Model training methods, equipment monitoring methods, monitoring models and electronic equipment

Abstract

The exemplary embodiments of this disclosure relate to model training methods, equipment monitoring methods, monitoring models, and electronic devices. The method includes: extracting overall trend features and frequency domain distribution features from training data, where the training data includes time-series data related to wafer processes; generating fused features based on the overall trend features and frequency domain distribution features; encoding the training data to generate coded data; and decoding the coded data based on the fused features to generate prediction data. The embodiments of this disclosure can improve the efficiency of equipment monitoring, reduce false alarms and missed fault reports, thereby increasing the yield of products manufactured by the equipment, reducing material waste, and lowering production costs.

Description

Technical Field

[0001] The embodiments of this disclosure primarily relate to integrated circuits, and more specifically, to model training methods, device monitoring methods, monitoring models, and electronic devices. Background Technology

[0002] Manufacturing semiconductor devices (e.g., chips) often involves hundreds of process steps. In each process step, numerous sensors monitor the production status of the equipment (e.g., the machine) used to produce semiconductor devices to prevent equipment operating data from drifting out of the acceptable operating window due to some malfunction, thereby producing defective products and wasting materials.

[0003] Existing fault monitoring and classification methods require system manufacturers and equipment engineers to set specification lines based on experience for equipment operating data from sensors, and then monitor whether the equipment's operating status is abnormal by determining whether the operating data exceeds the specification lines (i.e., determining whether the monitoring data drifts out of the acceptable window). Summary of the Invention

[0004] According to exemplary embodiments of this disclosure, a monitoring scheme is provided to at least partially overcome the above or other potential defects.

[0005] In a first aspect of this disclosure, a model training method is provided. The method includes: extracting overall trend features and frequency domain dimensionality distribution features from training data, the training data including time-series data related to wafer process technology; generating fused features based on the overall trend features and frequency domain dimensionality distribution features; encoding the training data to generate encoded data; and decoding the encoded data based on the fused features to generate prediction data.

[0006] In some embodiments, extracting the overall trend features and frequency domain dimension distribution features of the training data includes: determining the statistical parameter values ​​of a predetermined type of training data; and generating statistical feature vectors based on each statistical parameter value as overall trend features.

[0007] In some embodiments, extracting the frequency domain dimension distribution features of training data includes: converting the training data into a frequency domain complex sequence; determining relevant spectral features based on the complex sequence; and generating frequency domain dimension distribution features based on each spectral feature.

[0008] In some embodiments, generating fused features based on overall trend features and frequency domain dimensional distribution features includes: stacking and merging the overall trend features and frequency domain dimensional distribution features of the training data using deep multi-mode feature fusion to obtain the fused features.

[0009] In some embodiments, iteratively training the model based on training data and prediction data to generate a monitoring model with optimal model parameters includes: optimizing the model parameters of the monitoring model through multiple iterations using a stochastic gradient descent optimization algorithm.

[0010] In some embodiments, optimizing the model parameters of the monitoring model using the stochastic gradient descent optimization algorithm through multiple iterations includes: calculating the reconstruction error between the training data and the prediction data of the monitoring model in each iteration; and determining the corresponding model parameters as the optimal model parameters in response to the reconstruction error being lower than a predetermined threshold.

[0011] In some embodiments, another portion of the timing data related to the wafer process includes verification run data, which is divided into time intervals with the training data. The method further includes: providing the verification run data to a trained monitoring model to output verification prediction data; and verifying the accuracy of the trained monitoring model based on a comparison of the verification run data and the verification prediction data.

[0012] In some embodiments, the model training method further includes standardizing a portion of the running data to generate standardized data with a consistent data structure, which is then used as input to the monitoring model.

[0013] In some embodiments, the running data comes from multiple sub-running processes during device operation, and the standardization of the running data to generate standardized data with a consistent data structure includes: performing time-series alignment processing and missing value filling processing on the running data so that the data points of the running data are aligned based on time points, wherein the number of time-series aligned data points in each sub-running process is consistent and there are no missing values.

[0014] In some embodiments, the runtime data includes sub-time series data corresponding to each sub-process, wherein performing time-series alignment and missing value filling on the runtime data includes: calculating the average of all data points for the sub-time series data of missing data points in the sub-processes of multiple sub-running processes; and using the average to fill in the missing data points.

[0015] In some embodiments, encoding training data to generate encoded data includes: extracting features from the training data using an encoder to generate a compressed feature vector.

[0016] In some embodiments, decoding encoded data based on fused features to generate prediction data includes: concatenating the fused features with a compressed feature vector to generate concatenated features; and reconstructing the concatenated features by a decoder to generate prediction data.

[0017] In some embodiments, the method further includes: iteratively training the model based on the training data and the prediction data to generate a monitoring model with optimal model parameters.

[0018] In a second aspect of this disclosure, an equipment monitoring method is provided. The method includes: generating predictive data from a monitoring model based on collected equipment operating data, wherein the operating data includes time-series data related to wafer process collected at specific time points; and determining the operating status of the equipment based on the difference between the operating data and the predictive data.

[0019] In some embodiments, the monitoring model is trained according to the method described in the first aspect.

[0020] In some embodiments, determining the operating status of a device based on the difference between operational data and predicted data includes: calculating a standard score corresponding to each time point based on the predicted data and operational data; comparing the calculated standard score with a predetermined threshold; and determining the operating status of the device based on the comparison result.

[0021] In some embodiments, calculating the standard score corresponding to each time point based on the predicted data and the operational data includes: calculating the root mean square deviation corresponding to each time point based on the predicted data and the operational data to obtain reconstruction error data; calculating the mean and standard deviation corresponding to each time point based on the reconstruction error data; and calculating the standard score based on the mean and standard deviation.

[0022] In some embodiments, the equipment monitoring method further includes: standardizing the operating data to generate standardized data with a consistent data structure, which is then used as input to the monitoring model.

[0023] In some embodiments, the running data comes from multiple sub-running processes during device operation, and the standardization of the running data to generate standardized data with a consistent data structure includes: performing time-series alignment processing and missing value filling processing on the running data so that the data points of the running data are aligned based on time points, wherein the number of data points in each sub-running process is consistent and does not include missing values.

[0024] In some embodiments, the runtime data includes sub-running data corresponding to each sub-process, wherein performing time-series alignment and missing value filling on the runtime data includes: in the sub-processes of multiple sub-running processes, calculating the average of all data points for the sub-running data with missing data points; and using the average to fill in the missing data points.

[0025] In a third aspect of this disclosure, a monitoring model trained according to the method of the first aspect of this disclosure is provided.

[0026] In a fourth aspect of this disclosure, an electronic device is provided. The electronic device includes a processor; and a memory coupled to the processor, the memory having instructions stored therein, which, when executed by the processor, cause the device to perform the methods according to the first and second aspects of this disclosure.

[0027] In a fifth aspect of this disclosure, a computer-readable storage medium is provided. A computer program is stored on this computer-readable storage medium, which, when executed by a processor, implements the methods of the first and second aspects of this disclosure. In a sixth aspect of this disclosure, a prediction model is provided. The prediction module includes: a statistical feature extraction module configured to extract overall trend features of input data, the input data including time-series data related to wafer process technology; a frequency domain feature extraction module configured to extract frequency domain dimensionality distribution features of the input data; a fusion module configured to generate fused features based on the overall trend features and the frequency domain dimensionality distribution features; an encoder configured to encode the input data to generate encoded data; and a decoder configured to decode the encoded data based on the fused features to generate the prediction data. In some embodiments, the model further includes a judgment module configured to determine the operating state of the device based on the difference between the input data and the predicted data, wherein the process-related timing data of the wafer is the operating data of the device.

[0028] The solution of this disclosure, which replaces the equipment engineer in setting specification lines, can improve the efficiency and accuracy of equipment monitoring, thereby increasing the yield of the produced semiconductor devices, reducing material waste, and lowering production costs.

[0029] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0030] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0031] Figure 1 This diagram illustrates equipment operating data with specification lines set by engineers in a conventional monitoring method.

[0032] Figure 2 A schematic diagram illustrating an example environment in which embodiments of the present disclosure can be implemented is shown;

[0033] Figure 3A flowchart of a model training method according to some embodiments of the present disclosure is shown;

[0034] Figure 4 A flowchart of a model training method according to some other embodiments of the present disclosure is shown;

[0035] Figure 5 A schematic flowchart of a device monitoring method according to some embodiments of the present disclosure is shown;

[0036] Figure 6 A schematic flowchart of a device monitoring method according to some embodiments of the present disclosure is shown;

[0037] Figure 7 A flowchart illustrating device operating data with predetermined thresholds set according to some embodiments of the present disclosure is shown;

[0038] Figure 8 A block diagram of a computing device capable of implementing several embodiments of the present disclosure is shown. Detailed Implementation

[0039] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0040] In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0041] Fault Detection and Classification (FDC) is a technology used for real-time monitoring and analysis of equipment operating data during production. FDC systems generate massive amounts of data, with each device and sensor collecting data at rates ranging from 1 to 10 values ​​per second. How to effectively utilize this vast amount of data to accurately monitor and alarm when equipment malfunctions has been a continuous focus of the industry.

[0042] As mentioned earlier, to monitor the status of semiconductor manufacturing equipment, FDC system manufacturers or equipment engineers need to set specification limits based on experience. The monitoring method involves determining whether operating data exceeds these limits to monitor for abnormal equipment status. This method also includes calculating statistical values ​​of the operating data, such as minimum, maximum, average, and standard deviation, before setting the specification limits. Furthermore, engineers set specification limits for critical steps. Because statistical values ​​cannot reflect the real-time status of the equipment, it cannot be determined whether the equipment is operating normally even if the statistical values ​​are within the specification limits. Therefore, the manual setting of specification limits by engineers consumes significant manpower and time. Additionally, since the specification limits are fixed while the operating data changes over time, this monitoring method cannot detect anomalies within the specification limits that do not conform to the over-time trend of the operating data.

[0043] Figure 1 This diagram illustrates equipment operating data where specifications have been set by engineers using conventional monitoring methods. (For example...) Figure 1 As shown, the solid lines 102 represent statistical values ​​of multiple operating data points when the equipment performs the same process repeatedly. The vertical axis represents the values ​​collected by sensors; a single machine may have dozens or even hundreds of sensors, each with different physical meanings. These sensors can be, for example, pressure sensors, temperature sensors, humidity sensors, etc. The dashed lines 104 show the specification lines manually set by engineers corresponding to key steps (e.g., S3-1 to S3-13). From... Figure 1 As can be seen, in this critical step, one operating data curve, 106, deviates significantly from the other operating data, but this curve does not exceed the set specification line. Therefore, this operating data curve actually indicates that the equipment is in an abnormal state, but the existing monitoring methods cannot detect this abnormal state. This will lead to lower monitoring reliability and accuracy, resulting in a decrease in the yield of semiconductor devices and an increase in production costs due to the inability to detect equipment abnormalities in a timely manner.

[0044] In view of this, this disclosure provides an improved solution.

[0045] According to some embodiments of this disclosure, a model training method is provided. The method includes: extracting overall trend features and frequency domain dimensional distribution features from training data, the training data including time-series data related to wafer fabrication processes; generating fused features based on the overall trend features and frequency domain dimensional distribution features; encoding the training data to generate encoded data; and decoding the encoded data based on the fused features to generate prediction data.

[0046] Other embodiments of this disclosure also provide a device monitoring method, which includes: generating predictive data from a monitoring model based on collected device operating data, wherein the operating data includes multiple data points collected at different time points; and determining the device's operating status based on the difference between the operating data and the predicted data. This monitoring method eliminates the need for engineers to manually set specification lines and considers both the overall trend information and local abrupt changes in the operating data, thereby overcoming the problems of low efficiency, low reliability, and low accuracy of known solutions.

[0047] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings. First, refer to... Figure 2 , Figure 2 A schematic diagram of an example environment 200 that can be implemented according to embodiments of the present disclosure is shown. Figure 2 As shown, the example environment 200 includes a computing device 210 and a client 220.

[0048] In some embodiments, the computing device 210 can interact with the client 220. For example, the computing device 210 can receive input messages from the client 220 and output feedback messages to the client 220. In some embodiments, the input messages from the client 220 can be device operating data. The computing device 210 can perform corresponding mathematical operations on the operating data and output the corresponding operation results to the client 220.

[0049] In some embodiments, the computing device 210 may include, but is not limited to, a personal computer, a server computer, a handheld or laptop device, a mobile device (such as a mobile phone, a personal digital assistant PDA, a media player, etc.), a consumer electronics product, a minicomputer, a mainframe computer, cloud computing resources, etc.

[0050] It should be understood that the description of the structure and functionality of example environment 200 for illustrative purposes only is not intended to limit the scope of the subject matter described herein. The subject matter described herein may be implemented in different structures and / or functionalities. This environment is merely illustrative and is not intended to limit the application environment of the embodiments of this disclosure.

[0051] To more clearly explain the principles of this disclosure, reference will be made below. Figures 3 to 8 Let's describe it in more detail.

[0052] Figure 3 A flowchart of a model training method 300 according to some embodiments of the present disclosure is shown.

[0053] At box 302, the overall trend features and local trend features of the training data are extracted, and the training data includes the device's operating data.

[0054] In some embodiments, training data may be or include at least a portion of the collected operational data of the device. Training data may be currently collected data or previously collected operational data stored in memory; this disclosure does not impose any limitations on this.

[0055] In some embodiments, the device's operational data includes data from the device over a time period, meaning the operational data is time-series. For example, in some embodiments, the training data can be time-series data of wafer samples, mainly including control parameters set during wafer fabrication, such as radio frequency power, chamber pressure, gas flow rate, and gas velocity.

[0056] In some embodiments, operational data includes multiple data points collected at specific time points. Operational data can indicate data describing the operating status of the equipment throughout the entire process of manufacturing a semiconductor device, such as temperature and pressure. In some embodiments, training data used to train a monitoring model should be a type of operational data. For example, the training data might correspond to the same equipment (e.g., a chamber of a machine tool), the same physical quantity, and the same process. In some embodiments, the equipment may be a machine tool used to manufacture chips. In some embodiments, the multiple data points collected at specific time points may be data collected by sensors at fixed sampling time intervals.

[0057] For example, engineers can filter the database to select the machine, chamber, sensor name, and time range information to be analyzed based on their analytical needs. For instance, they might need to analyze data for site 1, machine equipment A, chamber equipment A_chamber B, recipe C, process time period 2024-01-01 to 2024-06-01, parameter name sensor D. The database can be used to retrieve parameter data collected within that chamber during that time period. This data records each wafer entering chamber equipment A_chamber B of machine equipment A for processing, using recipe C during processing, with a sensor sampling interval of 1 second per reading. Assuming a processing time of 1 minute, this would provide data from 60 time points for sensor sensor D.

[0058] In some embodiments, the collected data can be divided into two parts according to time intervals: one part serves as model training and validation data for model parameter tuning; the other part serves as a model test set for testing the model's recognition performance. For example, data from January to April 2024 can be used as the training and validation set, and data from May 2024 as the test set. It should be understood that the embodiments of this disclosure are not limited thereto, and the collected data can be processed in other ways as needed.

[0059] The models trained by some embodiments of this disclosure can encode and decode runtime data to obtain the overall trend of runtime data over time and information on local mutations.

[0060] At box 304, iterative training is performed based on the overall change trend features and the local change trend features to generate a monitoring model with the optimal fusion coefficient.

[0061] In some embodiments, iterative training based on overall and local trend features may include integrating the overall and local trend features of the training data to obtain training data with integrated features. For example, multi-modal prediction fusion may be used to stack and merge the overall and local trend features of the training data to obtain training data with integrated features. Feature stacking and merging are mentioned in this embodiment; however, it should be understood that the embodiments of this disclosure are not limited thereto, and various variations can be made as needed.

[0062] In some embodiments, the outputs of the two aforementioned modules can be fed into the multi-mode prediction fusion module. The optimization objective can be set to minimize the reconstruction error of the training samples. The reconstruction error is the mean squared error calculated for each sample point of the input sequence and the predicted output sequence. The model parameters can be optimized using a stochastic gradient descent (SGD) optimization algorithm through multiple iterations. The parameter tuning process during model training can be implemented using existing neural network training frameworks, such as PyTorch, and it should be understood that this disclosure is not limited thereto.

[0063] By fusing global and local perception features, we can better handle complex anomaly recognition tasks such as sudden shifts, gradual drifts, and periodic fluctuations.

[0064] The fused features are integrated features that exhibit both the overall and local trends of the training data. As mentioned earlier, the relevant parameters mentioned in the examples can be varied as needed, for example, manually specified as required. This disclosure does not impose any limitations on this.

[0065] The fused feature data can be viewed as a predicted sequence obtained by encoding and decoding the original input data. The model can learn the data patterns of normal samples and the correlation information between the beginning and end of the sequence from the training dataset, and can perform sequence prediction and reconstruction very well.

[0066] In some embodiments, iterative training based on overall and local trend features to generate a monitoring model with optimal model parameters (also known as optimal fusion coefficients) may include: optimizing the model parameters of the monitoring model through multiple iterations using a stochastic gradient descent optimization algorithm.

[0067] In some embodiments, optimizing the model parameters of the monitoring model using the Stochastic Gradient Descent (SGD) optimization algorithm through multiple iterations may include: calculating the reconstruction error between the training data and the prediction data of the monitoring model in each iteration; and determining the corresponding model parameters as the optimal fusion coefficients in response to the reconstruction error being lower than a predetermined threshold.

[0068] In some embodiments, stochastic gradient descent optimization can be implemented using an SGD model. In some embodiments, the number of iterations can be set to 30, the batch size to 16, and the initial learning rate to 0.001. In embodiments of this disclosure, the above parameters can be manually specified as needed. This disclosure does not impose any limitations on this.

[0069] In some embodiments, calculating the reconstruction error between the training data and the predicted data of the monitoring model in each iteration may include: calculating the mean squared error for each sample point of the training data and the predicted data. That is, calculating the mean squared error for each sample point separately.

[0070] During training, the model continuously outputs predicted sequences. Through iterative training, the model's parameters are continuously optimized until the optimal parameters are obtained, at which point the model training is complete.

[0071] In some embodiments, another portion of the operational data includes validation operational data, which is divided into time intervals with the training data. The model training method 300 further includes: providing the validation operational data to the trained monitoring model to output validation prediction data; and verifying the accuracy of the trained monitoring model based on a comparison of the validation operational data and the validation prediction data.

[0072] In some embodiments, the model training method 300 may further include: standardizing a portion of the running data to generate standardized data with a consistent data structure, which is then used as input to the monitoring model. This prevents data point shifts corresponding to the same time point, prevents model training failure, and improves the accuracy of the trained model.

[0073] In some embodiments, the operational data originates from multiple sub-operational processes during the operation of the device. Data points collected by different sensors may have different formats; in such cases, these data points can be standardized. Standardizing the operational data to generate standardized data with a consistent data structure may include: performing time-series alignment and missing value filling on the operational data, so that the data points are aligned based on time points. After processing, the number of data points in each sub-operational process is consistent and there are no missing values.

[0074] In some embodiments, runtime data includes sub-run data corresponding to each sub-process. Timing alignment and missing value filling of the runtime data may include: calculating the average of all data points for the missing data points in the sub-run data of multiple sub-run processes; and using the average to fill in the missing data points.

[0075] Figure 4 A flowchart of a model training method according to some other embodiments of the present disclosure is shown.

[0076] At box 402, extract the overall trend features and frequency domain dimension distribution features of the training data. The training data includes time-series data related to the wafer process.

[0077] In combination Figure 4 In the described embodiments, the training data may be the same as the training data described in the foregoing embodiments, and will not be described in detail here.

[0078] In some embodiments, extracting the overall trend features and frequency domain distribution features of the training data may include: determining statistical parameter values ​​of a predetermined type for the training data, such as the mean, standard deviation, maximum value, minimum value, range, bias, first-order difference standard deviation, energy, etc.; and generating statistical feature vectors based on each statistical parameter value as overall trend features. Statistical feature extraction can effectively characterize the overall trend, shape, and volatility of a sequence, resulting in a feature expression describing the stability of the curve trend.

[0079] In some embodiments, extracting the frequency domain dimension distribution features of the training data may include: converting the training data into a frequency domain complex sequence, for example, by using a fast Fourier transform; determining relevant spectral features based on the complex sequence, which may include, as mentioned later, the dominant frequency amplitude, dominant frequency, spectral bandwidth, etc.; and then generating frequency domain dimension distribution features based on each spectral feature.

[0080] In some embodiments, mathematical statistical calculations can be used to extract statistical features, thereby deriving the overall trend information of the curve. The overall trend features can characterize the overall trend, shape, and volatility of the data.

[0081] Frequency domain distribution features can characterize the distribution pattern of data in the frequency domain. Frequency domain feature extraction can reveal the distribution pattern of data in the frequency domain. Specifically, this includes two aspects: First, certain anomalies are often accompanied by a shift in the main frequency. For example, when equipment experiences mechanical failure, the spectrum changes from a concentrated low-frequency energy distribution to a discrete high-frequency energy distribution. Second, some semiconductor manufacturing processes may exhibit amplitude jitter or slight phase shifts in signal amplitude, but this is within the normal production process. Frequency domain features are phase-insensitive, effectively improving the problem of false alarms caused by start-time shifts in wafer processing. For example, considering the characteristics of FDC data, signals such as pressure, RF, temperature, and valve switching signals have periodic signal components. Abnormal wafers typically exhibit increased high-frequency energy or abnormal energy shifts in specific frequency bands; therefore, frequency domain features help identify periodic anomalies or frequency energy anomalies.

[0082] In some embodiments, the input data can be fed into the statistical feature extraction module and the frequency domain feature extraction module respectively for extraction. This will be further described below.

[0083] In some embodiments, the training data can be aligned as described in the foregoing embodiments, and then the aligned data can be input into the model. The time-series data of a single wafer sample is denoted as... Where L is the aligned sequence length and 1 represents the number of sensors (taking a single sensor as an example). Represents a real number.

[0084] Can be Perform calculations to generate a set of vectors based on statistical features. The following parameters are statistically analyzed: mean, standard deviation, maximum value, minimum value, range, deviation, first-order difference standard deviation, and energy.

[0085] For an input sequence L=100 The calculation process is shown below, and the final statistical feature vector is:

[0086] .in =8 represents the value 8. This represents an 8-row, 1-column real matrix.

[0087] The calculation formulas for each parameter are as follows:

[0088] (1) Mean ;

[0089] (2) Standard deviation ;

[0090] (3) Maximum value ;

[0091] (4) Minimum value ;

[0092] (5) Range ;

[0093] (6) Skewness ;

[0094] (7) First-order difference standard deviation ;

[0095] (8) Signal energy

[0096] In some embodiments, frequency domain feature extraction can be performed as follows:

[0097] For an input sequence L=100 Performing a Fast Fourier Transform (FFT) on the sequence yields the complex result C. Since the input is a real signal, its spectrum is symmetrical. Here, we only need to take the first half of the single-sided spectrum to calculate its amplitude A, with a length of... ,in , For the real part of the complex number, For the imaginary part of a complex number, This is the frequency value corresponding to the k-th frequency component. Based on single-sided amplitude spectrum calculations: dominant frequency amplitude, dominant frequency, high-frequency energy, spectral centroid, spectral bandwidth, spectral flatness, and spectral entropy are calculated as follows. The final frequency domain feature vector is:

[0098] This represents the value 7.

[0099] (1) Main frequency amplitude ;

[0100] (2) Main frequency , representing the frequency value corresponding to the maximum frequency amplitude;

[0101] (3) High-frequency energy (calculate the frequency band after 30% Nyquist frequency) ; This indicates a spectrum index.

[0102] (4) Spectral centroid ;

[0103] (5) Spectrum bandwidth ;

[0104] (6) Spectral flatness ;

[0105] (7) Spectral Entropy ;

[0106] It should be noted that the above formula is general, and the parameters not specifically described have meanings known in the field.

[0107] For an original sequence of length L=100, performing an FFT on the sequence yields C, which is represented by 100 complex values. Due to the symmetry of the spectrum, only half of the spectrum needs to be studied, M=L / 2=50. The frequencies at these 50 points are Nyquist frequencies, which are composed of 50 frequencies in the spectrum. Calculate the spectral index Kh=50 corresponding to the Nyquist frequency at 30%. 30% = 15.

[0108] At box 404, a fused feature is generated based on the overall trend features and the frequency domain dimension distribution features.

[0109] In some embodiments, generating fused features based on overall trend features and frequency domain dimensional distribution features may include: stacking and merging the overall trend features and frequency domain dimensional distribution features of the training data using deep multi-modal feature fusion to obtain training data with integrated features. For example, the stacking and merging can be performed in the same manner as described above. It should be understood that the embodiments of this disclosure are not limited thereto, and other methods may be used to generate fused features as needed.

[0110] At box 406, the training data is encoded to generate encoded data.

[0111] In some embodiments, encoding training data to generate encoded data may include: extracting features from the training data using an encoder to generate a compressed feature vector.

[0112] The encoder can extract features from the original time-series samples and is good at capturing local dependency information between adjacent points within the sequence.

[0113] At box 408, the encoded data is decoded based on the fused features to generate the predicted data.

[0114] In some embodiments, decoding encoded data based on fused features to generate prediction data may include: concatenating the fused features with a compressed feature vector to generate concatenated features; and reconstructing the concatenated features by a decoder to generate prediction data.

[0115] Time-domain statistical characteristics can reflect the stability and macroscopic trends of a process. For example, changes in energy often foreshadow shifts in the plasma discharge state. Frequency-domain characteristics can characterize the periodic behavior and mechanical properties of equipment. In semiconductor equipment, RF power supply anomalies often first manifest as amplitude or spectral shifts at specific frequencies.

[0116] Adding perceptual data from different dimensions of statistical and spectral distributions to traditional time-series samples can significantly enhance the model's ability to identify complex process fluctuations and its robustness. In other words, decoding encoded data based on fused features offers significant technical advantages during the decoding process. Firstly, it increases constraints and improves signal reconstruction quality. In conventional encoder-decoder systems, the decoder relies solely on the encoder's compressed features for reconstruction, which can easily lead to detail loss due to over-compression. Adding time-frequency fusion information as a supervisory signal guides the decoder to ensure that the generated signal maintains a high degree of consistency with actual normal samples in both time-domain morphology and statistical characteristics, improving the accuracy and realism of the reconstruction. Secondly, it amplifies the difference between normal and abnormal samples. When anomalies occur, although the morphology of the original time-series signal may change slightly, its statistical properties or spectral distribution often shift significantly. The model is trained on normal samples, and for normal samples, the statistical-frequency fusion features and encoder features have a certain consistency and coupling. However, for abnormal samples, the encoder's time-domain signal and the statistical-frequency fusion signal conflict, leading to reconstruction failure and increasing the reconstruction error, thus facilitating the identification of abnormal data.

[0117] In some embodiments, the model is iteratively trained based on training data and prediction data to generate a monitoring model with optimal model parameters.

[0118] In some embodiments, iteratively training the model based on training data and prediction data to generate a monitoring model with optimal model parameters may include: optimizing the model parameters of the monitoring model through multiple iterations using a stochastic gradient descent optimization algorithm.

[0119] In some embodiments, optimizing the model parameters of the monitoring model using the stochastic gradient descent optimization algorithm through multiple iterations includes: calculating the reconstruction error between the training data and the prediction data of the monitoring model in each iteration; and determining the corresponding model parameters as the optimal model parameters in response to the reconstruction error being lower than a predetermined threshold.

[0120] In some embodiments, the process of iteratively training the model based on training data and prediction data can be similar to the previous process. Figure 3 The process described is the same, so it will not be detailed here.

[0121] The trained monitoring model can be used to analyze operational data to monitor the operating status of equipment. (This will be discussed later.) Figure 5 Describe it.

[0122] Figure 5 A flowchart of a device monitoring method 500 according to some embodiments of the present disclosure is shown.

[0123] At box 502, predictive data is generated by the monitoring model based on the collected operational data of the equipment, wherein the operational data includes multiple data points collected at different times.

[0124] In some embodiments, the monitoring model is trained according to the model training method provided in this disclosure.

[0125] At box 504, the operating status of the device is determined based on the difference between the operating data and the predicted data.

[0126] For anomalous samples that differ significantly from normal samples in the feature space, the model cannot effectively capture the sequence association information of the anomalous samples, resulting in larger sequence prediction errors. This characteristic can be used to determine whether a sample is normal.

[0127] In some embodiments, the model-predicted sequence and the original input sequence can be fed into the reconstruction error threshold determination module to calculate the difference between the two sequences, i.e., the mean squared error is calculated for each sample point. A standard score (z-score) is calculated for the reconstruction error of each time point of the sample, and anomalies are determined by applying a threshold.

[0128] In some embodiments, determining the operating status of a device based on the difference between operational data and predicted data may include: calculating a standard score for each time point based on the predicted data and operational data; comparing the calculated standard score with a predetermined threshold. This comparison may include comparing the absolute value of the standard score with the threshold. The operating status of the device is then determined based on the result of the comparison.

[0129] In some embodiments, if the standard score is greater than a predetermined threshold, it indicates an anomaly in the data, thereby determining that the device is in an abnormal operating state; if the standard score is less than or equal to the predetermined threshold, it indicates that the data is normal, thereby determining that the device is in a normal operating state. Alternatively, in other embodiments, if the standard score is greater than or equal to the predetermined threshold, it is determined that the device is in an abnormal operating state; if the standard score is less than the predetermined threshold, it is determined that the device is in a normal operating state. It is understood that those skilled in the art can set corresponding judgment criteria according to actual needs to determine the comparison result corresponding to the device's operating state. In some embodiments, the predetermined threshold can be between 0.5 and 1.5. Furthermore, those skilled in the art can set specific predetermined thresholds according to actual needs. The predetermined threshold reflects the tolerance for deviations in the operating data; the larger the predetermined threshold, the greater the tolerance and the less sensitive the monitoring; conversely, the smaller the predetermined threshold, the lower the tolerance and the more sensitive the monitoring. This disclosure does not limit this.

[0130] In some embodiments, calculating the standard score for each time point based on the predicted data and the operational data may include: calculating the root mean squared error for each time point based on the predicted data and the operational data to obtain reconstruction error data; calculating the mean and standard deviation for each time point based on the reconstruction error data; and calculating the standard score based on the mean and standard deviation.

[0131] In some embodiments, standard score The calculation can be performed using the following formula (1): (1) in, This represents the j-th data point in the i-th reconstruction error. This represents the mean of the j-th data point in each reconstruction error. This represents the overall standard deviation of the j-th data in each reconstruction error.

[0132] In some embodiments, for example, there are three training data sets, each with 5 data points. The model predicts the corresponding prediction data. Based on the training data and the prediction data, the corresponding reconstruction errors are calculated, for example, x1=[76, 70, 35, 77, 55], x2=[93, 61, 38, 64, 95], and x3=[56, 21, 46, 71, 88], i.e., i=3, j=5. The reconstruction error data x1, x2, and x3 can be used to calculate... , , , and The values ​​are 75, 50.67, 39.67, 70.67, and 79.33 respectively. , , , and The figures are 15.12, 21.29, 4.64, 5.31 and 17.44 respectively.

[0133] Calculate Z1=[0.066, 0.91, -1.01, 1.19, -1.395], Z2=[1.19, 0.49, -0.36, -1.25, 0.898], and Z3=[-1.26, -1.39, 1.36, 0.06, 0.497] using formula (1). Set a predetermined threshold of 1.3. Compare the above standard scores Z1, Z2, and Z3 with the predetermined threshold of 1.3. It can be found that among Z1, Z2, and Z3, Z... 15 Z 32 and Z 33The absolute value of the standard score is greater than 1.3. In this embodiment, this indicates that the corresponding running data X 15 X 32 and X 33 abnormal.

[0134] Those skilled in the art will understand that the amount of training data, reconstruction error, standard score, and threshold mentioned in the examples are illustrative. This disclosure does not impose any limitations in these regard.

[0135] In some embodiments, the runtime data originates from multiple sub-running processes during device operation. Standardizing the runtime data to generate standardized data with a consistent data structure may include: performing time-series alignment and missing value imputation on the runtime data, ensuring that data points in the runtime data are aligned based on time points. The number of data points in each sub-running process is consistent and does not include missing values.

[0136] In some embodiments, runtime data includes sub-run data corresponding to each sub-process. Timing alignment and missing value filling of the runtime data may include: calculating the average of all data points for the missing data points in the sub-run data of multiple sub-run processes; and using the average to fill in the missing data points.

[0137] See below Figure 7 The method of the embodiments of this disclosure is further described.

[0138] Figure 6 A schematic flowchart of a device monitoring method according to some other embodiments of the present disclosure is shown.

[0139] At box 602, input the original sequence, such as the timing data of a single wafer sample mentioned earlier, denoted as... .

[0140] At box 604, frequency domain feature extraction is performed. Frequency domain feature extraction can be implemented by a frequency domain feature extraction module. The extracted frequency domain features are as described in the previous embodiments and will not be repeated here.

[0141] At box 606, statistical feature extraction is performed. Statistical feature extraction can be implemented by the statistical feature extraction module.

[0142] As mentioned earlier, the statistical feature extraction module can extract features using mathematical statistical calculations, that is, it can extract the overall trend information of the curve. The frequency domain feature extraction module is used to extract the distribution pattern of data in the frequency domain dimension. Frequency domain features have phase insensitivity, which can effectively improve the false alarm problem caused by the start time offset in the wafer processing process.

[0143] At box 608, deep multi-mode feature fusion is performed on the acquired operational data with overall trend characteristics and the data with frequency domain dimension distribution characteristics.

[0144] For example, the extracted statistical and frequency domain features can be fed into a deep multi-modal feature fusion module for processing. The fusion module can employ an MLP-Mixer architecture, which is a stacked module composed of alternating token-mixing MLPs and channel-mixing MLPs. It lacks convolutional or recurrent structures, making it more computationally efficient and suitable for learning non-linear relationships between features.

[0145] In some embodiments, the feature vectors of the two modalities can be stacked to form a preliminary fusion feature matrix. The concatenated features are then fed into the MLP-Mixer module. For example, the parameters for this module are set as follows: number of patches S=15, hidden layer dimension D=64, total number of layers L_mixer=4, and output dimension... The final feature fusion module outputs a feature dimension of 1. , recorded as .

[0146] At box 610, the encoder extracts features from the training data to generate compressed feature vectors.

[0147] The training data is fed into the encoder module as input. As mentioned earlier, the time-series data of a single wafer sample is denoted as... , has dimensions Where L is the aligned sequence length and 1 represents the number of sensors (taking a single sensor as an example). For example, the encoder can be constructed using three stacked fully connected networks to achieve information compression and key information extraction from the original sequence signal. The dimension settings of each fully connected layer are shown below, and the final encoded feature vector dimension is... , recorded as It should be understood that encoders are not limited to being constructed using a stack of three fully connected networks, but can be constructed in other ways as needed.

[0148] At box 612, the fused features are concatenated with the compressed feature vector to generate concatenated features; the decoder reconstructs the concatenated features to generate prediction data.

[0149] The decoder module can be viewed as the inverse process of the encoder, realizing the decompression and reconstruction of information. Unlike a conventional encoder-decoder, this module incorporates contextual information from statistical frequency domain fusion to guide the decoder's reconstruction of the encoded features. Here, the previously output statistical and frequency domain fused features are... With encoder output features Perform vector concatenation to obtain concatenated features. The spliced ​​features are then fed into the decoder for sequence reconstruction. The decoder can consist of three fully connected layers, with a structure similar to the encoder. The output sequence of the decoder is the same length as the original sequence.

[0150] ;

[0151] .

[0152] The above embodiments can obtain a recovered sequence from the original input data after sequence encoding and decoding.

[0153] At box 614, a reconstruction error threshold is determined. As mentioned earlier, the optimization objective can be set to minimize the reconstruction error of the training samples. The reconstruction error is the mean squared error calculated for each sample point of the input sequence and the predicted output sequence. The stochastic gradient descent (SGD) optimization algorithm can be used to optimize the model parameters through multiple iterations. The parameter tuning process for model training can be implemented using existing neural network training frameworks, such as PyTorch, and it should be understood that this disclosure is not limited to this. The previous content regarding model training and testing using reconstruction error also applies to... Figure 7 The process is shown in the diagram, so it will not be described in detail here.

[0154] The equipment monitoring method 600 also includes comparing a standard score with a predetermined threshold. Then, based on the comparison result, it is determined whether the equipment's operating state is abnormal. The determined result, along with the corresponding data point for the abnormality, is output. As mentioned earlier, the comparison result can correspond to a specific equipment operating state, which will not be elaborated further here.

[0155] In box 616, output the monitoring results. For example, the results can show whether the device is operating normally or abnormally, and output specific abnormal data points.

[0156] Figure 6 The content shown can be executed by the monitoring model provided in the embodiments of this disclosure, and achieve the technical effects of the monitoring model mentioned above.

[0157] In some embodiments of this disclosure, the feature-guided autoencoder method integrates global, cross-modal deep features into the sequence autoencoder (encoder, decoder), which can better handle complex anomaly recognition tasks such as sudden shifts, gradual drifts, and periodic fluctuations.

[0158] Figure 7 A schematic diagram of device operating data with predetermined thresholds set according to an embodiment of the present disclosure is shown.

[0159] like Figure 7 As shown, with Figure 1 Similarly, multiple solid lines 702 show statistical values ​​of multiple operating data points when the device performs the same process multiple times. A line 704 located below and spaced apart from these solid lines represents a "specification line" with a predetermined threshold. Figure 1 In contrast, this specification line is not a fixed value but a variable one, and it reflects the trend of data changes. Therefore, the monitoring method provided in this disclosure has higher accuracy than traditional methods. Furthermore, as mentioned above, this monitoring method does not require engineers to manually set the specification line, thus improving monitoring efficiency. Figure 7 The horizontal and vertical axes can be compared with Figure 1 Meaning the same thing ( Figure 7 The values ​​of the horizontal axis have been omitted.

[0160] See Table 1 below, which shows the method of this disclosure and a comparison of metrics between the isolated forest and ARIMA methods. Table 1:

[0161] As can be seen from Table 1, the method of this disclosure embodiment has the highest score (F1-Score) and the lowest false alarm rate.

[0162] In the solution of this disclosure, unsupervised monitoring of equipment anomalies can be achieved without engineers manually setting specifications. This solves the problem that each device's sensors require manual setting of specification lines for monitoring. Compared to manually setting fixed specification lines, the solution of this disclosure can effectively monitor abnormal fluctuations in the curve within the specification lines, improving monitoring accuracy.

[0163] It should be understood that the embodiments shown in the accompanying drawings are merely illustrative of some embodiments of this disclosure and are not intended to limit this disclosure. Embodiments of this disclosure may also have various other forms.

[0164] An electronic device is also disclosed in the embodiments of this disclosure. The electronic device includes: a processor; and a memory coupled to the processor, the memory having instructions stored therein, the instructions causing the device to perform actions when executed by the processor, the actions including: extracting overall trend features and local trend features of training data, the training data including the device's operating data; and performing iterative training based on the overall trend features and local trend features to generate a monitoring model with optimal model parameters.

[0165] This disclosure also discloses a prediction model in its embodiments. The prediction model includes: a Long Short-Term Memory (LSTM) network configured to extract overall trend features of input data, including device operating data; a temporal convolutional network configured to extract local trend features of the input data; a prediction fusion module configured to generate prediction data based on the overall and local trend features; and a judgment module configured to determine the device's operating state based on the difference between the input data and the prediction data. In some embodiments, the prediction fusion module stacks and merges the overall and local trend features to generate a fused feature. The fused feature can then be used for prediction (e.g., through weighted fusion) to generate prediction data. In some embodiments, the prediction fusion module may be composed of a fully connected network. It should be understood that this disclosure is not limited thereto.

[0166] Additionally or alternatively, an electronic device includes: a processor; and a memory coupled to the processor, the memory having instructions stored therein, which, when executed by the processor, cause the device to perform actions, including: generating predictive data from a monitoring model based on collected operational data of the device, the operational data including multiple data points collected at time points; and determining the operating state of the device based on the difference between the operational data and the predictive data.

[0167] In other embodiments, the electronic device includes: a processor; and a memory coupled to the processor, the memory having instructions stored therein, which, when executed by the processor, cause the device to perform the following actions: extracting overall trend features and frequency domain dimensionality distribution features of training data, the training data including time-series data related to wafer process; generating fused features based on the overall trend features and frequency domain dimensionality distribution features; encoding the training data to generate encoded data; and decoding the encoded data based on the fused features to generate prediction data.

[0168] The embodiments of this disclosure also disclose a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a model training method and / or a device monitoring method according to embodiments of this disclosure.

[0169] In some embodiments of this disclosure, the feature-guided autoencoder method integrates global, cross-modal deep features into the sequence autoencoder (encoder, decoder), which can better handle complex anomaly recognition tasks such as sudden shifts, gradual drifts, and periodic fluctuations.

[0170] Figure 8 A schematic block diagram of an example device 800 that can be used to implement embodiments of the present disclosure is shown. For example, Figure 3 The model training method 300 shown is... Figure 4The model training method 400 shown is... Figure 5 The device monitoring method 500 and / or shown Figure 6 The device monitoring method 600 shown can be implemented by device 800. As shown, device 800 includes a central processing unit (CPU) 810, which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) 820 or loaded from storage unit 880 into random access memory (RAM) 830. RAM 830 can also store various programs and data required for the operation of device 800. CPU 810, ROM 820, and RAM 830 are interconnected via bus 840. Input / output (I / O) interface 850 is also connected to bus 840.

[0171] Multiple components in device 800 are connected to I / O interface 850, including: input unit 860, such as keyboard, mouse, etc.; output unit 870, such as various types of monitors, speakers, etc.; storage unit 880, such as disk, optical disk, etc.; and communication unit 890, such as network card, modem, wireless transceiver, etc. Communication unit 890 allows device 800 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0172] Processing unit 810 executes the various methods and processes described above, such as method 300, method 400, and / or method 500. For example, in some embodiments, method 300, method 400, and / or methods 500 and 600 may be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 880. In some embodiments, part or all of the computer program may be loaded and / or installed on device 800 via ROM 820 and / or communication unit 890. When the computer program is loaded into RAM 830 and executed by CPU 810, one or more steps of methods 300, method 400, and / or methods 500 and 600 described above may be performed. Alternatively, in other embodiments, CPU 810 may be configured to execute methods 300, method 400, and / or methods 500 and 600 by any other suitable means (e.g., by means of firmware).

[0173] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0174] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0175] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0176] Furthermore, although the operations are described in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0177] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A model training method, comprising: Extract the overall trend characteristics and frequency domain dimension distribution characteristics of the training data, which includes time-series data related to the wafer process; A fusion feature is generated based on the overall change trend characteristics and the frequency domain dimension distribution characteristics; The training data is encoded to generate encoded data; as well as The encoded data is decoded based on the fusion features to generate prediction data.

2. The method according to claim 1, wherein extracting the overall trend features and frequency domain dimension distribution features of the training data includes: Determine the statistical parameter values ​​for the predetermined type of the training data; as well as Statistical feature vectors are generated based on the values ​​of various statistical parameters to serve as the overall trend features.

3. The method according to claim 1, wherein extracting the frequency domain dimension distribution features of the training data includes: The training data is converted into a frequency domain complex sequence; The relevant spectral features are determined based on the frequency domain complex sequence; as well as The frequency domain dimension distribution features are generated based on each spectral feature.

4. The method according to claim 1, wherein generating the fused features based on the overall change trend features and the frequency domain dimension distribution features includes: The overall trend features and frequency domain dimensionality distribution features of the training data are stacked and merged using deep multimodal feature fusion to obtain the fused features.

5. The method according to claim 1, wherein iteratively training the model based on the training data and the prediction data to generate a monitoring model with optimal model parameters comprises: The model parameters of the monitoring model are optimized through multiple iterations using the stochastic gradient descent optimization algorithm.

6. A method for monitoring equipment, comprising: Predictive data is generated by a monitoring model based on the collected equipment operating data, wherein the operating data includes time-series data related to the wafer process collected at specific points in time; as well as The operating status of the device is determined based on the difference between the operating data and the predicted data.

7. A monitoring model trained by the method according to any one of claims 1 to 5.

8. A prediction model, comprising: The statistical feature extraction module is configured to extract the overall trend features of the input data, which includes time-series data related to the wafer process. The frequency domain feature extraction module is configured to extract the frequency domain dimensionality distribution features of the input data; The fusion module is configured to generate fusion features based on the overall change trend features and the frequency domain dimension distribution features; An encoder is configured to encode the input data to generate encoded data; as well as The decoder is configured to decode the encoded data based on the fused features to generate the predicted data.

9. An electronic device, comprising: processor; as well as A memory coupled to the processor, the memory having instructions stored therein, which, when executed by the processor, cause the device to perform the method according to any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method according to any one of claims 1-6.

Images