Wafer polishing process health state detection method, device, equipment and medium
By dividing the multidimensional time series of the wafer polishing process into subsequences and extracting global and local temporal features as well as cross-dimensional features, the problem of low computational efficiency in existing technologies is solved, and efficient and accurate health status detection is achieved.
Patent Information
- Application Number
- CN202511383747.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-25
- Publication Date
- 2025-11-14
AI Technical Summary
In existing technologies, the large amount of data generated during wafer polishing leads to low computational efficiency of deep learning models in feature extraction and classification, affecting the efficiency and accuracy of health status detection.
The multidimensional time series is divided into multiple subsequences, and the temporal features of each subsequence are extracted. The health status is determined by global temporal features, local temporal features and cross-dimensional features, which reduces computational complexity and improves detection efficiency.
By reducing the probability of parameter explosion during feature extraction, the efficiency and accuracy of health status detection in the wafer polishing process are improved, enabling accurate identification of abnormal states.
Smart Images
Figure CN120941284A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of wafer polishing technology, and in particular to a method, apparatus, equipment and medium for detecting the health status of wafer polishing processes. Background Technology
[0002] Maintaining the health of the wafer polishing process is crucial for ensuring its quality and stability. Monitoring the health of the wafer polishing process allows for the timely identification of normal and abnormal states, enabling prompt action to address any issues.
[0003] In related technologies, the main approach is to utilize process parameters, sensor data, and equipment operation data involved in wafer polishing to form a multidimensional time series. By extracting features from this multidimensional time series for classification, the health status of the wafer polishing process can be identified and detected. Typically, deep learning models are used to extract the temporal features of the multidimensional time series. However, the large amount of data involved in wafer polishing results in a high number of parameters and low computational efficiency for deep learning models when extracting and classifying features from the multidimensional time series, thus affecting the efficiency of detecting the health status of the wafer polishing process. Summary of the Invention
[0004] This application provides a method, apparatus, electronic device, and storage medium for detecting the health status of a wafer polishing process, in order to improve the detection efficiency and accuracy of the health status of the wafer polishing process.
[0005] In a first aspect, embodiments of this application provide a method for detecting the health status of a wafer polishing process, including: Acquire process data during the wafer polishing process; the process data includes process parameter data, sensor data, and equipment operation data; Based on process data, a multidimensional time series of the wafer polishing process is established. The multidimensional time series is divided into multiple subsequences; where the dimension of each subsequence is the same as the dimension of the multidimensional time series. The temporal features of each subsequence are extracted separately, and the global and local temporal features of the multidimensional time series are determined based on the temporal features of each subsequence. Extract cross-dimensional features from multidimensional time series; Based on global timing features, local timing features, and cross-dimensional features, the health status of the wafer polishing process is determined.
[0006] In one possible implementation, the temporal features of each subsequence are extracted separately, including: A first preset vector is embedded before the first column of each subsequence; wherein, the first preset vector is a random variable with a length of 1 and a dimension that is the same as the dimension of the multidimensional time series; The subsequences embedded with the first preset vector are respectively input into the preset first encoding module to obtain the first encoding sequence of each subsequence output by the first encoding module; wherein, the first encoding module is trained based on different subsequences embedded with the first preset vector and the corresponding first encoding sequences; Extract the vector corresponding to the first column in each first encoded sequence to obtain the temporal features of each subsequence.
[0007] In one possible implementation, the global temporal features of the multidimensional time series are determined based on the temporal features of each subsequence, including: According to the order of each subsequence in the multidimensional time series, the temporal features of each subsequence are concatenated to obtain the temporal feature sequence; A second preset vector is embedded before the first column of the time series feature sequence; wherein, the second preset vector is a random variable with a length of 1 and the same dimension as the dimension of the multidimensional time series; The temporal feature sequence embedded with the second preset vector is input into the preset second encoding module to obtain the second encoding sequence output by the second encoding module; wherein, the second encoding module is trained based on different temporal feature sequences embedded with the second preset vector and the corresponding second encoding sequences; Extract the vector corresponding to the first column in the second encoded sequence to obtain the global temporal features.
[0008] In one possible implementation, local temporal features of the multidimensional time series are determined based on the temporal characteristics of each subsequence, including: Based on the temporal characteristics of each subsequence, determine the feature weight corresponding to each temporal characteristic; Based on each time series feature and its corresponding feature weight, the local time series features of the multidimensional time series are determined.
[0009] In one possible implementation, the health status of the wafer polishing process is determined based on global timing features, local timing features, and cross-dimensional features, including: Global temporal features, local temporal features, and cross-dimensional features are concatenated to obtain a concatenated matrix; The splicing matrix is input into a preset classification module to obtain the health status of the wafer polishing process output by the classification module; the classification module is trained based on the splicing matrix and the corresponding health status under different health statuses.
[0010] In one possible implementation, before dividing the multidimensional time series into multiple subsequences, the following is also included: A linear transformation is performed on a multidimensional time series to obtain a transformed time series; wherein the dimension of the transformed time series is greater than or equal to the dimension of the multidimensional time series. Transform the time series as a new multidimensional time series.
[0011] In one possible implementation, the process data includes time series data from multiple dimensions; Based on process data, a multidimensional time series of the wafer polishing process was established, including: Noise processing and normalization are performed on the time series data of each dimension in the process data to obtain normalized data for each dimension. A multidimensional time series is obtained based on the normalized data of each dimension.
[0012] Secondly, embodiments of this application provide a wafer polishing process health status detection device, comprising: The acquisition module is used to acquire process data during the wafer polishing process; the process data includes at least one of process parameter data, sensor data, and equipment operation data. The module is used to build a multi-dimensional time series of the wafer polishing process based on process data; The partitioning module is used to divide a multidimensional time series into multiple subsequences; wherein the dimension of each subsequence is the same as the dimension of the multidimensional time series. The determination module is used to extract the temporal features of each subsequence separately, and determine the global and local temporal features of the multidimensional time series based on the temporal features of each subsequence; The extraction module is used to extract cross-dimensional features from multidimensional time series. The detection module is used to determine the health status of the wafer polishing process based on global time-series features, local time-series features, and cross-dimensional features.
[0013] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.
[0014] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.
[0015] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.
[0016] The beneficial effects of the embodiments in this application compared with the prior art are: This application embodiment acquires process data during the wafer polishing process, including at least one of process parameter data, sensor data, and equipment operation data. This allows for comprehensive acquisition of various data points during the wafer polishing process, facilitating monitoring of the process's health status. By using this process data, a multidimensional time series of the wafer polishing process is established, allowing for the combination of data points in chronological order to form structured data. Dividing the multidimensional time series into multiple subsequences reduces its length. Extracting the temporal features of each subsequence reduces the probability of parameter explosion during feature extraction. Based on the temporal features of each subsequence... This method identifies global and local temporal features of a multidimensional time series. Based on the temporal features of subsequences, it effectively extracts the temporal features of the multidimensional time series during the wafer polishing process, fully considering the different temporal features of the multidimensional time series, reducing computational complexity, and improving feature extraction efficiency. By extracting cross-dimensional features of the multidimensional time series, the correlation between various data points during the wafer polishing process can be obtained. Finally, through global, local, and cross-dimensional features, the health status of the wafer polishing process can be determined. This method fully considers the different spatial and temporal characteristics of various data points during the wafer polishing process during classification, improving the accuracy and efficiency of health status determination during the wafer polishing process. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is an application scenario diagram of the wafer polishing process health status detection method provided in the embodiments of this application; Figure 2 This is a schematic diagram of the wafer polishing process health status detection device provided in the embodiments of this application; Figure 3 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation
[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.
[0020] The inventors of this application have discovered that the wafer polishing process involves a large amount of data, resulting in long multidimensional time series. Consequently, when using deep learning models to detect the health status of the wafer polishing process, the large number of model parameters leads to low computational efficiency, affecting the efficiency of detecting the health status of the wafer polishing process.
[0021] Furthermore, the data obtained during the wafer polishing process exhibit dimensional heterogeneity, with different dimensions potentially displaying nonlinear or asynchronous correlations. Additionally, there are characteristics of class imbalance and similarity in time series patterns; time series within the same category show significant differences in form, while different categories may only exhibit subtle differences in local features.
[0022] When classifying and detecting the health status of wafer polishing processes using dynamic time warping expansion, the temporal morphological differences between different variables are aligned, and the minimum alignment path distance is calculated as a similarity measure. However, this method has high time complexity, is extremely slow for long sequences, and is sensitive to noise, making it difficult to quickly and accurately obtain the health status of the wafer polishing process. When classifying and detecting the health status of wafer polishing processes using random convolution kernel transformation, it is necessary to use random convolution kernels to transform the time series to extract various statistics of the transformed features (such as maximum, minimum, mean, and variance). However, it is weak in extracting long-range dependent features and struggles to accurately extract the global complete features of the wafer polishing process. When classifying and detecting the health status of wafer polishing processes using similarity prototype matching, it is necessary to find local features with class discriminative power in the time series formed by the wafer polishing process, performing an exhaustive search, which is time-consuming and has low classification computation efficiency.
[0023] To improve the efficiency and accuracy of health status detection in the wafer polishing process, this embodiment divides the multidimensional time series formed by various data points during the wafer polishing process into multiple subsequences. Temporal features are extracted from these subsequences, and then the extracted time series are used to further extract global and local temporal features of the wafer polishing process. This reduces the data length and eliminates the need to directly extract temporal features from the original multidimensional time series, thereby improving computational speed and feature extraction efficiency. Furthermore, the global and local temporal features are concatenated with the extracted cross-dimensional features of the wafer polishing process. The resulting concatenated matrix is used for classification, simultaneously considering the time-varying features of the wafer polishing process and the correlations between different types of data, thus accurately classifying and detecting the health status of the wafer polishing process.
[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.
[0025] See Figure 1 The document illustrates a flowchart of the wafer polishing process health status detection method provided in an embodiment of the present invention, which is described in detail below: Step 101: Obtain process data during the wafer polishing process; process data includes at least one of process parameter data, sensor data, and equipment operation data.
[0026] In this embodiment, appropriate process data can be selected for acquisition based on the health status of the wafer polishing process to be detected. For example, to detect the overall health status of the wafer polishing process, all process parameter data, sensor data, and equipment operation data can be acquired to comprehensively and accurately detect the health status of the wafer polishing process. Similarly, to detect the health status of the wafer in the wafer polishing process, wafer-related process parameter data can be acquired to determine whether defects exist on the wafer surface and their specific types, achieving accurate detection of the wafer's health status. Likewise, to detect the health status of the equipment in the wafer polishing process, equipment-related sensor data and equipment operation data can be acquired to detect whether the equipment is faulty and its specific fault type, achieving accurate detection of the equipment's health status.
[0027] The aforementioned process parameter data, sensor data, and equipment operation data are data from different aspects involved in the wafer polishing process, which together constitute the basic data of the entire chain of wafer polishing process health status.
[0028] Among them, the process parameter data refers to the data corresponding to the process parameters set during the wafer polishing process, which may include polishing head pressure data, stage pressure data, polishing head speed, wafer speed, slurry speed, and polishing time, etc.
[0029] The sensing data refers to the data from sensors at key locations in the chemical mechanical polishing equipment during the wafer polishing process. This data may include polishing pad temperature sensing data, spindle temperature sensing data, spindle vibration sensing data, and polishing table vibration sensing data.
[0030] The equipment operation data is the operating status data of the chemical mechanical polishing equipment, which can directly reflect the operating status of the equipment. It can include spindle motor current and voltage, motor power, equipment operating time, wafer adsorption vacuum degree, etc.
[0031] Step 102: Based on the process data, establish a multidimensional time series of the wafer polishing process.
[0032] Here, the scattered data are integrated into a structured multidimensional time series, which unifies the timestamps of the three types of data and ensures that different dimensions of data at the same point in time correspond to the same process time, so as to facilitate subsequent processing of the structured multidimensional time series.
[0033] Optionally, the process data includes time-series data in multiple dimensions. Specifically, process parameter data, sensor data, and equipment operation data each include time-series data in multiple dimensions, i.e., various types of data. To establish a multi-dimensional time series for the wafer polishing process based on the process data, one can first perform noise processing and normalization on the time-series data of each dimension of the process data to obtain normalized data for each dimension; then, based on the normalized data for each dimension, the multi-dimensional time series is obtained.
[0034] In this embodiment, time-series data for each dimension of the process data can be transformed into frequency-domain data. Noise frequencies are filtered out using a fixed-hertz high-pass filter, and then inversely transformed back into a time series, thus achieving filtering and noise reduction of the time-series data. The frequency-domain data conversion can be achieved using Fourier transform.
[0035] Since the units and distributions of time series data differ in each dimension, zero-mean normalization can be performed on the time series data for each dimension separately, setting the mean of all dimensions to 0 and the variance to 1. The specific formula is as follows: In the formula, Represents the first in time series data One data point, Represents the normalized result The corresponding value, express The mean of the time series data in which it is located. express The variance of the time series data.
[0036] Here, each type of data is treated as one dimension in a multidimensional time series, and a multidimensional time series is formed by aligning the process times of each type of data.
[0037] For example, suppose there is a total C This data has a time duration of [number]. L Then multidimensional time series X The dimension is C , length is L Specifically, it can be expressed as .
[0038] Step 103: Divide the multidimensional time series into multiple subsequences; wherein the dimension of each subsequence is the same as the dimension of the multidimensional time series.
[0039] In this embodiment, considering the low computational efficiency of directly using long sequences for feature extraction and state detection, the multidimensional time series is divided into multiple subsequences. The division of these subsequences is based on time length, while the dimension remains unchanged; that is, the dimension of the subsequences is the same as the dimension of the multidimensional time series.
[0040] Here, the lengths of the subsequences can be the same. Accordingly, a sliding window can be used to truncate the multidimensional time series. The dimensions of the sliding window are... C , length is l Step size is s ,in, s ≤ l < L For example, if the length of a multidimensional time series is 100, the sliding window length is 10, and the step size is 5, then the multidimensional time series can be divided into 19 subsequences. As another example, if the length of a multidimensional time series is 52, the sliding window length is 10, and the step size is 5, then the last sliding window is shorter than 10. The last 7 data points can be used as the last subsequence. This subsequence can be padded with zeros, or the step size of the last sliding window can be changed to ensure that the last sliding window contains 10 data points.
[0041] Step 104: Extract the temporal features of each subsequence, and determine the global and local temporal features of the multidimensional time series based on the temporal features of each subsequence.
[0042] In this embodiment, extracting the temporal features of each subsequence separately can shorten each subsequence to the corresponding temporal features, reduce data length and data volume, thereby ensuring that the computational complexity can be reduced during subsequent temporal feature extraction, thus improving computational efficiency.
[0043] In this context, the temporal features of each subsequence can be represented by a column vector, where the value corresponding to each dimension of the column vector represents the temporal features of the data in the corresponding dimension of the subsequence.
[0044] By extracting the temporal features of all subsequences, global temporal features of the multidimensional time series can be obtained, capturing the long-range temporal dependencies throughout the entire wafer polishing process. Since global temporal features are extracted from the temporal features of all subsequences, there is no need to utilize the original multidimensional time series, reducing the amount of data to be processed and thus improving the efficiency of global temporal feature extraction.
[0045] By extracting local temporal features from the temporal features of subsequences, we can focus on important features in the multidimensional time series and obtain more comprehensive and accurate information about the wafer polishing process.
[0046] Step 105: Extract cross-dimensional features of the multidimensional time series.
[0047] In this embodiment, in addition to extracting global and local time-series features of the multidimensional time series, cross-dimensional features of the multidimensional time series are also extracted. The time-series features can indicate the characteristics of how various parameters change over time during the wafer polishing process, while the cross-dimensional features can indicate the correlation between various parameters during the wafer polishing process.
[0048] Optionally, a feature extraction module consisting of convolutional layers and average pooling layers can be used to extract cross-dimensional features of multidimensional time series.
[0049] Here, by inputting a multidimensional time series into a convolutional layer, we can obtain the dimensionality-related feature map output by the convolutional layer. For example, the convolutional kernel of the convolutional layer can be... h ×1, stride of 1. Convolution kernel. h ×1 h This represents the size of the convolution window along the dimensional direction, with 1 representing the window size along the temporal direction. Convolution along the dimensional direction captures local correlations between adjacent or nearby dimensions.
[0050] The dimensionality-related feature maps output from the convolutional layers are then fed into the average pooling layer to compress their length. For example, performing average pooling in the time direction compresses L time steps in each dimension into a single value, ultimately yielding cross-dimensional features.
[0051] Step 106: Determine the health status of the wafer polishing process based on global timing features, local timing features, and cross-dimensional features.
[0052] In this embodiment, by using global time-series features, local time-series features, and cross-dimensional features, the characteristics of the changes of various data generated during the wafer polishing process over time and the correlation characteristics between various data can be obtained comprehensively and accurately. This allows for a comprehensive consideration of the wafer polishing process and accurate determination of the health status of the wafer polishing process.
[0053] Here, the overall health status of the wafer polishing process can include normal status, consumable abnormalities (such as abnormal polishing pad wear), process parameter abnormalities (such as sudden pressure changes), wafer quality abnormalities (such as wafer surface defects), and equipment malfunctions (such as polishing fluid blockage). Health status detection can be performed on wafer surface defects, including both normal and surface defect states, with surface defect states further encompassing various specific defect types. Similarly, the health status of the wafer polishing equipment can be detected, including both normal and fault states, with fault states further encompassing various specific fault types.
[0054] This application embodiment acquires process data during the wafer polishing process, including at least one of process parameter data, sensor data, and equipment operation data. This allows for comprehensive acquisition of various data points during the wafer polishing process, facilitating monitoring of the process's health status. By using this process data, a multidimensional time series of the wafer polishing process is established, allowing for the combination of data points in chronological order to form structured data. Dividing the multidimensional time series into multiple subsequences reduces its length. Extracting the temporal features of each subsequence reduces the probability of parameter explosion during feature extraction. Based on the temporal features of each subsequence... This method identifies global and local temporal features of a multidimensional time series. Based on the temporal features of subsequences, it effectively extracts the temporal features of the multidimensional time series during the wafer polishing process, fully considering the different temporal features of the multidimensional time series, reducing computational complexity, and improving feature extraction efficiency. By extracting cross-dimensional features of the multidimensional time series, the correlation between various data points during the wafer polishing process can be obtained. Finally, through global, local, and cross-dimensional features, the health status of the wafer polishing process can be determined. This method fully considers the different spatial and temporal characteristics of various data points during the wafer polishing process during classification, improving the accuracy and efficiency of health status determination during the wafer polishing process.
[0055] In some embodiments, extracting the temporal features of each subsequence can be achieved by: firstly embedding a first preset vector before the first column of each subsequence; wherein the first preset vector is a random variable with a length of 1 and a dimension the same as the dimension of the multidimensional time series; inputting the subsequences embedded with the first preset vector into a preset first encoding module to obtain the first encoding sequence of each subsequence output by the first encoding module; wherein the first encoding module is trained based on different subsequences embedded with the first preset vector and the corresponding first encoding sequences; and then extracting the vector corresponding to the first column in each first encoding sequence to obtain the temporal features of each subsequence.
[0056] In this embodiment, each subsequence is processed separately. A first preset vector is embedded before the first column of the subsequence, which can add a column of markers to the subsequence. The first preset vector is a learnable vector. After encoding by the first encoding module, the first encoded sequence corresponding to the subsequence can be obtained. The value of each dimension in the first column of the first encoded sequence, i.e., the column where the marker is located, can indicate the feature corresponding to the corresponding dimension in the subsequence.
[0057] The first preset vector can be a random variable of length 1 with the same dimension as the multidimensional time series. Taking the multidimensional time series as an example... C × L Taking a matrix as an example, the subsequence is of dimension 1. C , length is l The first preset vector is C A vertical vector of size ×1. The values in the first preset vector are randomly generated, following a normal distribution with a variance of 1. Correspondingly, the dimension of the subsequence embedded in the first preset vector is... C , length is l +1.
[0058] For example, the original first The subsequence is represented as Correspondingly, the subsequence embedded with the first preset vector can be represented as ,in, This represents the first preset vector for embedding.
[0059] Here, the first preset vector is used as the token to set the feature anchor point of the subsequence. After complex processing by the first encoding module, the token can integrate the temporal information of all time steps of the subsequence during the encoding process, reducing the probability of losing key information. Finally, by extracting the vector corresponding to the first column from the first encoded sequence, that is, the vector corresponding to the position where the first preset vector was previously embedded, the temporal features of each subsequence can be obtained. .
[0060] Alternatively, the first encoding module can be the encoder in a Transformer. The Transformer-Encoder captures the dependencies between different time steps within a sequence through a self-attention mechanism, and at the same time, it performs nonlinear transformations on the features of each time step through a feedforward neural network to enhance the feature representation capability. This enables it to capture long-range temporal dependencies more accurately, and it also has higher parallel computing efficiency.
[0061] In some embodiments, determining the global temporal features of a multidimensional time series based on the temporal features of each subsequence can be achieved by: firstly concatenating the temporal features of each subsequence according to their order in the multidimensional time series to obtain a temporal feature sequence; then embedding a second preset vector before the first column of the temporal feature sequence; wherein the second preset vector is a random variable with a length of 1 and the same dimension as the multidimensional time series; inputting the temporal feature sequence with the embedded second preset vector into a preset second encoding module to obtain a second encoding sequence output by the second encoding module; wherein the second encoding module is trained based on different temporal feature sequences with the embedded second preset vector and the corresponding second encoding sequences; finally, extracting the vector corresponding to the first column in the second encoding sequence to obtain the global temporal features.
[0062] In this embodiment, by splicing the temporal features of each subsequence according to the order of each subsequence in the multidimensional time series, the temporal sequence of each subsequence in the multidimensional time series can be preserved, and the features of the scattered subsequences can be integrated into a continuous feature sequence of the entire multidimensional time series. This avoids the loss of key long-range temporal information, so that the obtained temporal feature sequence can be used to accurately extract effective global temporal features in the future.
[0063] For example, the first The temporal characteristics of each subsequence can be Multidimensional time series The temporal feature sequence of each subsequence can be .
[0064] Here, a second preset vector is embedded before the first column of the time-series feature sequence, which adds another column of markers to the time-series feature sequence. After encoding by the second encoding module, the second encoded sequence corresponding to the time-series feature sequence can be obtained. The value of each dimension in the first column of the second encoded sequence, i.e., the column containing the markers, can indicate the feature corresponding to the corresponding dimension in the time-series feature sequence. Since the time-series feature sequence is composed of the time-series features of subsequences of a multi-dimensional time series, the value of each dimension in the column containing the markers can also indicate the feature corresponding to the corresponding dimension in the multi-dimensional time series, thus obtaining the global time-series features.
[0065] Here, the second preset vector is a learnable vector, which can be a random variable of length 1 with the same dimension as the multidimensional time series. For example, the multidimensional time series is... C × L The matrix, the second preset vector is C A vertical vector of ×1. The values in the second preset vector are randomly generated, following a normal distribution with a variance of 1. Divided into multidimensional time series... n Taking a subsequence as an example, the dimension of the temporal feature sequence embedded with the second preset vector is... C , length is n +1.
[0066] The second preset vector serves as the token, setting the feature anchor points of the temporal feature sequence. After complex processing by the second encoding module, this token can fuse the temporal information in the temporal features of each sub-sequence, accurately and comprehensively capturing key information in the wafer polishing process.
[0067] Finally, by extracting the vector corresponding to the first column from the second encoded sequence, that is, the vector corresponding to the position of the previously embedded second preset vector, the global time series features of the multidimensional time series can be obtained, which is also the global time series features in the wafer polishing process.
[0068] The second encoding module mentioned above can also be the encoder in the transformer, i.e., the Transformer-Encoder. Through its self-attention mechanism, it aggregates the information of all time steps in the temporal features of all subsequences, indirectly aggregating the information of all time steps in the original multidimensional time series. Thus, the vector corresponding to the first column in the second encoded sequence can completely and accurately represent the global temporal features of the original multidimensional time series.
[0069] In some embodiments, determining the local time series features of a multidimensional time series based on the time series features of each subsequence can be achieved by: first determining the feature weights corresponding to each time series feature based on the time series features of each subsequence; and then determining the local time series features of the multidimensional time series based on each time series feature and its corresponding feature weights.
[0070] In this embodiment, the feature weight corresponding to each time series feature is calculated based on the time series features of each subsequence. The contribution of each time series feature can be taken into account. The greater the contribution, the more important the local information contained in the subsequence is for the local state judgment of the multidimensional time series.
[0071] Among them, multidimensional time series are divided into n Taking a subsequence as an example, the formula for calculating the feature weight can be expressed as: In the formula, Indicates the first The feature weights corresponding to the temporal features of each subsequence. Indicates trainable n A vector of size 1, Indicates trainable n A matrix of size C, Indicates the first Temporal characteristics of individual subsequences Represents the transpose of a vector or matrix. This represents the hyperbolic tangent activation function.
[0072] Local time series features only consider whether the local feature appears, not the order in which the feature appears. Therefore, the formula for calculating local time series features can be expressed as: .
[0073] This involves scaling the temporal features of subsequences according to their feature weights. Higher feature weights amplify the corresponding temporal feature data, while lower feature weights weaken it. Then, by summing the calculation results for all subsequences, feature aggregation is achieved, yielding local temporal features. The aggregated result is dominated by the temporal features of the higher-weighted subsequences.
[0074] In traditional methods for processing long sequences, directly extracting local features from the entire sequence can easily dilute local anomalous features due to the abundance of data in normal time periods. This embodiment, however, proactively selects temporal features that contribute more to the subsequences through the aforementioned feature weighting method. Weighting these features amplifies them, reducing the probability of dilution and providing effective local features for subsequent classification, thereby improving classification accuracy.
[0075] In some embodiments, determining the health status of the wafer polishing process based on global temporal features, local temporal features, and cross-dimensional features can be achieved by: firstly concatenating the global temporal features, local temporal features, and cross-dimensional features to obtain a concatenation matrix; then inputting the concatenation matrix into a preset classification module to obtain the health status of the wafer polishing process output by the classification module; wherein the classification module is trained based on the concatenation matrix under different health statuses and the corresponding health status.
[0076] In this embodiment, by horizontally splicing global time series features, local time series features, and cross-dimensional features, three independent features can be formed into a splicing matrix containing a multi-dimensional time series, that is, the full information of the wafer polishing process.
[0077] Assuming that the global temporal features are 64×1 vectors, the local temporal features are 64×1 vectors, and the cross-dimensional features are 64×1 vectors, then the concatenation matrix is a 64×3 matrix.
[0078] By concatenating the three types of features, information from each type of feature can be preserved, ensuring that the information is not compressed. This allows for the subsequent acquisition of rich feature details based on the concatenation matrix, thereby improving classification accuracy.
[0079] Here, the classification module is used to classify the health status of the wafer polishing process. The classification module can directly output the health status of the wafer polishing process, or it can output the probability of each health status. Based on the health status with the highest probability, the health status of the wafer polishing process is obtained.
[0080] The classification module can consist of fully connected layers and softmax layers. Fully connected layers can map the concatenated matrix to the desired dimension. Furthermore, fully connected layers can also concatenate global temporal features, local temporal features, and cross-dimensional features to obtain the concatenated matrix.
[0081] The softmax layer transforms the input into a probability distribution, obtaining the probability for each class, and thus outputting the probability distribution for each class. Assume there are a total of... k Categories, namely k Given a certain health state, the probability distribution of its output can be expressed as: , Indicates the first The probability values for each health state are calculated. Finally, the health state corresponding to the highest probability value can be output.
[0082] Optionally, the wafer polishing process health status detection model includes a first encoding module, a second encoding module, a feature extraction module, and a classification module. By training the wafer polishing process health status detection model, the first encoding module, the second encoding module, the feature extraction module, and the classification module can be trained.
[0083] The wafer polishing process health status detection model can be simplified as follows: , The parameters to be learned in the wafer polishing process health status detection model, and the mapping relationship between the multidimensional time series and the output health status can be expressed as follows: The training can be performed using multi-dimensional time series corresponding to wafer polishing processes under different health states and their corresponding health states.
[0084] During training, it is necessary to ensure that the model's output values (i.e., the probabilities of each health state) are accurate. Compared to the label value (the actual probability of each health state). The probability distributions are as similar as possible, where It is the one-hot encoding of the tag. Assume there are a total of... For a given health state, the output value of the model corresponding to a sample can be represented as: The label value can represent... Here, the cross-entropy loss function can be used to evaluate the output value. With tag value Similarity between them: ; In the formula, Represents the cross-entropy loss function. Indicates the number of samples. The number of types of health status. Indicates the first The output value of the nth sample is in the nth The probability of a state. Indicates the first The label value of the nth sample is in the nth position. The probability of a certain state.
[0085] During training, the learning rate is adjusted using a linear warm-up technique. During the warm-up period, the learning rate is linearly increased from 0 to the initial preset learning rate, and then linearly decreased to the set minimum value. The optimizer can be the AdamW optimizer, combined with an L2 regularization term to minimize the parameters to be learned, effectively avoiding overfitting.
[0086] The process of updating the parameters to be learned can be as follows: ; ; In the formula, Indicates the parameters to be learned. Indicates L2 regularization terms. The coefficients of the L2 regularization term are represented. This represents the learning rate, a parameter that is used in each iteration. Update according to the formula above.
[0087] Set a maximum number of training iterations for the model. When the maximum number of training iterations is reached, the iteration terminates, and the optimal accuracy corresponds to the model's parameters. These are the optimal parameters, thus forming the final wafer polishing process health status detection model.
[0088] Alternatively, the first encoding module, the second encoding module, the feature extraction module, and the classification module can be trained separately.
[0089] The first encoding module is trained on the Transformer-Encoder using subsequences embedded with a first preset vector under different health states and their corresponding first encoding sequences. The second encoding module is trained on the Transformer-Encoder using subsequences embedded with a second preset vector under different health states and their corresponding second encoding sequences. The feature extraction module is trained on convolutional and average pooling layers using multidimensional time series data under different health states and their corresponding cross-dimensional features. The classification module is trained based on the concatenation matrix under different health states and their corresponding health states.
[0090] In some embodiments, before dividing the multidimensional time series into multiple subsequences, a linear transformation can be performed on the multidimensional time series to obtain a transformed time series; wherein the dimension of the transformed time series is greater than or equal to the dimension of the multidimensional time series; and the transformed time series is used as a new multidimensional time series.
[0091] In this embodiment, considering that cross-dimensional features require the extraction of dependencies between dimensions, the multi-dimensional time series can be linearly transformed to allow sufficient cross-interaction between dimensions, ensuring that cross-dimensional features can be accurately extracted.
[0092] Furthermore, the types of data acquired in wafer polishing processes are relatively limited, resulting in low-dimensionality, low-information, and insufficient inter-channel dependencies in multidimensional time series. Consequently, the performance of transformer-encoder architectures is generally less than ideal. Therefore, a linear transformation can be applied to multidimensional time series to increase their dimensionality, such as from dimension C to dimension D, where D ≥ C.
[0093] Here, multidimensional time series transformations can be achieved through fully connected layers. Each new feature is a linear combination of all the old dimensions, completing the initial fusion of cross-dimensional dependencies. The fully connected layer can be viewed as a 1×1 convolutional layer with independent time steps.
[0094] This application embodiment acquires process data during the wafer polishing process, including at least one of process parameter data, sensor data, and equipment operation data. This allows for comprehensive acquisition of various data points during the wafer polishing process, facilitating monitoring of the process's health status. By using this process data, a multidimensional time series of the wafer polishing process is established, allowing for the combination of data points in chronological order to form structured data. Dividing the multidimensional time series into multiple subsequences reduces its length. Extracting the temporal features of each subsequence reduces the probability of parameter explosion during feature extraction. Based on the temporal features of each subsequence... This method identifies global and local temporal features of a multidimensional time series. Based on the temporal features of subsequences, it effectively extracts the temporal features of the multidimensional time series during the wafer polishing process, fully considering the different temporal features of the multidimensional time series, reducing computational complexity, and improving feature extraction efficiency. By extracting cross-dimensional features of the multidimensional time series, the correlation between various data points during the wafer polishing process can be obtained. Finally, through global, local, and cross-dimensional features, the health status of the wafer polishing process can be determined. This method fully considers the different spatial and temporal characteristics of various data points during the wafer polishing process during classification, improving the accuracy and efficiency of health status determination during the wafer polishing process.
[0095] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0096] The following are device embodiments of this application. For details not described in detail, please refer to the corresponding method embodiments described above.
[0097] Figure 2 A schematic diagram of the wafer polishing process health status detection device provided in an embodiment of this application is shown. For ease of explanation, only the parts related to the embodiment of this application are shown, and are described in detail below: like Figure 2 As shown, the wafer polishing process health status detection device 20 includes: The acquisition module 21 is used to acquire process data during the wafer polishing process; the process data includes at least one of process parameter data, sensor data, and equipment operation data; Module 22 is used to build a multi-dimensional time series of the wafer polishing process based on process data; The partitioning module 23 is used to divide the multidimensional time series into multiple subsequences; wherein the dimension of each subsequence is the same as the dimension of the multidimensional time series. The determination module 24 is used to extract the temporal features of each subsequence, and determine the global and local temporal features of the multidimensional time series based on the temporal features of each subsequence. Extraction module 25 is used to extract cross-dimensional features of multidimensional time series; The detection module 26 is used to determine the health status of the wafer polishing process based on global timing features, local timing features and cross-dimensional features.
[0098] In one possible implementation, the determining module 24 is specifically used for: A first preset vector is embedded before the first column of each subsequence; wherein, the first preset vector is a random variable with a length of 1 and a dimension that is the same as the dimension of the multidimensional time series; The subsequences embedded with the first preset vector are respectively input into the preset first encoding module to obtain the first encoding sequence of each subsequence output by the first encoding module; wherein, the first encoding module is trained based on different subsequences embedded with the first preset vector and the corresponding first encoding sequences; Extract the vector corresponding to the first column in each first encoded sequence to obtain the temporal features of each subsequence.
[0099] In one possible implementation, the determining module 24 is specifically used for: According to the order of each subsequence in the multidimensional time series, the temporal features of each subsequence are concatenated to obtain the temporal feature sequence; A second preset vector is embedded before the first column of the time series feature sequence; wherein, the second preset vector is a random variable with a length of 1 and the same dimension as the dimension of the multidimensional time series; The temporal feature sequence embedded with the second preset vector is input into the preset second encoding module to obtain the second encoding sequence output by the second encoding module; wherein, the second encoding module is trained based on different temporal feature sequences embedded with the second preset vector and the corresponding second encoding sequences; Extract the vector corresponding to the first column in the second encoded sequence to obtain the global temporal features.
[0100] In one possible implementation, the determining module 24 is specifically used for: Based on the temporal characteristics of each subsequence, determine the feature weight corresponding to each temporal characteristic; Based on each time series feature and its corresponding feature weight, the local time series features of the multidimensional time series are determined.
[0101] In one possible implementation, the detection module 26 is specifically used for: Global temporal features, local temporal features, and cross-dimensional features are concatenated to obtain a concatenated matrix; The splicing matrix is input into a preset classification module to obtain the health status of the wafer polishing process output by the classification module; the classification module is trained based on the splicing matrix and the corresponding health status under different health statuses.
[0102] In one possible implementation, the building module 22 is also used for: A linear transformation is performed on a multidimensional time series to obtain a transformed time series; wherein the dimension of the transformed time series is greater than or equal to the dimension of the multidimensional time series. Transform the time series as a new multidimensional time series.
[0103] In one possible implementation, the process data includes time series data from multiple dimensions; Module 22 includes components for: Noise processing and normalization are performed on the time series data of each dimension in the process data to obtain normalized data for each dimension. A multidimensional time series is obtained based on the normalized data of each dimension.
[0104] Figure 3 This is a schematic diagram of the electronic device provided in an embodiment of this application. For example... Figure 3 As shown, the electronic device 30 of this embodiment includes a processor 31 and a memory 32. The memory 32 stores a computer program 33. When the processor 31 executes the computer program 33, it implements the steps in the various method embodiments described above. Alternatively, when the processor 31 executes the computer program 33, it implements the functions of each module / unit in the various device embodiments described above.
[0105] For example, computer program 33 may be divided into one or more modules / units, which are stored in memory 32 and executed by processor 31 to complete this application. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 33 in electronic device 30.
[0106] Electronic device 30 may include, but is not limited to, processor 31 and memory 32. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 30 and does not constitute a limitation on electronic device 30. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.
[0107] The processor 31 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.
[0108] The memory 32 can be an internal storage unit of the electronic device 30, such as a hard disk or RAM of the electronic device 30. The memory 32 can also be an external storage device of the electronic device 30, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 30. Furthermore, the memory 32 can include both internal and external storage units of the electronic device 30. The memory 32 is used to store computer programs and other programs and data required by the electronic device. The memory 32 can also be used to temporarily store data that has been output or will be output.
[0109] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.
[0110] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.
[0111] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.
[0112] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.
[0113] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.
[0114] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A method for detecting the health status of a wafer polishing process, characterized in that, include: Acquire process data during the wafer polishing process; The process data includes at least one of process parameter data, sensor data, and equipment operation data; Based on the process data, a multidimensional time series of the wafer polishing process is established. The multidimensional time series is divided into multiple subsequences; wherein the dimension of each subsequence is the same as the dimension of the multidimensional time series. The temporal features of each subsequence are extracted, and the global and local temporal features of the multidimensional time series are determined based on the temporal features of each subsequence. Extract cross-dimensional features from the multidimensional time series; The health status of the wafer polishing process is determined based on the global timing features, the local timing features, and the cross-dimensional features.
2. The method for detecting the health status of wafer polishing process according to claim 1, characterized in that, The extraction of temporal features for each subsequence includes: A first preset vector is embedded before the first column of each subsequence; wherein the first preset vector is a random variable with a length of 1 and a dimension that is the same as the dimension of the multidimensional time series; The subsequences embedded with the first preset vector are respectively input into the preset first encoding module to obtain the first encoding sequence of each subsequence output by the first encoding module; wherein, the first encoding module is trained based on different subsequences embedded with the first preset vector and the corresponding first encoding sequences; Extract the vector corresponding to the first column in each first encoded sequence to obtain the temporal features of each subsequence.
3. The method for detecting the health status of wafer polishing process according to claim 1, characterized in that, Based on the temporal characteristics of each subsequence, the global temporal characteristics of the multidimensional time series are determined, including: According to the order of each subsequence in the multidimensional time series, the temporal features of each subsequence are concatenated to obtain a temporal feature sequence; A second preset vector is embedded before the first column of the time-series feature sequence; wherein the second preset vector is a random variable with a length of 1 and a dimension that is the same as the dimension of the multidimensional time series; The temporal feature sequence embedded with the second preset vector is input into the preset second encoding module to obtain the second encoding sequence output by the second encoding module; wherein, the second encoding module is trained based on different temporal feature sequences embedded with the second preset vector and the corresponding second encoding sequences; Extract the vector corresponding to the first column in the second encoded sequence to obtain the global temporal features.
4. The method for detecting the health status of wafer polishing process according to claim 1, characterized in that, Based on the temporal characteristics of each subsequence, the local temporal characteristics of the multidimensional time series are determined, including: Based on the temporal characteristics of each subsequence, determine the feature weight corresponding to each temporal characteristic; Based on each time series feature and its corresponding feature weight, the local time series features of the multidimensional time series are determined.
5. The method for detecting the health status of wafer polishing process according to any one of claims 1 to 4, characterized in that, The determination of the health status of the wafer polishing process based on the global timing features, the local timing features, and the cross-dimensional features includes: The global temporal features, the local temporal features, and the cross-dimensional features are concatenated to obtain a concatenated matrix; The splicing matrix is input into a preset classification module to obtain the health status of the wafer polishing process output by the classification module; wherein, the classification module is trained based on the splicing matrix under different health statuses and the corresponding health status.
6. The method for detecting the health status of a wafer polishing process according to any one of claims 1 to 4, characterized in that, Before dividing the multidimensional time series into multiple subsequences, the method further includes: A linear transformation is performed on the multidimensional time series to obtain a transformed time series; wherein the dimension of the transformed time series is greater than or equal to the dimension of the multidimensional time series. The transformed time series is used as a new multidimensional time series.
7. The method for detecting the health status of a wafer polishing process according to any one of claims 1 to 4, characterized in that, The process data includes time series data in multiple dimensions; The step of establishing a multidimensional time series of the wafer polishing process based on the process data includes: The time series data of each dimension in the process data are subjected to noise processing and normalization processing respectively to obtain normalized data of each dimension; The multidimensional time series is obtained based on the normalized data of each dimension.
8. A device for detecting the health status of a wafer polishing process, characterized in that, include: The acquisition module is used to acquire process data during the wafer polishing process; The process data includes at least one of process parameter data, sensor data, and equipment operation data; The construction module is used to establish a multi-dimensional time series of the wafer polishing process based on the process data; A partitioning module is used to divide the multidimensional time series into multiple subsequences; wherein the dimension of each subsequence is the same as the dimension of the multidimensional time series. The determination module is used to extract the temporal features of each subsequence separately, and determine the global temporal features and local temporal features of the multidimensional time series based on the temporal features of each subsequence; An extraction module is used to extract cross-dimensional features of the multidimensional time series; The detection module is used to determine the health status of the wafer polishing process based on the global timing features, the local timing features, and the cross-dimensional features.
9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.