Polishing solution monitoring method and monitoring system

Abstract

The invention discloses a polishing solution monitoring method and monitoring system, and belongs to the technical field of polishing solutions. The polishing solution monitoring method comprises the following steps that 1, component data of a polishing solution along with the using process and mark data corresponding to the component data are obtained, the component data comprise the concentration of each component, and the mark data comprise the PH value, spectral information and temperature data; 2, time temperature data, time PH data and time spectrum information of the to-be-detected polishing solution are obtained and serve as prediction information; in the technical scheme provided by the invention, when the effective components in the polishing solution are monitored, the polishing solution is not sampled, the prediction information of the polishing solution in the use process is collected, and then the component data is predicted by using the prediction model, so that in the monitoring process, the efficiency is higher, and the change of the components in the polishing solution can be monitored more quickly.

Description

Technical Field

[0001] The present application relates to the technical field of polishing liquid, and in particular to a polishing liquid monitoring method. Background Art

[0002] Polishing fluid is widely used in various manufacturing and processing fields, such as semiconductor manufacturing, automobile manufacturing, CNC processing, etc. The composition of the polishing fluid will change with the increase in the number of uses, thereby reducing the effect of the polishing fluid. Therefore, the polishing fluid needs to be replaced in time according to the actual production requirements for the polishing fluid. Before replacing the polishing fluid, it is necessary to accurately detect the composition of the polishing fluid.

[0003] The traditional polishing liquid detection method generally analyzes the composition of the polishing liquid through chemical quantitative experiments to obtain the composition content of the polishing liquid. This polishing liquid component analysis method requires filtering, separating, extracting, and quantifying the collected polishing liquid samples to obtain the content of each component in the polishing liquid. Therefore, the composition detection of the polishing liquid requires a lot of time and cost.

[0004] In practice, monitoring the composition of the polishing liquid during its use has always been a difficult problem in the industry. For the replacement of the polishing liquid, the commonly used technical means is to detect the composition of the polishing liquid once the polishing liquid has been used for the expected number of times or the expected time. If the composition of the polishing liquid meets the requirements, continue to use it. If not, stop using it and replace the polishing liquid. This method cannot continuously monitor the changes in the composition of the polishing liquid during its use. Therefore, it is easy for the composition of the polishing liquid to no longer meet the polishing requirements, but the polishing liquid that cannot meet the processing requirements is still used for polishing operations, resulting in poor polishing effects and defective products. Summary of the invention

[0005] The content of this application is used to introduce concepts in a brief form, which will be described in detail in the detailed implementation section below. The content of this application is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to limit the scope of the technical solution claimed for protection.

[0006] As a first aspect of the present application, in order to solve the technical problems mentioned in the above background technology section, some embodiments of the present application provide a polishing liquid monitoring method, comprising the following steps:

[0007] Step 1: Obtaining component data of the polishing liquid during its use and marking data corresponding to the component data, wherein the component data includes the concentration of each component, and the marking data includes pH value, spectral information and temperature data;

[0008] Step 2: Obtaining time-temperature data, time-PH data and time-spectrum information of the polishing liquid to be tested as prediction information;

[0009] Step 3: Build a component prediction model, use the prediction information as the input of the prediction model, and the component data as the output, and cross-train the prediction model;

[0010] Among them, the prediction model includes LSTM neural network, convolutional neural network, fully connected layer and loop optimization layer;

[0011] The time-temperature data and the time-PH data are input into the LSTM neural network to obtain the first high-level features;

[0012] The spectral information is input into the convolutional neural network to obtain the second high-level features;

[0013] The first high-level feature and the second high-level feature are input into a fully connected layer to generate a fused feature;

[0014] The fused features are input to the loop optimization layer to generate component data;

[0015] Step 4: Use the cross-trained prediction model to predict the composition of the polishing liquid.

[0016] In the technical solution provided by the present application, when monitoring the effective components of the polishing liquid, the polishing liquid sample is not subjected to "filtration, separation, extraction, quantification" and other methods to extract various components in the polishing liquid, and then calculate the content of various components in the polishing liquid. Instead, the label data is used as the prediction information, the component data is used as the label, and the prediction model is trained with a large amount of prediction information and labels, and the prediction model is used to find the intrinsic connection between the prediction information and the label. Therefore, when monitoring the components of the polishing liquid, the component data of the polishing liquid can be directly predicted with the easily accessible label data, and then the changes of the components in the polishing liquid can be obtained as the use process of the polishing liquid, and then the polishing liquid can be replaced in time to avoid defective products during polishing. At the same time, the prediction model constructed in this solution includes LSTM neural network and convolutional neural network, which are respectively input with time-related sequence data and image data, wherein the LSTM neural network can capture the information before and after the time-temperature data and the time-PH data to obtain the change characteristics of the components dissolved in the polishing liquid, and the convolutional neural network can capture the corresponding image features to obtain the change characteristics of the suspended matter in the polishing liquid that affects the light scattering.

[0017] Furthermore, the polishing liquid includes: aluminum oxide powder, CTAB, hydrochloric acid, sodium hydroxide, and deionized water.

[0018] There are many types of substances in the polishing liquid, and the content of each substance is somewhat disordered. If the content of each substance is taken into account in the prediction model, it will inevitably lead to high accuracy requirements for the prediction model, and more feature connections need to be considered, resulting in high model construction costs. To this end, this application provides the following technical solutions:

[0019] Furthermore, step 1 includes the following steps:

[0020] Step 11: Obtain in advance the same type of polishing liquid, the content of each component during use, and the corresponding pH value and temperature data;

[0021] Step 12: According to the use requirements of the polishing liquid, set the sampling range and prediction accuracy of each component, and generate the marking points of each component according to the sampling range and prediction accuracy of each component.

[0022] In the technical solution provided in the present application, the sampling range and prediction accuracy of each component are set, so that invalid data on the concentration of each component in the polishing liquid is removed, so that the prediction model needs to consider fewer results, reducing the prediction model's demand for accuracy.

[0023] When measuring spectral information, it will be affected by other factors. To this end, this application provides the following solutions:

[0024] Furthermore, the spectral information is acquired by performing the following steps:

[0025] S1: Obtain a polishing liquid sample, irradiate the polishing liquid sample with a laser of a fixed wavelength, and collect a spectrum of the laser passing through the polishing liquid sample;

[0026] S2: According to the light intensity signal distribution in the spectrum, the spectrum is intercepted to obtain the area of ​​interest;

[0027] S3: median filtering is performed on the region of interest to obtain a filtered spectrum;

[0028] S4: binarizing the filtered spectrum, and then extracting the center of the filtered spectrum;

[0029] S5: extract the maximum spot and saturation spot of the filtered spectrum, mark the center of the circle, the maximum spot and the saturation spot in the region of interest, and use the region of interest as spectral information.

[0030] In the technical solution provided in the present application, when collecting spectral information, the spectrum is first intercepted to remove a lot of irrelevant information, and then median filtering is performed to reduce noise interference and increase the effectiveness of information; the maximum light spot, saturated light spot and center of the circle are marked in the area of ​​interest, which can reflect the scattering of the polishing liquid sample to the spectrum, thereby reflecting the dispersion of suspended matter in the polishing liquid, and then reflecting the concentration of the suspended matter.

[0031] The spectral information is affected by the optical path of the laser in the polishing liquid, so there will be some dark spots. These dark spots are mainly caused by the laser being blocked by the suspended matter in the polishing liquid during the scattering process. The distribution of these dark spots has no regularity and cannot reflect the concentration of the suspended matter in the polishing liquid. However, it will affect the reading of the spectral information. To this end, the present application provides the following technical solutions:

[0032] Furthermore, the window of the median filter is 3×3.

[0033] In the technical solution provided in the present application, a median filter of a 3×3 window is set, so for each pixel point, it is the median of the light intensity of the surrounding 8 points plus its own light intensity, so that special dark spots in the spectral information can be accurately filtered out, but the distribution of low-value light intensity in the spectral information will not be blurred, thereby increasing the ability to read spectral information.

[0034] The spectral information of the polishing liquid is affected by many aspects of information, especially the spectral information of the edge part of the spectral information. Because the laser is reflected and absorbed many times, there is a certain disorder. Therefore, it is difficult to distinguish the scattering characteristics of the laser. To this end, the present application provides the following technical solutions:

[0035] Furthermore, when extracting the center of the filtered spectrum, the threshold of the binarization process is the median of the intensity range.

[0036] In the technical solution provided in the present application, in order to accurately reflect the influence of the suspended matter concentration in the polishing liquid on laser scattering, binarization processing is performed using the median of the intensity range and the center position of the circle is located. Therefore, the center position of the circle can reflect the range distribution of laser scattering in the polishing liquid, making it easier to extract the relationship between the suspended matter concentration in the polishing liquid and laser scattering.

[0037] Furthermore, when extracting the maximum light spot, the threshold of the binarization process is the maximum value of the intensity of the background area within the intensity range.

[0038] In this solution, the "maximum value of the intensity in the background area" is used as the maximum light spot, so it can accurately reflect the scattering range of the laser irradiated in the polishing liquid.

[0039] Furthermore, when extracting the saturated light spot, the threshold of the binarization process is three quarters of the maximum value of the intensity range.

[0040] In this scheme, “three quarters of the maximum intensity range” is used to extract the saturated light spot, so the saturated light spot is the central area after the laser is scattered in the polishing liquid, which makes it easier for the prediction model to learn and locate the center position of the spectral information.

[0041] In the technical solution provided in the present application, the center of the filter spectrum, the maximum spot, and the saturated spot can accurately reflect the scattering effect of the polishing liquid sample on the spectrum, from which effective information and light transmission intensity can be distinguished.

[0042] Furthermore, the convolutional network model includes:

[0043] Input layer: The input layer has three channels, which are used to input the center mask of the region of interest and the filter spectrum, the mask of the region of interest and the maximum spot, and the saturated spot mask of the region of interest and the filter spectrum;

[0044] Convolutional layer: used to perform convolution operations on the information input by the input layer;

[0045] Pooling layer, which performs pooling operation on the output of the convolutional layer and outputs a feature map;

[0046] The fully connected layer classifies the feature map to generate the second high-level features.

[0047] In the technical solution provided in the present application, when generating the second high-level feature, an input layer with three channels is selected, and the input information is the center mask of the filtered spectrum, the mask of the region of interest and the maximum spot, and the saturated spot mask of the region of interest and the filtered spectrum. In this way, the convolutional network model can integrate the relevant features of the center, the maximum spot and the saturated spot when generating the second high-level feature.

[0048] As a second aspect of the present application, the present application provides a polishing liquid monitoring system, comprising:

[0049] The information acquisition module is used to acquire the marking information; the monitoring module is used to monitor the concentration of each component of the polishing liquid based on the above-mentioned polishing liquid monitoring method. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The drawings constituting a part of this application are used to provide a further understanding of this application, so that other features, purposes and advantages of this application become more obvious. The illustrative embodiment drawings and their descriptions of this application are used to explain this application and do not constitute an improper limitation on this application.

[0051] In addition, throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that the components and elements are not necessarily drawn to scale.

[0052] In the attached picture:

[0053] Figure 1 Flow chart of the polishing liquid monitoring method.

[0054] Figure 2 This is the effect of laser passing through the polishing liquid sample. DETAILED DESCRIPTION

[0055] Embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are only for exemplary purposes and are not intended to limit the scope of protection of the present application.

[0056] It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0057] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0058] Reference Figure 1 , the polishing liquid monitoring method comprises the following steps:

[0059] Step 1: Obtaining the composition data of the polishing liquid during its use and the marking data corresponding to the composition data, wherein the composition data includes the concentration of each component, and the marking data includes pH value, spectral information and temperature data.

[0060] In metal processing, polishing liquid can make metal products exceed their original gloss and improve the surface quality of the product. When the polishing liquid is configured, the concentration of each component is determined. As the polishing liquid is used, the components in the polishing liquid will gradually change under various chemical and physical reactions until they lose their function.

[0061] Polishing liquid generally includes suspended matter and soluble matter. For example, in a polishing liquid mainly composed of aluminum oxide powder, CTAB, hydrochloric acid, sodium hydroxide, and deionized water, aluminum oxide powder is suspended matter, and CTAB, hydrochloric acid, and sodium hydroxide are soluble matter. Although there is no necessary chemical or physical reaction relationship between suspended matter and soluble matter, during the use of the polishing liquid, the concentration of suspended matter and soluble matter changes synchronously with the polishing process.

[0062] In order to avoid defective products during processing, it is necessary to monitor the concentration of each component in the polishing liquid and replace the polishing liquid in time. During the use of the polishing liquid, the concentration of each component in the polishing liquid will change, and the corresponding temperature and pH value of the polishing liquid will also change. At the same time, after the concentration of each component in the polishing liquid changes, the spectral information will also change. The pH value and temperature are affected by the soluble matter inside the polishing liquid, while the spectral information is mainly affected by the suspended matter in the polishing liquid, that is, the pH value and temperature will reflect the concentration information of the soluble matter in the polishing liquid, and the spectral information will affect the information of the insoluble matter in the polishing liquid. The reason is that the soluble matter will affect the pH value of the polishing liquid, and the chemical reaction generated during the use of the polishing liquid will also affect the temperature of the polishing liquid, causing the temperature of the polishing liquid to change. The insoluble matter in the polishing liquid will be dispersed in the polishing liquid in a suspended manner. After being irradiated by the laser, the laser and the suspended matter will scatter, absorb and other information, thereby generating spectral information.

[0063] In this way, the pH value and temperature can reflect the concentration information of soluble matter in the polishing liquid, and the spectral information can reflect the concentration information of suspended matter in the polishing liquid.

[0064] Furthermore, the marking data in this solution includes pH value, spectral information and temperature data, which utilizes the intrinsic relationship between this information and the main components in the polishing liquid.

[0065] Because the temperature is affected by the ambient temperature and heat dissipation, the present solution requires that the polishing liquid be monitored during continuous use to avoid the situation where the polishing liquid cannot reflect the heat absorption and heat release information during use when it is used at intervals.

[0066] Furthermore, step 1 includes the following steps:

[0067] Step 11: Obtain in advance the same type of polishing liquid, the content of each component during use, and the corresponding pH value and temperature data.

[0068] There are many types of polishing liquids, each with different ingredients and reactions, so when predicting polishing liquids, each type of polishing liquid needs to be trained separately.

[0069] After the polishing liquid is configured, the effective ingredients in the polishing liquid will change with the reaction. For each stage, the content of each component, pH value, temperature data and spectral information can be extracted. Among them, the component content requires sampling the polishing liquid, and then conducting the corresponding component test in the laboratory. The pH value and temperature data are directly monitored by sensors. In this way, the change information of the component content of the polishing liquid under normal use, as well as the pH value and temperature corresponding to the change information of the component content can be obtained.

[0070] For example: The initial composition information of the polishing liquid is a 1 ; The pH value at this time is a 2 , temperature is a 3 , the spectral information is a 4 After one hour of use, the composition information of the polishing liquid changes from a 1 Became b 1 , the pH value becomes b 2 , the temperature becomes b 3 , the spectral information becomes b 4 The pH value, temperature, spectral information, and component information are continuously recorded as the polishing liquid is used. From the time the polishing liquid is configured to the time it cannot be used, the content of each component and the pH value and temperature data corresponding to the content of each component can be obtained, that is, multiple data groups can be obtained, each of which includes pH value, temperature, spectral information, and component information.

[0071] Step 11 is to obtain the changes in the content, pH value, temperature and spectral information of a large amount of polishing liquid during use. The content of each component is obtained through standard chemical quantitative test detection, and the pH value, temperature and spectral information are directly obtained through sensors or corresponding equipment. The specific acquisition methods are all prior art and will not be further elaborated here.

[0072] The core idea of ​​this solution is: for each polishing liquid, a large number of data sets are obtained in advance through step 11. These data sets are actually training samples for the prediction model. Among them, the component content is the information that the prediction model needs to predict, and the pH value, temperature and spectral information are the input information of the prediction model.

[0073] Step 12: According to the use requirements of the polishing liquid, set the sampling range and prediction accuracy of each component, and generate the marking points of each component according to the sampling range and prediction accuracy of each component.

[0074] During the use of the polishing liquid, the more frequently the component content is tested, the denser the corresponding data, the more training samples there are, and the more the relationship between the component content and the labeled data can be demonstrated. In subsequent training, the accuracy of the prediction model will be higher. For example, assuming that after a certain polishing liquid is configured, its life span is 1 hour under specific working conditions. If it is only tested three times within this hour, only 3 data groups can be obtained; if it is tested 10 times, 10 test groups can be obtained. The more frequent the test, the more cumbersome the component content test work is, but the more times the test is performed, the more data groups there are, and the better the training effect of the prediction model.

[0075] Therefore, it is necessary to control the sampling density when collecting data in step 11 to control the difficulty of obtaining training samples. The reasons for this design are as follows:

[0076] Assume that the total number of training samples is set to 1000. If these 1000 training samples are all extracted from a polishing liquid use cycle. Then this polishing liquid can extract 1000 data groups from the beginning of use to the final unusability. These 1000 data groups can well reflect the relationship between the composition information and the labeled data during the use of the polishing liquid. However, these 1000 data groups are only the detection results of a polishing liquid use cycle, and cannot well represent all polishing liquids. For this reason, if these 1000 training samples are all extracted from 10 polishing liquid use cycles, then in each of these 10 polishing liquid use cycles, there are only 100 data groups. Compared with the 1000 data groups in the previous use cycle, although the total number of data groups is reduced, these 1000 training samples involve 10 polishing liquid use cycles, and their representativeness is greater. For this reason, it is necessary to control the data extraction density, and try to involve more polishing liquid use cycles under the same data group extraction difficulty. For this reason, in practice, it is necessary to set it according to the use requirements of the polishing liquid.

[0077] For example, when a polishing liquid with the active ingredient "aluminum oxide powder" is selected, the concentration of the substance must be higher than 10%, and the initial concentration of the substance is 20% during configuration. After a workpiece is polished, the maximum consumption of aluminum oxide powder is 1%. In this way, in order to avoid polishing the workpiece with the polishing liquid when the content of aluminum oxide powder is lower than 10%, it is necessary to collect the data set in step 11. The difference in the content of aluminum oxide powder between two adjacent sampling points needs to be less than 1%, so that the prediction model can achieve an error of less than 1% in prediction accuracy as much as possible, and then the concentration of aluminum oxide powder in the polishing liquid can be monitored when it is 11% to 12%, avoiding insufficient concentration of aluminum oxide powder during polishing. Therefore, the extraction density when collecting data in step 11 is related to the error accuracy. If the error needs to be lower than 1%, the extraction density also needs to be lower than 1%.

[0078] Thus, in step 11, when extracting the relevant data group, it is necessary to ensure that the density of the obtained content of each component and the corresponding pH value, temperature data and spectral information is lower than the concentration change of "aluminum oxide powder" by 1%. For example, during the use of the polishing liquid, the aluminum oxide powder is consumed by an average of 0.5% per hour, so when collecting information, it is necessary to collect the content of each component and the corresponding pH value, temperature data and spectral information at least once an hour, and the density needs to be lower than the concentration change of "aluminum oxide powder" by 1%. In the obtained content of each component and the corresponding pH value, temperature data and spectral information, the content (concentration) change value of aluminum oxide powder in the component content will be around 0.5% (lower than 1%). At this time, the component content can be extracted as needed, so that the content (concentration) change value of aluminum oxide powder in the adjacent component content is around 0.9%.

[0079] In essence, step 11 is a process of obtaining denser information, while step 12 is a process of extracting denser information. Therefore, the marking points in step 12 are pre-set data points. For example, the data points of 11%, 12%, ... 20% of aluminum oxide powder can be set as marking points. In practice, if the component content data extracted from a batch of polishing liquid only contain data points of 11.1%, 12.1%, ... 20.1%, and do not contain data points of 11%, 12%, ... 20%, then the pre-set data points of 11%, 12%, ... 20% are replaced by data points of 11.1%, 12.1%, ... 20.1%. If the number of data groups obtained from the component content data extracted from a batch of polishing liquid is higher than the number of data points, delete the redundant data points to make the number of data points the same as the pre-set number.

[0080] Step 2: Obtain the time-temperature data, time-PH data and time-spectrum information of the polishing liquid to be tested as prediction information. Step 2 mainly uses the prediction model to collect the required information after the prediction model training is completed. The prediction information here is actually the input information of the prediction model in step 1.

[0081] In both step 1 and step 2, spectral information needs to be collected. The method for obtaining spectral information includes the following steps:

[0082] S1: Obtain a polishing liquid sample, irradiate the polishing liquid sample with a laser of a fixed wavelength, and collect a spectrum of the laser passing through the polishing liquid sample.

[0083] Since the influence of natural light needs to be avoided when obtaining spectral information, a sample of the polishing liquid needs to be obtained. In practice, the polishing liquid can be directly obtained from a container (water tank) of the polishing liquid manually or by machine, or the sample of the polishing liquid can be automatically obtained by a pipeline system.

[0084] After obtaining a sample of the polishing liquid, put the sample of the polishing liquid into a test tube, irradiate the test tube with a laser, and collect spectral information after the laser passes through the test tube. The principle is: when the laser is irradiated on pure water, the laser will basically pass through the pure water to form a light spot slightly larger than the initial diameter of the laser. When the laser is irradiated on the polishing liquid, scattering and absorption phenomena will occur, and these phenomena are closely related to the concentration of the corresponding substances in the polishing liquid. When the laser passes through the polishing liquid, it will interact with the particles or molecules in the polishing liquid, causing changes in the propagation direction and intensity of the light, so that the laser will be diffused, and finally a light spot much larger than the initial diameter of the laser will be formed. This light spot is the spectral information. Therefore, the spectral information is more related to the content of each component in the polishing liquid.

[0085] In practice, because the soluble particles in the polishing liquid are small and evenly dispersed, they have little effect on the scattering of the laser. However, the suspended matter in the polishing liquid has a large particle size and is more likely to scatter and reflect the laser noise, affecting the spectral information. In order to increase the correlation between the spectral information and the concentration of the suspended matter, in this scheme, the frequency of the laser is f, and the absorption peak of the suspended matter is v, f = c*v, where c is the speed of light. For example, the absorption peak of the Al-O bond vibration peak in aluminum oxide is approximately 400-600cm -1 Within the range, the laser frequency f = 400cm -1 ×3×10 10 cm / s=1.2×10 13 Hz.

[0086] In this scheme, the frequency of the laser is made equal to the absorption peak of the suspended matter. Therefore, the light intensity in the final spectrum has a greater correlation with the corresponding concentration of the suspended matter. When the concentration of the suspended matter changes, the corresponding changes can be more clearly obtained in the spectrum. At the same time, the concentration changes of other substances will reduce the impact on the light intensity in the spectrum.

[0087] In S1, the thickness (optical path) of the polishing liquid sample also needs to be controlled, that is, the polishing liquid should not be too thick, so that after the laser passes through the polishing liquid, most of the laser energy is scattered and consumed, and no effective spectral information can be detected. Therefore, for each polishing liquid, the thickness of the polishing liquid sample needs to be controlled in advance. It is mainly judged based on the detected intensity (light intensity). In this scheme, the light intensity is controlled between 0.1 and 0.2 of the initial light intensity to allow the laser to be fully dispersed in the polishing liquid and avoid other problems. Figure 2 As shown, Figure 2 This is a diagram showing the scattering effect of infrared light irradiating onto the polishing liquid sample.

[0088] S2: According to the light intensity signal distribution in the spectrum, the spectrum is intercepted to obtain the area of ​​interest.

[0089] The pixels at the edge of the spectrum are information unrelated to spectral scattering, so the spectrum needs to be intercepted to obtain the region of interest. When obtaining the region of interest, it is mainly judged based on the change rate of intensity (light intensity), and the area with an intensity change rate lower than the first preset value is set as the region of interest.

[0090] For example, in a spectrum graph, the intensity increases from 100 to 1000 from the center to the edge. The pixels at the edge with an intensity less than 100 have no usable information because the intensity has not changed. Therefore, the area with increased intensity is regarded as the area of ​​interest.

[0091] S3: Perform median filtering on the region of interest to obtain a filtered spectrum.

[0092] The region of interest is the middle position of the spectral information, but because there is still a lot of noise in the spectral information, a median filter is required. The window of the median filter is 3×3. During the median filter, for each pixel point, the original pixel value is replaced by the median of the pixel values ​​in its neighborhood (3×3 range). In this scheme, the reason for using a 3×3 window is mainly to avoid the influence of some special dark spots on the spectral information. The 3×3 window is actually the pixels directly adjacent to the pixel point. The intensity of the center point is replaced by the intensity of these pixels, which has the least impact on the spectral information and can effectively eliminate special dark spot information.

[0093] S4: Binarize the filtered spectrum and then extract the center of the filtered spectrum.

[0094] S5: extract the maximum spot and saturation spot of the filtered spectrum, mark the center of the circle, the maximum spot and the saturation spot in the region of interest, and use the region of interest as spectral information.

[0095] S4 and S5 are mainly used to obtain effective information in the filter spectrum. The effective information here mainly refers to the center of the filter spectrum, the maximum spot and the saturated spot. When extracting the above three information, the binarization threshold is mainly used. After the binarization threshold, information extraction is performed.

[0096] When obtaining the center of the filtered spectrum, the threshold of the binarization process is the median of the intensity range. For example, in the filtered spectrum, the intensity range of all pixels is 1000-2000, and the median of the intensity range is 1500. Set the pixel grid with an intensity value (light intensity) exceeding 1500 in the filtered spectrum to 1, and set the pixel grid with an intensity value (light intensity) below 1500 in the filtered spectrum to 0. For the pixel points set to 1, find its center point and use the center point as the center of the filtered spectrum. When obtaining the maximum light spot and the saturated light spot, the scheme is the same, except that the critical point for the binarization process is different. Specifically, when obtaining the maximum light spot, the threshold of the binarization process is the maximum value of the intensity of the background area in the intensity range.

[0097] The background area refers to the edge area in the filter spectrum. In the filter spectrum, the intensity of the central area is very high and the intensity changes significantly, while the edge of the filter spectrum will show a low intensity and a slow intensity change. The region of interest is the part of the filter spectrum where the intensity change is lower than the second preset value. For example, the preset value is that the intensity change of 10 adjacent pixel grids is less than 100, which is used as the boundary of the background area.

[0098] After obtaining the boundary line of the background area, the maximum value of the background area intensity can be obtained, and then the pixels exceeding this value are set to 1, and those less than this value are set to 0. For the pixel points set to 1, they are regarded as the maximum light spot.

[0099] When extracting saturated light spots, the threshold of the binarization process is an intensity range of three quarters of the maximum intensity in the filtered spectrum. For example, in the filtered spectrum, the intensity range of all pixels is 1000-2000, then the pixels exceeding 1750 are set to 1, and those less than 1750 are set to 0. For the pixels set to 1, they are regarded as saturated light spots.

[0100] Therefore, in this scheme, the maximum spot is the area in the region of interest that is most affected by the scattering of the polishing liquid, and the saturated spot is the core area of ​​laser scattering.

[0101] Step 3: Build a component prediction model, use the prediction information as the input of the prediction model, and the component data as the output, and cross-train the prediction model.

[0102] Among them, the prediction model includes LSTM neural network, convolutional neural network, fully connected layer and loop optimization layer.

[0103] The time-temperature data and the time-PH data are input into the LSTM neural network to obtain the first high-level feature; the spectral information is input into the convolutional neural network to obtain the second high-level feature; the first high-level feature and the second high-level feature are input into the fully connected layer to generate fused features; the fused features are input into the loop optimization layer to generate component data.

[0104] Among them, the LSTM neural network includes:

[0105] Input network, used to input pH value and temperature data.

[0106] LSTM Layer: Contains one or more LSTM units. Each LSTM unit has three main gates: Forget Gate, Input Gate, and Output Gate, as well as a cell state to store long-term information.

[0107] Fully Connected layer: outputs the first high-level features.

[0108] The convolutional network model includes:

[0109] Input layer: The input layer has three channels, which are used to input the center mask of the region of interest and the filter spectrum, the mask of the region of interest and the maximum spot, and the saturated spot mask of the region of interest and the filter spectrum;

[0110] The convolution layer is used to perform convolution operations on the information input by the input layer;

[0111] Pooling layer, which performs pooling operation on the output of the convolutional layer and outputs a feature map;

[0112] The fully connected layer classifies the feature map to generate the second high-level features.

[0113] LSTM neural network and convolutional network model are mainly used to extract high-level features related to the content of each component. Therefore, when building a prediction model, it is necessary to first train the convolutional network model and LSTM neural network. The training method is to use the pH value and temperature data as the input information of the LSTM neural network, the spectral information as the input information of the convolutional network model, and then use the content of each component as the supervision information of the LSTM neural network and the convolutional network model. The LSTM neural network outputs the first high-level feature, and the convolutional network model outputs the second high-level feature. In this way, the first high-level feature and the second high-level feature will have a greater correlation with the content of each component.

[0114] The cycle optimization layer is actually a support vector machine model. In practice, the concentration (content) of each substance in the polishing liquid is not disordered, but has certain rules. For example, if the concentration of aluminum oxide powder is measured, it can actually show the reaction state of the current polishing liquid. The concentrations of CTAB, hydrochloric acid, and sodium hydroxide are basically the same.

[0115] Therefore, the loop optimization layer used in this scheme is to use a support vector machine to regress the fusion features to the corresponding component information. That is, in the loop optimization layer, the label is the component information, and the prediction information (input information) is the labeled data. Among them, the types of component information are limited, and the types of component information are mainly related to the prediction accuracy set in step 12. For example, a polishing liquid with an effective ingredient of "aluminum oxide micropowder" is selected. The concentration of the substance must be higher than 10%. The initial concentration of the substance is 20% during configuration, and the error is less than 1%. Then the sample types are 11%, 12%,...20%. As for the remaining components, in fact, after determining the content of aluminum oxide micropowder, the content of the remaining components is basically determined. Therefore, the types of component information in this scheme are limited. In this way, the loop optimization layer in this scheme is actually based on the support vector machine to distinguish the fusion features and regress the fusion features to the corresponding component information.

[0116] Specifically, the training process of the loop optimization layer is as follows:

[0117] S01: Initialize the weight vector w, bias term b, deviation vector e, and regularization parameter λ;

[0118] S02: Set the objective function:

[0119] T represents the transpose of a vector;

[0120] Set up constraints:

[0121]

[0122] y i represents the true output of the i-th sample, Indicates that the input vector X i Nonlinear function mapped to high-dimensional space, e i represents the prediction error of the i-th sample;

[0123] Set up the Lagrangian function:

[0124]

[0125] i and j represent the index of the sample, a i represents the Lagrange multiplier, L represents the Lagrange function, and K represents the kernel function;

[0126] S03: Find the partial derivative of the Lagrangian function and set it equal to zero to obtain a set of optimization conditions; the conditions include:

[0127]

[0128] γe i -ai =0;

[0129]

[0130] S04: By arranging the above optimization conditions, a linear equation system is obtained;

[0131] I represents the identity matrix, represents an m×m matrix, where m is the number of training samples, Elements is the kernel function K(x i , x j ) value, indicating the training sample x i and x j The Euclidean distance between .

[0132] S05: Use Gaussian elimination method to solve the above linear equations to obtain the parameters of the model, including the weight vector w, the bias term b and the Lagrange multiplier vector a;

[0133] S06 uses the training set data to evaluate the trained model and optimize the model.

[0134] Through the above scheme, the loop optimization layer can be optimized.

[0135] Step 4: Use the cross-trained prediction model to predict the composition of the polishing liquid.

[0136] Embodiment 2: The present invention further provides a polishing liquid monitoring system. The polishing liquid monitoring system comprises: an information acquisition module for acquiring marking information and a monitoring module for monitoring the concentration of each component of the polishing liquid based on the polishing liquid monitoring method of embodiment 1.

[0137] The above description is only some preferred embodiments of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present application is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above invention concept. For example, the above features are replaced with (but not limited to) the technical features with similar functions disclosed in the embodiments of the present application to form a technical solution.

Claims

1. A polishing liquid monitoring method, characterized in that: The steps include: Step 1: Obtaining component data of the polishing liquid during its use and marking data corresponding to the component data, wherein the component data includes the concentration of each component, and the marking data includes pH value, spectral information and temperature data; Step 2: Obtaining time-temperature data, time-PH data and time-spectrum information of the polishing liquid to be tested as prediction information; Step 3: Build a component prediction model, use the prediction information as the input of the prediction model, and the component data as the output, and cross-train the prediction model; Among them, the prediction model includes LSTM neural network, convolutional neural network, fully connected layer and loop optimization layer; The time-temperature data and the time-PH data are input into the LSTM neural network to obtain the first high-level features; The spectral information is input into the convolutional neural network to obtain the second high-level features; The first high-level feature and the second high-level feature are input into a fully connected layer to generate a fused feature; The fused features are input to the loop optimization layer to generate component data; Step 4: Use the cross-trained prediction model to predict the composition of the polishing liquid.

2. The polishing liquid monitoring method according to claim 1, characterized in that: Step 1 includes the following steps: Step 11: Obtain in advance the same type of polishing liquid, the content of each component during use, and the corresponding pH value and temperature data; Step 12: According to the use requirements of the polishing liquid, set the sampling range and prediction accuracy of each component, and generate the marking points of each component according to the sampling range and prediction accuracy of each component.

3. The polishing liquid monitoring method according to claim 1, characterized in that: The following steps are performed when acquiring spectral information: S1: Obtain a polishing liquid sample, irradiate the polishing liquid sample with a laser of a fixed wavelength, and collect a spectrum of the laser passing through the polishing liquid sample; S2: According to the light intensity signal distribution in the spectrum, the spectrum is intercepted to obtain the area of ​​interest; S3: median filtering is performed on the region of interest to obtain a filtered spectrum; S4: Binarize the filtered spectrum and extract the center of the filtered spectrum; S5: extract the maximum spot and saturation spot of the filtered spectrum, mark the center of the circle, the maximum spot and the saturation spot in the region of interest, and use the region of interest as spectral information.

4. The polishing liquid monitoring method according to claim 3, characterized in that: The window of the median filter is 3×3.

5. The polishing liquid monitoring method according to claim 3, characterized in that: When extracting the center of the filtered spectrum, the threshold of the binarization process is the median of the intensity range.

6. The polishing liquid monitoring method according to claim 3, characterized in that: When extracting the maximum light spot, the threshold of the binarization process is the maximum value of the intensity of the background area in the intensity range.

7. The polishing liquid monitoring method according to claim 3, characterized in that: When extracting the saturated light spot, the threshold of the binarization process is three quarters of the maximum value of the intensity range.

8. The polishing liquid monitoring method according to claim 1, characterized in that: The convolutional network model includes: Input layer: The input layer has three channels, which are used to input the center mask of the region of interest and the filter spectrum, the mask of the region of interest and the maximum spot, and the saturated spot mask of the region of interest and the filter spectrum; The convolution layer is used to perform convolution operations on the information input by the input layer; Pooling layer, which performs pooling operation on the output of the convolutional layer and outputs a feature map; The fully connected layer classifies the feature map to generate the second high-level features.

9. A polishing liquid monitoring system, characterized in that: include: An information acquisition module, used to acquire marking information; The monitoring module monitors the concentration of each component of the polishing liquid based on the polishing liquid monitoring method according to any one of claims 1 to 8.

Images