Data collection method, learning method, and abnormality prediction method

Abstract

This data collection method is for collecting sensing data during operation of a device, the method including: (a) a step in which one or more virtual devices (4) corresponding on a one-to-one basis with actual devices (3) are prepared; and (b) a step in which the virtual devices (4) are made to operate under a plurality of operating conditions that can be executed by the actual devices (3), and in which virtual sensing data under each of the operating conditions is collected. The virtual devices (4) each implement an operation equivalent to one of the actual devices (3). The collection of sensing data, operation results, and the like using the virtual devices (4) makes it possible to efficiently collect data such as sensing data and operation results for devices under a plurality of operating conditions without increasing the number of times the actual devices (3) are activated.

Description

Data collection method, learning method, and anomaly prediction method

[0001] The present invention relates to a data collection method for efficiently collecting sensing data of a device with respect to its operating conditions, and a method for predicting an abnormality in the device.

[0002] When setting new operating conditions for various types of equipment, it is necessary to confirm in advance whether or not abnormalities will occur under those operating conditions. For example, in the case of an equipment that produces a product, whether or not the product maintains the desired quality, or whether or not abnormal operation of the equipment occurs, such as an abnormally high temperature inside the equipment. Furthermore, when setting new operating conditions, it is desirable to obtain the operation results of the equipment under a number of operating conditions in order to search for the best conditions.

[0003] In this way, running a large number of operating conditions using actual equipment to predict abnormalities or optimize operating conditions is costly in terms of time and money, and in some cases can cause problems such as a large environmental load.

[0004] Japanese Patent Application Laid-Open No. 2022-185394

[0005] Patent Document 1 discloses a technique for calculating virtual sensor data and virtual process result data using a process simulation model running within a semiconductor manufacturing device. In this way, by using a simulation model that mimics the operation of the device, it is possible to easily check a large number of operating conditions.

[0006] However, in actual devices, when various sensing data are collected during device operation, bias in the sensing data for each device may occur due to differences between devices, environmental factors such as device installation conditions, aging deterioration such as operating time of each device, etc. Therefore, even if appropriate operating conditions are found for one device, it is not guaranteed that similar results will be obtained for other devices of the same type.

[0007] For this reason, simply constructing a simulation model and testing operating conditions using the simulation model, as in Patent Document 1, does not necessarily result in the same results being obtained for multiple actual devices. In other words, it is most preferable to test multiple operating conditions for multiple devices with different machine differences and environments.

[0008] An object of the present invention is to provide a technique for efficiently collecting data such as sensing data and operation results of a device under a plurality of operating conditions.

[0009] In order to solve the above problem, the first invention is a data collection method for collecting sensing data during operation of a device, comprising: a) a step of preparing at least one virtual device that corresponds one-to-one with a real device; and b) a step of operating the virtual device under multiple operating conditions that can be executed by the real device, and collecting virtual sensing data under each of the operating conditions, wherein each of the virtual devices realizes operation equivalent to one of the real devices.

[0010] A second aspect of the present invention is the data collection method of the first aspect of the present invention, further comprising the step of: c) operating the actual device under at least one of the operating conditions and collecting actual sensing data.

[0011] A third aspect of the present invention is the data collection method of the second aspect of the present invention, wherein the number of the virtual devices operated in the step b) is greater than the number of the real devices operated in the step c).

[0012] A fourth invention is a data collection method of the second or third invention, in which step c) is performed before step b), and in step c) the actual sensing data is collected for the operating conditions in which an abnormality was detected, and in step b) the operating conditions in which an abnormality was detected in step c) are applied to the virtual device to collect the virtual sensing data.

[0013] The fifth invention is a method for learning a machine learning model that analyzes the operation of a device, comprising: a) a step of preparing at least one virtual device that corresponds one-to-one to a real device; b) a step of operating the virtual device under a plurality of operating conditions that can be executed by the real device and collecting virtual sensing data for each of the operating conditions; and p) a step of learning a machine learning model using the operating conditions and the virtual sensing data in step b), wherein each of the virtual devices realizes operation equivalent to one of the real devices.

[0014] A sixth invention is a learning method of the fifth invention, further comprising a step c) before step p), of operating the real device under at least one of the operating conditions and collecting actual sensing data, and in step p), learning of a machine learning model is performed using the operating conditions and the virtual sensing data in step b) and the operating conditions and the actual sensing data in step c).

[0015] A seventh aspect of the present invention is the learning method of the sixth aspect of the present invention, wherein the number of the virtual devices operated in the step b) is greater than the number of the real devices operated in the step c).

[0016] An eighth invention is a learning method of the sixth or seventh invention, in which step c) is performed before step b), and in step c) the actual sensing data is collected for the operating conditions under which an abnormality was detected, and in step b) the operating conditions under which an abnormality was detected in step c) are applied to the virtual device to collect the virtual sensing data.

[0017] A ninth invention is a learning method of the sixth or seventh invention, wherein in step b), for each of the operating conditions, the virtual sensing data and virtual result data relating to a virtual resultant object obtained by operation under the operating conditions are collected; in step c), for each of the operating conditions, the actual sensing data and actual result data relating to an actual resultant object obtained by operation under the operating conditions are collected; and in step p), the machine learning model is trained using the virtual result data and the actual result data as training data.

[0018] A tenth invention is a learning method of the ninth invention, wherein the virtual result data and the actual result data indicate whether or not there is an abnormality in the virtual result and the actual result, and the machine learning model predicts whether or not an abnormality will occur in the result from the input operating conditions.

[0019] An eleventh invention is an abnormality prediction method for predicting the possibility of an abnormality occurring in a resultant product from the operating conditions of an apparatus, comprising the steps of: S) training the machine learning model using the learning method of the tenth invention to generate a trained model; and T) inputting the operating conditions into the trained model generated by step S) to predict whether an abnormality will occur in the resultant product.

[0020] According to the first to eleventh aspects of the present invention, it is possible to efficiently collect sensing data and operation results of the device under a plurality of operating conditions while suppressing the number of times the actual device is operated.

[0021] In particular, according to the fifth to tenth inventions, a highly accurate trained machine learning model can be obtained using efficiently collected data.

[0022] In particular, according to the tenth and eleventh inventions, it is possible to use a highly accurate trained machine learning model to predict whether an abnormality will occur in the resulting product of the device.

[0023] FIG. 1 is a block diagram showing the configuration of an analysis system. FIG. 2 is a flowchart showing the flow of data collection processing, learning processing, and prediction processing. FIG. 3 is a block diagram illustrating the operating conditions performed by a real device and an analysis device during data collection processing, and the data obtained. FIG. 4 is a perspective view showing an exposure device, which is an example of a device to be analyzed. FIG. 5 is a schematic diagram showing the structure of a printing device, which is an example of a device to be analyzed.

[0024] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings. Note that the components described in the embodiment are merely examples and are not intended to limit the scope of the present invention. In the drawings, the dimensions and numbers of each part may be exaggerated or simplified as necessary to facilitate understanding.

[0025] <1. Analysis System> FIG. 1 is a block diagram showing the conceptual configuration of an analysis system 1 according to one embodiment of the present invention. The analysis system 1 is a system for collecting sensing data and result data for each operating condition of a predetermined type of device and analyzing the obtained data. As shown in FIG. 1, the analysis system 1 includes an analysis device 2, a real device 3, and multiple virtual devices 4. The analysis device 2, the real devices 3, and the virtual devices 4 are connected to each other via IoT devices and a network. In the example of FIG. 1, the analysis system 1 includes two real devices 3 and four virtual devices 4. The real devices 3 and the virtual devices 4 are all devices of the same type.

[0026] The analysis device 2 collects sensing data and result data for each operating condition from the real device 3 and the virtual device 4, and analyzes the obtained data. The analysis device 2 is configured, for example, by a computer having a processor such as a CPU, a memory such as RAM, and a storage unit 20 such as a hard disk drive. The storage unit stores computer programs and various data for executing data collection processing and analysis processing. The analysis device 2 can collect and analyze operating conditions, sensing data, and result data in the real device 3 and the virtual device 4 by reading the computer programs and various data from the storage unit to the memory and having the processor perform arithmetic processing in accordance with the computer programs and data.

[0027] As described above, the analysis device 2 includes the storage unit 20. The analysis device 2 also includes a learning unit 21 and a prediction unit 22 as processing units realized on software.

[0028] The analysis device 2 receives input from the real device 3 the operating conditions, real sensing data, and evaluation results when the real device 3 is operated. The analysis device 2 also receives input from the virtual device 4 the operating conditions, virtual sensing data, and evaluation results when the virtual device 4 is operated. These pieces of data are then stored in the storage unit 20 as a database DB.

[0029] The learning unit 21 uses the database DB to perform machine learning on the machine learning model M. This machine learning model M is a machine learning model that, for devices of the same type as the real device 3 and the virtual device 4, has operating conditions as input variables and predicted values ​​of sensing data and predicted values ​​of result data as output variables.

[0030] The prediction unit 22 uses the machine learning model M that has undergone machine learning processing in the learning unit 21 to predict the sensing data and result data from the operating conditions.

[0031] The real device 3 is a real device of a predetermined type. In this embodiment, the real device 3 includes a first real device 31 and a second real device 32.

[0032] The virtual device 4 is a device in a virtual space 40 that is created to resemble a specific type of device. The virtual space 40 is created on a computer. This virtual space 40 may be created on the computer that constitutes the analysis device 2, or may be created on another computer or cloud that is connected to the analysis device 2 via a network. In this embodiment, the virtual devices 4 include a first virtual device 41, a second virtual device 42, a third virtual device 43, and a fourth virtual device 44.

[0033] The virtual device 4 is a so-called digital twin that faithfully reproduces the real device 3 in a virtual space 40. The first virtual device 41 is a digital twin of the first real device 31. Therefore, the first virtual device 41 reproduces the positional deviation of sensors due to machine differences in the first real device 31, and operations that do not appear in other devices of the same type. Similarly, the second virtual device 42 is a digital twin of the second real device 32.

[0034] This analysis system 1 does not include real devices corresponding to the third virtual device 43 and the fourth virtual device 44. However, the third virtual device 43 and the fourth virtual device 44 are digital twins that faithfully reproduce actual real devices that actually exist. The third virtual device 43 and the fourth virtual device 44 may be, for example, digital twins generated by a manufacturer before shipping of real devices that are to be shipped to users different from the users of the first real device 31 and the second real device. They may also be digital twins of real devices that the user owns at other locations.

[0035] In this way, the virtual device 4 may include virtual devices corresponding to real devices other than the real device 3 from which the user can directly obtain sensing data and result data. This allows data relating to more devices to be obtained, and data in which the influence of machine differences has been smoothed can be collected as a database DB.

[0036] 2. Data Collection Processing, Learning Processing, and Prediction Processing Next, a description will be given of data collection processing, learning processing, and prediction processing using the analysis system 1 according to one embodiment of the present invention. Fig. 2 is a flowchart showing the flow of data collection processing, learning processing, and prediction processing in the analysis system 1. Fig. 3 is a block diagram illustrating the operating conditions performed by the real device 3 and the virtual device 4 during data collection processing, and the data obtained.

[0037] 2, in the analysis system 1, when performing the data collection process for training the machine learning model M and the training process for the machine learning model M, first, a real device 3 and a virtual device 4 are prepared (step S101). At this time, the prepared virtual device 4 corresponds one-to-one with the real device. The real device corresponding to the virtual device 4 may be a real device other than the real device 3 prepared in step S101. In this analysis system 1, in step S101, two real devices 31 and 32, virtual devices 41 and 42 corresponding to the two real devices 31 and 32, and virtual devices 43 and 44 corresponding to the other real devices are prepared.

[0038] Next, the real device 3 is operated under a plurality of operating conditions. Then, for each operating condition, actual sensing data detected by various sensors provided in the real device 3 is collected (step S102). Here, as shown in FIG. 3 , in this embodiment, the first real device 31 is operated under a plurality of first operating conditions C1 to obtain a plurality of sets of first actual sensing data Sr11. Furthermore, the second real device 32 is operated under a plurality of second operating conditions C2 to obtain a plurality of sets of second actual sensing data Sr22. The plurality of first operating conditions C1 and the plurality of second operating conditions C2 may all be different, all be the same, or may be partially the same.

[0039] Then, an evaluation result for the operation of the actual device 3 in step S102 is obtained (step S103). This evaluation result may be, for example, a determination of whether or not an abnormality occurred during operation of the actual device 3 based on the actual sensing data obtained in step S102. Furthermore, if there is an actual result obtained by operation of the actual device 3, this evaluation result may be an evaluation result of the actual result. For example, if the actual result is a printed matter, the evaluation result may be whether or not the printed matter is good. Furthermore, the evaluation result is not limited to a binary value such as good or bad, and may be indicated by a numerical value or the like.

[0040] In step S103, in this embodiment, as shown in FIG. 3, an evaluation result Rr11 is obtained when the first real device 31 is operated under the first operating condition C1, and an evaluation result Rr22 is obtained when the second real device 32 is operated under the second operating condition C2.

[0041] Then, the operating conditions, actual sensing data, and evaluation results for the operation of the real device 3 are transferred to the analysis device 2 and written to the database DB in the storage unit 20, thereby updating the database DB (step S104). That is, multiple sets of the first operating conditions C1, the first actual sensing data Sr11, and the first evaluation results Rr11 are transferred from the first real device 31 to the analysis device 2. Also, multiple sets of the second operating conditions C2, the second actual sensing data Sr22, and the second evaluation results Rr22 are transferred from the second real device 32 to the analysis device 2.

[0042] Next, the virtual device 4 is operated in the virtual space 40 using the same operating conditions C1 and C2 as the real device 3. Then, for each operating condition, virtual sensing data detected by various virtual sensors provided in the virtual device 4 is collected (step S105). At this time, each of the virtual devices 41, 42, 43, and 44 is operated in the virtual space using both the first operating condition C1 used for the first real device 31 and the second operating condition C2 used for the second real device 32. In this way, virtual sensing data for each virtual device 4 is obtained.

[0043] 3 , in this embodiment, the first virtual device 41 is operated in the virtual space 40 under the first operating condition C1 to obtain virtual sensing data Sv11. The first virtual device 41 is operated in the virtual space 40 under the second operating condition C2 to obtain virtual sensing data Sv12. Similarly, the second virtual device 42 is operated in the virtual space 40 under the first operating condition C1 and the second operating condition C2 to obtain virtual sensing data Sv21 and Sv22. The third virtual device 43 is operated in the virtual space 40 under the first operating condition C1 and the second operating condition C2 to obtain virtual sensing data Sv31 and Sv32. The fourth virtual device 44 is operated in the virtual space 40 under the operating condition C1 and the second operating condition C2 to obtain virtual sensing data Sv41 and Sv42.

[0044] Then, an evaluation result for the operation of the virtual device 4 in step S105 is obtained (step S106). For example, if the evaluation result of the actual result is obtained in step S103, the evaluation result obtained in step S106 will be an evaluation of the virtual result.

[0045] In this embodiment, as shown in FIG. 3, an evaluation result Rv11 is obtained when the first virtual device 41 is operated under the first operating condition C1, an evaluation result Rv12 is obtained when the first virtual device 41 is operated under the second operating condition C2, an evaluation result Rv21 is obtained when the second virtual device 42 is operated under the first operating condition C1, an evaluation result Rv22 is obtained when the second virtual device 42 is operated under the second operating condition C2, an evaluation result Rv31 is obtained when the third virtual device 43 is operated under the first operating condition C1, an evaluation result Rv32 is obtained when the third virtual device 43 is operated under the second operating condition C2, an evaluation result Rv41 is obtained when the fourth virtual device 44 is operated under the first operating condition C1, and an evaluation result Rv42 is obtained when the fourth virtual device 44 is operated under the second operating condition C2.

[0046] Then, the operating conditions, virtual sensing data, and evaluation results for the operation of the virtual device 4 are transferred to the analysis device 2 and written to the database DB in the storage unit 20, thereby updating the database DB (step S107). That is, multiple combinations of the operating conditions C1 and C2, the virtual sensing data Sv11, Sv12, Sv21, Sv22, Sv31, Sv32, Sv41, and Sv42, and the evaluation results Rv11, Rv12, Rv21, Rv22, Rv31, Rv32, Rv41, and Rv42 obtained in step S106 are transferred from the virtual space 40 to the analysis device 2.

[0047] In this way, the database DB stores combinations of “operating conditions, real sensing data, and evaluation results” collected by the real device 3 and combinations of “operating conditions, virtual sensing data, and evaluation results” collected using the virtual device 4.

[0048] Next, the learning unit 21 performs a learning process for the machine learning model M using the database DB obtained in this way (step S108). Specifically, machine learning is performed for the machine learning model M using the operating conditions as input variables and the sensing data (actual sensing data or virtual sensing data) and the evaluation results as teacher data for the output variables. Note that the output variable of the machine learning model M may be only the sensing data or only the evaluation results.

[0049] The evaluation results do not necessarily have to be the learning target of the machine learning model M. In that case, the evaluation result acquisition processes of steps S103 and S106 described above can be omitted, and the database DB does not have to include the evaluation results. In that case, the machine learning model M uses the operating conditions as input variables and outputs predicted sensing data.

[0050] The learning unit 21 passes the trained machine learning model M thus obtained to the prediction unit 22. As a result, after the machine learning process of step S108 is completed, the prediction unit 22 can predict predicted sensing data and predicted evaluation results from operating conditions using the trained machine learning model M. That is, for a new operating condition, the user can obtain predicted sensing data and predicted evaluation results for the operating condition using the trained machine learning model M (step S109).

[0051] At this time, if the evaluation result indicates the presence or absence of an abnormality, in step S109, new operating conditions are input to the generated trained machine learning model M, and it is possible to predict whether an abnormality will occur in the resulting product.

[0052] Furthermore, when adjusting the operating conditions, the user creates multiple operating conditions by gradually changing each item of the operating conditions, using the predicted sensing data and predicted evaluation results obtained in step S109 as reference. These operating conditions are then input into the trained machine learning model M to search for more appropriate operating conditions. By easily obtaining predicted sensing data and predicted evaluation results using the trained machine learning model M in the prediction unit 22, it is possible to reduce the time and financial costs involved in trial and error using multiple operating conditions in the actual device 3. At the same time, environmental impact can be reduced because unnecessary operation of the actual device 3 is not performed.

[0053] When training the machine learning model M, the greater the number of combinations of operating conditions included in the database DB that serves as training data and the corresponding sensing data and evaluation results, the higher the accuracy of the machine learning model M. Furthermore, considering differences between actual devices, collecting training data using as many devices as possible reduces bias due to differences between devices. However, preparing a large number of actual devices is practically difficult. Furthermore, running a large number of operating conditions using actual devices is costly in terms of time and money, and in some cases, it can cause problems such as a large environmental load.

[0054] Therefore, in this analysis system 1, sensing data for each operating condition is collected in a virtual space 40 using virtual devices 41 and 42 that are reproductions of real devices 31 and 32 that can actually be operated by the user, and virtual devices 43 and 44 that are reproductions of real devices that cannot be operated by the user but actually exist. This makes it possible to collect a large amount of learning data more efficiently than when only the real device 3 is used. As a result, the accuracy of the machine learning model M is improved.

[0055] 4 is a perspective view showing an exposure apparatus 9, which is an example of an apparatus to be analyzed by the analysis apparatus 2 according to the first embodiment of the present invention. The exposure apparatus 9 is a substrate processing apparatus that processes substrates W, and is an apparatus that irradiates light onto the upper surface of the substrate W on which a layer of photosensitive material (photosensitive layer) such as resist has been formed, to draw a pattern. The substrate W is, for example, a semiconductor substrate, a printed circuit board, a substrate for a color filter, a glass substrate for a flat panel display used in a liquid crystal display device or a plasma display device, or a substrate for an optical disk.

[0056] As shown in FIG. 4, the exposure apparatus 9 includes a base 91 , a stage 92 , a stage moving mechanism 93 , a gantry 94 , an exposure unit 95 , and a control unit 90 .

[0057] The base 91 has a rectangular shape when viewed from above, and supports the stage 92, the stage moving mechanism 93, and the gantry 94 from below.

[0058] The stage 92 has a support stage 921, an X stage 923, and a Y stage 925. The support stage 921 supports the substrate W. The X stage 923 rotatably supports the support stage 921 below the support stage 921. The Y stage 925 supports the X stage 923 and a sub-scanning mechanism 932 below the X stage 923.

[0059] The stage moving mechanism 93 is a mechanism that moves the support stage 921 in the main scanning direction (Y direction), the sub-scanning direction (X axis direction), and the rotation direction (rotation direction around the Z axis: θ axis direction). The stage moving mechanism 93 has a sub-scanning mechanism 932, a main scanning mechanism 934, and a rotation mechanism 935.

[0060] The sub-scanning mechanism 932 moves the X stage 923 in the X direction, which is the sub-scanning direction, relative to the Y stage 925. The main scanning mechanism 934 moves the Y stage 925 in the Y direction, which is the main scanning direction, relative to the base 91. The rotation mechanism 935 is provided on the X stage 923, and rotates the support stage 921 about a rotation axis extending in the Z direction. The sub-scanning mechanism 932, main scanning mechanism 934, and rotation mechanism 935 operate based on control commands from the control unit 90.

[0061] The gantry 94 is fixed to the base 91. The gantry 94 has two support columns 941 whose lower portions are fixed to the base 91, and a beam 943 that connects the upper portions of the two support columns 941. The two support columns 941 extending in the Z direction are spaced apart in the X direction. The gantry 94 is a member that supports the exposure unit 95. Each exposure head 951 of the exposure unit 95, which will be described later, is fixed to the beam 943 of the gantry 94.

[0062] The exposure unit 95 has one or more exposure heads 951. In this example, five exposure heads 951 are arranged along the X-axis direction. Each exposure head 951 has a spatial light modulator 950. The spatial light modulator 950 modulates laser light based on strip data corresponding to the drawing pattern.

[0063] The exposure unit 95 has a light irradiation section 953. The light irradiation section 953 irradiates the exposure heads 951 with laser light. The light irradiation section 953 is housed, for example, inside the gantry 94. The laser light emitted from the light irradiation section 953 is irradiated onto the spatial light modulator 950 of each exposure head 951.

[0064] The spatial light modulator 950 spatially modulates the laser light emitted from the light irradiation unit 953 on a channel-by-channel basis, and reflects necessary light that contributes to pattern drawing and unnecessary light that does not contribute to pattern drawing in different directions. Spatially modulating light means changing the spatial distribution (amplitude, phase, polarization, etc.) of light. The exposure head 951 irradiates the modulated laser light onto the substrate W moving directly below the exposure head 951. This exposes the drawing pattern onto the unprocessed substrate W.

[0065] The control unit 90 controls the light irradiation unit 953 of the exposure unit 95 to cause the light irradiation unit 953 to emit a line beam of light toward the exposure head 951. The control unit 90 also controls the stage movement mechanism 93 to move the support stage 921 relative to the exposure head 951 in the Y direction, which is the main scanning direction, and the X direction, which is the sub-scanning direction. Then, based on the drawing recipe stored in the storage unit 900 and position information of the support stage 921, the control unit 90 controls the spatial light modulator 950 of the exposure head 951. As a result, pattern light corresponding to the drawing pattern is formed, and the pattern light is emitted from the exposure head 951.

[0066] The drawing recipe describes, in a predetermined data format, for example, pattern data indicating a drawing pattern to be formed on the substrate W and various conditions for drawing (such as the amount of light emitted from the exposure unit 95 and the moving speed of the support stage 921). The pattern data is, for example, data obtained by rasterizing CAD (Computer Aided Design) data generated using CAD, and position information on the substrate W to be irradiated with light is recorded in pixel units.

[0067] When exposing a substrate W in the exposure device 9, the Y stage 925 is moved in the Y direction to perform main scanning movement, in which the substrate W supported on the support stage 921 is moved in the Y direction (main scanning direction). The exposure head 951 then irradiates the substrate W moving in the Y direction with pattern light. After the main scanning movement, the X stage 923 is moved in one direction in the X direction by the width of the pattern light to perform sub-scanning movement, in which the substrate W is moved. The exposure device 9 exposes the photosensitive layer of the substrate W by repeatedly performing such main scanning movement and sub-scanning movement.

[0068] When such an exposure apparatus 9 is the subject of analysis in the analysis system 1 and the analysis apparatus 2, the operating conditions, sensing data, and evaluation results are, for example, as follows: (Examples of operating conditions) Drawing pattern included in the drawing recipe Light amount emitted from the exposure unit 95 included in the drawing recipe Moving speed of the support stage 921 included in the drawing recipe Type of substrate W Ambient temperature of the exposure apparatus 9 (Examples of sensing data) Detected values ​​of temperature sensors provided at various locations in the exposure apparatus 9 Detected values ​​of vibration sensors provided at predetermined locations in the exposure apparatus 9 (Examples of evaluation results) Presence or absence of abnormalities in the detected values ​​of the temperature sensors Presence or absence of abnormalities in the detected values ​​of the vibration sensors Good / bad exposure result of the substrate W

[0069] 4. Example of Target Device / Printing Device FIG. 5 is a schematic diagram showing the structure of a printing device 8, which is an example of a device to be analyzed by the analysis device 2 according to the first embodiment of the present invention. The printing device 8 prints an image on the surface of a long strip-shaped substrate B (printing medium) by ejecting ink droplets from multiple heads 821-824 toward the substrate B while transporting the substrate B. The substrate B may be printing paper or a resin film. The substrate B may also be a metal foil or glass substrate. As shown in FIG. 5, the printing device 8 includes a transport mechanism 81, a printing unit 82, a drying unit 83, an ink supply unit 84, and a control unit 80.

[0070] The transport mechanism 81 is a mechanism that transports the substrate B in a transport direction along its longitudinal direction. The transport mechanism 81 of this embodiment has an unwinding section 811, multiple transport rollers 812, and a winding section 813. The substrate B is unwound from the unwinding section 811 and transported along a transport path formed by the multiple transport rollers 812. Each transport roller 812 rotates around an axis extending in a direction perpendicular to the transport direction, thereby guiding the substrate B to the downstream side of the transport path. After transport, the substrate B is collected in the winding section 813. In addition, tension is applied to the substrate B in the transport direction. This prevents the substrate B from sagging or wrinkling during transport.

[0071] The printing unit 82 is a processing unit that ejects ink droplets (hereinafter referred to as "ink droplets") onto the substrate B transported by the transport mechanism 81. The printing unit 82 of this embodiment has a first head 821, a second head 822, a third head 823, and a fourth head 824. The first head 821, the second head 822, the third head 823, and the fourth head 824 are arranged at intervals along the transport direction of the substrate B. The substrate B is transported below the four heads 821 to 824 with its printing surface facing upward. Each of the heads 821 to 824 ejects ink droplets of each color, K (black), C (cyan), M (magenta), and Y (yellow), which are color components of a multicolor image, from multiple nozzles toward the upper surface of the substrate B.

[0072] The drying unit 83 is located downstream in the transport direction of the heads 821 to 824 and blows heated gas toward the substrate B to dry the ink adhering to the substrate B. Note that if the ink used is UV-curable ink, the drying unit 83 may be replaced by an ultraviolet irradiation unit that irradiates the ink with ultraviolet light to cure the ink.

[0073] The ink supply unit 84 is a mechanism that supplies ink to each of the heads 821 to 824 of the printing unit 82. The ink supply unit 84 of this embodiment has a first ink supply unit 841, a second ink supply unit 842, a third ink supply unit 843, and a fourth ink supply unit 844. The first ink supply unit 841 supplies K ink to the first head 821. The second ink supply unit 842 supplies C ink to the second head 822. The third ink supply unit 843 supplies M ink to the third head 823. The fourth ink supply unit 844 supplies Y ink to the fourth head 824. The ink supply units 841 to 842 of each color adjust the temperature of the ink supplied to each of the heads 821 to 824 to be within a predetermined range.

[0074] The control unit 80 is an information processing device for controlling each unit of the printing device 8. The control unit 80 is communicatively connected to the above-mentioned transport mechanism 81, the four heads 821-824, the drying unit 83, and the four color ink supply units 841-844. Manuscript data is input to the control unit 80 from an external device. The manuscript data is image data to be printed by the printing device 8 to obtain a printed product. The control unit 80 controls the operation of the transport mechanism 81 and the four heads 821-824 based on the manuscript data. This allows the printing process in the printing device 8 to proceed. The control unit 80 also controls the operation of the ink supply process in the ink supply units 841-844 in accordance with the consumption of ink in the printing process by the four heads 821-824.

[0075] The operating conditions, sensing data, and evaluation results when such a printing device 8 is the subject of analysis in the analysis system 1 and the analysis device 2 are, for example, as follows: (Examples of operating conditions) Input data Type of substrate B Transport speed of substrate B in transport mechanism 81 Set temperature of ink in ink supply units 841 to 844 (Examples of sensing data) Detected values ​​of temperature sensors provided at various locations on the printing device 8 Detected values ​​of vibration sensors provided at predetermined locations on the printing device 8 Detected values ​​of tension sensors that measure the tension of substrate B provided on the transport path (Examples of evaluation results) Presence or absence of abnormalities in the detected values ​​of the temperature sensor Presence or absence of abnormalities in the detected values ​​of the vibration sensor Presence or absence of abnormalities in the detected values ​​of the tension sensor Good / bad print image printed on substrate B

[0076] 5. Modifications Although the embodiments have been described above, the present invention is not limited to the above and various modifications are possible.

[0077] In the above embodiment, all of the operating conditions C1 and C2 used in the real device 3 are applied to each of the virtual devices 4 to collect virtual sensing data, but the present invention is not limited to this. The operating conditions for running the virtual devices 4 in the virtual space 40 may be a portion of the operating conditions used in the real device 3, or different operating conditions may be applied to each virtual device 4. Furthermore, the operating conditions applied to the virtual devices 4 may include operating conditions that are not used in the real device 3.

[0078] In the above embodiment, when collecting sensing data, both the real device 3 and the virtual device 4 are operated under a plurality of operating conditions, and real sensing data and virtual sensing data are collected. However, the present invention is not limited to this. Operating conditions that can be executed by the real device 3 may be executed only by the virtual device 4, and only virtual sensing data may be collected.

[0079] When anomaly detection is performed in the analysis process using the machine learning model M, it is preferable to apply the operating conditions under which an anomaly was detected in the real device 3 in steps S102 and S103 to all virtual devices 4 in steps S104 and S105 to collect virtual sensing data. Generally, the operating conditions of the real device 3 are set so that the device operates within a range in which anomalies are unlikely to occur, making it difficult to collect a large amount of data under which anomalies are detected. For this reason, it is preferable to apply the operating conditions under which an anomaly occurred in the real device 3 to all virtual devices 4 in order to ensure as much data as possible.

[0080] Although the present invention has been described in detail, the above description is merely illustrative in all respects and does not limit the present invention. It is understood that countless variations not illustrated can be envisioned without departing from the scope of the present invention. The configurations described in the above embodiments and variations can be combined or omitted as appropriate as long as they are not mutually inconsistent.

[0081] 1: Analysis system 2: Analysis device 3, 31, 32: Real device 4, 41, 42, 43, 44: Virtual device 8: Printing device 9: Exposure device 20: Memory unit 21: Learning unit 22: Prediction unit 40: Virtual space C1, C2 Operating conditions M: Machine learning model Rr11, Rr22: Evaluation results Rv11, Rv12, Rv21, Rv22: Evaluation results Rv31, Rv32, Rv41, Rv42: Evaluation results Sr11, Sr22: Real sensing data Sv11, Sv12, Sv21, Sv22: Virtual sensing data Sv31, Sv32, Sv41, Sv42: Virtual sensing data

Claims

1. A data collection method for collecting sensing data during operation of a device, comprising: a) a step of preparing at least one virtual device that corresponds one-to-one with a real device; and b) a step of operating the virtual device under a plurality of operating conditions that can be executed by the real device, and collecting virtual sensing data under each of the operating conditions, wherein each of the virtual devices realizes an operation equivalent to that of one of the real devices.

2. A data collection method according to claim 1, further comprising the step of: c) operating said real device under at least one of said operating conditions and collecting real sensing data.

3. A data collection method according to claim 2, wherein the number of virtual devices operated in step b) is greater than the number of real devices operated in step c).

4. A data collection method according to claim 2 or claim 3, wherein step c) is performed before step b), and in step c) the actual sensing data is collected for the operating conditions under which an abnormality was detected, and in step b) the operating conditions under which an abnormality was detected in step c) are applied to the virtual device to collect the virtual sensing data.

5. A method for training a machine learning model that analyzes the operation of a device, comprising: a) preparing at least one virtual device that corresponds one-to-one with a real device; b) operating the virtual device under multiple operating conditions that can be executed by the real device and collecting virtual sensing data for each of the operating conditions; and p) training a machine learning model using the operating conditions and the virtual sensing data in step b), wherein each of the virtual devices realizes operation equivalent to that of one of the real devices.

6. A learning method according to claim 5, further comprising: c) a step prior to step p) of operating the real device under at least one of the operating conditions and collecting real sensing data; and in step p), learning of a machine learning model is performed using the operating conditions and the virtual sensing data in step b) and the operating conditions and the real sensing data in step c).

7. A learning method according to claim 6, wherein the number of virtual devices operated in step b) is greater than the number of real devices operated in step c).

8. A learning method according to claim 6 or claim 7, wherein step c) is performed before step b), and in step c) the actual sensing data is collected for the operating conditions under which an abnormality is detected, and in step b) the operating conditions under which an abnormality is detected in step c) are applied to the virtual device to collect the virtual sensing data.

9. A learning method according to claim 6 or 7, wherein in step b), for each of the operating conditions, the virtual sensing data and virtual result data relating to a virtual resultant object obtained by operation under the operating condition are collected; in step c), for each of the operating conditions, the actual sensing data and actual result data relating to an actual resultant object obtained by operation under the operating condition are collected; and in step p), the machine learning model is trained using the virtual result data and the actual result data as training data.

10. A learning method as described in claim 9, wherein the virtual result data and the actual result data indicate whether or not there is an abnormality in the virtual result and the actual result, and the machine learning model predicts whether or not an abnormality will occur in the result from the input operating conditions.

11. An anomaly prediction method for predicting the possibility of an abnormality occurring in a resultant product from the operating conditions of a device, comprising: S) a step of training the machine learning model using the learning method described in claim 10 to generate a trained model; and T) a step of inputting the operating conditions into the trained model generated by step S) and predicting whether an abnormality will occur in the resultant product.

Images