Grinding performance determination device and grinding performance determination method

Abstract

Provided are a grinding performance determination device and a grinding performance determination method capable of determining grinding performance with good accuracy. The grinding performance determination method includes: an extraction step of obtaining an AE signal output from an AE sensor (104) provided to a grinding machine (100) and processing the AE signal to extract a micro-feature amount of the AE signal; and a determination step of inputting the micro-feature amount extracted in the extraction step to an AI model learned with the micro-feature amount of the AE signal for training as input and outputting information indicating a grinding performance of the grinding machine.

Description

Technical Field

[0001] This invention relates to a grinding performance determination apparatus and a grinding performance determination method, and more particularly to a technique for determining the grinding performance of grinding machines used in semiconductor manufacturing processes. Background Technology

[0002] Traditionally, semiconductor wafers (e.g., silicon wafers, compound semiconductor wafers (silicon carbide (SiC), gallium arsenide (GaAs), etc.)) have been used as materials for semiconductor integrated circuits. In semiconductor manufacturing processes, semiconductor wafers are produced by slicing and grinding (thinning) ingots of semiconductor material. There are also cases where grinding processes, such as chamfering the edges, are performed on the semiconductor wafers manufactured in this way.

[0003] In the aforementioned grinding process of semiconductor wafers, a grinding machine equipped with a grinding stone and a spindle on which the grinding stone is mounted is used. The grinding process of semiconductor wafers is performed by bringing the grinding stone, which rotates under the action of the spindle, into contact with the cutting part of the semiconductor wafer.

[0004] In grinding machines like those described above, grinding performance can be reduced due to dulling of the grinding stone, clogging of the orifices, etc. As a process to maintain the grinding performance of such grinding stones, dressing (or shaping, hereinafter also referred to as dressing) is performed.

[0005] To maintain the grinding performance of a grinding machine, it is necessary to perform grinding stone overhaul at the appropriate time. Determining when to perform overhaul (overhaul determination) involves, for example, measuring the condition of the grinding stone and determining the necessity of overhaul at the measured time. However, the condition of the grinding stone is difficult to measure directly. Therefore, conventional overhaul determinations have relied on the user's experience and intuition, such as the usage status of the grinding machine or the number of semiconductor wafers processed. Consequently, the results of overhaul determinations can vary depending on the user's skill, making it difficult to consistently perform overhaul at the optimal time.

[0006] Due to the aforementioned circumstances, efforts are being made to stably determine the finishing process based on objective criteria. For example, Patent Document 1 discloses a grinding state monitoring device for monitoring the state of a workpiece being ground by a grinding stone and a finishing state monitoring device for monitoring the state of a grinding stone being finished by a finishing tool. In Patent Document 1, an Acoustic Emission Signal (AE) is detected, representing the breaking sound generated in relation to the breaking of abrasive grains or the like constituting the grinding stone. Based on this AE signal, the grinding state and finishing state of the grinding stone, such as grinding burns, pore blockage, sharpness of the grinding stone, and the condition of the grinding stone's circumferential surface, are monitored. Furthermore, in Patent Document 1, the original AE signal generated in the grinding stone is calculated by subtracting the frequency-analyzed workpiece-side AE ​​signal from the frequency-analyzed grinding stone-side AE ​​signal.

[0007] Prior art literature

[0008] Patent documents

[0009] Patent Document 1: Japanese Patent Application Publication No. 2000-233369 Summary of the Invention

[0010] The problem that the invention aims to solve

[0011] As mentioned above, in Patent Document 1, the original AE signal generated in the grinding stone is obtained based on the frequency-analyzed AE signal on the grinding stone side and the AE signal on the workpiece side, thereby improving the accuracy of determining the grinding state and dressing state of the grinding stone. However, it is difficult to accurately determine the change (decrease) in the grinding performance of the grinding machine simply by using the AE signal generated in the grinding stone as is.

[0012] The present invention was made in view of the following circumstances, and its purpose is to provide a grinding performance determination device and a grinding performance determination method that can accurately determine grinding performance.

[0013] Solution for solving the problem

[0014] The grinding performance determination apparatus of the first aspect of the present invention includes: an extraction unit that acquires an AE signal output from an AE sensor provided on a grinding machine and processes the AE signal to extract microscopic feature quantities of the AE signal; and a determination unit that inputs the microscopic feature quantities extracted by the extraction unit into an AI model obtained by learning with the microscopic feature quantities of the AE signal used for training as input, and outputs information representing the grinding performance of the grinding machine.

[0015] The grinding performance determination device of the second embodiment of the present invention, based on the first embodiment, calculates the macroscopic feature quantity of the AE signal by obtaining the moving average of the AE signal, and extracts the microscopic feature quantity by subtracting the macroscopic feature quantity from the AE signal.

[0016] The grinding performance determination device of the third embodiment of the present invention is based on the second embodiment. The AI ​​model is obtained by learning with macroscopic and microscopic feature quantities as inputs. The determination unit takes the macroscopic and microscopic feature quantities extracted by the extraction unit as inputs and outputs information representing the grinding performance of the grinding machine.

[0017] The grinding performance determination device of the fourth embodiment of the present invention is based on the second embodiment. The extraction unit calculates the peak value and variance of the amplitude histogram obtained based on microscopic feature quantities. The AI ​​model is obtained by learning with the peak value and variance of the amplitude histogram as input. The determination unit takes the peak value and variance of the amplitude histogram extracted by the extraction unit as input and outputs information representing the grinding performance of the grinding machine.

[0018] The grinding performance determination device of the fifth aspect of the present invention, based on any one of the first to fourth aspects, includes at least one of the following information indicating grinding performance: information indicating the surface roughness of the grinding object after grinding by the grinding machine, information indicating the state of the grinding stone of the grinding machine, and information indicating the timing of dressing the grinding stone.

[0019] The grinding performance determination method of the sixth aspect of the present invention includes: an extraction step, which obtains an AE signal output from an AE sensor installed on a grinding machine and processes the AE signal to extract microscopic feature quantities of the AE signal; and a determination step, which inputs the microscopic feature quantities extracted in the extraction step into an AI model obtained by learning with the microscopic feature quantities of the AE signal used for training as input, and outputs information representing the grinding performance of the grinding machine.

[0020] Invention Effects

[0021] According to the present invention, by learning the microscopic features of the AE signal as input, the grinding performance of the grinding machine can be determined with good accuracy. Attached Figure Description

[0022] Figure 1 This is a block diagram illustrating the relationship between the grinding performance determination device and the grinding machine according to the first embodiment of the present invention.

[0023] Figure 2 This is a block diagram showing the grinding performance determination device.

[0024] Figure 3 This is a flowchart illustrating the learning method of the grinding performance assessment device.

[0025] Figure 4 This is a diagram illustrating the order in which the training dataset for the grinding performance evaluation device was created.

[0026] Figure 5This is a diagram (chart) used to illustrate unsupervised learning in a grinding performance assessment device.

[0027] Figure 6 This is a flowchart illustrating a method for determining grinding performance based on a grinding performance determination device.

[0028] Figure 7 This is a diagram illustrating the order in which the training dataset for the grinding performance determination device according to the second embodiment of the present invention is created.

[0029] Figure 8 This is a diagram (graph) illustrating the unsupervised learning of the grinding performance determination device according to the second embodiment of the present invention. Detailed Implementation

[0030] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0031] [First Implementation]

[0032] (Grinding performance assessment device)

[0033] Figure 1 This is a block diagram illustrating the relationship between the grinding performance determination device and the grinding machine according to the first embodiment of the present invention.

[0034] Grinding machine 100 is an apparatus that performs grinding operations by bringing a grinding stone 102 mounted on and rotating on a spindle into contact with a grinding object (e.g., a semiconductor wafer). Figure 1 As shown, the grinding machine 100 includes a grinding stone 102, an AE sensor 104, and a vibration sensor 106. It should be noted that, for the sake of simplicity, the spindle, motor, semiconductor wafer, and other components of the grinding machine 100 are omitted from the illustrations.

[0035] The AE sensor (sound sensor) 104 is a sensor used to detect elastic waves (sound signals) emitted when the grinding stone 102 comes into contact with the grinding object or dressing tool. The AE sensor 104 is, for example, located near the grinding stone 102, near the spindle, or near the semiconductor wafer of the grinding object in the grinding machine 100. The AE sensor 104 includes, for example, a piezoelectric element and converts the elastic wave into an electrical signal (hereinafter referred to as an AE signal) which is output to the controller of the grinding machine 100. The frequency band detectable by the AE sensor 104 is, in an example, tens of kHz to several MHz, and the AE sensor 104 is used for the detection of anomalies in the high-frequency (ultrasonic) region of elastic waves.

[0036] Vibration sensor 106 is a sensor used to detect vibrations (vibration accelerations) caused by the movement of a drive unit such as a rotating grinding stone 102. The frequency band that can be detected by vibration sensor 106 is lower than that of AE sensor 104, and vibration sensor 106 is used to detect abnormalities that generate vibrations in the low-frequency region (e.g., imbalance of grinding stone 102).

[0037] The grinding performance determination device 10 of this embodiment acquires the AE signal detected by the AE sensor 104 installed on the grinding machine 100, and outputs information indicating the change in grinding performance based on the AE signal.

[0038] Figure 2 This is a block diagram showing the grinding performance determination device 10. (Example) Figure 2 As shown, the grinding performance evaluation device 10 includes a processor 12, a memory 14, a storage device 16, and an interface (I / F) 18. For example, a personal computer (PC) or workstation can be used as the grinding performance evaluation device 10. It should be noted that the grinding performance evaluation device 10 can also function as the controller of the grinding machine 100, or it can be a separate device other than the controller of the grinding machine 100.

[0039] The processor 12 is a device that performs calculations and controls in the grinding performance determination device 10, and may include, for example, a CPU (Central Processing Unit). It should be noted that, for the purpose of implementing the learning described later, the processor 12 may also include, in addition to or in place of a CPU, a GPU (Graphics Processing Unit), a GPGPU (General Purpose GPU), or a TPU (Tensor Processing Unit). The processor 12 is an example of the extraction and determination unit in an embodiment of the present invention.

[0040] The memory 14 is a storage device (main storage device) used as a working area when the processor 12 performs operations, such as RAM (Random Access Memory).

[0041] Storage device 16 is a storage device (auxiliary storage device) that stores the program and data used for the calculation or control of grinding performance determination device 10, such as an HDD (Hard Disk Drive) or SSD (Solid State Drive). The AI ​​analysis software and AE signal data, described later, are stored in storage device 16.

[0042] Interface (I / F) 18 is a device that serves as a medium for the input and output of signals or data between the grinding performance assessment device 10 and an external device. I / F 18 serves as a medium for the input and output of signals between the grinding performance assessment device 10 and an external device, for example, via a wired or wireless communication mechanism such as LAN (Local Area Network) or USB (Universal Serial Bus).

[0043] The input unit 30 is a device connected to the grinding performance determination device 10 and receiving input from the user to the grinding performance determination device 10. The input unit 30 includes, for example, a keyboard for text input, a pointing device (mouse, touch panel, etc.) for inputting position information on the display.

[0044] The output unit 32 is the output device of the grinding performance determination device 10. The output unit 32 includes, for example, a display (LCD (Liquid Crystal Display) or similar device for displaying images).

[0045] The analog AE signal detected by the AE sensor 104 installed on the grinding machine 100 is converted into a digital AE signal by the A / D (Analog to Digital) converter 50 and input to the grinding performance determination device 10 via the I / F 18. The digital AE signal is stored in the storage device 16.

[0046] It should be noted that the A / D converter 50 may also be included in the grinding performance determination device 10, or it may be a device other than the grinding performance determination device 10 (for example, it may also be included in the grinding machine 100 or its controller).

[0047] The processor 12 reads AE signal data and AI (Artificial Intelligence) analysis software from the storage device 16 to learn and obtain a learned AI model. It should be noted that the AE signal data used during learning is not limited to data collected from the grinding machine 100, the object of the grinding performance evaluation device 10. For example, data obtained from a grinding machine 100 that is the same as or similar to the grinding machine 100, or data created specifically for learning purposes, can also be used. Furthermore, this learning can also be performed before the grinding machine 100 leaves the factory.

[0048] During grinding, the processor 12 continuously or at predetermined time intervals acquires the AE signals detected by the AE sensor 104 during grinding. The processor 12 inputs the AE signals during grinding into the learned AI model and outputs information indicating changes in the grinding performance of the grinding machine 100. This information includes, for example, information indicating the surface roughness of the workpiece after grinding by the grinding stone 102, information indicating the condition of the grinding stone 102, or information indicating when to perform maintenance. This output is displayed, for example, on the display of the output unit 32. Thus, the user can identify when the grinding performance of the grinding machine 100 has decreased and when the grinding stone 102 should be maintained (or replaced).

[0049] It should be noted that the AE signal obtained from the grinding machine 100 during the grinding process can also be used to relearn the AI ​​model of the grinding performance determination device 10.

[0050] (Learning about grinding performance evaluation devices)

[0051] Figure 3 This is a flowchart illustrating the learning method of the grinding performance evaluation device 10. It should be noted that the following learning can be performed before the grinding performance evaluation device 10 leaves the factory or during setup, or it can be performed after the grinding machine 100 has been running.

[0052] First, the processor 12 collects the AE signal (step S10). The AE signal during learning can be a signal obtained by the processor 12 from the AE sensor 104 of the grinding machine 100, which is the object of the grinding performance determination device 10, or a signal obtained from a grinding machine 100 that is the same as or similar to the grinding machine 100 and stored in the storage device 16 in advance, or a signal created for learning purposes.

[0053] Next, the processor 12 processes the AE signal collected in step S10 to create a dataset for learning (training) the AI ​​model (step S12).

[0054] Figure 4 This diagram illustrates the order in which the training dataset for the grinding performance evaluation device 10 was created. It should be noted that... Figure 4 In the diagram, the amplitude of short-period sounds is simplified using a wider line.

[0055] In this embodiment, the processor 12 processes the raw data (RAW data) of the collected AE signal data and extracts the macroscopic and microscopic features of the AE signal.

[0056] like Figure 4As shown, when the horizontal axis is plotted as time, the AE signal data (raw data) becomes a combination of microscopic vibrations with relatively short periods (equivalent to the thickness of a wide line) and macroscopic changes with longer time periods than the microscopic vibrations.

[0057] First, the processor 12 calculates a moving average of the AE signal data (raw data). That is, the AE signal data is smoothed by calculating the average (arithmetic mean) of the AE signal over a constant range along the time axis while shifting the range. It should be noted that the width of the range used to calculate the moving average is experimentally determined based on the characteristics of the grinding machine 100, etc. From this, macroscopic feature quantities representing macroscopic changes in the distribution of sound are extracted from the AE signal data (raw data).

[0058] Then, the processor 12 extracts the microscopic features corresponding to the microscopic changes in sound by subtracting the macroscopic features from the AE signal data (raw data).

[0059] Following the above sequence, a training dataset is created that uses the macroscopic and microscopic features of the AE signal as input.

[0060] Next, the processor 12 inputs the aforementioned training dataset into the AI ​​parsing software for learning (step S14) and obtains the learned AI model (step S16). The learned AI model is saved in the storage device 16.

[0061] In step S14, unsupervised learning is performed using the macroscopic and microscopic features of the AE signal as input.

[0062] Figure 5 This is a diagram (chart) used to illustrate the unsupervised learning of the grinding performance determination device 10. Figure 5 The horizontal and vertical axes of the graph represent the macroscopic and microscopic characteristics of the AE signal, respectively. The pattern of the plotted points indicates the surface roughness (arithmetic mean roughness Ra (Japanese Industrial Standard JIS B 0601:1994)) after grinding. Figure 5 The graph shown illustrates the correlation between the macroscopic and microscopic characteristics of the AE signal and the surface roughness after grinding.

[0063] In this implementation, clustering (Isolation Forest parsing) is performed as unsupervised learning. Isolation Forest parsing is a method of data analysis (clustering) using decision trees, which infers outliers based on the distance (depth) of the data to the isolated points.

[0064] exist Figure 5In the example shown, the points with coarser arithmetic mean roughness Ra are distributed in region A1 on the graph. Therefore, the vicinity of region A1 can be isolated and judged as "presumed reduced grinding performance", "coarser roughness", "presumed grinding stone deterioration" or "approaching overhaul time".

[0065] As mentioned above, by using the macroscopic and microscopic features of the AE signal as input for unsupervised learning, a fully trained AI model for determining grinding performance can be obtained.

[0066] (Determination of grinding performance using AI models)

[0067] Figure 6 This is a flowchart illustrating a method for determining grinding performance based on the grinding performance determination device 10.

[0068] First, the processor 12 acquires the AE signal output from the AE sensor 104 during the grinding process based on the grinding machine 100 (step S30). In step S30, the processor 12 may continuously accept the input of the AE signal or accept the input of the AE signal at predetermined time intervals.

[0069] Next, the processor 12 processes the AE signal collected in step S30 to create a dataset for determining grinding performance (step S32: extraction step). In step S32, similar to step S12 above, the AE signal data (raw data) is processed to obtain the macroscopic and microscopic features of the AE signal.

[0070] Next, the processor 12 inputs the dataset used to determine the grinding performance into the learned AI model (step S34) and outputs information representing the grinding performance of the grinding machine 100 (step S36: determination step).

[0071] According to this embodiment, by using the macroscopic and microscopic features of the AE signal as input for unsupervised learning (clustering, isolated forest analysis), the grinding performance of the grinding machine 100 can be determined with good accuracy.

[0072] [Second Implementation]

[0073] In the first embodiment, unsupervised learning is performed by using the macroscopic and microscopic features of the AE signal as input, but unsupervised learning can also be performed using the microscopic features.

[0074] Figure 7This diagram illustrates the sequence of creating the training dataset for the grinding performance evaluation device 10. In this embodiment, the training dataset is created by further processing the microscopic features. It should be noted that the method for extracting the microscopic features is the same as in the first embodiment, and therefore its description is omitted.

[0075] In this embodiment, the processor 12 from Figure 7 The chart of microscopic characteristic quantities shown on the left is created. Figure 7 The amplitude histogram shown on the right is an example of a graph of microscopic features. This amplitude histogram is obtained by plotting the number of amplitudes within a specific range on the horizontal axis and the number of occurrences (counts) of each amplitude within that range on the vertical axis. In this embodiment, after the processor 12 calculates such an amplitude histogram, the peak value P1 in the amplitude histogram is taken as microscopic feature 1 (described later). Figure 8 The horizontal axis), and the variance in the amplitude histogram is used as the microscopic characteristic quantity 2 (described later). Figure 8 (vertical axis).

[0076] Figure 8 This is a diagram (chart) used to illustrate the unsupervised learning of the grinding performance determination device 10. Figure 8 The horizontal and vertical axes of the graph represent the peak value and variance of the amplitude histogram, respectively. The pattern of the plotted points indicates the surface roughness (arithmetic mean roughness Ra) of the ground surface. That is, Figure 8 The graph shown illustrates the correlation between the peak value and variance in the amplitude histogram, which is an example of a microscopic characteristic quantity, and the roughness of the ground surface.

[0077] In this embodiment, clustering (isolated forest parsing) is performed as unsupervised learning, similar to the first embodiment.

[0078] exist Figure 8 In the example shown, the points with coarser arithmetic mean roughness Ra are distributed in region A2 on the graph. Therefore, the vicinity of region A2 can be isolated and judged as "presumed reduced grinding performance", "coarser roughness", "presumed grinding stone deterioration" or "approaching overhaul time".

[0079] As described above, by using the peak value and variance of the amplitude histogram, which are examples of microscopic features, as input for unsupervised learning, a fully trained AI model for judging grinding performance can be obtained.

[0080] It should be noted that, in the above embodiments, a camera for capturing images of the surface of the grinding stone 102, or a measuring device (e.g., a white interference microscope) for measuring the surface roughness of the semiconductor wafer being ground, may also be provided on the grinding machine 100 side. Therefore, the grinding performance determination device 10 can verify the accuracy of the determination result by comparing the surface image of the grinding stone 102 or the measurement result of the surface roughness of the semiconductor wafer being ground with the determination result of the grinding performance.

[0081] Alternatively, the grinding stone 102 can be balanced based on the detection signals from the AE sensor 104 or the vibration sensor 106.

[0082] Alternatively, in the above embodiment, the AE signal generated in the grinding stone 102 can be calculated by subtracting the AE signal (raw data) on the grinding object side from the AE signal (raw data) on the grinding stone 102 side, and then the AE signal can be processed (steps S12 and S32).

[0083] It should be noted that in the above embodiments, a semiconductor wafer was described as an example of a grinding object (processing object) based on the grinding machine 100, but the examples of grinding objects are not limited to semiconductor wafers. The determination of grinding performance in the above embodiments can also be applied to the grinding of other materials (e.g., hard and brittle materials, composite materials, metals, ceramics, glass, minerals such as sapphire, wood, resin, etc.).

[0084] Explanation of reference numerals in the attached figures

[0085] 10: Grinding performance assessment device; 12: Processor; 14: Memory; 16: Storage device; 18: Interface (I / F); 30: Input unit; 32: Output unit; 50: A / D converter; 100: Grinding machine; 102: Grinding stone; 104: AE sensor; 106: Vibration sensor.

Claims

1. A grinding performance determination device, wherein, The grinding performance determination device includes: An extraction unit acquires an AE signal output from an AE sensor mounted on a grinding machine, and processes the AE signal to extract its microscopic features; and The determination unit inputs the micro-features extracted by the extraction unit into the AI ​​model obtained by learning from the micro-features of the AE signal used for training, and outputs information representing the grinding performance of the grinding machine.

2. The grinding performance determination device according to claim 1, wherein, The extraction unit calculates the macroscopic features of the AE signal by taking the moving average of the AE signal, and extracts the microscopic features by subtracting the macroscopic features from the AE signal.

3. The grinding performance determination device according to claim 2, wherein, The AI ​​model is obtained by learning from the macroscopic and microscopic features as inputs. The determination unit takes the macroscopic feature quantity and the microscopic feature quantity extracted by the extraction unit as input, and outputs information representing the grinding performance of the grinding machine.

4. The grinding performance determination device according to claim 2, wherein, The extraction unit calculates the peak value and variance of the amplitude histogram derived from the microscopic features. The AI ​​model is obtained by learning from the peak value and variance of the amplitude histogram as input. The determination unit takes the peak value and variance of the amplitude histogram extracted by the extraction unit as input, and outputs information representing the grinding performance of the grinding machine.

5. The grinding performance judging device according to any one of claims 1 to 4, wherein, Information indicating the grinding performance includes at least one of the following: information indicating the surface roughness of the object being ground by the grinding machine, information indicating the condition of the grinding stone of the grinding machine, and information indicating the timing of dressing the grinding stone.

6. A method for determining grinding performance, wherein, The grinding performance determination method includes: The extraction step involves acquiring the AE signal output from an AE sensor mounted on a grinding machine, and processing the AE signal to extract its microscopic features; and The determination step involves inputting the micro-features extracted in the extraction step into the AI ​​model, which is obtained by learning from the micro-features of the AE signal used for training, and outputting information representing the grinding performance of the grinding machine.

Images