A method for on-line monitoring of a polishing process state of a silicon carbide wafer

Abstract

The application discloses a kind of on-line monitoring methods of silicon carbide wafer polishing processing state, comprising the following steps: collecting the periodic vibration signal data of silicon carbide wafer grinding polishing complete working condition, constructs silicon carbide wafer grinding polishing material removal characteristic data set;Establishment processing test working condition library, and vibration signal test value is put into matched working condition library;Variational modal decomposition is carried out to vibration signal, and IMF main component energy spectrum is obtained;Actual processing characteristic index is constructed;Particle filtering state estimation model is established;New characteristic index is constructed by particle point, and this iteration realizes silicon carbide wafer polishing processing state on-line monitoring.The original vibration signal is decomposed by variation modal in the application, noise can be effectively removed, main energy spectrum is obtained, so that subsequent analysis is more accurate, and anti-interference ability is enhanced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of high-hardness and high-brittle material silicon carbide wafer processing, in particular to a processing state monitoring technology, and more particularly to a silicon carbide wafer polishing processing state online monitoring method based on particle filtering of vibration signals. BACKGROUND

[0002] As a high-hardness and high-brittle material, silicon carbide is widely used in the field of semiconductor chips, and the requirement for the surface engineering processing quality thereof is getting higher and higher. In the polishing processing of silicon carbide wafers, the most important link besides the material removal mechanism process is the monitoring link of the processing state, which restricts the realization of automatic control of the processing process and other problems. At present, the quality of the polished workpiece depends on offline testing after shutdown, which seriously leads to a large waste of processing time and cost. Therefore, the online monitoring of the polishing processing state of the silicon carbide wafer has important engineering value.

[0003] At present, there are mainly two methods of mechanism analysis and shallow machine learning analysis in the field of online monitoring. The main disadvantages of mechanism analysis are that mathematical and physical modeling is difficult, the monitoring precision is low, and the operability is not strong. Shallow machine learning uses a data-driven method to learn and analyze based on the collected physical signals, which greatly improves the precision, but the extraction and processing of characteristic values are complicated and difficult, and it is difficult to handle complex monitoring models.

[0004] In summary, the monitoring of the polishing processing state of the silicon carbide wafer plays a crucial role in the surface quality, but the current monitoring method has the disadvantages of complex processing process, poor anti-interference ability, and low precision, which need to be solved urgently. SUMMARY

[0005] In order to overcome the disadvantages of the existing online monitoring method of the processing state, such as complex processing process, poor anti-interference ability, and low precision, the present application provides an online monitoring method of the polishing processing state of a silicon carbide wafer based on particle filtering.

[0006] The online monitoring method of the polishing processing state of the silicon carbide wafer comprises the following steps:

[0007] A: Collecting periodic vibration signal data of the complete working condition of silicon carbide wafer grinding and polishing, and constructing a silicon carbide wafer grinding and polishing material removal characteristic data set;

[0008] B: Establishing a processing test working condition library, and putting the vibration signal test value into the matched working condition library;

[0009] C: Performing variational modal decomposition on the vibration signal to obtain an IMF principal component energy spectrum;

[0010] D: Constructing an actual processing characteristic index;

[0011] E: Establishing a particle filter state estimation model;

[0012] F: New feature index construction is carried out by particle points, and online monitoring of the polishing processing state of the silicon carbide wafer is realized through iteration.

[0013] Further, step B further comprises the following steps:

[0014] B1 According to the processing test outline, the operating conditions are established for each step of processing test according to the operating conditions;

[0015] B2 Real-time comparison of vibration signal test value, when matching changes, the current vibration signal is put into the re-matching condition library.

[0016] Further, step C further comprises the following steps:

[0017] C1 After variational mode decomposition, Winer filter is used for denoising;

[0018] C2 Set the limited bandwidth parameter and the center angular frequency initialization to obtain the estimated each center angular frequency;

[0019] C3 According to different center angular frequencies, IMF main component energy spectrum is obtained, and this signal is used as the input of the next step.

[0020] Further, the IMF main component energy spectrum is each intrinsic mode function.

[0021] Further, step D forms an adaptive feature index through three steps of benchmark learning, offset tolerance calculation and feature value construction, and carries out normalization processing to obtain the actual feature index of the processing state.

[0022] Further, the forming step of the adaptive feature index in step D is:

[0023] D1 In the benchmark learning stage, the IMF main component energy spectrum spectrum of the first n times of acquisition processing is calculated, and then the mean value of the n times of signal spectrum is taken as the benchmark value;

[0024] D2 In the offset tolerance calculation stage, the frequency spectrum diagram of the continuous 5 times of acquisition signals is calculated, and the offset tolerance between the mean value frequency spectrum of the five times of signals and the benchmark frequency spectrum is analyzed respectively;

[0025] D3 The feature values of the time period are obtained by extracting the features of different condition libraries through steps D1 and D2, and the feature indexes are respectively constituted.

[0026] Further, step E comprises the following steps:

[0027] E1 Analyzing the experimental data feature index in step D;

[0028] E2 selects a double exponential model as the degradation model for fitting training, and determines the initial value of the model parameter of the observation equation through the data fitting result;

[0029] E3 selects the training data with good fitting effect in step E2, and substitutes the data into the model for parameter initialization.

[0030] Further, step F includes the following steps:

[0031] F1 real-time monitoring data are used to update the model parameters and new characteristic indexes (processing state indexes) in real time;

[0032] F2 updates and adjusts the distribution of the model parameters.

[0033] Further, step F3 is further included, which iteratively calculates through the value of the new characteristic index.

[0034] Compared with the prior art, the present application has the following beneficial effects:

[0035] The present application can effectively remove noise and obtain the main energy spectrum by performing variational modal decomposition on the original vibration signal, so that the subsequent analysis is more accurate and the anti-interference ability is enhanced.

[0036] The present application proposes an adaptive characteristic index formed by three steps of benchmark learning, offset tolerance calculation and eigenvalue construction, and uses particle filtering to iteratively update the parameters to realize online monitoring, effectively solving the problem of poor adaptability of traditional eigenvalues and improving the precision. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flowchart of the silicon carbide wafer polishing processing state online monitoring method based on particle filtering of the present application. DETAILED DESCRIPTION

[0038] The present application will be further described below with reference to the accompanying drawings.

[0039] As shown in Figure 1 A silicon carbide wafer polishing processing state online monitoring method includes the following steps:

[0040] First, the silicon carbide wafer grinding and polishing material removal feature data set is constructed, and the periodic vibration original signal data of the complete working condition of the silicon carbide wafer grinding and polishing is collected.

[0041] According to the processing test outline, all working condition libraries are established according to the running conditions, the vibration signal test values are put into the successfully matched working condition library, and real-time comparison is performed, when the matching changes, the current vibration signal is put into the re-matched working condition library, and the vibration signal of each working condition is provided for subsequent feature extraction and monitoring identification.

[0042] The vibration signal is decomposed by variational modal decomposition, denoised by Wiener filter, and the estimated center angular frequency is initialized by setting the finite bandwidth parameter and center angular frequency. Then, the intrinsic mode function (IMF) principal component energy spectrum is obtained according to different center angular frequencies, and this signal is taken as the input of the next step.

[0043] In the benchmark learning stage, the IMF principal component energy spectrum spectrum of the first n acquisition processing is calculated, and then the mean value of the n signal spectrum is taken as the benchmark value.

[0044] In the offset tolerance calculation stage, the spectrum of the continuous 5 acquisition signals is calculated, and then the offset tolerance between the mean spectrum of the five signals and the benchmark spectrum is analyzed respectively.

[0045] Through the above two steps, the feature values of different working condition libraries are extracted, the characteristic indexes are formed respectively, and the actual characteristic indexes of the processing state are obtained by normalization processing.

[0046] A particle filter state estimation model is constructed, the double exponential model is selected as the degradation model for data fitting through the analysis of the characteristic indexes of the experimental data, and the training data is substituted into the model for parameter initialization.

[0047] According to the real-time monitoring data, the model parameters and new feature indexes (processing state indexes) are updated in real time, the model parameter distribution is adjusted, and finally the online monitoring of the polishing processing state of the silicon carbide wafer is realized through iterative calculation of the new feature index values.

[0048] The specific implementation is as follows:

[0049] First, the periodic vibration signal data of the silicon carbide wafer grinding and polishing complete working condition is collected, and the silicon carbide wafer grinding and polishing material removal characteristic data set is constructed;

[0050] According to the running condition, all working condition libraries are established, which are equivalent to n file libraries. During normal operation, the running condition is monitored, the stable processing state is compared with the processing state in the working condition library, the vibration signal test value is put into the successfully matched working condition library, and the processing state is monitored in real time. When the match changes, the current vibration signal is put into the re-matched working condition library, so that each stable working condition has a working condition database, which provides the vibration signal of each working condition for subsequent feature extraction and life prediction.

[0051] The vibration signal f is decomposed by variational modal decomposition to obtain the intrinsic mode function (IMF) principal component energy spectrum, wherein the multi-component signal f to be decomposed is composed of k IMF components u k , and the center frequency of each IMF corresponds to ω kThe intrinsic mode function u k (t) is

[0052] u k (t) = A h (t) cos [Φ h (t)]

[0053] where A k (t) represents the instantaneous amplitude of u k (t), ω k = Φ k (t) represents the instantaneous frequency of u k (t).

[0054] The actual machining feature index is constructed in three steps: benchmark learning, offset tolerance calculation, and feature value construction. In the benchmark learning stage, the frequency spectrum of the reducer is first calculated for n times of signal acquisition, and then the mean value of the n times of signal spectrum is calculated as the benchmark value.

[0055]

[0056] In the formula, X i is the frequency spectrum of the i-th acquired signal, is the benchmark value.

[0057] In the offset tolerance calculation stage, the frequency spectrum of the continuous 5 times of acquired signals is calculated, and then the offset tolerance between the mean value frequency spectrum of the 5 times of signals and the benchmark frequency spectrum is analyzed respectively. The specific steps are as follows

[0058]

[0059]

[0060] ······

[0061]

[0062] T k = ∑|s i |(i = 1, 2, 3,..., k)

[0063] where S k is the offset tolerance value of the k-th variation spectrum in the frequency spectrum, T k is the feature value at time k, which is formed by superposition and summation of the variation spectrum.

[0064] The feature values of the working condition library are extracted respectively to obtain the feature values of the time period, and the machining state index is constructed. The feature values of each time period are accumulated and normalized to obtain the actual machining state index T of the test piece, wherein the normalization is realized by transforming the accumulated feature value data to map the data to the range of [i, j] (default is [0, 1])

[0065]

[0066] Wherein, max is the maximum value in the characteristic value, min is the minimum value in the characteristic value.

[0067] The particle filter state estimation model is established, and the dynamic change process of the system can be represented by state transition equation and observation equation

[0068] x k =f k (x k-1 ,v k-1 )

[0069] z k =h k (x k ,n k )

[0070] Wherein, x k is the system state, z k is the system observation value, f k is the state transfer function, h k is the observation function, v k-1 is the process noise, and n k is the observation noise.

[0071] Based on the material removal characteristics of silicon carbide processing, a double exponential model is used as the degradation model, and the observed data is substituted into the model for parameter initialization, and the data fitting method is the least square function method, wherein the double exponential observation equation is

[0072] z k =ae bt +ce dt

[0073] Wherein a, c are model degradation weights, b, d are degradation factors, t is a sampling point, and a, b, c, d initial value model parameters are determined by data fitting.

[0074] According to the real-time monitoring data, the model parameters and the new feature index (processing state index) are updated in real time, and the model parameter distribution is adjusted, and finally the online monitoring of the polishing processing state of the silicon carbide wafer is realized through the iterative calculation of the new feature index value.

[0075] The original vibration signal is decomposed by variational modal decomposition, noise can be effectively removed, main energy spectrum is obtained, subsequent analysis is more accurate, and anti-interference ability is enhanced.

[0076] The application proposes a self-adaptive feature index formed based on three steps of benchmark learning, offset tolerance calculation and feature value construction, and adopts particle filtering to iteratively update parameters to realize online monitoring, effectively solves the problem of poor adaptability of traditional feature values, and improves the precision.

Claims

1. A method for on-line monitoring of a polishing process state of a silicon carbide wafer, characterized by, It comprises the following steps: A: Collecting the periodic vibration signal data of the silicon carbide wafer grinding and polishing complete working condition, and constructing the silicon carbide wafer grinding and polishing material removal characteristic data set; B: Establishing a processing test working condition library, and putting the vibration signal test value into the matched working condition library; C: Performing variational modal decomposition on the vibration signal to obtain the IMF principal component energy spectrum; D: Constructing the actual processing characteristic index; E: Establishing a particle filtering state estimation model; F: Constructing a new characteristic index through particle points, and realizing online monitoring of the silicon carbide wafer polishing processing state through iteration; Step D forms an adaptive characteristic index through three steps of benchmark learning, offset tolerance calculation, and characteristic value construction, and performs normalization processing to obtain the actual characteristic index of the processing state; The forming step of the adaptive characteristic index in step D is: D1: In the benchmark learning stage, the IMF principal component energy spectrum spectrum of the first n times of acquisition processing is calculated, and then the average of the n times of signal spectrum is taken as the benchmark value; D2: In the offset tolerance calculation stage, the spectrum graph of the continuous 5 times of acquisition signals is calculated, and the offset tolerance between the average spectrum of the five times of signals and the benchmark spectrum is analyzed respectively; D3: Through steps D1 and D2, the characteristic values of different working condition libraries in this time period are obtained, and the characteristic indexes are formed respectively.

2. The method of claim 1, wherein the polishing process of the silicon carbide wafer is monitored on-line. Step B further comprises the following steps: B1: According to the processing test outline, the working condition library is established for each step of processing test according to the running working condition; B2: Comparing the vibration signal test value in real time, when the match changes, the current vibration signal is put into the re-matched working condition library.

3. The method of claim 1, wherein the polishing process of the silicon carbide wafer is monitored on-line. Step C further comprises the following steps: C1: After variational modal decomposition, Winer filtering is used for denoising; C2: Setting the limited bandwidth parameter and the center angle frequency initialization to obtain the estimated center angle frequency of each; C3: According to different center angle frequencies, the IMF principal component energy spectrum is obtained, and this signal is taken as the input of the next step.

4. The method according to claim 1 or 3, wherein the method is characterized by, The IMF principal component energy spectrum is each intrinsic mode function.

5. The method of claim 1, wherein the polishing process of the silicon carbide wafer is monitored on-line. Step E comprises the following steps: E1: Analyzing the experimental data characteristic index in step D; E2: Selecting a double exponential model as its degradation model for fitting training, and determining the initial value of the model parameter of the observation equation through the data fitting result; E3: Selecting the training data fitted in step E2, and substituting the data into the model for parameter initialization.

6. The method of claim 5, wherein the polishing process of the silicon carbide wafer is monitored on-line. Step F comprises the following steps: F1: Real-time monitoring data are used to update the model parameters and the new characteristic index in real time; F2: The model parameter distribution is updated and adjusted.

7. The method of claim 6, wherein the step of monitoring the polishing process of the silicon carbide wafer is performed by measuring the thickness of the silicon carbide wafer. It further comprises step F3: iterative calculation is performed through the new characteristic index value.

Images