Ejection pressure monitoring device and method, recording medium and coating device

Abstract

The present application provides a kind of to be able to detect the unexpected unknown abnormality of ejection pressure suitable for ejection pressure, and method, recording medium and coating device.Ejection pressure measuring unit (911) obtains pressure data.Feature quantity calculation unit (913) calculates the feature quantity of pressure data.Anomaly degree calculation unit (915) uses the unsupervised learning model (M1) of the output indicating the degree of distribution deviating from normal feature quantity as input, calculates the feature quantity calculated by feature quantity calculation unit Abnormal degree.Anomaly probability calculation unit (917) uses the supervised learning model (M2) of the output indicating the probability belonging to specific anomaly (A), (B) as input, calculates the anomaly probability calculated by feature quantity calculation unit (913) Feature quantity.Anomaly judging unit (919) uses the anomaly degree calculated by anomaly degree calculation unit (915) and the anomaly probability calculated by anomaly probability calculation unit to judge the anomaly of pressure data.

Description

Technical Field

[0001] The subject matter disclosed in this specification relates to a jet pressure monitoring device, a jet pressure monitoring method, a recording medium, and a coating apparatus. Background Technology

[0002] The manufacturing process of flat panel displays utilizes a device called a coating machine. Driven by a pump, the coating machine sprays a processing liquid from a slit nozzle, coating the entire substrate with the liquid. In recent years, with the increasing demand for higher-quality products, such coating machines aim to apply the processing liquid in a manner that creates a uniform film thickness across the entire substrate. While maintaining a constant film thickness uniformity, monitoring the spray pressure (discharge flow rate) of the processing liquid during production is crucial. For example, Patent Document 1 achieves optimization of parameters related to the spray pressure.

[0003] However, it is difficult to detect minute anomalies in the ejection pressure obtained from pressure sensors using a simple threshold. As a general technique, most solutions propose methods and systems for anomaly detection using machine learning.

[0004] For example, Patent Document 2 proposes a system capable of identifying the type (pattern) of anomalies based on provided sensor information. This system, in addition to enabling the development of specific strategies such as component replacement, can also function as a fault prediction system, capable of detecting subtle changes (signs) that may develop into faults.

[0005] Patent Document 1: Japanese Patent Application Publication No. 2020-040046

[0006] Patent Document 2: Japanese Patent Application Publication No. 2022-125288

[0007] As a system for predicting faults, Patent Document 2 proposes methods using unsupervised learning and supervised learning. However, because unsupervised learning is a method of learning data from a normal state and evaluating the degree of deviation of newly measured data from the normal state, it can detect anomalies but cannot identify the type of anomaly. Furthermore, supervised learning, by learning from training data pre-labeled with anomaly types, can identify anomaly types; however, because all anomalies must be assigned to a pre-conceived anomaly type during learning, it cannot handle unknown anomalies. Summary of the Invention

[0008] The purpose of this invention is to provide a technique that can appropriately detect unforeseen and unknown anomalies in ejection pressure.

[0009] To address the aforementioned issues, a first approach is a jet pressure monitoring device comprising: a pressure data acquisition unit that acquires pressure data representing the pressure change over time within a nozzle from which the treatment fluid is jetted; a feature quantity calculation unit that calculates feature quantities from the pressure data; an anomaly calculation unit that, using the feature quantities as input, calculates the anomaly degree of the feature quantities calculated by the feature quantity calculation unit using an unsupervised learning model that outputs anomaly degree, where the anomaly degree represents the degree to which the feature quantities deviate from a normal distribution; an anomaly probability calculation unit that, using the feature quantities as input, calculates the anomaly probability of the feature quantities calculated by the feature quantity calculation unit using a supervised learning model that outputs anomaly probability, where the anomaly probability represents the probability of belonging to a specific anomaly; and an anomaly judgment unit that uses the anomaly degree calculated by the anomaly quantity calculation unit and the anomaly probability calculated by the anomaly probability calculation unit to determine whether the pressure data is abnormal.

[0010] The second approach is based on the ejection pressure monitoring device of the first approach, wherein the unsupervised learning model is obtained by unsupervised learning using only the feature quantities of a normal plurality of the pressure data as input data.

[0011] The third approach is based on the ejection pressure monitoring device of the first or second approach, wherein the unsupervised learning model outputs Mahalanobis distance as the anomaly degree.

[0012] The fourth approach is based on the ejection pressure monitoring device of the first or second approach, wherein the supervised learning model outputs the probability of a specific anomaly belonging to a plurality of categories.

[0013] The fifth approach is a jet pressure monitoring device, comprising: a pressure data acquisition unit for acquiring pressure data representing the change in pressure within a nozzle dispensing treatment fluid over time; a feature quantity calculation unit for calculating feature quantities of the pressure data; an anomaly degree calculation unit for calculating an anomaly degree, representing the degree of deviation from the normal feature quantity, based on the feature quantity calculated by the feature quantity calculation unit; an anomaly probability calculation unit for calculating an anomaly probability based on the feature quantity calculated by the feature quantity calculation unit, using a supervised learning model that outputs an anomaly probability, wherein the anomaly probability represents the probability of belonging to a specific anomaly; and an anomaly judgment unit for judging the anomaly of the pressure data using the anomaly degree calculated by the anomaly degree calculation unit and the anomaly probability calculated by the anomaly probability calculation unit.

[0014] The sixth method is a method for monitoring ejection pressure, comprising: step a), acquiring pressure data representing the pressure change over time within a nozzle from which the treatment fluid is ejected; step b), calculating a feature quantity of the pressure data; step c), using the feature quantity as input, and employing an unsupervised learning model that outputs anomaly degree, calculating the anomaly degree of the feature quantity calculated in step b), wherein the anomaly degree represents the degree of deviation from the normal distribution of the feature quantity; step d), using the feature quantity as input, and employing a supervised learning model that outputs anomaly probability, calculating the anomaly probability of the feature quantity calculated in step b), wherein the anomaly probability represents the probability of belonging to a specific anomaly; and step e), using the anomaly degree calculated in step c) and the anomaly probability calculated in step d), determining the anomaly of the pressure data.

[0015] The seventh method is a method for monitoring ejection pressure, comprising: step A), acquiring pressure data representing the change in pressure within a nozzle ejecting treatment fluid over time; step B), calculating a feature quantity of the pressure data; step C), calculating an anomaly degree, representing the degree of deviation from the normal feature quantity, for the feature quantity calculated by step B); step D), using the feature quantity as input and a supervised learning model that outputs an anomaly probability, calculating the anomaly probability for the feature quantity calculated by step B), the anomaly probability representing the probability of belonging to a specific anomaly; and step E), using the anomaly degree calculated by step C) and the anomaly probability calculated by step D), determining the anomaly of the pressure data.

[0016] The eighth method is a recording medium containing a computer program that can be executed by a computer, wherein the computer program causes the computer to perform the ejection pressure monitoring method described in the sixth or seventh method.

[0017] The ninth embodiment is a coating apparatus comprising: a substrate holding section for holding a substrate; a nozzle for spraying a treatment liquid onto the substrate held by the substrate holding section; a pressure sensor for measuring the pressure within the nozzle; and a spray pressure monitoring device as described in any one of the first, second, and fifth embodiments.

[0018] According to methods one through nine, when anticipated anomalies have been assessed, anomalies can be identified by calculating the anomaly probability. Furthermore, even in the event of an unforeseen, unknown anomaly, the anomaly can be appropriately detected by calculating the anomaly degree.

[0019] According to the second method of the ejection pressure monitoring device, since it can learn only the distribution of normal characteristic quantities, it can appropriately detect anomalies.

[0020] According to the third-party ejection pressure monitoring device, since it can calculate the degree of anomaly by considering the correlation between the various dimensions of the characteristic quantities represented by the pressure data, it can more accurately capture the characteristic patterns or abnormal movements of the pressure data.

[0021] According to the ejection pressure monitoring device of the fourth method, since it is able to calculate the probability of a plurality of specific anomalies, it is possible to determine which of the plurality of anomalies the pressure data belongs to. Attached Figure Description

[0022] Figure 1 This is a schematic diagram showing the overall structure of the coating apparatus according to the embodiment.

[0023] Figure 2 This is a diagram showing the structure of the coating liquid supply mechanism.

[0024] Figure 3 It is a graph representing the pressure waveform of the ejection pressure.

[0025] Figure 4 This is a block diagram illustrating an example of the configuration of a control unit.

[0026] Figure 5 This is a diagram used to illustrate an example of a characteristic quantity.

[0027] Figure 6 This is a diagram used to illustrate other examples of characteristic quantities.

[0028] Figure 7 This is a block diagram representing the control unit that performs machine learning.

[0029] Figure 8 This is a diagram representing the flow of machine learning processing performed by the control unit.

[0030] Figure 9 This is a block diagram representing a control unit that monitors for anomalies in pressure data.

[0031] Figure 10 This is a diagram showing the process of monitoring the ejection pressure performed by the control unit.

[0032] Figure 11 It is a diagram that conceptually represents the process of calculating anomalies (Mahalanobis distance).

[0033] Figure 12 It is a diagram that conceptually represents the output of the anomaly probability calculation unit.

[0034] Figure 13 This is a diagram that conceptually represents an example of a comprehensive judgment of anomalies using anomaly degree and anomaly probability.

[0035] Explanation of reference numerals in the attached figures

[0036] 1 Coating device

[0037] 51. Chuck Mechanism (Substrate Holding Section)

[0038] 52 Adsorption and Travel Control Mechanism

[0039] 7. Coating facilities

[0040] 71 nozzles

[0041] 86 pressure sensor

[0042] 9. Control Unit (Ejection Pressure Monitoring Device)

[0043] 91 Computing Department

[0044] 911 Ejection Pressure Measurement Unit (Pressure Data Acquisition Unit)

[0045] 913 Characteristic Quantity Calculation Department

[0046] 915 Anomaly Calculation Department

[0047] 917 Anomaly Probability Calculation Department

[0048] 919 Anomaly Detection Department

[0049] 931 Computer Program

[0050] M1 unsupervised learning model

[0051] M2 supervised learning model Detailed Implementation

[0052] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. Furthermore, the constituent elements described in these embodiments are merely illustrative, and the scope of the present invention is not limited thereto. In the accompanying drawings, for ease of understanding, the dimensions or quantities of various parts may be exaggerated or simplified as needed.

[0053] Figure 1 This diagram schematically illustrates the overall structure of the coating apparatus 1 according to the embodiment. The coating apparatus 1 is a substrate processing apparatus that applies a coating liquid to the upper surface Sf of a substrate S. The substrate S is, for example, a glass substrate for a liquid crystal display device. Alternatively, the substrate S can be various substrates to be processed for electronic devices, such as semiconductor wafers, photomasks, plasma displays, magneto-optical disks (GODs), organic electroluminescent substrates (OLEDs), solar cells, silicon substrates, other flexible substrates, and printed circuit boards. The coating apparatus 1 is, for example, a slot coater.

[0054] exist Figure 1In this document, to illustrate the configuration of the elements of the coating apparatus 1, an XYZ coordinate system is defined. The transport direction of the substrate S is designated as the "X direction". The direction in which the substrate S travels in the X direction (downstream of the transport direction) is designated as the +X direction, and the opposite direction (upstream of the transport direction) is designated as the -X direction. Furthermore, the direction orthogonal to the X direction is designated as the Y direction, and the direction orthogonal to both the X and Y directions is designated as the Z direction. In the following description, the Z direction is defined as the vertical direction, and the X and Y directions are defined as the horizontal directions. Within the Z direction, the +Z direction is the upward direction, and the -Z direction is the downward direction.

[0055] The coating apparatus 1, facing the +X direction, includes, in sequence, an input conveyor 100, an input transfer unit 2, a floating stage unit 3, an output transfer unit 4, and an output conveyor 110. The input conveyor 100, input transfer unit 2, floating stage unit 3, output transfer unit 4, and output conveyor 110 form a transport path for the substrate S to pass through. Additionally, the coating apparatus 1 also includes a substrate transport unit 5, a coating mechanism 7, a coating liquid supply mechanism 8, and a control unit 9.

[0056] The substrate S is transported from the upstream device of the coating apparatus 1 to the input conveyor 100. The input conveyor 100 includes a roller conveyor 101 and a rotary drive mechanism 102. The rotary drive mechanism 102 rotates each roller of the roller conveyor 101. By rotating each roller of the roller conveyor 101, the substrate S is transported downstream (in the +X direction) in a horizontal orientation. "Horizontal orientation" means that the main surface (the surface with the largest area) of the substrate S is parallel to the horizontal plane (XY plane).

[0057] The input transfer unit 2 includes a roller conveyor 21 and a rotation and lifting drive mechanism 22. The rotation and lifting drive mechanism 22 rotates each roller of the roller conveyor 21 and raises and lowers the roller conveyor 21. By rotating the roller conveyor 21, the substrate S is transported downstream in a horizontal position (+X direction). In addition, by raising and lowering the roller conveyor 21, the position of the substrate S in the Z direction is changed. The substrate S is transferred from the input conveyor 100 to the floating worktable unit 3 via the input transfer unit 2.

[0058] like Figure 1As shown, the floating worktable section 3 is generally flat. The floating worktable section 3 is divided into three parts along the X direction. Facing the +X direction, the floating worktable section 3 has, in sequence, an inlet floating worktable 31, a coating worktable 32, and an outlet floating worktable 33. The upper surfaces of the inlet floating worktable 31, the coating worktable 32, and the outlet floating worktable 33 are located on the same plane. The floating worktable section 3 also has a lifting pin drive mechanism 34, a floating control mechanism 35, and a lifting drive mechanism 36. The lifting pin drive mechanism 34 causes the plurality of lifting pins disposed on the inlet floating worktable 31 to rise and fall. The floating control mechanism 35 supplies compressed air to the inlet floating worktable 31, the coating worktable 32, and the outlet floating worktable 33 to raise and fall the substrate S. The lifting drive mechanism 36 causes the outlet floating worktable 33 to rise and fall.

[0059] Multiple ejector holes are arranged in a matrix on the upper surfaces of the inlet floating stage 31 and the outlet floating stage 33, ejecting compressed air supplied by the floating control mechanism 35. When compressed air is ejected from each ejector hole, the substrate S floats upward relative to the floating stage 3. Consequently, the lower surface Sb of the substrate S separates from the upper surface of the floating stage 3 and is supported in a horizontal position. The distance (floatation amount) between the lower surface Sb of the substrate S and the upper surface of the floating stage 3 when the substrate S is in the floating state is, for example, 10 μm to 500 μm.

[0060] On the upper surface of the coating stage 32, nozzles for compressed air supplied by the lifting control mechanism 35 and suction nozzles for suction gas are alternately arranged in the X and Y directions. The lifting control mechanism 35 controls the amount of compressed air ejected from the nozzles and the amount of air drawn from the suction nozzles. This precisely controls the amount of levitation of the substrate S relative to the coating stage 32, ensuring that the position of the upper surface Sf of the substrate S above the coating stage 32 in the Z direction is a predetermined value. Furthermore, based on the detection results of the sensors 61 or 62 (described later), the control unit 9 calculates the amount of levitation of the substrate S relative to the coating stage 32. Preferably, the amount of levitation of the substrate S relative to the coating stage 32 can be adjusted with high precision via airflow control.

[0061] The substrate S, which is fed into the floating worktable section 3, is propelled in the +X direction by the roller conveyor 21 and transported onto the inlet floating worktable 31. The inlet floating worktable 31, the coating worktable 32, and the outlet floating worktable 33 support the substrate S in a floating state. The floating worktable section 3 can, for example, adopt the structure described in Japanese Patent No. 5346643.

[0062] The substrate transport unit 5 is positioned below the floating worktable unit 3. The substrate transport unit 5 includes a chuck mechanism 51 and a suction and travel control mechanism 52. The chuck mechanism 51 has a suction pad (not shown) disposed on the suction member. The chuck mechanism 51 abuts the suction pad against the periphery of the lower surface Sb of the substrate S, thereby supporting the substrate S from below. The suction and travel control mechanism 52 applies negative pressure to the suction pad, thereby adsorbing the substrate S onto the suction pad. Furthermore, the suction and travel control mechanism 52 causes the substrate transport unit 5 to reciprocate in the X direction.

[0063] With the lower surface Sb of the substrate S positioned above the upper surface of the floating stage 3, the chuck mechanism 51 holds the substrate S. The substrate S is held horizontally by the buoyancy applied by the floating stage 3 while its periphery is held by the chuck mechanism 51.

[0064] like Figure 1 As shown, the coating apparatus 1 has a sensor 61 for measuring the thickness of the substrate. The sensor 61 is positioned near the roller conveyor 21. The sensor 61 detects the position of the upper surface Sf of the substrate S held by the chuck mechanism 51 in the Z direction. Furthermore, by positioning the chuck (not shown) in a state where the substrate S is not held directly below the sensor 61, the sensor 61 can detect the position of the adsorption surface of the upper surface of the adsorption member in the vertical Z direction.

[0065] The chuck mechanism 51 holds the substrate S that has been loaded into the floating stage 3 and moves it in the +X direction. As a result, the substrate S is transported from above the inlet floating stage 31 to above the outlet floating stage 33 via the coating stage 32. Then, the substrate S is moved from the outlet floating stage 33 to the output transfer section 4.

[0066] The output transfer unit 4 moves the substrate S from a position above the outlet floating stage 33 toward the output conveyor 110. The output transfer unit 4 includes a roller conveyor 41 and a rotation and lifting drive mechanism 42. The rotation and lifting drive mechanism 42 rotates the roller conveyor 41 and moves it up and down in the Z direction. By rotating each roller of the roller conveyor 41, the substrate S moves in the +X direction. Furthermore, by raising and lowering the roller conveyor 41, the substrate S is displaced in the Z direction.

[0067] The output conveyor 110 includes a roller conveyor 111 and a rotary drive mechanism 112. The output conveyor 110 transports the substrate S in the +X direction by rotating the rollers of the roller conveyor 111, thus moving the substrate S out of the coating apparatus 1. Furthermore, the input conveyor 100 and the output conveyor 110 are part of the coating apparatus 1. However, the input conveyor 100 and the output conveyor 110 can also be assembled into an apparatus different from the coating apparatus 1.

[0068] The coating mechanism 7 applies a coating liquid to the upper surface Sf of the substrate S. The coating mechanism 7 is positioned above the transport path of the substrate S. The coating mechanism 7 has a nozzle 71. The nozzle 71 is a slit nozzle with a slit-shaped outlet on its lower surface. The nozzle 71 is connected to a positioning mechanism (not shown). The positioning mechanism positions the nozzle 71 at a coating position above the coating worktable 32. Figure 1 The position shown by the solid line in the middle moves between the maintenance position described later. The coating liquid supply mechanism 8 is connected to the nozzle 71. The coating liquid supply mechanism 8 supplies coating liquid to the nozzle 71, thereby spraying the coating liquid from the spray outlet disposed on the lower surface of the nozzle 71.

[0069] Figure 2 This diagram shows the structure of the coating liquid supply mechanism 8. The coating liquid supply mechanism 8 includes a pump 81, piping 82, a coating liquid replenishment unit 83, piping 84, an on / off valve 85, a pressure sensor 86, and a drive unit 87. The pump 81 is the supply source for supplying coating liquid to the nozzle 71, supplying the coating liquid through volume change. The pump 81 can be, for example, the bellows pump described in Japanese Patent Application Publication No. 10-61558. Figure 2 As shown, the pump 81 has a flexible tube 811 that can elastically expand and contract in the radial direction. One end of the flexible tube 811 is connected to the coating liquid replenishment unit 83 via a piping 82. The other end of the flexible tube 811 is connected to the nozzle 71 via a piping 84.

[0070] Pump 81 has a bellows 812 that can freely and elastically deform in the axial direction. Bellows 812 has a small bellows section 813, a large bellows section 814, a pump chamber 815, and a working disc section 816. Pump chamber 815 is disposed between flexible pipe 811 and bellows 812. An incompressible medium is sealed in pump chamber 815. Working disc section 816 is connected to drive unit 87.

[0071] The coating fluid replenishment unit 83 includes a storage tank 831 for storing the coating fluid. The storage tank 831 is connected to the pump 81 via a piping 82. An on / off valve 833 is installed on the piping 82. The on / off valve 833 opens and closes according to instructions from the control unit 9. If the on / off valve 833 is open, coating fluid can be replenished from the storage tank 831 to the flexible conduit 811 of the pump 81. Conversely, if the on / off valve 833 is closed, the replenishment of coating fluid from the storage tank 831 to the flexible conduit 811 of the pump 81 can be restricted.

[0072] Pipe 84 is connected to the output side of pump 81. An on / off valve 85 is installed on pipe 84. The on / off valve 85 opens and closes according to instructions from control unit 9. By opening and closing the on / off valve 85, it is possible to switch between dispensing coating liquid to nozzle 71 and stopping the dispensing of coating liquid to nozzle 71. A pressure sensor 86 is disposed on pipe 84. The pressure sensor 86 detects the pressure (dispensing pressure) of the coating liquid dispensed to nozzle 71 and outputs a signal indicating the detected pressure value to control unit 9.

[0073] Figure 3 It is a graph representing the pressure waveform of the ejection pressure. Figure 3 In the diagram, the horizontal axis represents time, and the vertical axis represents pressure value. In the coating apparatus 1, by adjusting various parameters of the movement of the specified working disc 816 (acceleration time, stabilization speed, stabilization speed time, deceleration time, etc.), appropriate optimization is performed, that is, the pressure waveform of the coating liquid sprayed from the nozzle 71 is made approximately the ideal pressure waveform.

[0074] like Figure 3 As shown, the spray pressure is measured from the time the coating liquid begins to spray from nozzle 71 until the time the coating liquid stops spraying from nozzle 71. Figure 3 In the example shown, the spray pressure at the moment nozzle 71 begins to spray the coating liquid (ta) and the spray pressure at the moment nozzle 71 ends to spray the coating liquid (te) are the initial pressure Pi. However, the pressures at the start and end of spraying are not always consistent with the initial pressure Pi. Figure 3 As shown, the ejection period is divided into the rising period T1, the transition period T2, the stable period T3, and the falling period T4.

[0075] The rising period T1 is the period from the moment when the coating liquid supply mechanism 8 starts spraying coating liquid from the nozzle 71 (that is, the moment when the coating liquid supply mechanism 8 starts moving the working disc 816) to the moment when the spray pressure reaches the target pressure Pt. In other words, if the coating liquid starts spraying from the nozzle 71 at moment ta, the spray pressure increases from the initial pressure Pi to the target pressure Pt from moment ta to moment tb.

[0076] The transition period T2 is the period from time tb to time tc after a predetermined vibration decay period. This vibration decay period is the period required for the ejection pressure to stabilize over time. For example, the vibration decay period can be set by the user through input operation on the input device 97 and stored in the storage unit 93.

[0077] The stabilization period T3 is the period from time tc to time td (i.e., the time when the coating liquid supply mechanism 8 begins to reduce the ejection pressure) from the target speed. In other words, from time tc to time td, the coating liquid supply mechanism 8 moves the working disc 816 at a constant speed, and begins to decelerate the working disc 816 at time td. Furthermore, during the stabilization period T3, the ejection pressure is essentially stable at the target pressure Pt. However, even during the stabilization period T3, the change in ejection pressure over time still includes minor fluctuations, and the ejection pressure may be greater than or less than the target pressure Pt.

[0078] The descent period T4 is the period from time td to time te when the coating liquid supply mechanism 8 stops spraying coating liquid from nozzle 71 (i.e., time te when the coating liquid supply mechanism 8 stops the working disc 816). In other words, the spray pressure decreases to the initial pressure Pi from time td to time te, and at time te, nozzle 71 stops spraying coating liquid.

[0079] like Figure 1 and Figure 2 As shown, sensor 62 is positioned at nozzle 71, from which coating liquid is supplied by coating liquid supply mechanism 8. Sensor 62 detects the height of substrate S in the Z direction non-contactly. Sensor 62 is electrically connected to control unit 9. Based on the detection result of sensor 62, control unit 9 measures the distance (separation distance) between the floating substrate S and the upper surface of coating stage 32. Then, control unit 9 adjusts the coating position of nozzle 71 using positioning mechanism based on the measured separation distance. Furthermore, sensor 62 can be an optical sensor or an ultrasonic sensor.

[0080] The coating mechanism 7 includes a nozzle cleaning standby unit 72. The nozzle cleaning standby unit 72 performs prescribed maintenance on the nozzles 71 positioned in the maintenance position. The nozzle cleaning standby unit 72 includes a roller 721, a cleaning section 722, and a roller groove 723. The nozzle cleaning standby unit 72 cleans the nozzles 71 and forms accumulated liquid, thereby adjusting the nozzle outlet of the nozzles 71 to a state suitable for coating processing. Furthermore, in the coating apparatus 1, in order to evaluate the spray pressure applied to the coating liquid, a simulated spraying of coating liquid from the nozzles 71 is performed while the nozzles 71 are positioned in the maintenance position.

[0081] Figure 4 This is a block diagram illustrating an example structure of the control unit 9. The control unit 9 controls the operation of various elements of the coating apparatus 1. The control unit 9 includes an arithmetic unit 91, a storage unit 93, a display 95, and an input device 97. For example, a desktop, laptop, or tablet computer can be used as the control unit 9. As described later, the control unit 9 functions as a spray pressure monitoring device to monitor for abnormalities in the spray pressure.

[0082] The arithmetic unit 91 is a processor composed of a CPU (Central Processing Unit). The storage unit 93 is composed of temporary storage devices such as RAM (Random Access Memory) and non-temporary auxiliary storage devices such as HDD (Hard Disk Drive) and SDD (Solid State Drive).

[0083] The display 95 is a device for displaying information to the user; specifically, it is an LCD display, etc. The input device 97 is a device for receiving user input operations; it is a mouse and keyboard, etc.

[0084] Storage unit 93 stores computer program 931. Computer program 931 is provided by recording medium M. That is, recording medium M records computer program 931 in a readable manner using computer, i.e., control unit 9. Recording medium M is, for example, a USB (Universal Serial Bus) memory, a DVD (Digital Versatile Disc) or other optical disc, disk, etc.

[0085] The arithmetic unit 91 executes the computer program 931, thereby functioning as the ejection control unit 910, the ejection pressure measurement unit 911, the characteristic quantity calculation unit 913, the anomaly degree calculation unit 915, the anomaly probability calculation unit 917, and the anomaly judgment unit 919.

[0086] The spray control unit 910 controls the operation (liquid delivery operation) of the pump 81 that supplies coating liquid to the nozzle 71 according to preset parameters.

[0087] The ejection pressure measuring unit 911 measures the ejection pressure. Specifically, the ejection pressure measuring unit 911 periodically acquires the ejection pressure measured by the pressure sensor 86 at a predetermined sampling period. The ejection pressure measuring unit 911 acquires the change in ejection pressure applied to the coating liquid over time (time series data) during the ejection of the coating liquid from the nozzle 71, and stores the acquired data as pressure data in the storage unit 93. The pressure data represents the pressure measured at each moment. The ejection pressure measuring unit 911 is an example of a pressure data acquisition unit.

[0088] The characteristic quantity calculation unit 913 derives the characteristic quantity from the ejection pressure measured by the ejection pressure measuring unit 911. For example, several characteristic quantities described in Japanese Patent Application Publication No. 2022-138109 can be used as the characteristic quantity.

[0089] Figure 5 This is a diagram illustrating an example of a characteristic quantity. In this example, the degree of overshoot generated during the rise of the ejection pressure is calculated as the characteristic quantity F1. Specifically, the characteristic quantity calculation unit 913 calculates the sign (positive or negative) of the second derivative Dif2 of the ejection pressure at time t11 when the ejection pressure reaches its maximum value Pmax. Then, the characteristic quantity calculation unit 913 calculates the time t12 when the sign of the second derivative changes twice from the sign at time t11. Then, the characteristic quantity calculation unit 913 calculates the characteristic quantity F1 of the change of ejection pressure with time during the initial oscillation period T2_s from time t11 to t12.

[0090] For example, the characteristic quantity calculation unit 913 selects the smaller of the minimum ejection pressure Pmin during the initial vibration period T2_s and the stable pressure Pm (the average ejection pressure during the stable period T3) as the target pressure Pg. Furthermore, the characteristic quantity calculation unit 913 can calculate the difference between the maximum pressure Pmax and the target pressure Pg (=Pmax-Pg) as the characteristic quantity F1.

[0091] Because the ejection pressure has an upward trend, the greater the overshoot of the ejection pressure change over time, the larger the characteristic quantity F1 will be.

[0092] Figure 6 This is a diagram illustrating another example of a characteristic quantity. In this example, the stability of the ejection pressure change over time during the transition period T2 is calculated as the characteristic quantity F2. Specifically, the characteristic quantity calculation unit 913 calculates the root mean square error (RMSE) (P_measure, Pm) of the average value of the ejection pressure during the transition period T2 and the ejection pressure during the stable period T3, i.e., the stable pressure Pm, as the characteristic quantity F2. The greater the oscillation (waveform vibration) of the ejection pressure change over time during the transition period T2, the larger the characteristic quantity F2 is.

[0093] The anomaly calculation unit 915 inputs the feature values ​​calculated by the feature value calculation unit 913 into the unsupervised learning model M1 to calculate the anomaly of the stress data.

[0094] The anomaly probability calculation unit 917 inputs the feature quantities calculated by the feature quantity calculation unit 913 into the supervised learning model M2 to calculate the probability of an anomaly. Specifically, stress data can be divided into normal states and one or more pre-defined abnormal states. The feature quantity calculation unit 913 calculates the probability of belonging to one or more pre-defined anomalies (hereinafter also referred to as "anomaly probability"). For example, during the learning of the supervised learning model M2, if two anomalies ("anomaly A" and "anomaly B") are pre-defined, the stress data is classified into normal states, states of anomaly A, and states of anomaly B. Then, the anomaly probability calculation unit 917 calculates the anomaly probability, which represents the probability of belonging to anomaly A or B.

[0095] The anomaly judgment unit 919 uses the anomaly degree calculated by the anomaly degree calculation unit 915 and the anomaly probability calculated by the anomaly probability calculation unit 917 to judge the anomaly of the pressure data.

[0096] <Learning Stage>

[0097] Figure 7 This is a block diagram representing control unit 9, which performs machine learning. Additionally, Figure 8 This is a diagram illustrating the flow of machine learning processing performed by control unit 9. Furthermore, Figure 8 and Figure 9The machine learning shown can be performed by a computer device other than control unit 9.

[0098] First, such as Figure 7 and Figure 8 As shown, in order to perform machine learning, a considerable amount of normal pressure data D1, a considerable amount of pressure data D2 belonging to anomaly A, and a considerable amount of pressure data D3 belonging to anomaly B are acquired in advance using the ejection pressure measurement unit 911. Figure 8 Step S11). The acquired pressure data D1 to D3 are stored in the storage unit 93. Then, the feature calculation unit 913 acquires the feature values ​​of each pressure data D1 to D3. Figure 8 Step S12).

[0099] Anomaly calculation unit 915 performs machine learning (unsupervised learning) to obtain unsupervised learning model M1. Figure 8 (Step S13). The anomaly calculation unit 915 saves the unsupervised learning model M1 (specifically, the parameters after learning) obtained through machine learning in the storage unit 93 (step S14).

[0100] In unsupervised learning, specifically, a set of features from a plurality of normal stress data points D1 is prepared as training data. Then, using this training data, machine learning based on the k-nearest neighbor algorithm is performed. The unsupervised learning model M1 takes the features as input and outputs an anomaly score, which represents the degree to which the distribution of features deviates from the normal stress data. While the anomaly score is preferably the Mahalanobis distance from the k-neighbors, it can also be the Euclidean distance or the Manhattan distance.

[0101] The anomaly probability calculation unit 917 performs machine learning (supervised learning) to obtain the supervised learning model M2. Figure 8 Step S15). Additionally, the anomaly probability calculation unit 917 stores the supervised learning model M2 (specifically, the learned parameters) obtained using machine learning in the storage unit 93. Figure 8 Step S16).

[0102] In supervised learning, firstly, features are prepared as input data for a plurality of normal stress data points, a plurality of abnormal stress data points (A), and a plurality of abnormal stress data points (B). Furthermore, the foundation of the supervised learning model M2 is, for example, a linear regression model. In supervised learning, the optimal parameters for predicting target values ​​(labels representing "normal," "abnormal A," and "abnormal B") are determined based on the input data. Specifically, to distinguish between the three states of normal, abnormal A, and abnormal B, the target value for normal is set to "0," the target value for abnormal A is set to "1," and the target value for abnormal B is set to "-1," and the model is trained to output the target value of the input data.

[0103] As a characteristic quantity, the rise period T1 in the figure can also be considered. However, when there are several patterns of ejection waveforms with different target pressures Pt, even under normal conditions, the rise period T1 will deviate due to the different patterns. Therefore, if the rise period T1 deviates, it is preferable not to use the rise period T1 as a characteristic quantity.

[0104] <Prediction Phase>

[0105] Figure 9 This is a block diagram of control unit 9, which monitors for anomalies in pressure data. Figure 10 This is a diagram showing the flow of the monitoring process for the ejection pressure performed by the monitoring control unit 9.

[0106] In this monitoring process, firstly, the ejection pressure measuring unit 911 of the control unit 9 acquires pressure data ( Figure 10 Step S21). Next, the feature calculation unit 913 calculates the feature quantities of the acquired pressure data ( Figure 10 Step S22).

[0107] The anomaly calculation unit 915 inputs the feature values ​​calculated by the feature calculation unit 913 into the unsupervised learning model M1 to calculate the anomaly ( Figure 10 Step S23). Additionally, the anomaly probability calculation unit 917 inputs the feature quantities calculated by the feature quantity calculation unit 913 into the supervised learning model M2 to calculate the probability of each anomaly (…). Figure 10 Step S24).

[0108] Figure 11 This is a conceptual diagram representing the calculation of outlier (Mahanobis distance). In Figure 11 In the example shown, the multiple hollow circles (“○”) represent the positions (distributions) of normal stress data used for machine learning in the feature space. The blank triangles (“△”) represent the positions of newly acquired stress data in the feature space. For example... Figure 11As shown, the anomaly calculation unit 915 calculates the Mahalanobis distance between the newly acquired pressure data (△) and the nearest normal pressure data (○) in the k-neighborhood as the anomaly. The anomaly is an index that indicates the degree to which the newly acquired pressure data (△) deviates from the normal pressure data (○).

[0109] Figure 12 This is a diagram conceptually representing the output of the anomaly probability calculation unit 917, and it illustrates the pressure waveforms for normal, anomaly A, and anomaly B. Figure 12 In the example shown, an overshoot condition where the ejection pressure rises excessively compared to the normal state is defined as an anomaly A. On the other hand, an anomaly B is defined as a condition where the ejection pressure does not rise sufficiently compared to the normal state, and the period during which the ejection pressure reaches the stable pressure Pm is prolonged.

[0110] The anomaly probability calculation unit 917 is configured to output a value between -1 and +1 for newly acquired stress data. As described above, in machine learning, the output value is 0 for normal, +1 for anomaly A, and -1 for anomaly B. Therefore, the closer the output value is to +1, the higher the probability of anomaly A; the closer the output value is to -1, the higher the probability of anomaly B.

[0111] return Figure 9 The anomaly determination unit 919 sends the anomaly degree calculated by the anomaly degree calculation unit 915 and the anomaly probability calculated by the anomaly probability calculation unit 917 (a value between -1 and 1) to the anomaly determination unit 919. The anomaly determination unit 919 compares the anomaly degree and the anomaly probability with a threshold to determine whether the newly obtained pressure data is normal or abnormal. Figure 10 Step S25).

[0112] Specifically, the anomaly detection unit 919 determines whether the anomaly level exceeds a predetermined threshold Th1 (see [reference]). Figure 11 If the anomaly level exceeds the threshold Th1, the anomaly judgment unit 919 determines that the pressure data is abnormal.

[0113] In addition, the anomaly detection unit 919 determines whether the anomaly probability exceeds a predetermined threshold. For example, such as... Figure 12 As shown, when determining whether a data point belongs to an anomaly A or an anomaly B, thresholds Th2 and Th3 are predetermined. Threshold Th2 is a value greater than 0 and less than 1, and threshold Th3 is a value greater than -1 and less than 0. If the anomaly probability exceeds threshold Th2 on the positive side (i.e., the anomaly probability value is greater than threshold Th2), the anomaly determination unit 919 determines that the new pressure data belongs to anomaly A. Conversely, if the anomaly probability exceeds threshold Th3 on the negative side (i.e., the anomaly probability is less than threshold Th3), the anomaly determination unit 919 determines that the new pressure data belongs to anomaly B.

[0114] For example, in Figure 12 In the example shown, because the anomaly probability of the first stress data point (data 1) does not exceed the thresholds Th2 and Th3, the stress data point is judged to be normal. Because the anomaly probability of the second stress data point (data 2) exceeds the threshold Th2 on the positive side, the stress data point is judged to be anomaly A. Because the anomaly probability of the third stress data point (data 3) exceeds the threshold Th3 on the negative side, the stress data point is judged to be anomaly B.

[0115] Furthermore, the anomaly detection unit 919 can also comprehensively use the results of anomaly detection based on anomaly degree and anomaly probability to make the final anomaly detection for the stress data. Figure 13 This is a diagram that conceptually represents an example of a comprehensive judgment of anomalies using both anomaly degree and anomaly probability. In Figure 13 In the diagram, the horizontal axis represents the probability of anomalies, and the vertical axis represents the degree of anomaly. When both the degree of anomaly and the probability of anomalies do not exceed the threshold (i.e., both are considered normal), the overall assessment of the pressure data is "normal" (equivalent to...). Figure 13 (The third quadrant in the model). Furthermore, when both the anomaly degree and the anomaly probability exceed the threshold (i.e., both are anomalous), the new stress data is judged as "abnormal." Conversely, when the anomaly probability is high and the anomaly degree is low, the new stress data is judged as "anomaly close to a normal distribution." Conversely, when the anomaly probability is low and the anomaly degree is high, the new stress data is judged as "unknown anomaly," i.e., an anomaly not anticipated during learning.

[0116] The anomaly detection unit 919 outputs the detection result obtained in step S25 to the outside. As an example, the anomaly detection unit 919 displays the detection result on the display 95 (step S26). The anomaly detection unit 919 can also print out the detection result using a printer. Alternatively, the anomaly detection unit 919 can output the detection result by illuminating a light or emitting an alarm sound using a speaker.

[0117] As described above, the control unit 9 (ejection pressure monitoring device) includes an ejection pressure measurement unit 911 as a pressure data acquisition unit, a feature quantity calculation unit 913, an anomaly calculation unit 915, an anomaly probability calculation unit 917, and an anomaly judgment unit 919. The ejection pressure measurement unit 911 acquires pressure data representing the change in pressure within the nozzle 71 from which the ejected treatment fluid is dispensed over time. The feature quantity calculation unit 913 calculates feature quantities from the pressure data. The anomaly calculation unit 915 takes the feature quantities as input and uses an unsupervised learning model M1 that outputs anomaly degree to calculate an anomaly degree for the feature quantities calculated by the feature quantity calculation unit 913, where the anomaly degree represents the degree to which the distribution of the feature quantities deviates from the normal distribution. The anomaly probability calculation unit 917 takes the feature quantities as input and uses a supervised learning model M2 that outputs anomaly probability to calculate an anomaly probability for the feature quantities calculated by the feature quantity calculation unit 913, where the anomaly probability represents the probability of belonging to a predetermined anomaly. The anomaly judgment unit 919 uses the anomaly degree calculated by the anomaly calculation unit 915 and the anomaly probability calculated by the anomaly probability calculation unit 917 to determine anomalies in the pressure data.

[0118] Additionally, the control unit 9 executes a pressure monitoring method. The pressure monitoring method includes: step a), acquiring pressure data representing the pressure change over time within the nozzle 71 from which the treatment fluid is ejected; step b), calculating a characteristic quantity of the pressure data; step c), using the characteristic quantity as input and an unsupervised learning model M1 that outputs anomaly degree, calculating anomaly degree for the characteristic quantity calculated in step b), the anomaly degree representing the degree of deviation from the normal distribution of the characteristic quantity; step d), using the characteristic quantity as input and a supervised learning model M2 that outputs anomaly probability, calculating anomaly probability for the characteristic quantity calculated in step b), the anomaly probability representing the probability of belonging to a specific anomaly; and step e), using the anomaly degree calculated in step c) and the anomaly probability calculated in step d), determining an anomaly in the pressure data.

[0119] Based on this structure, when judging pre-conceived anomalies, anomalies can be identified by calculating the anomaly probability. Furthermore, even in the event of unforeseen, unknown anomalies, the anomaly degree can be calculated to appropriately detect the unknown anomaly.

[0120] In addition, the unsupervised learning model M1 is obtained through unsupervised learning by using only the feature quantities of normal complex stress data as input data.

[0121] Based on this structure, since it is possible to learn only the distribution of normal feature quantities, it is possible to perform anomaly detection appropriately.

[0122] The unsupervised learning model M1 outputs Mahalanobis distance as anomaly score.

[0123] According to this structure, since the anomaly degree is calculated by considering the correlation between the dimensions of the feature quantities represented by the stress data, it is possible to capture the feature patterns or abnormal actions of the stress data more accurately.

[0124] The supervised learning model M2 outputs the probability of a specific anomaly belonging to a complex number of categories (e.g., anomalies A and B).

[0125] Based on this structure, since it is possible to calculate the probability of more than one type, it is possible to determine which of the many types of anomalies the pressure data belongs to.

[0126] <Other Implementation Methods>

[0127] In the above embodiment, the anomaly calculation unit 915 uses an unsupervised learning model M1 to calculate the anomaly degree. However, it is not mandatory to use the unsupervised learning model M1 to calculate the anomaly degree. For example, the anomaly calculation unit 915 can use a predetermined formula to calculate the difference between the normal feature quantity and the feature quantity of the newly obtained pressure data. This difference corresponds to the anomaly degree, which represents the degree of deviation from the normal feature quantity of the newly obtained pressure data. In addition, the normal feature quantity, which is the comparison object, can be defined by the feature quantities of a plurality of normal pressure data or by the feature quantity of a single normal pressure data.

[0128] The anomaly determination unit 919 compares the anomaly degree (difference) calculated by the anomaly degree calculation unit 915 with a preset threshold. Furthermore, if the anomaly degree exceeds the threshold, the anomaly determination unit 919 determines that the newly acquired pressure data is abnormal.

[0129] For example, Japanese Patent Application Publication No. 2012-098133 describes a method for calculating the similarity between the characteristic quantity of the pulse component under normal conditions and the characteristic quantity of the pulse component of the target device when detecting an abnormality in an inverter device based on the pulse component of the current waveform. Furthermore, Japanese Patent Application Publication No. 2006-026584 describes a method for calculating the difference between the voltage under normal conditions and the voltage of the target device when detecting an abnormality in ink ejection based on the voltage applied to the actuator that ejects the ink. These similarities and differences can also be used as the aforementioned degree of abnormality.

[0130] While the invention has been described in detail, the foregoing description is exemplary in all respects and the invention is not limited thereto. It is understood that many modifications not illustrated can be contemplated without departing from the scope of the invention. The structures described in the above embodiments and modifications can be appropriately combined or omitted, provided they do not contradict each other.

Claims

1. A coating apparatus, wherein, have: Substrate holding section, for holding the substrate; The nozzle sprays processing liquid onto the substrate held by the substrate holding portion; A pressure sensor measures the pressure inside the nozzle; and Ejection pressure monitoring device The ejection pressure monitoring device has the following features: The pressure data acquisition unit acquires pressure data representing the change of pressure within the nozzle over time via the pressure sensor. The feature calculation unit calculates the feature quantities of the pressure data; The anomaly calculation unit uses an unsupervised learning model that takes the feature quantity as input and outputs anomaly degree to calculate the anomaly degree of the feature quantity calculated by the feature quantity calculation unit. The anomaly degree represents the degree to which the feature quantity deviates from the normal distribution. The anomaly probability calculation unit uses a supervised learning model that takes the feature quantity as input and outputs the anomaly probability to calculate the anomaly probability on the feature quantity calculated by the feature quantity calculation unit. The anomaly probability represents the probability of belonging to a specific anomaly. as well as The anomaly determination unit acquires the anomaly degree calculated by the anomaly degree calculation unit and the anomaly probability calculated by the anomaly probability calculation unit. Based on a comparison of the thresholds of the anomaly degree and the anomaly probability, it determines whether the pressure data is normal, close to a normal distribution anomaly, an unknown anomaly, or an anomaly. The characteristic quantity is the degree of overshoot that occurs during the period from the moment the treatment liquid begins to be ejected from the nozzle until the ejection pressure reaches the target pressure, i.e., the rise in ejection pressure.

2. The coating apparatus according to claim 1, wherein, The unsupervised learning model is obtained by unsupervised learning using only the features of a complex number of normal stress data as input data.

3. The coating apparatus according to claim 1 or 2, wherein, The unsupervised learning model outputs Mahalanobis distance as the anomaly score.

4. The coating apparatus according to claim 1 or 2, wherein, The supervised learning model outputs the probability of a specific anomaly belonging to a complex number of categories.

5. A coating apparatus, wherein, have: Substrate holding section, for holding the substrate; The nozzle sprays processing liquid onto the substrate held by the substrate holding portion; A pressure sensor measures the pressure inside the nozzle; and Ejection pressure monitoring device The ejection pressure monitoring device has the following features: The pressure data acquisition unit acquires pressure data representing the change of pressure within the nozzle over time via the pressure sensor. The feature calculation unit calculates the feature quantities of the pressure data; Anomaly calculation unit calculates an anomaly degree, representing the degree to which the feature quantity deviates from the normal feature quantity, based on the feature quantity calculated by the feature quantity calculation unit. The anomaly probability calculation unit uses a supervised learning model that takes the feature quantity as input and outputs the anomaly probability to calculate the anomaly probability on the feature quantity calculated by the feature quantity calculation unit. The anomaly probability represents the probability of belonging to a specific anomaly. as well as The anomaly determination unit acquires the anomaly degree calculated by the anomaly degree calculation unit and the anomaly probability calculated by the anomaly probability calculation unit. Based on a comparison of the thresholds of the anomaly degree and the anomaly probability, it determines whether the pressure data is in a normal, near-normal distribution, unknown anomaly, or an abnormal state. The characteristic quantity is the degree of overshoot that occurs during the period from the moment the treatment liquid begins to be ejected from the nozzle until the ejection pressure reaches the target pressure, i.e., the rise in ejection pressure.

Images