Virtual measurement method and system for key quality indicators of TFT-LCD continuous correlation process
Patent Information
- Application Number
- CN202611516750.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-28
- Publication Date
- 2026-10-30
Smart Images
Figure CN123014820A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial process quality technology, and in particular to a virtual measurement method and system for key quality indicators in continuous correlated processes of TFT-LCD. Background Technology
[0002] The manufacturing process of TFT-LCD (Thin Film Transistor Liquid Crystal Display) typically consists of multiple consecutive and interconnected steps, each involving various process variables such as temperature, gas flow rate, liquid flow rate, power, and processing time. There is a clear cascading and synergistic relationship between different steps; the processing status and quality of preceding steps can affect the quality of subsequent steps and even the final product. Therefore, the key quality indicators of a target step are not only related to the process parameters of that step but may also have cross-correlation with the quality output of preceding steps.
[0003] In actual production processes, TFT-LCD process data exhibits characteristics such as high dimensionality, nonlinearity, and temporal sequence. Factors such as sensor fluctuations, equipment wear, and operating condition drift can also lead to missing, abnormal, or noisy data. Traditional physical measurement methods typically require actual testing equipment or experimental methods to obtain quality results, which is insufficient for rapid quality assessment. Virtual measurement technology, by utilizing historical production process data to establish a mapping relationship between process parameters and quality indicators, can predict key quality indicators using production process data, thereby reducing reliance on sample-by-sample physical measurements.
[0004] Existing virtual measurement methods often use the current process parameters of a single process as input, or simply concatenate variables from multiple processes, failing to fully utilize the quality correlations between consecutive processes. For manufacturing data with significant time-series dependencies, conventional machine learning models or simple convolutional models have limited ability to represent cross-timescale dependencies, and one-way recurrent networks also struggle to fully utilize the contextual information of preceding and following positions within an established time window.
[0005] Furthermore, the production equipment, process variables, and number of parameters used in different processing steps often vary, resulting in significant heterogeneity in the production data for each step. If a uniform and fixed set of input features is used to model different steps, redundant variables with low correlation to the quality prediction of the current step are easily introduced, and it is difficult to fully reflect the key process characteristics of each step.
[0006] The information disclosed in this background section is intended only to enhance the understanding of the general background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention
[0007] This invention provides a virtual measurement method and system for key quality indicators in the continuous correlation process of TFT-LCD, thereby effectively solving the problems in the background art.
[0008] To achieve the above objectives, the technical solution adopted by this invention is: a virtual measurement method for key quality indicators in a continuous correlation process of TFT-LCD, comprising the following steps:
[0009] Acquire production process data for at least two consecutive related processes in the TFT-LCD manufacturing process. The production process data includes process parameter data for each consecutive related process and a quality index sequence for at least one preceding related process. Perform process division and preprocessing on the production process data according to process identifiers and / or tool identifiers to form a process parameter sequence for each process.
[0010] For the target associated process for which key quality indicators are to be virtually measured, the process parameter sequence of the target associated process is aligned with the quality indicator sequence of the preceding associated process on a time-node basis, and then merged into the associated process input sequence of the target associated process in a parallel input channel manner; the associated process input sequence is then divided into time windows and normalized.
[0011] For the associated process input sequence, multi-scale temporal features are extracted through at least two dilated causal convolutional layers with different dilation rates. Bidirectional temporal modeling is performed on the multi-scale temporal features within the time window to obtain bidirectional temporal features. Temporal pattern features spanning multiple time nodes are extracted from the bidirectional temporal features, and attention weighting is performed based on the temporal pattern features to obtain weighted temporal features.
[0012] Using key quality indicators corresponding to historical production process data as supervisory labels, a virtual measurement model for key quality indicators is trained based on the weighted time-series features. The TFT-LCD production process data to be measured is processed in the same way as the training phase, and then input into the trained virtual measurement model for key quality indicators. The virtual measurement results of key quality indicators corresponding to the target associated process are output.
[0013] Furthermore, the preceding associated process is a sequential associated process that is located before the target associated process in the processing sequence, and whose process status or quality indicators have a data correlation with the key quality indicators of the target associated process.
[0014] The target associated process refers to the continuous associated processes that currently require virtual measurement of key quality indicators;
[0015] The quality index sequence is a sequence of key quality indicators of the preceding related processes formed according to the production sequence;
[0016] The key quality indicators are continuous numerical indicators that reflect the processing quality of the target-related process or the quality of the final product and must be obtained through physical measurement.
[0017] Furthermore, the process parameter data includes one or more of temperature, gas flow rate, liquid flow rate, power, and process time;
[0018] The step of dividing the production process data into processes based on process identifiers and / or tool identifiers includes: determining the corresponding process data group based on the tool identifier;
[0019] The preprocessing includes: deleting null value columns, single value columns, and duplicate columns from the production process data;
[0020] For zero-value data and / or missing-value data that are determined to be invalid, fill them in using the statistical values of non-zero valid data under the same process and tool category, based on the process and tool category to which the corresponding data belongs.
[0021] Furthermore, the sequence of process parameters for each step includes:
[0022] The importance of each candidate process parameter is evaluated, and a predetermined number of key process parameters are selected from the candidate process parameters of the corresponding process based on the ranking of the importance of each process. This is to construct the process parameter sequence of the corresponding process.
[0023] The number and / or composition of candidate process parameters differ for different processes, and the composition of the key process parameter set corresponding to each process varies with the process; the feature importance evaluation is performed using the random forest algorithm.
[0024] Further, let the target associated process be the i-th process, and let the sequence of process parameters for the target associated process be X. i The quality index sequence of the preceding associated process is Y. i-1 Then the associated process input sequence Z i Satisfy Z i =[X i ,Y i-1 The input vector z of the associated process input sequence at time node t i,t =[x i,t , y i-1,t ];
[0025] The time-node alignment is based on one or more of product identifier, batch identifier, and production sequence identifier to determine the correspondence between the same product or corresponding production samples.
[0026] The time window is divided into sliding time windows of length L.
[0027] Furthermore, the extraction of multi-scale temporal features through at least two dilated causal convolutional layers with different dilation rates is achieved using a temporal convolutional network;
[0028] The temporal convolutional network includes at least two cascaded dilated causal convolutional layers, each of which has a different dilation rate. Different dilation rates correspond to receptive fields at different time scales, so as to extract temporal dependency features at different time scales and form the multi-scale temporal features.
[0029] The temporal convolutional network employs causal convolution, and the output of each of the dilated causal convolutional layers at the current time node is determined only by the current time node and the input data located before the current time node.
[0030] The temporal convolutional network also includes a residual connection structure. For any dilated causal convolutional layer, the input of the dilated causal convolutional layer is added to the convolutional output of the dilated causal convolutional layer through a residual branch, and the added features are nonlinearly activated.
[0031] When the feature dimensions of the input and the convolution output are inconsistent, the feature dimensions of the input in the residual branch are adjusted by convolution mapping before the addition is performed.
[0032] Furthermore, the bidirectional temporal modeling of the multi-scale temporal features within the time window is implemented using a bidirectional long short-term memory network.
[0033] The bidirectional long short-term memory network includes a forward long short-term memory unit and a reverse long short-term memory unit. The multi-scale temporal features are input into the forward long short-term memory unit and the reverse long short-term memory unit respectively to obtain the forward hidden state sequence and the reverse hidden state sequence.
[0034] The reverse long short-term memory unit only processes the multi-scale temporal features observed within the established time window, and does not involve production data generated in the future when virtual measurement is performed;
[0035] The forward hidden state and the reverse hidden state at the same time node are spliced and fused to obtain the bidirectional hidden state at that time node, and the bidirectional hidden state at each time node constitutes the bidirectional temporal feature.
[0036] Furthermore, the step of extracting temporal pattern features across multiple time nodes from the bidirectional temporal features and performing attention weighting based on the temporal pattern features is achieved using a temporal pattern attention mechanism.
[0037] The implementation using the time-mode attention mechanism includes:
[0038] A hidden state matrix is constructed based on the hidden states of the bidirectional temporal features at multiple time nodes, and the hidden state matrix is divided into time series corresponding to each state variable according to the dimension of the state variable.
[0039] One-dimensional convolution is performed along the time dimension on the time series corresponding to each state variable to obtain the time pattern features corresponding to each state variable.
[0040] The degree of correlation between each of the time pattern features and the hidden state at the current time node in the hidden state matrix is determined to obtain the attention score corresponding to each of the time pattern features. The attention scores are then normalized to obtain the attention weight corresponding to each of the time pattern features.
[0041] Using the time pattern features as units, each time pattern feature is weighted and fused according to the attention weight to obtain a weighted time pattern feature, and the feature obtained by fusing the weighted time pattern feature with the bidirectional time series feature is used as the weighted time series feature.
[0042] Furthermore, for multiple target associated processes to be virtually measured for key quality indicators, the process parameter sequence of each target associated process and the quality indicator sequence of the preceding associated process corresponding to each target associated process are obtained respectively.
[0043] For each of the target associated processes, a corresponding associated process input sequence is constructed, and virtual measurement results of key quality indicators for the corresponding target associated processes are obtained based on each of the associated process input sequences.
[0044] The number of process parameters in the process parameter sequences of each of the target associated processes are different, and there is no hierarchical quality index transmission chain between the target associated processes.
[0045] This invention also includes a virtual measurement system for key quality indicators in continuous correlation processes of TFT-LCDs, the system comprising:
[0046] The data acquisition module is used to acquire production process data of at least two consecutive related processes in the TFT-LCD manufacturing process. The production process data includes process parameter data of each consecutive related process and a quality index sequence of at least one preceding related process. The data acquisition module is also used to divide and preprocess the production process data according to process identifiers and / or tool identifiers to form a process parameter sequence for each process.
[0047] The associated process input construction module is used to align the process parameter sequence of the target associated process with the quality indicator sequence of the preceding associated process on a time-node basis for the target associated process to be virtually measured for key quality indicators, and merge them into the associated process input sequence of the target associated process in a parallel input channel manner; the associated process input construction module is also used to perform time window division and normalization processing on the associated process input sequence.
[0048] The temporal feature processing module is used to extract multi-scale temporal features from the associated process input sequence through at least two dilated causal convolutional layers with different dilation rates, and to perform bidirectional temporal modeling on the multi-scale temporal features within the time window to obtain bidirectional temporal features; to extract temporal pattern features across multiple time nodes from the bidirectional temporal features, and to perform attention weighting based on the temporal pattern features to obtain weighted temporal features;
[0049] The virtual measurement module is used to train a virtual measurement model of key quality indicators based on the weighted time-series features, using key quality indicators corresponding to historical production process data as supervision labels. The virtual measurement module is also used to process the TFT-LCD production process data to be measured in a data processing method consistent with the training stage, input it into the trained virtual measurement model of key quality indicators, and output the virtual measurement results of key quality indicators corresponding to the target associated process.
[0050] The beneficial effects of this invention are as follows: by fusing the quality index information of preceding associated processes with the process parameter information of the target associated process itself, the virtual measurement input can simultaneously characterize the internal process state and the quality correlation between consecutive processes. Through multi-scale temporal feature extraction, local changes and long-term dependence features at different time scales are obtained, improving the characterization capability of complex TFT-LCD production time-series data. Attached Figure Description
[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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 recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0052] Figure 1 This is a schematic diagram of the overall process of a virtual measurement method for key quality indicators in a continuous associative process of TFT-LCD according to the present invention.
[0053] Figure 2 This is a schematic diagram illustrating the construction of the associated process input sequence for the continuous associated processes of the present invention;
[0054] Figure 3 This is a schematic diagram of the multi-scale temporal feature extraction structure based on temporal convolutional networks of the present invention;
[0055] Figure 4 This is a schematic diagram of the bidirectional temporal feature modeling structure based on a bidirectional long short-term memory network according to the present invention;
[0056] Figure 5 This is a schematic diagram of the temporal pattern extraction and attention weighting structure based on the temporal pattern attention mechanism of the present invention;
[0057] Figure 6 This is a schematic diagram of the data processing, model training, and virtual measurement process of key quality indicators in the TFT-LCD production process of this invention.
[0058] Figure 7 This is a schematic diagram of the structure of the computer device of the present invention. Detailed Implementation
[0059] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0060] Example 1:
[0061] like Figure 1 The following is a virtual measurement method for key quality indicators in a continuous correlation process of TFT-LCD, comprising the following steps:
[0062] Acquire production process data for at least two consecutive related processes in the TFT-LCD manufacturing process. The production process data includes process parameter data for each consecutive related process and a quality index sequence for at least one preceding related process. Divide and preprocess the production process data according to process identifiers and / or tool identifiers to form a process parameter sequence for each process.
[0063] For the target associated process for which key quality indicators are to be virtually measured, the process parameter sequence of the target associated process is aligned with the quality indicator sequence of the preceding associated process at each time node, and then merged into the associated process input sequence of the target associated process in a parallel input channel manner; the associated process input sequence is then divided into time windows and standardized.
[0064] For the input sequence of the associated process, multi-scale temporal features are extracted through at least two dilated causal convolutional layers with different dilation rates. Bidirectional temporal modeling is performed on the multi-scale temporal features within the time window to obtain bidirectional temporal features. Temporal pattern features across multiple time nodes are extracted from the bidirectional temporal features, and attention weighting is performed based on the temporal pattern features to obtain weighted temporal features.
[0065] Using key quality indicators corresponding to historical production process data as supervisory labels, a virtual measurement model for key quality indicators is trained based on weighted time-series features. The TFT-LCD production process data to be measured is processed in the same way as the data processing in the training stage, and then input into the trained virtual measurement model for key quality indicators. The virtual measurement results of key quality indicators corresponding to the target associated process are output.
[0066] By fusing quality index information from preceding associated processes with process parameter information from the target associated process itself, virtual measurement inputs can simultaneously characterize the internal process state and the quality correlation between consecutive processes. Multi-scale temporal feature extraction obtains local variations and long-term dependency features at different time scales, improving the characterization capability of complex TFT-LCD production time-series data.
[0067] In this embodiment, the preceding associated process is a continuous associated process that is located before the target associated process in the processing sequence, and whose process status or quality indicators have a data correlation with the key quality indicators of the target associated process.
[0068] Target-related processes refer to the continuous related processes that currently require virtual measurement of key quality indicators;
[0069] The quality indicator sequence is a sequence of key quality indicators of preceding related processes formed according to the production sequence;
[0070] Key quality indicators are continuous numerical indicators that reflect the processing quality of the target-related processes or the quality of the final product and must be obtained through physical measurement.
[0071] The process parameter data includes one or more of the following: temperature, gas flow rate, liquid flow rate, power, and process time.
[0072] The production process data is divided into processes based on process identifiers and / or tool identifiers, including: determining the corresponding process data groups based on tool identifiers;
[0073] Preprocessing includes: removing null, single-value, and duplicate columns from the production process data;
[0074] For zero-value data and / or missing-value data that are determined to be invalid, fill them in using the statistical values of non-zero valid data under the same process and tool category, based on the process and tool category to which the corresponding data belongs.
[0075] As a preferred embodiment of the above, the process parameter sequence for each step includes:
[0076] The importance of each candidate process parameter is evaluated, and a predetermined number of key process parameters are selected from the candidate process parameters of the corresponding process based on the ranking of the importance of each process. This is to construct the process parameter sequence of the corresponding process.
[0077] Among them, the number and / or composition of candidate process parameters differ in different processes, and the composition of the key process parameter set corresponding to each process varies with the process; the feature importance evaluation is performed using the random forest algorithm.
[0078] In this embodiment, the target associated process is denoted as the i-th process, and the sequence of process parameters for the target associated process is X. i The quality index sequence of the preceding related processes is Y. i-1 Then the associated process input sequence Z i Satisfy Z i =[X i ,Y i-1 The input vector z of the associated process input sequence at time node t. i,t =[x i,t , y i-1,t ];
[0079] Alignment at each time node is based on one or more of the product identifier, batch identifier, and production sequence identifier to determine the correspondence between the same product or corresponding production samples.
[0080] The time window is divided into sliding time windows of length L.
[0081] Multi-scale temporal features are extracted through at least two dilated causal convolutional layers with different dilation rates, which is implemented using a temporal convolutional network.
[0082] Temporal convolutional networks include at least two cascaded dilated causal convolutional layers, each with a different dilation rate. Different dilation rates correspond to receptive fields at different time scales, so as to extract temporal dependency features at different time scales and form multi-scale temporal features.
[0083] Temporal convolutional networks employ causal convolution, where the output of each dilated causal convolutional layer at the current time node is determined solely by the current time node and the input data preceding it.
[0084] Temporal convolutional networks also include residual connection structures. For any dilated causal convolutional layer, the input of the dilated causal convolutional layer is added to the convolutional output of the dilated causal convolutional layer through a residual branch, and the added features are nonlinearly activated.
[0085] When the feature dimensions of the input and the convolution output are inconsistent, the feature dimensions of the input in the residual branch are adjusted by convolution mapping before addition is performed.
[0086] As a preferred embodiment of the above, bidirectional temporal modeling of multi-scale temporal features is performed within a time window, which is achieved using a bidirectional long short-term memory network.
[0087] The bidirectional long short-term memory network includes forward long short-term memory units and reverse long short-term memory units. Multi-scale temporal features are input into the forward long short-term memory units and the reverse long short-term memory units respectively to obtain forward hidden state sequences and reverse hidden state sequences.
[0088] The reverse long short-term memory unit only processes the multi-scale temporal features observed within the established time window, and does not involve production data generated in the future when virtual measurement is performed;
[0089] By splicing and fusing the forward and reverse hidden states at the same time point, a bidirectional hidden state at that time point is obtained, and the bidirectional hidden states at each time point constitute a bidirectional temporal feature.
[0090] In this embodiment, temporal pattern features spanning multiple time nodes are extracted from bidirectional temporal features, and attention weighting is applied based on the temporal pattern features, which is achieved using a temporal pattern attention mechanism.
[0091] The implementation using a time-based attention mechanism includes:
[0092] A hidden state matrix is constructed based on the hidden states at multiple time points according to the bidirectional temporal features, and the hidden state matrix is divided into time series corresponding to each state variable according to the dimension of the state variable.
[0093] One-dimensional convolution is performed along the time dimension on the time series corresponding to each state variable to obtain the time pattern features corresponding to each state variable.
[0094] The degree of correlation between each temporal pattern feature and the hidden state at the current time node in the hidden state matrix is determined to obtain the attention score corresponding to each temporal pattern feature. Then, the attention score is normalized to obtain the attention weight corresponding to each temporal pattern feature.
[0095] Using time pattern features as units, each time pattern feature is weighted and fused according to attention weights to obtain weighted time pattern features. The feature obtained by fusing the weighted time pattern features with the bidirectional time series features is then used as the weighted time series feature.
[0096] As a preferred embodiment of the above, for multiple target associated processes to be virtually measured for key quality indicators, the process parameter sequence of each target associated process and the quality indicator sequence of the preceding associated process corresponding to each target associated process are obtained respectively.
[0097] For each target-related process, a corresponding associated process input sequence is constructed, and virtual measurement results of key quality indicators for the corresponding target-related process are obtained based on each associated process input sequence.
[0098] The number of process parameters in the process parameter sequences of each target-related process is different, and there is no hierarchical quality indicator transmission chain between each target-related process.
[0099] This embodiment also includes a virtual measurement system for key quality indicators in continuous correlation processes of TFT-LCD, the system comprising:
[0100] The data acquisition module is used to acquire production process data of at least two consecutive related processes in the TFT-LCD manufacturing process. The production process data includes process parameter data of each consecutive related process and a quality index sequence of at least one preceding related process. The data acquisition module is also used to divide and preprocess the production process data according to process identifiers and / or tool identifiers to form a process parameter sequence for each process.
[0101] The associated process input construction module is used to align the process parameter sequence of the target associated process with the quality indicator sequence of the preceding associated process on a time-node basis for the target associated process to be virtually measured for key quality indicators, and merge them into the associated process input sequence of the target associated process in a parallel input channel manner; the associated process input construction module is also used to perform time window division and normalization processing on the associated process input sequence;
[0102] The temporal feature processing module is used to extract multi-scale temporal features from the associated process input sequence through at least two dilated causal convolutional layers with different dilation rates, and to perform bidirectional temporal modeling on the multi-scale temporal features within a time window to obtain bidirectional temporal features; from the bidirectional temporal features, temporal pattern features across multiple time nodes are extracted, and attention weighting is performed based on the temporal pattern features to obtain weighted temporal features;
[0103] The virtual measurement module is used to train a virtual measurement model of key quality indicators based on weighted time-series features, using key quality indicators corresponding to historical production process data as supervision labels. The virtual measurement module is also used to process the TFT-LCD production process data to be measured in the same data processing method as the training stage, input it into the trained virtual measurement model of key quality indicators, and output the virtual measurement results of key quality indicators corresponding to the target associated process.
[0104] Example 2:
[0105] like Figure 1 As shown, this embodiment provides a virtual measurement method for key quality indicators in a continuous correlation process of TFT-LCD, including steps S1 to S6.
[0106] S1, acquire production process data corresponding to at least two consecutive related processes in the TFT-LCD manufacturing process. The data includes process parameter data for the corresponding processes and quality indicator data corresponding to at least one preceding related process. The production process data is historical production process data corresponding to each related process.
[0107] S2, the production process data is divided into processes based on process identifiers and / or tool identifiers, and the data for each process is cleaned and preprocessed. Preprocessing may include deleting null, single-value, and duplicate columns, as well as imputing zero and missing values. Furthermore, feature importance evaluation can be performed on candidate parameters for each process to obtain the corresponding set of key process parameters.
[0108] S3. For the target associated process, the process parameter sequence corresponding to the target associated process is sample-aligned and sequence-fused with the quality index sequence corresponding to the preceding associated process to form an associated process input sequence; then the associated process input sequence is divided according to a preset time window and normalization processing is performed.
[0109] S4. Multi-scale temporal feature extraction is performed on the associated process input sequence obtained in step S3, and bidirectional temporal modeling is performed on the extracted temporal features. In a preferred embodiment, a temporal convolutional network (TCN) is used for multi-scale temporal feature extraction, and a bidirectional long short-term memory network (BiLSTM) is used for bidirectional temporal modeling. TCN and BiLSTM are preferred implementations in this embodiment. The key point of the independent technical solution of this invention lies in sequentially performing multi-scale temporal feature extraction and bidirectional temporal modeling on the associated process input sequence.
[0110] S5. Extract time pattern features spanning multiple time nodes from the bidirectional temporal features obtained in step S4, determine attention weights based on the time pattern features, and perform weighted fusion of the corresponding temporal features. In a preferred embodiment, the temporal pattern attention mechanism (TPA) is used to achieve time pattern extraction and feature weighting.
[0111] S6 uses key quality indicators corresponding to historical production process data as supervisory labels to construct training samples and train a virtual measurement model. For the TFT-LCD production process data to be measured, the same or corresponding data processing rules as in the training stage are used for process division, data preprocessing, process parameter determination, construction of associated process inputs, time window division, and normalization processing, so that the data to be measured forms a data structure corresponding to the model input in the training stage, and then inputs it into the trained model to obtain the virtual measurement results of key quality indicators.
[0112] like Figure 6 As shown, the virtual measurement process for key quality indicators in this embodiment includes a training phase and a virtual measurement phase. In the training phase, historical TFT-LCD production data is sequentially divided into processes, cleaned, missing value handled, data normalized, key parameters for each process are selected, and a continuous, related process input sequence is constructed. The processed historical production data is then used to train the virtual measurement model. In the virtual measurement phase, the TFT-LCD production process data to be measured undergoes the corresponding data preprocessing and key process parameter extraction as in the training phase. A continuous, related process input sequence is constructed and input into the trained virtual measurement model to obtain the virtual measurement results for the key quality indicators.
[0113] like Figure 2 As shown, this embodiment further illustrates the process of constructing the associated process input sequence between consecutive associated processes. Let the process parameter vector of the i-th target associated process at time node t be:
[0114]
[0115] In the formula, Indicates the first The target-related processes at the time nodes The process parameter vector at the location; Indicates the first The target-related processes at the time nodes The first There are several process parameters, among which... ; Indicates the first The number of key process parameters included in the target-related process after feature screening.
[0116] The process parameter sequence X of the target associated process is composed of process parameter vectors at consecutive time nodes. i Let Y be the sequence of quality indicators for the preceding associated processes corresponding to the i-th target process. i-1 Then, after aligning the samples or production sequence, a related process input sequence is constructed:
[0117]
[0118] At time node t, the corresponding associated input vector can be represented as z. i,t =[x i,t , y i-1,t Applying a sliding time window of length L to the sequence, we can obtain:
[0119]
[0120] The time window data is then normalized to obtain the associated process input for subsequent time-series feature extraction. Sample alignment can be performed based on product identifiers, batch identifiers, production sequence, or other identifiers that can determine the relationship between the same product or corresponding production samples.
[0121] When the TFT-LCD manufacturing process includes multiple target correlation steps that require virtual measurement, the aforementioned construction process can be performed separately for each target correlation step. For example, for the i-th target correlation step, Z is constructed. i =[X i ,Y i-1 For another target-related process j, construct Z. j =[X j ,Y j-1 Each target-related process can form its own associated process input based on its own data structure. It is not necessary to require different target-related processes to use the same number of process parameters, nor is it necessary to require all TFT-LCD processes to form a process-by-process quality index recursive relationship.
[0122] like Figure 3 As shown, this embodiment uses a temporal convolutional network (TCN) to extract multi-scale temporal features from the normalized associated process input sequence. The TCN includes causal convolution, dilated convolution, and residual connection structures.
[0123] For a regular causal convolution, the output at time t depends only on the current time and the previous inputs, and can be represented as:
[0124]
[0125] After introducing the dilation rate d, the dilated causal convolution can be expressed as:
[0126]
[0127] In the formula, Indicates time node Convolution output at the point; and These represent the corresponding historical inputs used for ordinary causal convolution and dilated causal convolution, respectively. The convolution kernel is represented by the first... Each weight; Indicates the kernel size; Indicates the kernel position number; This represents the expansion rate. Different expansion rates can be set. , … It can expand the receptive field without significantly increasing the number of parameters, allowing a smaller expansion rate to focus on local short-term changes and a larger expansion rate to focus on dependency information over a longer time range, thus forming multi-scale temporal features.
[0128] To improve feature propagation and network training in deep temporal convolutional networks, residual connections are added to the dilated causal convolution module. Let the convolutional transformation be F(R) and the residual mapping be G(R), then the module output can be expressed as:
[0129]
[0130] In the formula, This represents the input to the dilated causal convolution module; Indicates convolution transformation; Represents residual mapping; Represents a nonlinear activation function; This indicates the output of the residual connection module.
[0131] When the input feature dimension and the convolution output dimension are the same, G(R) can use an identity mapping; when the two dimensions are different, the residual branch can be adjusted in dimension through convolution mapping. The TCN obtains the associated process input sequence Z. i The multi-scale temporal features output by the process are used as input to the BiLSTM.
[0132] like Figure 4 As shown, this embodiment performs bidirectional temporal modeling based on the obtained multi-scale temporal features. Let the TCN output sequence be F. i =[f i,1 ,f i,2 ,…,f i,T The data are fed into the forward LSTM and the reverse LSTM, respectively.
[0133]
[0134]
[0135] By fusing the forward hidden state and the reverse hidden state, we can obtain the bidirectional hidden state at time node t:
[0136]
[0137] In the formula, Indicates time node In the positive hidden state, Indicates time node The reverse hidden state at the location; This indicates that the temporal convolutional network is at time nodes. Multi-scale temporal features output at the location; and These represent forward long short-term memory units and reverse long short-term memory units, respectively. This represents a bidirectional hidden state obtained by concatenating the forward hidden state and the reverse hidden state.
[0138] The bidirectional temporal feature matrix H is composed of the bidirectional hidden states at each time point. i The reverse time-series processing only reverse-encodes the observed data within the established time window, and does not involve future production data that has not yet been generated at the time of actual prediction. Therefore, the bidirectional contextual relationships within the time window can be utilized without compromising the availability of the measured data.
[0139] like Figure 5 As shown, this embodiment performs temporal pattern extraction and attention weighting on the obtained bidirectional temporal features. A hidden state matrix H is constructed based on the hidden states generated by the BiLSTM at multiple time points. i And form the time series H corresponding to each state variable according to the dimension of state variables. i (j) .
[0140] Performing a one-dimensional convolution along the time dimension on the j-th state variable yields the corresponding temporal pattern features:
[0141]
[0142] Further based on the time pattern feature C j The degree of association between the current bidirectional hidden state and the attention score e is calculated. j The attention weight α is obtained after normalization. j Note that the specific scoring function for the weights can be a trainable function that can characterize the correlation between temporal pattern features and the current bidirectional hidden state.
[0143]
[0144] In the formula, Indicates the first Attention scores corresponding to each time pattern feature; This represents the normalization function used to normalize each attention score; Indicates the first Attention weights corresponding to each temporal pattern feature; The index represents the state variable or the corresponding time pattern feature.
[0145] The temporal pattern features are weighted and fused based on attention weights to obtain weighted temporal features:
[0146]
[0147] In the formula, Indicates the first Weighted time-series features corresponding to the process associated with the target; This indicates the number of state variables, that is, the number of time pattern features participating in the weighted fusion; Indicates the first Attention weights corresponding to each temporal pattern feature; Indicates the first A time pattern feature.
[0148] In one implementation, V may also be used. i After further fusion with the current bidirectional hidden state, the result is input into the fully connected prediction output unit to obtain the predicted values of key quality indicators. Compared with the method of assigning weights only to a single time node, the temporal pattern attention processing can perform differentiated weighting for change patterns formed across multiple time nodes.
[0149] This embodiment has at least the following beneficial effects:
[0150] First, by integrating the quality index information of the preceding associated process with the process parameter information of the target associated process itself, the virtual measurement input can simultaneously characterize the internal process status of the process and the quality correlation between successive processes.
[0151] Second, by extracting multi-scale time-series features, we can obtain local changes and long-term dependence features at different time scales, thereby improving the ability to characterize complex TFT-LCD production time-series data.
[0152] Third, by using bidirectional temporal modeling, dynamic dependency information within the established time window is extracted from both the forward and reverse directions to obtain a more complete temporal state representation.
[0153] Fourth, by extracting and weighting the time patterns in the bidirectional time series features, different time patterns can be fused in a differentiated manner based on their contribution to the prediction of key quality indicators.
[0154] Fifth, by determining key process parameters for different processes, we can adapt to the different characteristics of production data with varying numbers of original parameters and variable compositions in each process, and reduce the impact of irrelevant or redundant features on subsequent virtual measurements.
[0155] Sixth, the production data to be measured adopts the same data processing rules as the model training phase, so that the training phase and the virtual measurement phase maintain the corresponding input structure and data scale.
[0156] Example 3:
[0157] This embodiment uses historical production data from an actual TFT-LCD production line for offline experimental verification. The data contains 800 records, including process factors such as reactor temperature, gas flow rate, liquid flow rate, power, and process time, as well as quality prediction indicators. To protect the privacy of the production data, the column names and feature values have been anonymized, mainly including the ID column, tool column, and data column. The data is divided into thirteen groups according to the tool column identifier, corresponding to thirteen main processes.
[0158] The number of machines and the number of original input parameters differ across processes. Specifically, the number of machines for processes 1 through 13 are 6, 2, 2, 3, 2, 8, 3, 4, 9, 3, 1, 10, and 2, respectively; the corresponding number of input parameters are 56, 51, 2, 33, 12, 44, 42, 26, 45, 59, 10, 39, and 37, respectively. This demonstrates that different processes have significantly different dimensions of original data.
[0159] The raw data is cleaned and preprocessed, including removing null, single-value, and duplicate columns. For columns containing 0 or NA values, missing values are handled separately according to the process and tool category, preferably using the non-zero average of similar tools for filling.
[0160] For different processes, a random forest algorithm is used to evaluate the importance of candidate process parameters, and the importance of each process feature is independently ranked. In this embodiment, the top 6 parameters in importance from the candidate process parameters for each process are selected as the process parameter inputs for the corresponding process. This specific number is for the embodiment and does not constitute a limitation on the scope of protection of this invention.
[0161] For a target associated process for which corresponding quality indicators of preceding associated processes can be obtained, the key process parameter sequence of the target associated process is aligned with and merged with the quality indicator sequence of preceding associated processes to form an associated process input sequence, and then time window division and data normalization are performed. Since the publicly available data has been anonymized, this embodiment does not assign specific physical process names to each of the thirteen anonymous processes, nor does it require all thirteen processes to form a hierarchical quality indicator recursive chain.
[0162] In this embodiment, the first 700 records of the dataset are used as training data, and the last 100 records are used as test data for offline experimental verification. The model adopts a two-layer dilated causal convolutional structure, where the first layer has a kernel size of 3 and a dilation rate of 4, and the second layer has a kernel size of 5 and a dilation rate of 8; the BiLSTM uses 128 neurons; the time step is set to 30, the learning rate is set to 0.0001, and the batch size is set to 128. The above parameters are all specific implementation parameters of this embodiment.
[0163] After model training is complete, new TFT-LCD production process data to be measured is processed according to the corresponding data processing flow used in the training phase. This includes dividing the process into steps according to the same process identification rules, preprocessing according to the data cleaning and missing value handling rules determined in the training phase, extracting input variables using the set of key process parameters already determined for the corresponding process, fusing the quality index information of the preceding related processes according to the same associated process input construction method, and performing data transformation using the time window rules and normalization rules or parameters determined in the training phase. After the above processing, the resulting input data to be measured is fed into the trained TCN-BiLSTM-TPA virtual measurement model to obtain the corresponding virtual measurement results of key quality indicators.
[0164] This embodiment uses Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²) to evaluate prediction performance. Offline comparison results using TFT-LCD data show differences in prediction performance among Random Forest, Extra Trees Regressor, Gradient Boosting Regressor, CNN, CNN-LSTM, and TCN-BiLSTM-TPA models. The TCN-BiLSTM-TPA model achieves an MAE of 0.1034, an RMSE of 0.6093, and an R² of 0.9382 on this dataset.
[0165] Ablation experiments further compared LSTM, BiLSTM, TCN-LSTM, TCN-BiLSTM, BiLSTM-TPA, and TCN-BiLSTM-TPA. The results show that, under this offline dataset and corresponding experimental settings, the combined model including TCN, BiLSTM, and TPA achieved a better overall prediction index.
[0166] The above experiments are used to verify the feasibility and predictive performance of the technical solution of the present invention on historical TFT-LCD production data. They do not imply that the method has been deployed online on actual production lines, nor do they consider the specific data scale, training and testing division method, and hyperparameter values as necessary limitations of the independent technical solution of the present invention.
[0167] Table 1. Prediction results of key quality indicators for TFT-LCD using different models.
[0168] Random Forest 0.4172 1.7956 0.5139 Extra Trees Regressor 0.3801 1.5778 0.6483 Gradient Boosting Regressor 0.3623 1.2062 0.6997 CNN 0.2148 0.9219 0.8016 CNN-LSTM 0.1759 0.8541 0.8647 TCN-BiLSTM-TPA 0.1034 0.6093 0.9382
[0169] Table 2 Ablation experimental results of different network structures
[0170] LSTM 0.2347 0.7354 0.8137 BiLSTM 0.2069 0.7137 0.8474 TCN-LSTM 0.1352 0.6481 0.8995 TCN-BiLSTM 0.1246 0.6339 0.9103 BiLSTM-TPA 0.1408 0.6505 0.8906 TCN-BiLSTM-TPA 0.1034 0.6093 0.9382
[0171] Please see Figure 7 The diagram shows a structural schematic of a computer device provided in an embodiment of this application. An embodiment of this application provides a computer device 400, including a processor 410 and a memory 420. The memory 420 stores a computer program executable by the processor 410. When the computer program is executed by the processor 410, it performs the method described above.
[0172] This application embodiment also provides a storage medium 430, on which a computer program is stored, and the computer program is executed by a processor 410 to perform the above method.
[0173] The storage medium 430 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0174] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. "A plurality of" means two or more, unless otherwise explicitly specified.
[0175] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0176] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0177] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.
[0178] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.
[0179] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, when implemented in hardware, any one or a combination of the following techniques known in the art can be used: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0180] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.
[0181] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.
Claims
1. A virtual measurement method for key quality indicators in a continuous correlation process of TFT-LCD, characterized in that, The method includes: Acquire production process data for at least two consecutive related processes in the TFT-LCD manufacturing process. The production process data includes process parameter data for each consecutive related process and a quality index sequence for at least one preceding related process. Perform process division and preprocessing on the production process data according to process identifiers and / or tool identifiers to form a process parameter sequence for each process. For the target associated process for which key quality indicators are to be virtually measured, the process parameter sequence of the target associated process is aligned with the quality indicator sequence of the preceding associated process on a time-node basis, and then merged into the associated process input sequence of the target associated process in a parallel input channel manner; the associated process input sequence is then divided into time windows and normalized. For the associated process input sequence, multi-scale temporal features are extracted through at least two dilated causal convolutional layers with different dilation rates. Bidirectional temporal modeling is performed on the multi-scale temporal features within the time window to obtain bidirectional temporal features. Temporal pattern features spanning multiple time nodes are extracted from the bidirectional temporal features, and attention weighting is performed based on the temporal pattern features to obtain weighted temporal features. Using key quality indicators corresponding to historical production process data as supervisory labels, a virtual measurement model for key quality indicators is trained based on the weighted time-series features. The TFT-LCD production process data to be measured is processed in the same way as the training phase, and then input into the trained virtual measurement model for key quality indicators. The virtual measurement results of key quality indicators corresponding to the target associated process are output.
2. The virtual measurement method for key quality indicators in the continuous correlation process of TFT-LCD according to claim 1, characterized in that, The preceding associated process is a sequential associated process that precedes the target associated process in the processing sequence, and whose process status or quality indicators have a data correlation with the key quality indicators of the target associated process. The target associated process refers to the continuous associated processes that currently require virtual measurement of key quality indicators; The quality index sequence is a sequence of key quality indicators of the preceding related processes formed according to the production sequence; The key quality indicators are continuous numerical indicators that reflect the processing quality of the target-related process or the quality of the final product and must be obtained through physical measurement.
3. The virtual measurement method for key quality indicators in the continuous correlation process of TFT-LCD according to claim 2, characterized in that, The process parameter data includes one or more of the following: temperature, gas flow rate, liquid flow rate, power, and process time. The step of dividing the production process data into processes based on process identifiers and / or tool identifiers includes: determining the corresponding process data group based on the tool identifier; The preprocessing includes: deleting null value columns, single value columns, and duplicate columns from the production process data; For zero-value data and / or missing-value data that are determined to be invalid, fill them in using the statistical values of non-zero valid data under the same process and tool category, based on the process and tool category to which the corresponding data belongs.
4. The virtual measurement method for key quality indicators in the continuous correlation process of TFT-LCD according to claim 1, characterized in that, The sequence of process parameters for each step includes: The importance of each candidate process parameter is evaluated, and a predetermined number of key process parameters are selected from the candidate process parameters of the corresponding process based on the ranking of the importance of each process. This is to construct the process parameter sequence of the corresponding process. The number and / or composition of candidate process parameters differ for different processes, and the composition of the key process parameter set corresponding to each process varies with the process; the feature importance evaluation is performed using the random forest algorithm.
5. The virtual measurement method for key quality indicators in the continuous correlation process of TFT-LCD according to claim 1, characterized in that, Let the target associated process be the i-th process, and let the sequence of process parameters for the target associated process be X. i The quality index sequence of the preceding associated process is Y. i-1 Then the associated process input sequence Z i Satisfy Z i =[X i ,Y i-1 The input vector z of the associated process input sequence at time node t i,t =[x i,t , y i-1,t ]; The time-node alignment is based on one or more of product identifier, batch identifier, and production sequence identifier to determine the correspondence between the same product or corresponding production samples. The time window is divided into sliding time windows of length L.
6. The virtual measurement method for key quality indicators in the continuous correlation process of TFT-LCD according to claim 1, characterized in that, The extraction of multi-scale temporal features is achieved by using at least two dilated causal convolutional layers with different dilation rates, and is implemented using a temporal convolutional network. The temporal convolutional network includes at least two cascaded dilated causal convolutional layers, each of which has a different dilation rate. Different dilation rates correspond to receptive fields at different time scales, so as to extract temporal dependency features at different time scales and form the multi-scale temporal features. The temporal convolutional network employs causal convolution, and the output of each of the dilated causal convolutional layers at the current time node is determined only by the current time node and the input data located before the current time node. The temporal convolutional network also includes a residual connection structure. For any dilated causal convolutional layer, the input of the dilated causal convolutional layer is added to the convolutional output of the dilated causal convolutional layer through a residual branch, and the added features are nonlinearly activated. When the feature dimensions of the input and the convolution output are inconsistent, the feature dimensions of the input in the residual branch are adjusted by convolution mapping before the addition is performed.
7. The virtual measurement method for key quality indicators in the continuous correlation process of TFT-LCD according to claim 1, characterized in that, The multi-scale temporal features are bidirectionally modeled within the time window, which is achieved using a bidirectional long short-term memory network. The bidirectional long short-term memory network includes a forward long short-term memory unit and a reverse long short-term memory unit. The multi-scale temporal features are input into the forward long short-term memory unit and the reverse long short-term memory unit respectively to obtain the forward hidden state sequence and the reverse hidden state sequence. The reverse long short-term memory unit only processes the multi-scale temporal features observed within the established time window, and does not involve production data generated in the future when virtual measurement is performed; The forward hidden state and the reverse hidden state at the same time node are spliced and fused to obtain the bidirectional hidden state at that time node, and the bidirectional hidden state at each time node constitutes the bidirectional temporal feature.
8. The virtual measurement method for key quality indicators in the continuous correlation process of TFT-LCD according to claim 1, characterized in that, The step of extracting temporal pattern features across multiple time nodes from the bidirectional temporal features and performing attention weighting based on the temporal pattern features is achieved using a temporal pattern attention mechanism. The implementation using the time-mode attention mechanism includes: A hidden state matrix is constructed based on the hidden states of the bidirectional temporal features at multiple time nodes, and the hidden state matrix is divided into time series corresponding to each state variable according to the dimension of the state variable. One-dimensional convolution is performed along the time dimension on the time series corresponding to each state variable to obtain the time pattern features corresponding to each state variable. The degree of correlation between each of the time pattern features and the hidden state at the current time node in the hidden state matrix is determined to obtain the attention score corresponding to each of the time pattern features. The attention scores are then normalized to obtain the attention weight corresponding to each of the time pattern features. Using the time pattern features as units, each time pattern feature is weighted and fused according to the attention weight to obtain a weighted time pattern feature, and the feature obtained by fusing the weighted time pattern feature with the bidirectional time series feature is used as the weighted time series feature.
9. The virtual measurement method for key quality indicators in the continuous correlation process of TFT-LCD according to claim 1, characterized in that, For multiple target associated processes to be virtually measured for key quality indicators, obtain the process parameter sequence of each target associated process and the quality indicator sequence of the preceding associated process corresponding to each target associated process. For each of the target associated processes, a corresponding associated process input sequence is constructed, and virtual measurement results of key quality indicators for the corresponding target associated processes are obtained based on each of the associated process input sequences. The number of process parameters in the process parameter sequences of each of the target associated processes are different, and there is no hierarchical quality index transmission chain between the target associated processes.
10. A virtual measurement system for key quality indicators in a continuous correlation process of TFT-LCD, characterized in that, The system includes: The data acquisition module is used to acquire production process data of at least two consecutive related processes in the TFT-LCD manufacturing process. The production process data includes process parameter data of each consecutive related process and a quality index sequence of at least one preceding related process. The data acquisition module is also used to divide and preprocess the production process data according to process identifiers and / or tool identifiers to form a process parameter sequence for each process. The associated process input construction module is used to align the process parameter sequence of the target associated process with the quality indicator sequence of the preceding associated process on a time-node basis for the target associated process to be virtually measured for key quality indicators, and merge them into the associated process input sequence of the target associated process in a parallel input channel manner; the associated process input construction module is also used to perform time window division and normalization processing on the associated process input sequence. The temporal feature processing module is used to extract multi-scale temporal features from the associated process input sequence through at least two dilated causal convolutional layers with different dilation rates, and to perform bidirectional temporal modeling on the multi-scale temporal features within the time window to obtain bidirectional temporal features; to extract temporal pattern features across multiple time nodes from the bidirectional temporal features, and to perform attention weighting based on the temporal pattern features to obtain weighted temporal features; The virtual measurement module is used to train a virtual measurement model of key quality indicators based on the weighted time-series features, using key quality indicators corresponding to historical production process data as supervision labels. The virtual measurement module is also used to process the TFT-LCD production process data to be measured in accordance with the data processing method consistent with the training stage, input it into the trained virtual measurement model of key quality indicators, and output the virtual measurement results of key quality indicators corresponding to the target associated process.