Processing device and method for generating trained model
The processing device uses a neural network-based trained model to accurately determine the state of components and objects, addressing imaging variability and enhancing processing efficiency by simplifying image processing.
Patent Information
- Application Number
- JP2023077742
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-10
- Publication Date
- 2024-11-21
AI Technical Summary
Existing processing devices face challenges in accurately determining the state of components and objects due to variations in imaging conditions, leading to non-uniform images and increased effort and cost for image processing, which hinders smooth processing.
A processing device equipped with a neural network-based trained model that outputs a determination result for the state of components and objects, utilizing machine learning to handle variations in imaging conditions and extract features without complex image processing.
Enables precise and efficient determination of component and object states, allowing for timely detection of abnormalities and preventing inappropriate processing.
Smart Images

Figure 2024162341000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a processing device that performs processing such as machining on an object, and a method for generating a trained model that determines the state of the processing device. [Background technology]
[0002] A wafer on which a plurality of devices are formed is divided and diced into individual chips, thereby manufacturing a plurality of chips (device chips) each including a device. A package substrate is formed by mounting a plurality of device chips on a predetermined substrate and covering and sealing the mounted device chips with a resin layer (mold resin). A package substrate is divided and diced into individual chips, thereby manufacturing a plurality of chips (packaged devices) each including a plurality of packaged device chips. The device chips and package devices are incorporated into various electronic devices such as mobile phones and personal computers.
[0003] When an object (workpiece) such as a wafer or package substrate is made into chips, the object is processed by various processing devices. For example, the object is processed by processing devices such as a cutting device or a laser processing device, and divided into a plurality of chips. Furthermore, before dividing the object, the object is processed and thinned by processing devices such as a grinding device or a polishing device, thereby obtaining thin chips. Furthermore, before and after processing the object, the object may be cleaned using a cleaning device, or inspected using an inspection device.
[0004] Processing devices such as processing devices, cleaning devices, and inspection devices are composed of various components such as a holding table that holds an object, a transport unit that transports the object, and a processing unit that performs a predetermined process on the object. The operation of each component of the processing device is controlled by a controller, so that the object is subjected to a predetermined process. For example, Patent Document 1 discloses a cutting device that cuts an object with a cutting blade while adjusting the position of the cutting blade according to the amount of positional deviation between the holding table and the object. [Prior art documents] [Patent documents]
[0005] [Patent Document 1] JP 2006-93333 A Summary of the Invention [Problem to be solved by the invention]
[0006] When processing objects such as wafers and package substrates in processing devices such as processing devices, cleaning devices, and inspection devices, the state of the components constituting the processing device and the object being processed by the processing device may be monitored so that a series of processing operations of the object by the processing device proceeds appropriately. For example, the processing device is equipped with a holding table, and the object is held by the table when processing (processing, cleaning, etc.) of the object. Then, while the processing device is in operation, the state of the holding table (presence or absence of damage, etc.) and the state of the object held by the holding table (presence or absence of positional deviation, etc.) are inspected at a predetermined timing. This makes it possible to prevent processing, cleaning, etc. from being performed on the object when the object is not being held properly by the holding table.
[0007] The state of the processing device's components or objects is determined based on an image acquired by, for example, capturing an image of the components or objects with an imaging unit (camera). However, due to variations in the position and angle of the components or objects at the time of capturing images, variations in the imaging characteristics of the imaging unit, variations in the amount of light and color of the lighting that illuminates the components or objects, and the like, the image patterns, shading, and the like contained in the images may vary. In addition, the images of the components may change due to deterioration of the components over time (scratches, discoloration, and the like).
[0008] As described above, in a processing device in which a component or an object is imaged under various imaging conditions, the images of the component or object obtained are often not uniform, and it is difficult to determine the state of the component or object with high accuracy by uniform image processing. On the other hand, if individual image processing is performed according to the imaging conditions, the effort and cost spent on reconfiguring the image processing algorithm and optimizing parameters increases. In addition, when actually determining the state of the component or object, the time required for image processing is extended, which hinders smooth processing of the object by the processing device.
[0009] The present invention has been made in consideration of such problems, and aims to provide a processing device and a method for generating a trained model that can determine the state of a component or object easily and with high accuracy. [Means for solving the problem]
[0010] According to one aspect of the present invention, there is provided a processing device for processing an object, the processing device comprising: a holding table for holding the object; an imaging unit for imaging the holding table; and a controller including a trained model constructed by machine learning so as to output, when an image of the holding table or an image of the holding table and the object acquired by the imaging unit is input, a determination result indicating the state of the holding table or a determination result indicating the state of the holding table and the object.
[0011] Preferably, the trained model is a neural network including an input layer to which the image is input and an output layer to which the determination result is output. Also, preferably, the processing device further includes an alarm unit that alarms the determination result when the determination result indicates an abnormality.
[0012] According to another aspect of the present invention, there is provided a method for generating a trained model for determining the state of a processing device that processes an object, the method including: an image collection step of collecting, as training images, images of each state of a holding table provided in the processing device, or images of the holding table and the object held by the holding table; and a learning step of generating, by machine learning using the training images, a trained model that, when an image showing the holding table or an image showing the holding table and the object is input, outputs a determination result indicating the state of the holding table or a determination result indicating the state of the holding table and the object.
[0013] In addition, preferably, in the image collection step, a plurality of first learning images corresponding to a normal state and a plurality of second learning images corresponding to an abnormal state are obtained as the learning images by imaging the holding table, or the holding table and the object. Effect of the Invention
[0014] A processing device according to one aspect of the present invention includes a trained model configured by machine learning to output a determination result indicating the state of a component and / or object when an image of a component and / or object of the processing device is input. This makes it possible to appropriately extract features from images of various components and / or objects without performing complex image processing, and to easily and accurately determine the state of the component and / or object. [Brief description of the drawings]
[0015] [Figure 1] FIG. [Diagram 2] FIG. [Diagram 3] FIG. [Figure 4] FIG. 2 is a block diagram showing a controller. [Diagram 5] 1 is a flowchart showing a method for generating a trained model. [Figure 6] FIG. 2 is an image diagram showing a first learning image. [Figure 7] FIG. 7(A) is an image diagram showing a second learning image acquired when an object is held by an abnormal holding table, and FIG. 7(B) is an image diagram showing a second learning image acquired when an object is placed in an inappropriate position. [Figure 8] Figure 8(A) is an image diagram showing a second learning image acquired when an abnormal object is held by the holding table, and Figure 8(B) is an image diagram showing a second learning image acquired when an object of an inappropriate size is held by the holding table. [Figure 9] FIG. 1 is a block diagram showing a trained model generated by machine learning. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0016] Hereinafter, an embodiment according to one aspect of the present invention will be described with reference to the accompanying drawings. First, a configuration example of a processing device according to this embodiment will be described. FIG. 1 is a perspective view showing a cutting device 2. The cutting device 2 is a processing device that cuts an object with an annular cutting blade. In FIG. 1, the X-axis direction (processing feed direction, front-back direction, first horizontal direction) and the Y-axis direction (indexing feed direction, left-right direction, second horizontal direction) are perpendicular to each other. Also, the Z-axis direction (height direction, up-down direction, vertical direction) is perpendicular to the X-axis direction and the Y-axis direction.
[0017] The cutting device 2 includes a rectangular parallelepiped base 4 that supports or houses each of the components that make up the cutting device 2. A rectangular opening 4a is provided at a corner on the front end side of the base 4. A cassette support table 6 that is raised and lowered by a lifting mechanism (not shown) is provided inside the opening 4a.
[0018] A cassette 8 capable of accommodating a plurality of objects (workpieces) 11 to be processed by the cutting device 2 is placed on the upper surface of the cassette support table 6. The objects 11 correspond to workpieces to be cut by the cutting device 2. Note that in Fig. 1, only the outline of the cassette 8 is shown by a dashed line.
[0019] 2 is a perspective view showing an object 11. For example, the object 11 includes a disk-shaped substrate (wafer) 13 made of a semiconductor material such as single crystal silicon. The substrate 13 has a front surface (first surface) 13a and a back surface (second surface) 13b that are generally parallel to each other. The substrate 13 is divided into a plurality of rectangular regions by a plurality of streets (planned division lines) 15 that are arranged in a lattice pattern so as to intersect with each other.
[0020] Devices 17 such as ICs (Integrated Circuits), LSIs (Large Scale Integration), LEDs (Light Emitting Diodes), MEMS (Micro Electro Mechanical Systems) devices, etc. are formed on the front surface 13a side of the multiple regions partitioned by the streets 15. By dividing the substrate 13 along the streets 15, multiple chips (device chips) each including a device 17 are manufactured.
[0021] When the substrate 13 is processed by the cutting device 2, the substrate 13 is supported by a frame 19 for ease of handling (transporting, holding, etc.) the substrate 13. The frame 19 is an annular member made of a metal such as SUS (stainless steel), and a circular opening 19a is provided in the center of the frame 19, penetrating the frame 19 in the thickness direction. The diameter of the opening 19a is larger than the diameter of the substrate 13.
[0022] A circular sheet 21 is fixed to the substrate 13 and the frame 19. For example, the sheet 21 may be a tape including a circular film-like substrate and an adhesive layer (glue layer) provided on the substrate. The substrate is made of a resin such as polyolefin, polyvinyl chloride, or polyethylene terephthalate. The adhesive layer is made of an epoxy-, acrylic-, or rubber-based adhesive. The adhesive layer may be an ultraviolet-curable resin that is cured by irradiation with ultraviolet light.
[0023] With the substrate 13 placed inside the opening 19a of the frame 19, the center of the sheet 21 is attached to the back surface 13b of the substrate 13 and the outer periphery of the sheet 21 is attached to the frame 19, whereby the substrate 13 is supported by the frame 19 via the sheet 21. In this way, an object 11 (substrate unit, frame unit) including the substrate 13, frame 19 and sheet 21 is formed. Then, a plurality of objects 11 are stored in a cassette 8 (see FIG. 1).
[0024] There are no limitations on the type, material, shape, structure, size, etc. of substrate 13. For example, substrate 13 may be a substrate (wafer) made of a semiconductor other than silicon (GaAs, InP, GaN, SiC, etc.), sapphire, glass (quartz glass, borosilicate glass, etc.), ceramics, resin, metal, etc. There are also no limitations on the type, number, shape, structure, size, arrangement, etc. of devices 17, and devices 17 may not be formed on substrate 13.
[0025] Furthermore, the substrate 13 may be a package substrate such as a CSP (Chip Size Package) substrate or a QFN (Quad Flat Non-leaded package) substrate. For example, a package substrate is formed by mounting a plurality of device chips on a predetermined substrate and covering and sealing the mounted device chips with a resin layer (mold resin). By dividing and singulating the package substrate, a plurality of chips (packaged devices) each including a plurality of packaged device chips are manufactured.
[0026] 1, a rectangular opening 4b is provided on the side of the opening 4a, with the longitudinal direction aligned with the X-axis direction. A holding table (chuck table) 10 for holding an object 11 is provided inside the opening 4b. A moving mechanism 12 for moving the holding table 10 along the X-axis direction is connected to the holding table 10.
[0027] The moving mechanism 12 is, for example, a ball screw type moving mechanism, and includes an X-axis ball screw (not shown) arranged along the X-axis direction, and an X-axis pulse motor (not shown) that rotates the X-axis ball screw. The moving mechanism 12 also includes a flat table cover 14 provided to surround the holding table 10. At the front and rear of the table cover 14, bellows-shaped dust-proof and drip-proof covers 16 that are expandable and contractible along the X-axis direction are provided. The table cover 14 and the dust-proof and drip-proof covers 16 are set so as to cover the components of the moving mechanism 12 (X-axis ball screw, X-axis pulse motor, etc.) arranged inside the opening 4b.
[0028] The upper surface of the holding table 10 is a flat surface roughly parallel to the horizontal plane (XY plane) and constitutes a holding surface 10a that holds the object 11. The holding surface 10a is connected to a suction source (not shown) such as an ejector via a flow path (not shown), a valve (not shown), and the like provided inside the holding table 10.
[0029] A rotary drive source (not shown), such as a motor, is connected to the holding table 10 to rotate the holding table 10 around a rotation axis roughly parallel to the Z-axis direction. The holding table 10 also includes a plurality of clamps 18 provided around the periphery of the holding surface 10a. The clamps 18 grip and fix the end (frame 19) of the object 11.
[0030] A pair of guide rails 20 are provided above the front end of the opening 4b. The pair of guide rails 20 are disposed generally parallel to each other along the Y-axis direction, and approach and separate from each other along the X-axis direction. Each of the pair of guide rails 20 includes a support surface that supports the lower side of the end (frame 19) of the object 11, and a clamping surface that protrudes upward from the support surface and contacts the outer periphery of the object 11 (the outer periphery of the frame 19). The pair of guide rails 20 align the object 11 by clamping the object 11 placed on the support surface between the clamping surfaces.
[0031] A gate-shaped first support structure 22 is disposed on the upper surface of the base 4 so as to straddle the opening 4b. A rail 24 is fixed along the Y-axis direction on the surface (front surface) side of the first support structure 22. A transport unit (transport mechanism, transport device) 26 that transports the object 11 is attached to the rail 24.
[0032] The transport unit 26 includes a moving part (moving mechanism) 28 mounted on the rail 24, and a holding part (holding mechanism) 30 connected to the moving part 28 and holding the object 11. The transport unit 26 transports the object 11 between the cassette 8 and the pair of guide rails 20, and between the pair of guide rails 20 and the holding table 10.
[0033] The moving unit 28 moves the holding unit 30 in the Y-axis direction by moving along the rail 24. The moving unit 28 also includes an air cylinder. The air cylinder includes a cylinder case and a columnar piston rod housed in the cylinder case. The piston rod is disposed along the Z-axis direction, and the holding unit 30 is fixed to the lower end of the piston rod. When the air cylinder is operated, the piston rod moves (raise and lower) along the Z-axis direction together with the holding unit 30.
[0034] For example, the holding unit 30 includes a plurality of suction pads that hold the frame 19. The lower surfaces of the suction pads form a holding surface that suction-holds the upper surface side of the target object 11 (the upper surface side of the frame 19). The holding surface of the suction pad is connected to a suction source (not shown) such as an ejector via a flow path (not shown), a valve (not shown), and the like formed inside the holding unit 30. In addition, a gripping unit (gripping mechanism) 30a that grips the end (frame 19) of the target object 11 is provided at the end on the opening 4a side (cassette 8 side) of the holding unit 30.
[0035] Further, a rail 32 is fixed along the Y-axis direction on the surface (front surface) side of the first support structure 22. A transport unit (transport mechanism, transport device) 34 that transports the object 11 is attached to the rail 32.
[0036] The transport unit 34 includes a moving part (moving mechanism) 36 mounted on the rail 32, and a holding part (holding mechanism) 38 connected to the moving part 36 and holding the object 11. The transport unit 34 transports the object 11 between the holding table 10 and the cleaning unit 50, and between the cleaning unit 50 and the pair of guide rails 20. The configurations and functions of the moving part 36 and the holding part 38 are similar to those of the moving part 28 and the holding part 30 of the transport unit 26, respectively.
[0037] A gate-shaped second support structure 40 is disposed behind the first support structure 22 so as to straddle the opening 4b. Ball screw type moving mechanisms 42A and 42B are provided at both ends of the surface (front) side of the second support structure 40. Processing units (machining unit, cutting unit) 44A and 44B that perform cutting on the target object 11 are fixed to the lower parts of the moving mechanisms 42A and 42B, respectively.
[0038] Each of the processing units 44A and 44B includes a cylindrical spindle (not shown) arranged along the Y-axis direction. An annular cutting blade 46 is attached to the tip of the spindle, and a rotation drive source such as a motor is connected to the base end of the spindle. By rotating the spindle with the rotation drive source, the cutting blade 46 rotates around a rotation axis that is approximately parallel to the Y-axis direction.
[0039] For example, a hub-type cutting blade (hub blade) is used as the cutting blade 46. The hub blade has an annular hub base made of a metal such as an aluminum alloy, and an annular cutting edge formed along the outer periphery of the hub base. The cutting edge of the hub blade is formed by an electroformed grinding wheel containing abrasive grains made of diamond, cubic boron nitride (cBN), or the like, and a binder such as a nickel plating layer that fixes the abrasive grains.
[0040] However, a washer-type cutting blade (washer blade) may also be used as the cutting blade 46. A washer blade is composed only of an annular cutting edge containing abrasive grains and a binder made of metal, ceramics, resin, or the like that fixes the abrasive grains.
[0041] The moving mechanism 42A moves the processing unit 44A along the Y-axis direction and the Z-axis direction. This adjusts the position of the cutting blade 46 attached to the processing unit 44A in the Y-axis direction and the Z-axis direction. Similarly, the moving mechanism 42B moves the processing unit 44B along the Y-axis direction and the Z-axis direction. This adjusts the position of the cutting blade 46 attached to the processing unit 44B in the Y-axis direction and the Z-axis direction.
[0042] Imaging units 48 are provided at positions adjacent to the processing units 44A and 44B, respectively. The imaging units 48 include an image sensor such as a CCD (Charged-Coupled Devices) sensor or a CMOS (Complementary Metal-Oxide-Semiconductor) sensor, and capture an image of the object 11 held by the holding table 10. There is no limitation on the type of imaging unit 48, and for example, a visible light camera or an infrared camera is used. The image acquired by the imaging unit 48 is used for alignment between the object 11 and the cutting blade 46, etc.
[0043] A circular opening 4c is provided on the side of the opening 4b. A cleaning unit (cleaning mechanism, cleaning device) 50 for cleaning the object 11 is provided inside the opening 4c. The cleaning unit 50 includes a holding table (spinner table) 52 that holds and rotates the object 11, and a nozzle 54 that supplies a cleaning liquid (cleaning liquid) to the object 11 held by the holding table 52.
[0044] The upper surface of the holding table 52 is a flat surface roughly parallel to the horizontal plane (XY plane) and constitutes a holding surface 52a that holds the target object 11. Note that a detailed configuration example of the holding table 52 will be described later (see FIG. 3). The nozzle 54 supplies a cleaning liquid toward the holding table 52. As the cleaning liquid, a liquid such as pure water, a mixed fluid in which a liquid (such as pure water) and a gas (such as air) are mixed, or the like can be used.
[0045] The cutting device 2 is also equipped with an imaging unit for monitoring that images one or more components selected from the components (such as the base 4 to the cleaning unit 50) that make up the cutting device 2. Images of the components are acquired by imaging the predetermined components with the imaging unit, and the states of the components are monitored based on the images of the components.
[0046] For example, the imaging unit 56 is installed in the center of the first support structure 22. The imaging unit 56 captures an image of the holding table 10 from above, and obtains an image of a part or the whole of the holding table 10. Then, based on the image of the holding table 10 obtained by the imaging unit 56, the state of the holding table 10 is determined.
[0047] However, the imaging unit 56 can also image the object 11 held by the holding table 10 together with the holding table 10. In this case, images of a part or the whole of the holding table 10 and the object 11 are acquired. Then, based on the images of the holding table 10 and the object 11 acquired by the imaging unit 56, the states of the holding table 10 and the object 11 are determined.
[0048] Moreover, an imaging unit 60 is installed via a columnar support member 58 at an upper corner portion on the surface (front surface) side of the first support structure 22. A base end portion of the support member 58 is fixed to the surface of the first support structure 22, and the imaging unit 60 is attached to a tip portion of the support member 58.
[0049] The imaging unit 60 captures an image of the holding table 52 of the cleaning unit 50 from above, and obtains an image of a part or the whole of the holding table 52. Then, based on the image of the holding table 52 obtained by the imaging unit 60, the state of the holding table 52 is determined.
[0050] However, the imaging unit 60 can also image the object 11 held by the holding table 52 together with the holding table 52. In this case, images of a part or the whole of the holding table 52 and the object 11 are acquired. Then, based on the images of the holding table 52 and the object 11 acquired by the imaging unit 60, the states of the holding table 52 and the object 11 are determined.
[0051] Examples of the configuration and type of the imaging units 56, 60 are similar to those of the imaging unit 48. However, it is preferable that the imaging units 56, 60 are cameras capable of changing the imaging area and magnification. A moving mechanism for moving the imaging units 56, 60 may be connected to the imaging units 56, 60. This makes it possible to image any location of the holding table 10, 52 and / or the object 11 at any angle and magnification. The imaging units 56, 60 may also include lighting (light source) for illuminating the subject during imaging. By imaging the holding table 10, 52 and / or the object 11 while illuminating them with lighting, a clear image can be obtained.
[0052] However, there are no limitations on the configuration, number, imaging target, installation location, etc. of the imaging units mounted on the cutting device 2. For example, the cutting device 2 may be equipped with multiple imaging units that image the holding tables 10 and 52 from different angles. In addition, the cutting device 2 may be equipped with imaging units that image components other than the holding tables 10 and 52.
[0053] A cover 62 that covers the components of the cutting device 2 is provided on the upper side of the base 4. In Fig. 1, only the outline of the cover 62 is shown by a dashed line. A display unit (display section, display device) 64 that displays various information related to the cutting device 2 is provided on the side of the cover 62. In addition, an alarm unit (alarm section, alarm device) 66 that notifies the operator of information is provided on the upper part of the cover 62.
[0054] For example, a touch panel display is used as the display unit 64. In this case, the display unit 64 also functions as an input unit (input section, input device) for inputting information to the cutting device 2, and the operator can input information such as processing conditions to the cutting device 2 by touching the display unit 64. However, the input unit may be an electronic device such as a keyboard, a mouse, a transceiver, or the like provided separately and independently from the display unit 64.
[0055] As the notification unit 66, for example, an indicator light (warning light) is used. In this case, when an abnormality occurs in the cutting device 2, the notification unit 66 lights up or flashes to notify the operator of the abnormality. However, there is no limit to the type of notification unit 66. For example, the notification unit 66 may be a speaker that notifies the operator of information by sound or voice, or may be a transmitter that transmits information to the outside of the cutting device 2 by wire or wirelessly. Furthermore, when information to be notified to the operator is displayed on the display unit 64, the display unit 64 also functions as a notification unit.
[0056] The cutting device 2 includes a controller (control unit, control section, control device) 68 that controls the cutting device 2. The controller 68 is connected to each of the components constituting the cutting device 2 (the cassette support base 6, the holding table 10, the moving mechanism 12, the clamp 18, the transport unit 26, the transport unit 34, the moving mechanisms 42A and 42B, the processing units 44A and 44B, the imaging unit 48, the cleaning unit 50, the imaging unit 56, the imaging unit 60, the display unit 64, the notification unit 66, etc.). The controller 68 outputs a control signal to the components of the cutting device 2 to control the operation of each component.
[0057] For example, the controller 68 is configured by a computer. Specifically, the controller 68 includes a processing unit that performs processes such as calculations necessary for the operation of the cutting device 2, and a storage unit that stores various information (data, programs, etc.) used for the operation of the cutting device 2. The processing unit includes a processor such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The storage unit includes a memory such as a ROM (Read Only Memory) and a RAM (Random Access Memory).
[0058] When cutting the object 11 with the cutting device 2, a cassette 8 containing a plurality of objects 11 is set on the cassette support table 6, and the objects 11 are carried out one by one from the cassette 8. Specifically, the transport unit 26 grips an end (frame 19) of the object 11 contained in the cassette 8 with the gripping portion 30a, and moves along the Y-axis direction so as to move away from the cassette 8. As a result, the object 11 is carried out from the cassette 8 and placed on the pair of guide rails 20. Thereafter, the object 11 is sandwiched between the pair of guide rails 20, and the object 11 is aligned.
[0059] Next, the transport unit 26 suction-holds the upper surface side of the object 11 (the upper surface side of the frame 19) with multiple suction pads, and transports the object 11 to the holding table 10. As a result, the substrate 13 is placed on the holding surface 10a of the holding table 10 via the sheet 21. In addition, the frame 19 is gripped by the multiple clamps 18. In this state, when the suction force (negative pressure) of the suction source is applied to the holding surface 10a, the substrate 13 is suction-held by the holding table 10 via the sheet 21.
[0060] The object 11 held by the holding table 10 is cut by a cutting blade 46 attached to the processing unit 44A or the processing unit 44B. For example, the substrate 13 is cut by the cutting blade 46 attached to the processing unit 44A, whereby the substrate 13 is divided into a plurality of chips.
[0061] Specifically, first, the holding table 10 is rotated to align the length direction of a predetermined street 15 (see FIG. 2) with the X-axis direction. In addition, the position of the processing unit 44A in the Y-axis direction is adjusted so that the cutting blade 46 is positioned on an extension line of the predetermined street 15. Furthermore, the height of the processing unit 44A is adjusted so that the lower end of the cutting blade 46 is positioned lower than the back surface 13b of the substrate 13 (the upper surface of the sheet 21).
[0062] Then, while rotating the cutting blade 46, the holding table 10 is moved along the X-axis direction. As a result, the holding table 10 and the cutting blade 46 move relatively along the X-axis direction (processing feed), and the cutting blade 46 cuts into the substrate 13 along the streets 15. As a result, the substrate 13 is cut along the streets 15. Thereafter, the same procedure is repeated to cut the substrate 13 along all the streets 15. As a result, the substrate 13 is divided into a plurality of chips, each of which includes a device 17 (see FIG. 2).
[0063] When cutting of the object 11 is completed, the object 11 is transported by the transport unit 34 to the cleaning unit 50, and is suction-held by the holding table 52. Then, the object 11 is cleaned by supplying a cleaning liquid from the nozzle 54 toward the object 11 while rotating the holding table 52.
[0064] When cleaning of the object 11 is completed, the object 11 is placed on the pair of guide rails 20 by the transport unit 34. Then, the object 11 is sandwiched between the pair of guide rails 20, and the object 11 is aligned. Thereafter, the transport unit 26 grips the end (frame 19) of the object 11 with the gripping portion 30a, and moves along the Y-axis direction so as to approach the cassette 8. As a result, the object 11 is carried in and stored in the cassette 8.
[0065] As described above, while the cutting device 2 is in operation, the state of one or more components selected from the components constituting the cutting device 2 (base 4 to cleaning unit 50, etc.) is monitored. This allows an abnormality occurring in the components of the cutting device 2 to be quickly detected, and prevents the target object 11 from being subjected to inappropriate processing. The components to be monitored can be appropriately set while taking into consideration the possibility of abnormality occurrence, etc. As an example, a case in which the holding table 52 of the cleaning unit 50 is monitored by the imaging unit 60 will be described below.
[0066] 3 is a perspective view showing the holding table 52. The holding table 52 is a spinner table provided in the cleaning unit 50, and holds and rotates the object 11 when the object 11 is cleaned.
[0067] The holding table 52 includes a cylindrical frame (main body) 70 made of metal such as SUS (stainless steel), glass, ceramics, resin, or the like. A cylindrical recess 70b is provided in the center of the upper surface 70a side of the frame 70. A disk-shaped holding member 72 made of a porous material such as porous ceramics is fitted into the recess 70b of the frame 70. The holding member 72 includes a large number of pores that communicate from the upper surface to the lower surface of the holding member 72. The upper surface of the holding member 72 forms a circular suction surface 72a that sucks the object 11 when the object 11 is held by the holding table 52.
[0068] The upper surface 70a of the frame 70 and the suction surface 72a of the holding member 72 are disposed on approximately the same plane, constituting the holding surface 52a of the holding table 52. The holding surface 52a is connected to a suction source (not shown) such as an ejector via pores contained in the holding member 72, flow paths (not shown) provided inside the frame 70, valves (not shown), and the like. In addition, a rotation drive source (not shown) such as a motor that rotates the holding table 52 around a rotation axis approximately parallel to the Z-axis direction is connected to the holding table 52.
[0069] The holding table 52 also includes a plurality of clamps 74. The clamps 74 are provided around the holding surface 52a and grip and fix an end portion (frame 19, see FIG. 2) of the target object 11. Note that, although FIG. 3 shows an example in which four clamps 74 are arranged at approximately equal intervals (90° intervals) along the circumferential direction of the holding surface 52a, there is no limitation on the number and arrangement of the clamps 74.
[0070] For example, clamp 74 includes a support part 76 fixed to the outer periphery of frame body 70, and a pressing part 78 that presses an end (frame 19) of object 11. Pressing part 78 is attached to support part 76 in a state in which it can rotate around a rotation axis that is approximately parallel to the tangent direction of holding surface 52a, and functions as a pendulum.
[0071] The upper end of the holding portion 78 constitutes a contact portion that comes into contact with the end (frame 19) of the target object 11. Meanwhile, the lower end of the holding portion 78 serves as a weight. When the holding table 52 is not rotating, the upper end of the holding portion 78 is positioned radially outward of the holding surface 54a due to the weight of the lower end.
[0072] When cleaning the object 11 in the cleaning unit 50 (see FIG. 1), first, the object 11 is held by the holding table 52. Specifically, the object 11 shown in FIG. 2 is placed on the holding table 52 so that the front surface 13a of the substrate 13 is exposed upward and the back surface 13b (sheet 21) of the substrate 13 faces the holding surface 52a. In this state, when the suction force (negative pressure) of the suction source is applied to the holding surface 52a (suction surface 72a), the substrate 13 is sucked and held by the holding table 52 via the sheet 21.
[0073] Next, the rotary drive source is operated to rotate the holding table 52 at a predetermined rotation speed (e.g., 800 rpm). At this time, the holding portion 78 rotates due to centrifugal force acting on the lower end (weight) of the holding portion 78, and the upper end of the holding portion 78 comes into contact with the upper surface side of the frame 19 (see FIG. 2) of the target object 11. As a result, the upper end of the holding portion 78 presses the frame 19 toward the holding surface 52a side, and the frame 19 is fixed to the holding table 52 by the multiple clamps 74.
[0074] Then, while the holding table 52 is rotating, a cleaning liquid is supplied from the nozzle 54 (see FIG. 1) toward the object 11. This cleans the object 11, and foreign matter such as processing waste adhering to the object 11 is washed away.
[0075] In this embodiment, the holding table 52 and the object 11 held by the holding table 52 are imaged by the imaging unit 60 (see FIG. 1). As a result, an image is acquired showing the state of the object 11 held by the holding table 52. Then, based on the image acquired by the imaging unit 60, the states of the holding table 52 and the object 11 are determined.
[0076] However, there may be variations in the pattern, shading, and the like of the image contained in the image due to variations in the positions and angles of the holding table 52 and the object 11 during imaging, variations in the imaging characteristics of the imaging unit 60, and variations in the amount of light and color of the lighting that illuminates the holding table 52 and the object 11. Furthermore, the image of the holding table 52 may change due to deterioration over time (scratches, discoloration, and the like) of the holding table 52.
[0077] As described above, in the cutting device 2 in which the holding table 52 and the object 11 are imaged under various imaging conditions, the images of the holding table 52 and the object 11 obtained are often not uniform, and it is difficult to determine the state of the holding table 52 and the object 11 with high accuracy by uniform image processing. On the other hand, if individual image processing is performed according to the imaging conditions, the effort and cost required for reconfiguring the image processing algorithm and optimizing parameters increases. Furthermore, when the state of the holding table 52 and the object 11 is actually determined, the time required for image processing becomes longer, which hinders smooth processing of the object 11 by the cutting device 2.
[0078] Therefore, in this embodiment, a trained model configured by machine learning is used so that when images of the holding table 52 and the object 11 are input, a determination result indicating the state of the holding table 52 and the object 11 is output. This makes it possible to appropriately extract features from various images of the holding table 52 and the object 11 without performing complex image processing, and to easily and highly accurately determine the state of the holding table 52 and the object 11.
[0079] Fig. 4 is a block diagram showing the controller 68. In addition to a block showing the functional configuration of the controller 68, Fig. 4 also shows a schematic diagram of the holding table 52, the object 11 held by the holding table 52, the imaging unit 60, the display unit 64, and the notification unit 66. The states of the holding table 52 and the object 11 are determined by the controller 68.
[0080] The controller 68 includes a determination section 92 that determines the state of the holding table 52 based on an image (captured image) 90 of the holding table 52 and the object 11. When the holding table 52 holding the object 11 is imaged by the imaging unit 60, an image 90 showing a part or the whole of the holding table 52 and the object 11 is acquired and input to the determination section 92. Then, the determination section 92 determines the state of the holding table 52 and the object 11 based on the image 90, and outputs the determination result.
[0081] The controller 68 also includes a storage unit 94 capable of storing various types of information (data, programs, etc.), and a notification control unit 96 that controls the notification of information by the display unit 64 and / or the notification unit 66. The storage unit 94 corresponds to a computer-readable non-transitory recording medium including memories such as ROM and RAM. The determination result of the determination unit 92 is output to the storage unit 94 and the notification control unit 96.
[0082] The determination unit 92 includes a trained model 100 that determines the state of the holding table 52 and the object 11. The trained model 100 is configured by machine learning so as to output a determination result indicating the state of the holding table 52 and the object 11 when an image 90 is input. There is no limit to the type of trained model 100, and a support vector machine (SVM), logistic regression, k-nearest neighbor method, decision tree, neural network, etc. can be used. As an example, a case where the trained model 100 is a neural network NN will be described below.
[0083] The neural network NN is a hierarchical neural network, and includes an input layer 102 to which data is input, an output layer 104 to which data is output, and a plurality of hidden layers (intermediate layers) 106 provided between the input layer 102 and the output layer 104. The input layer 102, the output layer 104, and the hidden layer 106 each include a plurality of neurons (units, nodes). The neurons in the input layer 102 are connected to the neurons in the first hidden layer 106, and the neurons in the output layer 104 are connected to the neurons in the hidden layer 106, which is the final layer. The neurons in the hidden layer 106 are connected to the neurons in the input layer 102 or the hidden layer 106 of the previous layer, and the neurons in the output layer 104 or the hidden layer 106 of the subsequent layer.
[0084] There are no limitations on the number of neurons included in the input layer 102, output layer 104, and hidden layer 106, the activation function of each neuron, the connection relationships between each neuron, etc. Furthermore, there are no limitations on the number of hidden layers 106. A neural network NN that includes two or more hidden layers 106 can be called a deep neural network (DNN). Furthermore, learning of a deep neural network can be called deep learning.
[0085] The neural network NN is trained so that when an image 90 is input to the input layer 102, a determination result indicating the state of the holding table 52 and the object 11 is output from the output layer 104. There is no limitation on the training method of the neural network NN. A specific example of the training method of the neural network NN will be described later (see FIG. 9).
[0086] When determining the state of the holding table 52, the holding table 52 holding the object 11 is imaged by the imaging unit 60, and an image 90 representing the holding table 52 and the object 11 is acquired. Then, the image 90 is input to the neural network NN, and the states of the holding table 52 and the object 11 are determined by inference of the neural network NN. Specifically, a calculation using the image 90 as input data is sequentially executed in the input layer 102, hidden layer 106, and output layer 104, and data corresponding to the result of classifying the image 90 according to the states of the holding table 52 and the object 11 is output from the output layer 104. In this way, the states of the holding table 52 and the object 11 are automatically determined.
[0087] As described above, since the neural network NN handles the classification problem of the image 90, it is preferable to use a convolutional neural network (CNN) as the neural network NN. In this case, a convolutional layer, a pooling layer, a local contrast normalization (LCN) layer, a fully connected layer, etc. are provided as the hidden layer 106. Then, features of the image 90 are extracted by a convolution operation (generation of a feature map) in the convolutional layer and a pooling process in the pooling layer. Also, an operation for classifying the image 90 is performed in the fully connected layer. Then, a numerical value corresponding to the classification result of the image 90 is output from the output layer 104.
[0088] For example, when the judgment unit 92 judges whether the state of the storage table 52 and the object 11 is normal or abnormal (two-class classification), the output layer 104 can be configured with two neurons to which a softmax function is applied as an activation function. One neuron outputs a numerical value (normal output value) corresponding to the probability that the storage table 52 and the object 11 are normal, and the other neuron outputs a numerical value (abnormal output value) corresponding to the probability that the storage table 52 and / or the object 11 are abnormal.
[0089] The trained model 100 may be realized by either software or hardware. For example, the operations in the input layer 102, the output layer 104, and the hidden layer 106 of the neural network NN are described by a program, and the program is stored in the storage unit 94. When the program is read from the storage unit 94 and executed by the controller 68, the computer executes a process of inputting the image 90 of the holding table 52 to the trained model 100 and a process of executing the operations of the trained model 100. As a result, the states of the holding table 52 and the target object 11 are determined.
[0090] The judgment unit 92 also includes a judgment result processing unit 108 that processes the result of judgment by the trained model 100. For example, the judgment result processing unit 108 compares a normal output value and an abnormal output value output from the output layer 104 of the neural network NN. If the normal output value is greater than the abnormal output value, the judgment result processing unit 108 outputs a signal (normal judgment signal) indicating that the states of the holding table 52 and the object 11 are normal. On the other hand, if the abnormal output value is greater than the normal output value, the judgment result processing unit 108 outputs a signal (abnormal judgment signal) indicating that the states of the holding table 52 and the object 11 are abnormal. The normal judgment signal and the abnormal judgment signal correspond to the result of judgment by the judgment unit 92.
[0091] However, there is no restriction on the content of the processing by the judgment result processing unit 108. For example, the judgment result processing unit 108 may judge the state of the holding table 52 and the object 11 by comparing the normal output value and the abnormal output value with a reference value (lower limit value) previously stored in the storage unit 94. Specifically, when the normal output value is equal to or greater than the reference value, the judgment result processing unit 108 outputs a normal judgment signal. On the other hand, when the abnormal output value is equal to or greater than the reference value, the judgment result processing unit 108 outputs an abnormal judgment signal. The reference value to be compared with the normal output value and the abnormal output value can be appropriately set according to the required judgment accuracy. For example, the threshold value is set to 0.6 or more (60% or more), 0.8 or more (80% or more), or 0.9 or more (90% or more).
[0092] If both the normal output value and the abnormal output value are less than the threshold value, the image 90 used for the judgment may be acquired again. Specifically, after changing the imaging conditions of the holding table 52 and the object 11 (positions, angles, imaging area, magnification, etc. of the holding table 52 and the object 11, the imaging unit 60 captures images of the holding table 52 and the object 11 again, and acquires the image 90 again. Then, the judgment unit 92 judges the state of the holding table 52 and the object 11 based on the newly acquired image 90.
[0093] Furthermore, the judgment result processing unit 108 may output, as a judgment result, either or both of the normal output value and the abnormal output value input from the trained model 100. In this case, the judgment result processing unit 108 may perform a predetermined normalization process on the normal output value and the abnormal output value.
[0094] The image 90 and the information on the determination result output from the determination unit 92 are stored in a storage unit 94 in a linked state. As a result, the determination results on the states of the holding table 52 and the target object 11 are accumulated in the storage unit 94 together with the image 90.
[0095] Furthermore, when the result of the determination by the determination unit 92 is input to the notification control unit 96, the notification control unit 96 causes the display unit 64 and / or the notification unit 66 to notify the determination result as necessary. Specifically, the notification control unit 96 generates a control signal for causing the display unit 64 to display the determination result of the determination unit 92, and outputs the control signal to the display unit 64. The notification control unit 96 also generates a control signal for causing the notification unit 66 to notify the determination result of the determination unit 92, and outputs the control signal to the notification control unit 96. As a result, the determination result of the determination unit 92 is notified to the operator.
[0096] For example, the display unit 64 displays the states (normal or abnormal) of the holding table 52 and the object 11 using a character string (message), a code, a figure, a symbol, or the like. The display unit 64 may also display the probability of the states (normal or abnormal) of the holding table 52 and the object 11 using numbers, graphs, or the like. For example, a value corresponding to a normal output value output from the trained model 100 is displayed as a percentage on the display unit 64 as the probability that the holding table 52 and the object 11 are normal. A value corresponding to an abnormal output value output from the trained model 100 is displayed as a percentage on the display unit 64 as the probability that at least one of the holding table 52 and the object 11 is abnormal.
[0097] Furthermore, the display unit 64 may display the image 90 that is the subject of the judgment, in addition to the state or the probability of the holding table 52 and the object 11. In this case, the operator can simultaneously check and compare the image 90 and the judgment result. This allows the operator to check whether the judgment made by the trained model 100 is valid based on his or her own judgment.
[0098] When the determination result of the determination section 92 indicates that the holding table 52 and the object 11 are normal, a control signal is output from the controller 68 to each component of the cutting device 2, and the cutting device 2 continues to process (transport, process, clean, etc.) the object 11. In addition, a control signal is input from the notification control section 96 to the display unit 64, and information informing that the holding table 52 and the object 11 are normal is displayed on the display unit 64.
[0099] On the other hand, when the judgment result of the judgment unit 92 indicates that the state of the holding table 52 and the object 11 is abnormal, a control signal is output from the controller 68 to each component of the cutting device 2, and the processing of the object 11 by the cutting device 2 is interrupted. In addition, a control signal is input from the notification control unit 96 to the display unit 64 and the notification unit 66, and information notifying that the state of the holding table 52 and the object 11 is abnormal is displayed on the display unit 64, and the notification unit 66 lights up in a color or pattern notifying of the abnormality. As a result, the abnormality in the state of the holding table 52 and the object 11 is notified to the operator, and the operator can take measures according to the content of the abnormality.
[0100] There is no limit to the timing for determining the states of the holding table 52 and the object 11. For example, before and after cleaning of the object 11 by the cleaning unit 50 (see FIG. 1), the imaging unit 60 captures images of the holding table 52 and the object 11. Then, an image 90 acquired by the imaging is input to a determination unit 92, and the states of the holding table 52 and the object 11 are determined.
[0101] As described above, the cutting device 2 according to this embodiment is equipped with the trained model 100 capable of determining the state of the holding table 52 and the object 11 based on the image 90 of the holding table 52 and the object 11. This makes it possible to monitor the state of the holding table 52 and the object 11, and prevents the object 11 from being washed in a state in which the holding table 52 is not holding the object 11 properly. In addition, since the trained model 100 is configured by machine learning, it is possible to appropriately extract features from the image 90 of the holding table 52 and the object 11 captured under various imaging conditions, and to determine the state of the holding table 52 and the object 11.
[0102] There is no limitation on the method of mounting the trained model 100 on the cutting device 2. For example, the manufacturer of the cutting device 2 generates the trained model 100 and incorporates it into the controller 68, and then provides the cutting device 2 to a user.
[0103] In addition, the manufacturer of the cutting device 2 can provide the cutting device 2 equipped with the trained model 100 to the user after the fact by distributing the trained model 100 to a conventional cutting device (a cutting device not equipped with the trained model 100) used by the user. For example, the manufacturer of the cutting device 2 may distribute a program describing the trained model 100 to the cutting device 2 via a network connected to the cutting device 2 used by the user. In addition, the manufacturer of the cutting device 2 may store a program describing the trained model 100 in a recording medium (optical disk, flash memory, etc.) that is a computer-readable non-transitory recording medium, and provide the recording medium to the user of the cutting device 2.
[0104] By distributing the program using a network or a recording medium as described above, the trained model 100 can be introduced later into the cutting device 2 being used by the user. In addition, when the trained model 100 is updated, the manufacturer of the cutting device 2 may distribute a program describing the updated trained model 100 or data corresponding to the parameters of the updated trained model 100 using a network or a recording medium. In this way, the cutting device 2 equipped with the latest trained model 100 is provided to the user.
[0105] Next, a specific example of a trained model generation method according to the present embodiment will be described. Fig. 5 is a flowchart showing the trained model generation method. The trained model generation method according to the present embodiment includes an image collection step S1 for collecting training images, and a learning step S2 for generating a trained model by machine learning using the training images.
[0106] In the image collection step S1, images for each state of the holding table 52 and the object 11 are collected as learning images. For example, when the states of the holding table 52 and the object 11 are to be determined into two categories, normal and abnormal, the holding table 52 and the object 11 are imaged by the imaging unit 60, and a plurality of first learning images corresponding to the normal state and a plurality of second learning images corresponding to the abnormal state are collected as learning images.
[0107] The first learning image is collected by capturing images of the holding table 52 and the object 11 multiple times with the imaging unit 60 when the holding table 52 and the object 11 are in a normal state (first learning image collection step). Fig. 6 is an image diagram showing an image (first captured image, first learning image) 110 corresponding to the normal state.
[0108] When there is no abnormality in the holding table 52 and the object 11, the object 11 in a normal state in which the substrate 13 and the opening 19a of the frame 19 are arranged concentrically (with the same central position) is held by the holding table 52 in a normal state without any malfunction or other defect. When the holding table 52 and the object 11 in this state are imaged by the imaging unit 60 (see Figs. 1 and 4), an image 110 showing a part or the whole of the holding table 52 and the object 11 in a normal state is obtained.
[0109] Then, in the first learning image collection step, while changing the imaging conditions, the holding table 52 and the object 11 are imaged a plurality of times by the imaging unit 60. Note that the imaging conditions for the holding table 52 and the object 11 are not limited.
[0110] For example, the holding table 52 and the object 11 may be continuously imaged multiple times by the imaging unit 60 while rotating the holding table 52 at a low speed. This allows multiple images 110 showing the holding table 52 and the object 11 at different angles to be acquired. In addition, the holding table 52 and the object 11 may be imaged multiple times while changing the magnification or focal position of the imaging unit 60. This allows multiple images 110 with different magnifications or degrees of blurring to be acquired. Furthermore, if the imaging unit 60 is equipped with lighting that illuminates the subject, the holding table 52 and the object 11 may be imaged multiple times while changing the amount or color of light irradiated from the lighting to the holding table 52 and the object 11. This allows multiple images 110 with different brightness or saturation to be acquired.
[0111] As described above, in the first learning image collection step, a plurality of different images 110 (e.g., 1000 or more) are acquired by capturing images of the holding table 52 and the object 11 in a normal state a plurality of times. Then, the plurality of images 110 are classified (labeled) as images corresponding to the normal state, and stored in a predetermined database as first learning images.
[0112] The second learning image is acquired by capturing images of the holding table 52 and the object 11 multiple times with the imaging unit 60 when the holding table 52 and / or the object 11 are in an abnormal state (second learning image collection step). Examples of images (second captured images, second learning images) 120 corresponding to the abnormal state are shown in Figs. 7(A) to 8(B).
[0113] 7(A) is an image diagram showing an image 120 (image 120A) acquired when the object 11 is held by an abnormal holding table 52. For example, if some of the clamps 74 (the clamp 74 at the lower right in FIG. 7(A)) included in the holding table 52 are broken for some reason, the appearance of the broken clamps 74 will be different from the appearance of normal clamps 74. Therefore, when an image of the holding table 52 and the object 11 is captured in a state in which some of the clamps 74 are broken, an image 120A is acquired in which the appearance of the clamps 74 is different from that of image 110 (see FIG. 6).
[0114] Fig. 7(B) is an image diagram showing an image 120 (image 120B) acquired when the object 11 is placed in an inappropriate position. If there is a defect in the transport of the object 11 to the holding table 52, the object 11 may be placed in a position displaced from the position where it should be placed, for example, the frame 19 may be placed on some of the clamps 74 (the clamp 74 at the lower right in Fig. 7(B)). When the holding table 52 and the object 11 are imaged in this state, an image 120B is acquired in which the position of the object 11 and the state of some of the clamps 74 are different from those in the image 110 (see Fig. 6).
[0115] 8(A) is an image diagram showing an image 120 (image 120C) acquired when an abnormal object 11 is held by the holding table 52. For example, due to a misalignment of the substrate 13 or the frame 19 when the sheet 21 is fixed to the substrate 13 and the frame 19, the object 11 may be formed in a state in which the center position of the substrate 13 is misaligned with the center position of the opening 19a of the frame 19. In this case, when an image of the holding table 52 and the object 11 is captured, an image 120C is acquired in which the positional relationship between the substrate 13 and the frame 19 is different from that of the image 110 (see FIG. 6).
[0116] 8(B) is an image diagram showing an image 120 (image 120D) acquired when an object 11 of an inappropriate size is held by the holding table 52. For example, when an image of the holding table 52 and the object 11 is captured in a state in which the object 11, whose board 13 and frame 19 are smaller than those of the object that should be held by the holding table 52, is placed on the holding surface 52a, an image 120D is acquired in which the size of the object 11 is different from that of the image 110 (see FIG. 6).
[0117] As described above, in the second learning image collection step, the holding table 52 and the object 11 are imaged multiple times when the holding table 52 and / or the object 11 are in an abnormal state, thereby acquiring multiple different images 120 (e.g., 1000 or more) as shown in Figs. 7(A) to 8(B). Note that in the second learning image collection step, the holding table 52 and the object 11 may be imaged multiple times by the imaging unit 60 while changing the imaging conditions, as in the first learning image collection step. Then, the multiple images 120 are classified (labeled) as images corresponding to an abnormal state, and stored in a predetermined database as second learning images.
[0118] The multiple images 110 acquired in the first learning image collecting step correspond to the first learning images classified as a normal state. The multiple images 120 acquired in the second learning image collecting step correspond to the second learning images classified as an abnormal state. That is, by performing the image collecting step S1, multiple learning images classified according to the states of the holding table 52 and the object 11 are collected.
[0119] Although the above describes the case where learning images are obtained by capturing images of the holding table 52 and the object 11 with the imaging unit 60, there is no limitation on the method of collecting learning images. For example, the holding table 52 and the object 11 may be captured from different positions by a movable imaging unit 60 or a plurality of imaging units 60. Also, new learning images may be generated by performing a predetermined image processing on the images 110 and 120. Furthermore, the manufacturer of the cutting device 2 may collect learning images by having the user of the cutting device 2 provide images of the holding table 52 and the object 11.
[0120] Next, by machine learning using the learning images collected in the image collection step S1, a trained model 100 is generated that outputs a determination result indicating the state of the holding table 52 and the object 11 when an image of the holding table 52 and the object 11 is input (learning step S2). Fig. 9 is a block diagram showing the trained model 100 generated by machine learning.
[0121] In the learning step S2, supervised learning is performed using the learning images and the classification results of the learning images (the holding table 52 and the state of the object 11) as teacher data. Specifically, the images 110 and 120 are input as image data, and the holding table 52 and the state of the object 11 (normal / abnormal) are input as correct labels to the neural network NN.
[0122] Then, the parameters of the neural network NN (such as the weights and biases of the neurons) are updated so that the error between the output of the neural network NN and the correct label is reduced. As a result, the neural network NN is reconfigured so that when the image of the storage table 52 and the object 11 is input to the input layer 102, the neural network NN outputs the determination result of the storage table 52 and the state of the object 11 from the output layer 104. For example, the backpropagation method can be used as the learning algorithm. Note that when a CNN is used as the neural network NN, the filter value of the convolution layer, the weights and biases of the neurons of the fully connected layer, etc. are updated by learning using the images 110 and 120, the storage table 52, and the state of the object 11 as teacher data.
[0123] Through the above learning, a neural network NN capable of determining the states of the holding table 52 and the object 11 based on images of the holding table 52 and the object 11 is generated. Then, the trained neural network NN is mounted on the determination unit 92 (see FIG. 4) of the controller 68. In this way, a cutting device 2 capable of determining the states of the holding table 52 and the object 11 based on the image 90 is obtained.
[0124] The calculations for training the neural network NN may be executed by the controller 68, or may be executed by a large-scale, high-performance computer provided outside the cutting device 2. In the latter case, the trained neural network NN (or parameters of the trained neural network NN) generated outside the cutting device 2 is input to the controller 68 and stored.
[0125] In the above, a case has been described where the states of the holding table 52 and the object 11 are judged based on the image 90 showing the holding table 52 and the object 11 (see FIG. 4). However, the judgment unit 92 may judge only one of the holding table 52 and the object 11. For example, damage to the clamp 74 (see FIG. 7(A)), adhesion of a foreign object to the holding table 52, etc. can be detected by judging the state of the holding table 52 alone. In addition, misalignment between the substrate 13 and the frame 19 (see FIG. 8(A)), an error in the size of the object 11 (see FIG. 8(B)), etc. can be detected by judging the state of the object 11 alone.
[0126] Specifically, when determining the state of the holding table 52 alone, the imaging unit 60 captures an image of the holding table 52 alone or together with the object 11, and at least an image showing the holding table 52 is acquired. The trained model 100 is configured by machine learning so as to output a determination result showing the state of the holding table 52 when an image of the holding table 52 is input. In this case, images of the holding table 52 for each state are used as training images for training the trained model 100. For example, by capturing an image of the holding table 52 with the imaging unit 60, a first training image corresponding to the holding table 52 in a normal state and a plurality of second training images corresponding to the holding table 52 in an abnormal state are acquired.
[0127] On the other hand, when determining the state of the object 11 alone, the imaging unit 60 captures an image of the object 11 alone or together with the holding table 52, and at least an image showing the object 11 is acquired. The trained model 100 is configured by machine learning so as to output a determination result showing the state of the object 11 when an image of the object 11 is input. In this case, images of the object 11 for each state are used as training images for training the trained model 100. For example, by capturing an image of the object 11 with the imaging unit 60, a first training image corresponding to the object 11 in a normal state and a plurality of second training images corresponding to the object 11 in an abnormal state are acquired.
[0128] In addition, although the above describes a case where the trained model 100 judges the state of the holding table 52 and / or the object 11 in two stages (normal or abnormal) (see FIG. 4), the judgment content of the trained model 100 is not limited to this. For example, the trained model 100 may judge the state of the holding table 52 and / or the object 11 in three or more stages.
[0129] Furthermore, the trained model 100 may classify the state of the holding table 52 and / or the object 11 by type of abnormality (such as the location or content of the abnormality). In this case, the output layer 104 of the trained model 100 is configured with one neuron that outputs the probability that both the holding table 52 and the object 11 are normal, and multiple neurons that output the probability that at least one of the holding table 52 and the object 11 is abnormal for each type of abnormality. This makes it possible to determine the abnormality of the holding table 52 and / or the object 11 in more detail.
[0130] Furthermore, the determination unit 92 (see FIG. 4) of the controller 68 may include a plurality of trained models 100. In this case, the same image 90 may be input to the plurality of trained models 100, or different images 90 may be input to the plurality of trained models 100.
[0131] For example, a plurality of trained models 100 specialized for judging different types of abnormalities are mounted on the judgment unit 92. In this case, one image 90 is input to each of the plurality of trained models 100, and the state of the holding table 52 and / or the object 11 is judged for each type of abnormality by the plurality of trained models 100. In addition, when a plurality of types of images 90 with different imaging conditions (angles of the holding table 52 and the object 11, imaging area of the imaging unit 60, magnification, etc.) are acquired, a trained model 100 may be provided for each type of image 90.
[0132] Furthermore, although the above describes the case where the state of the holding table 52 and / or the target object 11 is determined, the determination unit 92 of the controller 68 (see FIG. 4) can also determine the state of components other than the holding table 52 among the components (base 4 to cleaning unit 50, etc.) that constitute the cutting device 2. In this case, an imaging unit that captures images of predetermined components other than the holding table 52 is mounted on the cutting device 2, and a trained model that determines the state of the predetermined components is mounted on the determination unit 92. The trained model then determines the state of the predetermined components based on the images of the predetermined components acquired by the imaging unit. Furthermore, the determination unit 92 may be equipped with a plurality of trained models that each determine the state of a different component.
[0133] For example, when the object 11 is processed by the processing units 44A and 44B, the object 11 is held by the holding table 10 (see FIG. 1). Therefore, an image showing the holding table 10 and / or the object 11 may be acquired by imaging the holding table 10 and / or the object 11 held by the holding table 10 with the imaging unit 56 (see FIG. 1).
[0134] In this case, the determination unit 92 (see FIG. 4) determines the state of the holding table 10 and / or the object 11 based on an image of the holding table 10 and / or the object 11. Specifically, a trained model configured by machine learning to output a determination result indicating the state of the holding table 10 and / or the object 11 when an image of the holding table 10 and / or the object 11 is input is mounted on the determination unit 92. Note that a method for generating a trained model for determining the state of the holding table 10 and / or the object 11 is similar to a method for generating a trained model 100 for determining the state of the holding table 10 and / or the object 11 (see FIG. 9).
[0135] As described above, the cutting device 2 according to this embodiment includes the trained model 100 configured by machine learning so that when an image of the components and / or object 11 of the cutting device 2 is input, the trained model 100 outputs a determination result indicating the state of the components and / or object 11. This makes it possible to appropriately extract features from images of various components and / or objects 11 without performing complex image processing, and to easily and highly accurately determine the state of the components and / or object 11.
[0136] In the above embodiment, the cutting machine 2 equipped with the trained model 100 has been described, but the trained model according to the present invention can also be installed in processing machines other than cutting machines. Examples of other processing machines that can be equipped with the trained model according to the present invention include grinding machines, polishing machines, laser processing machines, plasma processing machines, etc.
[0137] The grinding device includes a processing unit (grinding unit) that grinds an object (workpiece). The grinding unit includes a spindle, and an annular grinding wheel including multiple grinding stones is attached to the tip of the spindle. The object is ground by rotating the grinding wheel and bringing the grinding stones into contact with the object.
[0138] The polishing device is equipped with a processing unit (polishing unit) that polishes an object (workpiece). The polishing unit is equipped with a spindle, and a disk-shaped polishing pad is attached to the tip of the spindle. The object is polished by rotating the polishing pad and bringing it into contact with the object.
[0139] The laser processing device includes a processing unit (laser irradiation unit) that performs laser processing on an object (workpiece). For example, the laser irradiation unit includes a laser oscillator and a condenser that condenses the laser beam emitted from the laser oscillator. The laser beam emitted from the laser irradiation unit is condensed on the front, back, or inside of the object, thereby performing laser processing on the object.
[0140] The plasma processing apparatus includes a processing unit (gas supply unit) that supplies a gas in a plasma state to an object (workpiece). By supplying the gas in a plasma state from the gas supply unit to the object, the object is subjected to a plasma process such as plasma etching or film formation.
[0141] The trained model according to the present invention can also be mounted on processing devices other than the above-mentioned processing devices. Examples of other processing devices include a cleaning device that cleans an object (object to be cleaned), an inspection device that inspects an object (object to be inspected), a sheet attachment device that attaches a sheet to an object, and a sheet expansion device that applies an external force to the object by expanding a sheet attached to the object.
[0142] The various processing devices as described above are composed of multiple components (holding table, transport unit, processing unit, etc.). The processing device is equipped with an imaging unit that captures images of the predetermined components, and one or more trained models that determine the state of the components based on the images of the components captured by the imaging unit.
[0143] In addition, the structures, methods, etc. according to the above-described embodiments can be modified as appropriate without departing from the scope of the object of the present invention. [Explanation of symbols]
[0144] 11 Object (processing object) 13 Substrate (wafer) 13a Surface (first side) 13b Back side (2nd side) 15th Street (Planned division line) 17 Devices 19 Frames 19a aperture 21 sheets 2 Cutting equipment 4 Foundation 4a,4b,4c opening 6 Cassette Support Stand 8 Cassettes 10 Holding table (chuck table) 10a Holding surface 12 Moving mechanism 14 Table Cover 16 Dust-proof / water-proof cover 18 Clamp 20 Guide rail 22 First support structure 24 Rail 26 Transport unit (transport mechanism, transport device) 28 Moving part (moving mechanism) 30 Holding part (holding mechanism) 30a Gripping part (gripping mechanism) 32 Rail 34 Transport unit (transport mechanism, transport device) 36 Moving part (moving mechanism) 38 Holding part (holding mechanism) 40 Second support structure 42A,42B Moving mechanism 44A, 44B Processing unit (machining unit, cutting unit) 46 Cutting Blade 48 Imaging unit 50 Cleaning unit (cleaning mechanism, cleaning device) 52 Holding table (spinner table) 52a Holding surface 54 Nozzle 56 Imaging unit 58 Support member 60 Imaging unit 62 Cover 64 Display unit (display unit, display device) 66 Notification unit (notification section, notification device) 68 Controller (control unit, control section, control device) 70 Frame (main body) 70a top side 70b Recess 72 Retaining member 72a Suction surface 74 Clamp 76 Support part 78 Clamp 90 images (captured images) 92 Judgment section 94 Memory section 96 Notification control section 100 pre-trained models 102 Input Layer 104 Output layer 106 Hidden layer (intermediate layer) 108 Judgment result processing unit 110 Image (first captured image, first learning image) 120, 120A to 120D Images (second captured image, second learning image)
Claims
1. A processing device for processing an object, A holding table for holding the object; an imaging unit for imaging the holding table; a controller including a learned model configured by machine learning so as to output, when an image of the holding table or an image of the holding table and the object acquired by the imaging unit is input, a judgment result indicating the state of the holding table or a judgment result indicating the state of the holding table and the object.
2. 2. The processing device according to claim 1, wherein the trained model is a neural network including an input layer to which the image is input and an output layer to which the determination result is output.
3. 3. The processing device according to claim 1, further comprising an alarm unit that notifies the determination result when the determination result indicates an abnormality.
4. A method for generating a trained model for determining a state of a processing device that processes an object, comprising: an image collecting step of collecting images of a holding table provided in the processing device for each state, or images of the holding table and the object held by the holding table for each state, as learning images; a learning step of generating a trained model by machine learning using the learning image, which generates a trained model that outputs a judgment result indicating the state of the holding table or a judgment result indicating the state of the holding table and the object when an image showing the holding table or an image showing the holding table and the object is input.
5. The method for generating a trained model according to claim 4, characterized in that in the image collection step, a plurality of first learning images corresponding to a normal state and a plurality of second learning images corresponding to an abnormal state are obtained as the learning images by imaging the holding table, or the holding table and the object.
Citation Information
Patent Citations
Cutting method
JP2006093333A
