liquid applied to the divided portion P at the slit 207 (hereinafter referred to as predicted hydraulic pressure value) in response to input of the second manufacturing conditions 212, the first physical property information 213, and the second physical property information 214. The application thickness prediction model 223 outputs a predicted value of the application thickness of the liquid on the divided portion P (hereinafter referred to as predicted application thickness value) in response to input of the predicted tension value and the predicted hydraulic pressure value. As such, in the second embodiment as well, there are multiple types of predicted values, and prediction models are prepared for each of the multiple types of predicted values.

[0110] 28 , when predicting the quality of divided portion P1, the prediction unit 72 inputs first manufacturing conditions 211 and first physical property information 213_P1 of divided portion P1 into a tension prediction model 221. Then, the prediction unit 72 causes the tension prediction model 221 to output a predicted tension value 230_P1 of divided portion P1.

[0111] 29 , when predicting the quality of divided portion P1, the prediction unit 72 inputs second manufacturing conditions 212, first physical property information 213_P1 of divided portion P1, and second physical property information 214 to a hydraulic pressure prediction model 222. Then, the prediction unit 72 causes the hydraulic pressure prediction model 222 to output a hydraulic pressure prediction value 231_P1 of divided portion P1.

[0112] 30 , when predicting the quality of divided portion P1, prediction unit 72 inputs a predicted tension value 230_P1 output from tension prediction model 221 and a predicted hydraulic pressure value 231_P1 output from hydraulic pressure prediction model 222 to coating thickness prediction model 223. Then, prediction unit 72 causes coating thickness prediction model 223 to output a predicted coating thickness value 235_P1 for divided portion P1.

[0113] 28 to 30 show an example in which predicted tension value 230_P1, predicted hydraulic pressure value 231_P1, and predicted coating thickness value 235_P1 for divided portion P1 are output from each prediction model 221 to 223, but prediction unit 72 also causes predicted tension values ​​230_P2 to 230_P8 (not shown), predicted hydraulic pressure values ​​231_P2 to 231_P8 (not shown), and predicted coating thickness values ​​235_P2 to 235_P8 (see FIG. 31) for the remaining divided portions P2 to P8 to be output from each prediction model 221 to 223. Hereinafter, when there is no need to particularly distinguish between each of predicted coating thickness values ​​235_P1 to 235_P8 for divided portions P1 to P8, they may be referred to as predicted coating thickness value 235.

[0114] By determining the predicted coating thickness value 235 for each divided portion P in this manner, the predicted value group 240 of the second embodiment has, as an example, the contents shown in Fig. 31. That is, the predicted value group 240 has a predicted coating thickness value 235_P1 for divided portion P1, a predicted coating thickness value 235_P2 for divided portion P2, ..., a predicted coating thickness value 235_P7 for divided portion P7, and a predicted coating thickness value 235_P8 for divided portion P8.

[0115] The learning data for each of the prediction models 221 to 223 is data collected from divided portions P of the sheet 202 that were previously manufactured. Although not shown in the figure, the learning data for each of the prediction models 221 to 223 has the following content. That is, as learning input data, the learning data includes first learning manufacturing conditions corresponding to the first manufacturing conditions 211, second learning manufacturing conditions corresponding to the second manufacturing conditions 212, first learning physical property information corresponding to the first physical property information 213, and second learning physical property information corresponding to the second physical property information 214. The correct answer data includes an actual measurement value of the tension applied to the divided portion P in the slit 207, an actual measurement value of the pressure of the liquid applied to the divided portion P in the slit 207, and an actual measurement value of the applied thickness of the liquid on the divided portion P.

[0116] In addition, the learning data for each predictive model 221 to 223 is data collected from various types of sheets 202 manufactured in the past, such as a support 201 with a relatively large width and 10 or more divided parts P, or a support 201 with a relatively small width and less than 5 divided parts P.

[0117] As shown in FIG. 32 as an example, the objective function 245 of the second embodiment has a first term 246 and a second term 247. The first term 246 includes the difference between the predicted coating thickness value 235 of each divided portion P and the coating thickness set value 250 (see FIG. 33 ). More specifically, the first term 246 includes the sum of the squares of the value obtained by dividing the difference between the predicted coating thickness value 235 and the coating thickness set value 250 by the coating thickness set value 250. The first term 246 is a term obtained by multiplying this sum by a first weighting coefficient W1. The coating thickness set value 250 is common to each divided portion P, and therefore there is only one. The first term 246 is an example of a "term including a difference" according to the technology of the present disclosure. The coating thickness set value 250 is also an example of a "target value" according to the technology of the present disclosure.

[0118] The second term 247 is a regularization term for integrating the manufacturing conditions of each of the divided portions P into the constraints on the manufacturing conditions of the entire sheet 202. Here, the constraints on the manufacturing conditions of the entire sheet 202 are linearly connecting the control amounts (push amounts of the tension adjustment rollers 206) of each of the divided portions P. Therefore, the second term 247 is a term that linearly connects the control amounts (push amounts) of each of the tension adjustment rollers 206 of each of the divided portions P. The second term 247 is, for example, 1 minus the coefficient of determination obtained by linear regression analysis of the push amounts of the tension adjustment rollers 206 of each of the divided portions P. The second term 247 has a second weighting coefficient W2.

[0119] The first weighting coefficient W1 and the second weighting coefficient W2 are values ​​that sum to 1 (W1 + W2 = 1), similar to the first weighting coefficient W1 to the fourth weighting coefficient W4 in the first embodiment. For example, W1 = W2 = 0.5. The first weighting coefficient W1 and the second weighting coefficient W2 do not necessarily have to be the same value. For example, if emphasis is placed on the first term 246, that is, if the predicted coating thickness value 235 of each divided portion P is to be brought closer to the coating thickness set value 250, the first weighting coefficient W1 may be set to a value higher than the second weighting coefficient W2.

[0120] 33 , the preferred manufacturing conditions 251 of the second embodiment, which are derived by the derivation unit 73 by solving an optimization problem to find the manufacturing conditions that minimize the value of the objective function 245, include the control amount of the tension adjusting roller 206. As for the control amount of the tension adjusting roller 206, one value common to each divided portion P is derived by the action of the second term 247 of the objective function 245.

[0121] In the second embodiment, as described above, the input data 210 is input each time the support 201 is conveyed a predetermined length. Therefore, the prediction unit 72 predicts the predicted tension values ​​230_P1 to 230_P8, the predicted hydraulic pressure values ​​231_P1 to 231_P8, and the predicted coating thickness value 235, and the derivation unit 73 derives the control amount of the tension adjustment roller 206 as the preferable manufacturing condition 251 each time the support 201 is conveyed a predetermined length and the input data 210 is input. By continuously deriving the control amount of the tension adjustment roller 206 as the preferable manufacturing condition 251 in this manner, it is possible to apply liquid to the support 201, which is continuously conveyed, while the tension adjustment roller 206 is always applying a preferable tension to the support 201.

[0122] As described above, in the second embodiment, the product is a sheet 202 formed by applying a liquid to a conveyed elongated support 201. The divided portions P are portions obtained by dividing the sheet 202 along the width direction WD. The first manufacturing conditions 211 include a control amount for a tension adjustment roller 206 that is disposed upstream of a slit 207, which is the liquid application location, and that adjusts the tension applied to the support 201 at the slit 207. The predicted value includes the liquid application thickness (predicted application thickness value 235) at the divided portions P. This allows for various situations, such as a support 201 with a relatively large width and 10 or more divided portions P, and a support 201 with a relatively small width and fewer than five divided portions P. In other words, it is possible to derive generally suitable manufacturing conditions 251 regardless of the type of sheet 202. Note that the predicted value of the applied thickness of the liquid on the divided portion P is derived after deriving the predicted value of the tension applied to the divided portion P in the slit 207 and the predicted value of the hydraulic pressure applied to the divided portion P in the slit 207, but this is not limitative. The predicted applied thickness may be derived without deriving the predicted value of the tension and the predicted hydraulic pressure based on the first manufacturing conditions 211, etc.

[0123] Third Embodiment In a third embodiment, suitable production conditions for a culture tank 300 (see FIG. 34) are derived.

[0124] As an example, as shown in Fig. 34 , a culture medium 301 is stored in a culture tank 300. The capacity of the culture tank 300 is, for example, 50 L or more and 5000 L or less. Cells 302 are seeded in the culture tank 300, and the cells 302 are cultured in the culture medium 301. The cells 302 are an example of a "product" according to the technology of the present disclosure.

[0125] Cells 302 are antibody-producing cells established by inserting antibody genes into cells such as Chinese hamster ovary cells. Cells 302 produce immunoglobulins, i.e., antibodies, as a product during the culture process. Therefore, not only cells 302 but also antibodies are present in culture solution 301. The antibodies are, for example, monoclonal antibodies, and are the active ingredients of biopharmaceuticals.

[0126] The culture tank 300 is provided with a medium supply channel 303, a gas supply channel 304, an exhaust channel 305, a culture solution delivery / recovery channel 306, a sparger 307, a gas supply channel 308, and an agitator 309. The medium supply channel 303 is a channel for continuously supplying fresh medium into the culture tank 300. In other words, perfusion culture is performed in the culture tank 300. The gas supply channel 304 is a channel for supplying gas containing air and carbon dioxide from above. The exhaust channel 305 is a channel for exhausting the gas supplied from the gas supply channel 304 to the outside of the culture tank 300. The exhaust channel 305 is provided with an exhaust filter 310.

[0127] The culture medium delivery / recovery path 306 is connected to an inlet / outlet 312 of the cell removal filter 311. The culture medium delivery / recovery path 306 is a flow path for delivering the culture medium 301 in the culture tank 300 to the cell removal filter 311. The culture medium delivery / recovery path 306 is also a flow path for returning the culture medium 301 (concentrated liquid) from the cell removal filter 311 to the culture tank 300.

[0128] The sparger 307 is disposed at the bottom of the culture tank 300. The sparger 307 releases oxygen-containing gas supplied from a gas supply line 308 into the culture tank 300. The oxygen released from the sparger 307 dissolves in the culture solution 301 and aids in antibody production by the cells 302. The agitator blades 309 are rotated at a predetermined rotation speed by a motor or the like to agitate the culture solution 301 in the culture tank 300. This maintains the homogeneity of the culture solution 301 in the culture tank 300. In addition to the above, the culture tank 300 is provided with a flow path for a cell bleed process, in which a portion of the culture solution 301 is intentionally extracted. The agitator blades 309 may have multiple blades as shown in the figure, or may have a single disk-shaped blade; the shape is not particularly limited. Two or more agitator blades 309 may be disposed in the culture tank 300.

[0129] The cell removal filter 311 connected to the culture solution delivery / recovery path 306 has an internal filter membrane 313. The filter membrane 313 captures cells 302 and allows antibodies to pass through. The cell removal filter 311 removes cells 302 from the culture solution 301 using the filter membrane 313, for example, by tangential flow filtration (TFF), thereby obtaining a culture supernatant.

[0130] More specifically, the cell removal filter 311 has a diaphragm pump 315 having an elastic membrane 314 therein, and a degassing / air supply path 316. While air below the elastic membrane 314 is being degassed through the degassing / air supply path 316, the elastic membrane 314 is elastically deformed so as to adhere to the lower end of the diaphragm pump 315, causing the culture solution 301 in the culture tank 300 to flow into the cell removal filter 311 through the culture solution delivery / recovery path 306. Furthermore, while air is being supplied to the lower side of the elastic membrane 314 through the degassing / air supply path 316, the elastic membrane 314 is elastically deformed so as to adhere to the upper side of the diaphragm pump 315, causing the culture solution 301 (concentrated solution) that did not pass through the filter membrane 313 to be returned to the culture tank 300 through the culture solution delivery / recovery path 306.

[0131] The culture supernatant flows out from the outlet 317 of the cell removal filter 311. The culture supernatant mainly contains antibodies. The culture supernatant is sent to a downstream purification section (not shown), where it is subjected to various chromatography processes, virus inactivation processes, and the like, and is ultimately used as a drug substance for a biopharmaceutical.

[0132] 35 as an example, in the third embodiment, the quality of each of cubic divided parts P1, P2, P3, P4, ... obtained by three-dimensionally dividing the culture tank 300 is predicted. Here, one divided part P corresponds to a culture vessel 320, such as a petri dish, which is smaller than the culture tank 300 and has a capacity of, for example, several mL to several liters.

[0133] 36 as an example, input data 325 of the third embodiment has manufacturing conditions 326_P1 for divided portion P1, manufacturing conditions 326_P2 for divided portion P2, etc. Hereinafter, when there is no particular need to distinguish between manufacturing conditions 326_P1 for divided portion P1 and manufacturing conditions 326_P2 for divided portion P2, etc., they may be referred to as manufacturing conditions 326.

[0134] The production conditions 326 include culture conditions for the cells 302. Specifically, the production conditions 326 include the density (cell density) of the cells 302 in the divided portion P, the hydrogen ion exponent, oxygen concentration, temperature, medium concentration, and shear energy due to agitation by the agitator 309 in the divided portion P. All of these are values ​​predicted by a simulation using various parameters such as the capacity of the culture tank 300, the position of the culture medium supply channel 303, the position of the gas supply channel 304, the position of the sparger 307, the position of the agitator 309, the amount of medium supplied, the amount of gas supplied, and the rotation speed of the agitator 309. Note that the production conditions 326 may also include carbon dioxide concentration, etc., as appropriate. Furthermore, the input data 325 may also include physical property information of the cells 302, such as the size of the cells 302, and / or design information, such as the capacity of the culture tank 300.

[0135] As an example, as shown in Fig. 37 , the third embodiment uses a culture environment prediction model 330. The culture environment prediction model 330 is a machine learning model configured using, for example, a neural network. In other words, the culture environment prediction model 330 is an example of a "machine learning model" according to the technology of the present disclosure.

[0136] When predicting the quality of the divided portion P1, the prediction unit 72 inputs the manufacturing conditions 326_P1 into the culture environment prediction model 330. Then, the prediction unit 72 causes the culture environment prediction model 330 to output a culture environment prediction value 331_P1 of the divided portion P1.

[0137] 37 shows an example in which the culture environment predicted value 331_P1 of the divided portion P1 is output from the culture environment prediction model 330, but the prediction unit 72 similarly outputs culture environment predicted values ​​331_P2 (see FIG. 38) for the remaining divided portions P2, ... from the culture environment prediction model 330. Hereinafter, when there is no need to particularly distinguish between the culture environment predicted values ​​331_P1, 331_P2, ... of the divided portions P1, P2, ..., they may be referred to as the culture environment predicted value 331.

[0138] The culture environment predicted value 331 is a value that indicates the quality of the culture environment for the cells 302 based on the proliferation rate, survival rate, etc. of the cells 302, and takes a value between 0 and 100, for example. If the value is closer to 100, the culture environment for the cells 302 is good. The culture environment predicted value 331 is an example of an "index value" according to the technology of the present disclosure.

[0139] By determining the culture environment predicted value 331 for each divided portion P in this manner, the predicted value group 335 of the third embodiment has the contents as shown in Fig. 38 as an example. That is, the predicted value group 335 has a culture environment predicted value 331_P1 for divided portion P1, a culture environment predicted value 331_P2 for divided portion P2, and so on.

[0140] The learning data for the culture environment prediction model 330 is data collected from cells 302 previously produced using small-scale culture vessels 320. Although not shown in the figure, the learning data for the culture environment prediction model 330 has the following content. That is, as learning input data, it includes learning manufacturing conditions corresponding to the manufacturing conditions 326. And, as correct answer data, it includes actual measured values ​​of the quality of the culture environment for the cells 302 in the small-scale culture vessels 320.

[0141] Furthermore, the learning data for the culture environment prediction model 330 is data collected from various types of cells 302 that were previously produced using small-scale culture vessels 320, such as cells that produce antibody A and cells that produce antibody B, or cells that were cultured in a culture vessel 320 with a capacity of 50 mL and cells that were cultured in a culture vessel 320 with a capacity of 1 L.

[0142] As shown in FIG. 39 as an example, the objective function 340 of the third embodiment has a first term 341 and a second term 342. The first term 341 includes the difference between the culture environment predicted value 331 of each divided portion P and the culture environment set value 350 (see FIG. 40 ). More specifically, the first term 341 includes the sum of the squares of the value obtained by dividing the difference between the culture environment predicted value 331 and the culture environment set value 350 by the culture environment set value 350. The first term 341 is a term obtained by multiplying this sum by a first weighting coefficient W1. The culture environment set value 350 is common to each divided portion P, and therefore there is only one. The first term 341 is an example of a "term including a difference" according to the technology of the present disclosure. The culture environment set value 350 is also an example of a "target value" according to the technology of the present disclosure.

[0143] The second term 342 is a regularization term for integrating the production conditions of each divided portion P into the constraints on the production conditions of the entire cells 302 in the culture tank 300. The constraints on the production conditions of the entire cells 302 in the culture tank 300 are to make the production conditions of each divided portion P equal. Therefore, the second term 342 is a term for making the production conditions of each divided portion P equal. The second term 342 is, for example, the variance or standard deviation of the production conditions of each divided portion P. In this case, minimizing the variance or standard deviation can make the production conditions of each divided portion P equal. Note that "equal" not only refers to cases where the production conditions of each divided portion P are exactly the same, but also includes cases where the production conditions of each divided portion P fall within a predetermined difference range. The second term 342 has a second weighting coefficient W2.

[0144] The first weighting coefficient W1 and the second weighting coefficient W2 are values ​​that sum to 1 (W1 + W2 = 1), similar to the first weighting coefficient W1 to the fourth weighting coefficient W4 in the first embodiment. For example, W1 = W2 = 0.5. The first weighting coefficient W1 and the second weighting coefficient W2 do not necessarily have to be the same value. For example, if emphasis is placed on the first term 341, that is, if the culture environment predicted value 331 of each divided portion P is to be closer to the culture environment set value 350, the first weighting coefficient W1 may be set to a value higher than the second weighting coefficient W2.

[0145] As an example, as shown in FIG. 40 , in the third embodiment, the derivation unit 73 derives provisional suitable manufacturing conditions 351 for each divided portion P, such as provisional suitable manufacturing conditions 351_P1 for divided portion P1, ..., by solving an optimization problem to find manufacturing conditions that minimize the value of the objective function 340. The provisional suitable manufacturing conditions 351 include the hydrogen ion exponent, oxygen concentration, temperature, and culture medium concentration for each divided portion P. Here, it is assumed that the hydrogen ion exponent, oxygen concentration, temperature, and culture medium concentration are almost uniform across each divided portion P due to agitation by the agitator 309. On the other hand, it is assumed that the shear energy caused by agitation by the agitator 309 differs across each divided portion P.

[0146] As an example, as shown in FIG. 41 , the outlet unit 73 derives further suitable production conditions from the provisional suitable production conditions 351 for each divided portion P. FIG. 41 illustrates the process of deriving a suitable oxygen supply amount 356 to be supplied into the culture tank 300 via the sparger 307 from the oxygen concentration of the provisional suitable production conditions 351 for each divided portion P, as shown in Table 355. The outlet unit 73 derives the suitable oxygen supply amount 356 by simulating an oxygen supply amount that will cause the oxygen concentration of each divided portion P to match the oxygen concentration of the provisional suitable production conditions 351. The suitable oxygen supply amount 356 is an example of a "suitable production condition" according to the technology of the present disclosure. Instead of or in addition to the suitable oxygen supply amount 356, a suitable medium supply amount from the medium supply path 303, a suitable gas supply amount from the gas supply path 304, or a suitable rotation speed of the agitator blade 309 may be derived as a suitable production condition.

[0147] As described above, in the third embodiment, the product is the cells 302 cultured in the culture tank 300. The divided portion P is a portion obtained by dividing the culture tank 300. The production conditions 326 include the culture conditions of the cells 302, and the predicted value includes an index value (culture environment predicted value 331) that indicates the desirability of the culture environment for the cells 302. This makes it possible to accommodate various situations, such as the case of cells producing antibody A or cells producing antibody B. In other words, it becomes possible to derive generally suitable production conditions regardless of the type of cells 302.

[0148] The cells 302 are not limited to the illustrated antibody-producing cells. They may be pluripotent stem cells such as induced pluripotent stem (iPS) cells. As the machine learning model, multiple types of models may be prepared, such as a model for predicting the proliferation rate of the cells 302 in each divided portion P and a model for predicting the survival rate of the cells 302 in each divided portion P.

[0149] The information processing device 10 may be installed in the same facility as the extruder 12, etc., or may be installed in a data center independent of the facility where the extruder 12, etc. are installed.

[0150] In each of the above embodiments, the following various processors can be used as the hardware structure of processing units that perform various processes, such as the reception unit 70, the RW control unit 71, the prediction unit 72, the derivation unit 73, and the distribution control unit 74. The various processors include the CPU 57, which is a general-purpose processor that executes software (operation program 65) and functions as various processing units, as described above, as well as dedicated electrical circuits that are processors having a circuit configuration specifically designed to perform specific processes, such as a programmable logic device (PLD) that is a processor whose circuit configuration can be changed after manufacture, such as an FPGA (Field Programmable Gate Array), and an ASIC (Application Specific Integrated Circuit).

[0151] 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 and / or a combination of a CPU and an FPGA).Furthermore, multiple processing units may be configured with a single processor.

[0152] Examples of configuring multiple processing units with a single processor include, first, 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, a form in which a processor is used to realize the functions of the entire system including multiple processing units with a single IC (Integrated Circuit) chip, as typified by systems on chips (SoCs). In this way, various processing units are configured using one or more of the above-mentioned various processors as a hardware structure.

[0153] Furthermore, more specifically, the hardware structure of these various processors can be an electric circuit (circuitry) that combines circuit elements such as semiconductor elements.

[0154] From the above description, the technology described in the following supplementary paragraphs can be understood.

[0155] [Supplementary Item 1] An information processing device including a processor, the processor using a machine learning model that outputs a predicted value of the quality of a divided portion obtained by dividing an entire product in response to input of manufacturing conditions for the divided portion, inputting the manufacturing conditions of the divided portion into the machine learning model and causing the machine learning model to output the predicted value of the divided portion, and deriving suitable manufacturing conditions that result in a target value for the quality of the entire product by solving an optimization problem that finds manufacturing conditions that minimize a value of an objective function having a term including a difference between the predicted value of the divided portion and a target value for the quality of the divided portion. [Supplementary Item 2] The information processing device according to Supplementary Item 1, wherein the objective function has a regularization term that causes the manufacturing conditions of the divided portion to fit within constraints related to the manufacturing conditions of the entire product. [Supplementary Item 3] The information processing device according to Supplementary Item 2, wherein the regularization term is a term that smooths the manufacturing conditions of the divided portion. [Supplementary Item 4] The information processing device according to Supplementary Item 3, wherein the regularization term is a term that smooths the manufacturing conditions of the divided portion. [Supplementary Item 4] The information processing device according to Supplementary Item 3, wherein the regularization term is a sum of second-order derivatives of the manufacturing conditions of the divided portion. [Supplementary Item 5] The information processing device according to any one of Supplementary Items 1 to 4, wherein the term including the difference is a sum of values ​​obtained by dividing the difference by the target value of the divided portion. [Supplementary Item 6] The information processing device according to any one of Supplementary Items 1 to 5, wherein at least one of physical property information of materials constituting the product and design information of the product is also input to the machine learning model. [Supplementary Item 7] The information processing device according to any one of Supplementary Items 1 to 6, wherein there are a plurality of types of predicted values, and a machine learning model is prepared for each of the plurality of types of predicted values. [Appendix 8] An information processing device according to any one of appendixes 1 to 7, wherein the product is a flexible tube for an endoscope having a flexible tube base material and a resin layer covering the flexible tube base material, the resin layer being composed of an inner layer and an outer layer having different thickness ratios in the axial direction, the divided portions are portions obtained by dividing the flexible tube for an endoscope along the axial direction, the manufacturing conditions include an extrusion amount per unit time of a resin material for the inner layer and an extrusion amount per unit time of a resin material for the outer layer by an extrusion molding machine, and the predicted value includes an elasticity value of the divided portions, a thickness of the inner layer of the divided portions, and a thickness of the outer layer of the divided portions.[Supplementary Item 9] The information processing device according to any one of Supplementary Items 1, 2, and 5 to 7, wherein the product is a sheet obtained by applying a liquid to a transported long support, the divided portions are portions obtained by dividing the sheet along the width direction, the manufacturing conditions include a control amount of a tension adjustment roller that is arranged upstream of a position where the liquid is applied and adjusts the tension applied to the support at the application position, and the predicted value includes a coating thickness of the liquid in the divided portions. [Supplementary Item 10] The information processing device according to any one of Supplementary Items 1, 2, and 5 to 7, wherein the product is cells cultured in a culture vessel, the divided portions are portions obtained by dividing the culture vessel, the manufacturing conditions include culture conditions for the cells, and the predicted value includes an index value that represents a favorable culture environment for the cells. [Supplementary Item 11] A method for operating an information processing device, comprising: using a machine learning model that outputs a predicted value of the quality of a divided portion obtained by dividing an entire product in response to input of manufacturing conditions for the divided portion, inputting the manufacturing conditions for the divided portion into the machine learning model and causing the machine learning model to output the predicted value of the divided portion, and deriving suitable manufacturing conditions under which the quality of the entire product will be the target value by solving an optimization problem to find manufacturing conditions that minimize a value of an objective function having a term including a difference between the predicted value of the divided portion and a target value of the quality of the divided portion. [Supplementary Item 12] An operating program for an information processing device, causing a computer to execute processes, comprising: using a machine learning model that outputs a predicted value of the quality of a divided portion obtained by dividing an entire product in response to input of manufacturing conditions for the divided portion, inputting the manufacturing conditions for the divided portion into the machine learning model and causing the machine learning model to output the predicted value of the divided portion, and deriving suitable manufacturing conditions under which the quality of the entire product will be the target value by solving an optimization problem to find manufacturing conditions that minimize a value of an objective function having a term including a difference between the predicted value of the divided portion and a target value of the quality of the divided portion.

[0156] The technology of the present disclosure can be appropriately combined with the various embodiments and / or various modified examples described above. Furthermore, it is not limited to the above-described embodiments, and various configurations can be adopted without departing from the spirit of the present disclosure. Furthermore, the technology of the present disclosure extends not only to programs but also to storage media that non-temporarily store programs.

[0157] The above-described description and illustrations are a detailed explanation of the parts related to the technology of the present disclosure and are merely an example of the technology of the present disclosure. For example, the above description of the configuration, functions, actions, and effects is an explanation of an example of the configuration, functions, actions, and effects of the parts related to the technology of the present disclosure. Therefore, it goes without saying that unnecessary parts may be deleted, new elements may be added, or replacements may be made to the above-described description and illustrations within the scope of the gist of the technology of the present disclosure. Furthermore, to avoid confusion and facilitate understanding of the parts related to the technology of the present disclosure, the above-described description and illustrations omit explanations of common technical knowledge that do not require particular explanation to enable the implementation of the technology of the present disclosure.

[0158] In this specification, "A and / or B" is synonymous with "at least one of A and B." In other words, "A and / or B" means that it may be only A, only B, or a combination of A and B. Furthermore, in this specification, the same concept as "A and / or B" is also applied when three or more things are expressed by connecting them with "and / or."

[0159] All publications, patent applications, and technical standards mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent application, or technical standard was specifically and individually indicated to be incorporated by reference.

Claims

1. An information processing device comprising a processor, the processor uses a machine learning model that outputs a predicted value of the quality of a divided portion obtained by dividing an entire product in response to input of manufacturing conditions of the divided portion, inputs the manufacturing conditions of the divided portion into the machine learning model, causes the machine learning model to output the predicted value of the divided portion, and derives suitable manufacturing conditions that will result in the quality of the entire product being the target value by solving an optimization problem that finds manufacturing conditions that minimize the value of an objective function that has a term including the difference between the predicted value of the divided portion and a target value of the quality of the divided portion.

2. The information processing device according to claim 1, wherein the objective function has a regularization term for accommodating the manufacturing conditions of the divided parts within constraints related to the manufacturing conditions of the entire product.

3. The information processing device according to claim 2, wherein the regularization term is a term that smooths the manufacturing conditions of the divided portions.

4. The information processing device according to claim 3, wherein the regularization term is a sum of second-order differential values ​​of the manufacturing conditions of the divided portions.

5. The information processing device according to claim 1, wherein the term including the difference is the sum of values ​​obtained by dividing the difference by the target value of the divided portion.

6. The information processing device according to claim 1, wherein at least one of physical property information of the material that constitutes the product and design information of the product is also input to the machine learning model.

7. The information processing device according to claim 1, wherein there are a plurality of types of the predicted values, and the machine learning model is prepared for each of the plurality of types of the predicted values.

8. The information processing device according to claim 1, wherein the product is a flexible tube for an endoscope having a flexible tube base material and a resin layer covering the flexible tube base material, the resin layer being composed of an inner layer and an outer layer having different thickness ratios in the axial direction, the divided portions are portions obtained by dividing the flexible tube for an endoscope along the axial direction, the manufacturing conditions include the extrusion amount per unit time of the resin material of the inner layer by an extrusion molding machine and the extrusion amount per unit time of the resin material of the outer layer, and the predicted values ​​include the elasticity value of the divided portions, the thickness of the inner layer of the divided portions, and the thickness of the outer layer of the divided portions.

9. An information processing device as described in claim 1, wherein the product is a sheet formed by applying a liquid to a long support body being transported, the divided portions are portions obtained by dividing the sheet along the width direction, the manufacturing conditions include a control amount of a tension adjustment roller that is arranged upstream of the application point of the liquid and adjusts the tension applied to the support body at the application point, and the predicted value includes the application thickness of the liquid at the divided portions.

10. An information processing device according to claim 1, wherein the product is cells cultured in a culture tank, the divided parts are parts obtained by dividing the culture tank, the manufacturing conditions include culture conditions for the cells, and the predicted value includes an index value representing the desirability of the culture environment for the cells.

11. A method for operating an information processing device, comprising: using a machine learning model that outputs a predicted value of the quality of a divided portion obtained by dividing an entire product in response to input of manufacturing conditions for the divided portion; inputting the manufacturing conditions for the divided portion into the machine learning model and outputting the predicted value of the divided portion from the machine learning model; and deriving suitable manufacturing conditions that will result in the target quality of the entire product by solving an optimization problem that finds manufacturing conditions that minimize the value of an objective function having a term that includes the difference between the predicted value of the divided portion and a target value for the quality of the divided portion.

12. An operating program for an information processing device that causes a computer to execute processes including: using a machine learning model that outputs a predicted value of the quality of a divided portion obtained by dividing an entire product in response to input of manufacturing conditions for the divided portion; inputting the manufacturing conditions for the divided portion into the machine learning model and outputting the predicted value of the divided portion from the machine learning model; and deriving suitable manufacturing conditions that will result in the target quality of the entire product by solving an optimization problem that finds manufacturing conditions that minimize the value of an objective function having a term that includes the difference between the predicted value of the divided portion and the target value of the quality of the divided portion.

Images