A multi-point virtual measurement method, system, device and storage medium

Abstract

The present invention provides a multi-point virtual measurement method, system, device and storage medium, which relates to the field of virtual measurement. The method flow is: obtaining the operating parameters of the production machine and the measurement values ​​of the measurement machine; performing standardization processing, data filling, data splicing and data combination on the operating parameters of the production machine and the measurement values ​​of the measurement machine to obtain a sample set; constructing a Seq2seq model, wherein the Seq2seq model includes an encoder and a decoder; training the Seq2seq model based on the sample set to obtain a trained Seq2seq model; and realizing real-time multi-point virtual measurement based on the trained Seq2seq model. The present invention adopts the Seq2seq model to perform multi-point virtual measurement, fully extracts the time dimension dependency of the operating parameters, and captures the correlation information of different measurement points at the same time, solving the problems of low prediction accuracy, poor applicability and inability to effectively process time series data in existing virtual measurement.

Description

Technical Field

[0001] The present invention relates to the field of virtual measurement, and in particular to a multi-point virtual measurement method, system, device and storage medium. Background Art

[0002] In the panel production process, in order to control the quality of the final product, multiple measuring points of measuring machines are usually set up in the production chain, and the key characteristics of the panel such as film thickness and brightness value are measured by measuring machines. Abnormalities in the panel production process can be discovered in time to reduce the production of unqualified products and improve the product yield. However, traditional random inspections have the following disadvantages: (1) It is necessary to set up multiple measuring machines and frequently conduct product inspections, which greatly increases production costs; (2) The measurement process is time-consuming, resulting in a decrease in the overall efficiency of the production line; (3) The equipment of the measuring machine is expensive and the maintenance cost is high; (4) The measuring points of the measuring machine are limited and comprehensive inspection cannot be achieved. With the development of artificial intelligence technology, virtual measurement technology has gradually become an important means to solve the above problems. The existing technology mainly realizes virtual measurement by establishing a machine learning model for the operating parameters and measurement values ​​of the production machine. However, these methods have the following disadvantages: (1) The virtual measurement accuracy is not ideal, especially when virtual measurement is performed at multiple points simultaneously; (2) The machine learning model has poor generalization ability and is not adaptable to changes in production processes; (3) It is unable to effectively process time series data and ignores the time characteristics of process parameters. Summary of the invention

[0003] The present invention provides a multi-point virtual measurement method, system, device and storage medium, which solve the problems of low prediction accuracy, poor applicability and inability to effectively process time series data in existing virtual measurement.

[0004] In a first aspect, an embodiment of the present invention provides a multi-point virtual measurement method, the method comprising the following process:

[0005] Obtain the operating parameters of the production machine and the measurement values ​​of the measuring machine;

[0006] Standardize, fill, splice and combine the operating parameters of the production machine and the measurement values ​​of the measuring machine to obtain a sample set;

[0007] Constructing a Seq2seq model, wherein the Seq2seq model includes an encoder and a decoder;

[0008] The Seq2seq model is trained based on the sample set to obtain a trained Seq2seq model;

[0009] Real-time multi-point virtual measurement is achieved based on the trained Seq2seq model.

[0010] In the above embodiment, the present invention adopts the Seq2seq model to perform multi-point virtual measurement, and predicts the measurement values ​​of multiple measurement points of the product according to the operating parameters of multiple production machines, thereby reducing the number of actual measurements, reducing measurement costs, and improving product production efficiency.

[0011] As some optional implementations of the present application, the encoder and decoder of the Seq2seq model are respectively bidirectional LSTM.

[0012] In the above embodiment, the present invention adopts a bidirectional LSTM as an encoder, which can well capture the dependencies in the time series; and adopts a bidirectional LSTM as a decoder, which allows the correlation between the measurement points to be transmitted through the time step through stepwise regression, so that the correlation between the measurement points is implied in the output hidden state of the encoder.

[0013] As some optional implementations of the present application, the process of standardizing, filling, splicing and combining the operating parameters of the production machine and the measurement values ​​of the measurement machine is as follows:

[0014] Standardize the operating parameters of production machines according to the process;

[0015] Perform data splicing on the operating parameters of different processes, and fill the spliced ​​operating parameters with fixed time length and dimension to obtain the final operating parameters of the production machine;

[0016] The operating parameters of the final production machine and the measured values ​​of the measuring points of the measuring machine are combined to obtain training samples, and a sample set is constructed based on a number of training samples.

[0017] In the above-mentioned embodiment, in order to facilitate subsequent modeling, the present invention needs to perform standardization processing, data filling, data splicing and data combination on the operating parameters of the production machine and the measurement values ​​of the measurement machine.

[0018] As some optional implementations of the present application, the calculation formula for standardizing the operating parameters of the production machine according to the process is as follows:

[0019] in, Indicates Products in Each process corresponds to the operating parameters of the production machine. Indicates The mean value of the operating parameters of the production machine corresponding to each process, Indicates The standard deviation of the operating parameters of the production machine corresponding to each process.

[0020] In the above-mentioned embodiment, the present invention can solve the problem that the convergence of the model is affected due to the large difference in the values ​​of different operating parameters through standardization processing.

[0021] As some optional implementations of the present application, the structural principle of the Seq2seq model is as follows:

[0022] The operating parameters of the production machine are embedded in words through a one-dimensional convolutional neural network, and combined with a process encoding about time steps;

[0023] The encoder is used to represent and extract the operating parameters of the word embedding and process encoding to obtain the hidden layer state, memory cell state and feature output of each time step, and the feature vector output of all time steps is performed through the linear layer;

[0024] The decoder uses the hidden layer state and memory cell state of the encoder at the last time step as the initial state, and gradually predicts the measurement value of each point of the measuring machine through autoregression, and outputs the predicted value through the linear layer;

[0025] The feature vector output of the linear layer of the encoder and the predicted value output of the linear layer of the decoder are added through linear mapping by means of residual connection to obtain the final predicted value.

[0026] In the above embodiment, the present invention adopts an autoregressive method to capture the measurement point information and gradually predicts the measurement value of each measurement point, thereby further improving the accuracy of virtual prediction; and referring to the idea of ​​residual connection, the hidden layer output of the encoder and the output of the decoder are linearly added, thereby effectively improving the accuracy of virtual measurement.

[0027] As some optional implementations of the present application, the calculation formula for the decoder to gradually predict the measurement value of each point of the measuring machine by autoregression is as follows:

[0028] in, Indicates the point number of the measuring machine. and Respectively represent the decoder in The hidden layer state and memory cell state of time steps, Indicates that the decoder is The feature output of each time step, LSTM represents the decoder, and Linear represents the linear layer.

[0029] As some optional implementation methods of the present application, the process of training the Seq2seq model based on the sample set is as follows:

[0030] The Seq2seq model is trained based on the sample set, and the optimizer is used to optimize the model during the model training stage;

[0031] The Seq2seq model is selected and its performance is tested based on the sample set, and the root mean square error is used as the loss function to evaluate the model performance during the model training phase.

[0032] In the above embodiment, the present invention trains the Seq2seq model so that the Seq2seq model can fully extract the time dimension dependency of the operating parameters and capture the correlation information of different measurement points; and through model optimization and model performance evaluation, it can effectively prevent model overfitting and improve the generalization ability of the model.

[0033] As some optional implementations of the present application, the calculation formula for evaluating the performance of the model using the root mean square error as the loss function is as follows:

[0034] in, represents the loss function value, Indicates the quantity of products. Indicates the number of points of the measuring machine. Indicates Product No. The true value of the measurement value of a measuring machine, Indicates Product The predicted value of the measurement value of a measuring machine.

[0035] In the above embodiment, the present invention uses the mean square error as the loss function to evaluate the overall deviation between the model prediction value and the true value.

[0036] In a second aspect, the present invention provides a multi-point virtual measurement system, the system comprising:

[0037] A data acquisition unit, the data acquisition unit is used to acquire the operating parameters of the production machine and the measurement values ​​of the measurement machine;

[0038] A data processing unit, the data processing unit is used to perform standardization processing, data filling, data splicing and data combination on the operating parameters of the production machine and the measurement values ​​of the measurement machine to obtain a sample set;

[0039] A model building unit, wherein the model building unit is used to build a Seq2seq model, wherein the Seq2seq model includes an encoder and a decoder;

[0040] A model training unit, wherein the model training unit performs model training on the Seq2seq model based on the sample set to obtain a trained Seq2seq model;

[0041] A virtual measurement unit realizes real-time multi-point virtual measurement based on a trained Seq2seq model.

[0042] In a third aspect, the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the multi-point virtual measurement method when executing the computer program.

[0043] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements the multi-point virtual measurement method when executed by a processor.

[0044] The beneficial effects of the present invention are as follows:

[0045] 1. The present invention adopts the Seq2seq model to perform multi-point virtual measurement, fully extracts the time dimension dependency of the operating parameters, and simultaneously captures the correlation information of different measurement points, effectively improving the accuracy of virtual measurement.

[0046] 2. The present invention uses an autoregressive method to capture the measurement point information and gradually predicts the measurement value of each measurement point, thereby further improving the accuracy of virtual measurement.

[0047] 3. The present invention refers to the idea of ​​residual connection and linearly adds the hidden layer output of the encoder and the output of the decoder, which effectively improves the accuracy of virtual measurement.

[0048] 4. The present invention adopts a process coding method to fully utilize the process information of the operating parameters, effectively improves the semantic representation capability, and further improves the accuracy of virtual measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 is a flow chart of a multi-point virtual measurement method according to an embodiment of the present invention;

[0051] Figure 2 is a schematic diagram of preprocessing of operating parameters according to an embodiment of the present invention;

[0052] Figure 3 It is a schematic diagram of the structure of the Seq2seq model of an embodiment of the present invention;

[0053] Figure 4 It is a comparison chart of the prediction results of 4 products in Example 1 of the present invention. DETAILED DESCRIPTION

[0054] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] In order to solve the problems of low prediction accuracy, poor applicability and inability to effectively process time series data in existing virtual measurement, the present invention provides a multi-point virtual measurement method, please refer to Figure 1 , Figure 1 The flowchart of the multi-point virtual measurement method is as follows:

[0056] (1) Obtain the operating parameters of the production machine and the measurement values ​​of each point of the measurement machine.

[0057] In the production chain of panels, the operating parameters of each production machine, including process parameters and processing parameters, will be recorded during the processing of the product. The process parameters include but are not limited to process conditions such as temperature, pressure, and time; the processing parameters include but are not limited to equipment operating parameters, such as speed and power. After the production machine is processed, the key characteristic values ​​of multiple measurement points of the product will be measured on the measuring machine to obtain the measurement values ​​of multiple points.

[0058] Specifically, the operating parameters of the production machine are determined according to the actual selected production machine. The embodiment of the present invention does not limit this, and multiple parameters of multiple production machines can be selected according to the needs of the measuring machine. The measurement value of a measurement point of the measuring machine can be the measurement value of a coordinate position or the average measurement value of an area, and the measurement objects include but are not limited to film thickness, brightness, optical constants, etc. The embodiment of the present invention models the measurement values ​​of multiple measurement points of a product, so it is necessary to collect data information of multiple measurement points. Specifically, the data finally obtained is Products production machines, and the operating parameters selected for each production machine are operating parameters, each product The length of the time series data of the operating parameters of each production machine is , the true value of the measurement value of each product corresponding to the measuring machine is , Represents an array, Indicates the number of measuring points of the measuring machine.

[0059] (2) Preprocessing the operating parameters of the production machine and the measurement values ​​of the measuring machine to obtain a sample set; wherein the preprocessing includes standardization, data filling, data splicing and data combination.

[0060] In the embodiment of the present invention, the process of standardizing, filling and splicing the operating parameters of the production machine and the measured values ​​of each measuring point of the measuring machine is as follows:

[0061] (2.1) Since different production machines perform different processes, the operating parameters of different production machines may have different values. Large differences in the values ​​of different operating parameters will greatly affect the convergence of the model. Therefore, it is necessary to standardize the operating parameters of the production machines according to the process or machine to facilitate subsequent modeling.

[0062] Specifically, the calculation formula for standardizing the operating parameters of the production machine according to the process is as follows:

[0063]

[0064] in, Indicates Products in Each process corresponds to the operating parameters of the production machine. Indicates The mean value of the operating parameters of the production machine corresponding to each process, Indicates The standard deviation of the operating parameters of the production machine corresponding to each process.

[0065] (2.2) Since the processing time of each process is different and the number of operating parameters selected for each process is different, it is necessary to fill the operating parameters of the data splicing to the specified time length and specify dimensions , to obtain the final operating parameters of the production machine, recorded as , Indicates the number of running parameters, represents the time step, see Figure 2 , Figure 2 This is a schematic diagram of the operating parameter preprocessing. The "0" in the figure is the 0 after filling. Here, the final operating parameter is a vector, and the filling quantity of each data is inconsistent.

[0066] (2.3) The operating parameters of the final production machine and the measurement values ​​of the measurement points of the measurement machine are combined to obtain training samples, and a sample set is constructed based on several training samples.

[0067] (3) Constructing a Seq2seq model; wherein the Seq2seq model includes but is not limited to an encoder and a decoder.

[0068] In the production chain, the product is first processed on multiple production machines and then measured on the measuring machine; therefore, it is necessary to establish a model that uses multi-dimensional time series as input data (operating parameters of the production machine) and multi-point measurement values ​​as output data (measurement values ​​of the measuring machine). The embodiment of the present invention adopts a sequence model for modeling to fully extract the dependency of parameters in the time dimension. (T represents the time step and m represents the number of parameters). Bidirectional LSTM can be used for modeling. Bidirectional LSTM is a recurrent neural network that can capture the dependencies in time series well.

[0069] For the output of multiple point measurement values, the existing technology usually connects a fully connected layer after the LSTM output to map the LSTM output to the dimension of the measurement value. However, after the same product is processed through multiple steps, different points have spatial position relationships. Since there is a certain correlation between the measurement values ​​of different measurement points, a simple linear layer cannot capture this positional correlation.

[0070] Therefore, the embodiment of the present invention adopts a Seq2seq model. The encoder of the Seq2seq model extracts the representation of the operating parameters. The decoder also uses a bidirectional LSTM structure, and uses the hidden layer state of the encoder at the last time step as the initial state of the decoder. At the same time, the measurement value of each measurement point is gradually predicted through autoregression.

[0071] Specifically, the structural principle of the Seq2seq model is as follows:

[0072] (3.1) The operating parameters of the production machine (step1, step2, ..., step n ) is used for word embedding (Value Embedding), and the one-dimensional convolutional coding maps the original parameter dimension m to a higher dimension, so that the representation of each time step contains both process information and parameter information; and combined with a process encoding (step embedding) about the time step, please refer to Figure 3 , Figure 3 Schematic diagram of the structure of the Seq2seq model.

[0073] (3.2) The encoder (LSTM) is used to represent and extract the operating parameters of the word embedding and process encoding to obtain the hidden layer state, memory cell state and feature output of each time step, and the feature vector output of all time steps is performed through the linear layer (Linear).

[0074] (3.3) The decoder uses the hidden state and memory cell state of the encoder (LSTM) at the last time step as the initial state, and gradually predicts the measurement value of each measurement point (cell) of the measurement machine through autoregression, and outputs the predicted value through the linear layer (Linear); the decoder uses the stepwise regression of the bidirectional LSTM to transfer the correlation between the measurement points through the time step, so that the output hidden state of the encoder contains the correlation between the measurement points.

[0075] Specifically, the calculation formula of the decoder gradually predicting the measurement value of each measurement point of the measurement machine by autoregression is as follows:

[0076]

[0077] in, Indicates the serial number of the measuring machine point. and Respectively represent the decoder in The hidden layer state and memory cell state of time steps, Indicates that the decoder is The feature output of each time step, LSTM represents the encoder, and Linear represents the linear layer.

[0078] (3.4) Considering that the hidden state of the encoder at the last time step may forget part of the encoding information of the input data in the stepwise regression of the decoder, the embodiment of the present invention adds the feature vector output of the linear layer of the encoder and the prediction value output of the linear layer of the decoder through linear mapping by means of residual connection to obtain the final prediction value.

[0079] (4) Perform model training and model testing on the Seq2seq model based on the sample set to obtain a trained Seq2seq model.

[0080] In the embodiment of the present invention, the sample set is randomly divided into a training set, a validation set, and a test set according to a certain ratio. The training set is used for model training, the validation set is used for model selection, and the test set is used to evaluate the generalization performance of the model. The embodiment of the present invention divides the sample set in a ratio of 8:1:1, and the ratio can be determined according to actual data.

[0081] Specifically, the process of training the Seq2seq model based on the sample set is as follows:

[0082] (4.1) The Seq2seq model is trained based on the training set, and the Adam optimizer is used to optimize the model during the model training stage.

[0083] (4.2) Model selection and performance testing of the Seq2seq model are performed based on the validation set and test set. Considering that each product has multiple measurement points, the predicted value of each product needs to be evaluated. Therefore, the root mean square error is used as the loss function to evaluate the model performance during the model training stage.

[0084] Specifically, the calculation formula for evaluating the performance of the model using the root mean square error as the loss function is as follows:

[0085]

[0086] in, Indicates the quantity of products. Indicates the number of points of the measuring machine. Indicates Product No. The true value of the measurement value of a measuring machine, Indicates Product No. The predicted value of the measurement value of a measuring machine.

[0087] During the model training process, each model training process will be trained on the training set, and the model performance will be evaluated on the validation set. When the loss on the validation set no longer decreases, the model training is stopped, and the model with the best performance on the validation set is selected as the final model. In order to prevent the model from overfitting, the embodiment of the present invention adopts the following strategy:

[0088] ① Early stopping strategy: When the loss on the validation set does not decrease for multiple consecutive model training processes, the model training is terminated early. In the embodiment of the present invention, the validation threshold is set to 10, that is, when the loss on the validation set does not decrease for 10 consecutive model training processes, the model training is stopped.

[0089] ② Learning rate decay: As the model training progresses, the learning rate is gradually reduced so that the model can converge to a better local optimal solution. The embodiment of the present invention uses an Adam optimizer with decay, the initial learning rate is set to 0.001, and the learning rate decays to 0.1 times the original value every 50 training processes.

[0090] ③ Regularization: Add a Dropout layer to randomly discard some neurons to prevent the model from over-relying on certain features. The embodiment of the present invention sets the Dropout rate to 0.1.

[0091] Through the above strategies, the model can be effectively prevented from overfitting and the generalization ability of the model can be improved. After the model training is completed, the performance of the model is evaluated using the test set to ensure that the model has good generalization ability. The indicators for verifying the performance of the model in the embodiment of the present invention include but are not limited to root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). Root mean square error (RMSE) and mean absolute error (MAE) can evaluate the overall deviation between the model prediction value and the true value. The smaller the overall deviation, the better the model performance.

[0092] (5) Real-time virtual measurement based on the trained Seq2seq model.

[0093] After completing the model training, save the model with the best performance on the validation set locally. When making online predictions, follow these steps:

[0094] (5.1) Load the locally saved optimal model parameters, including the parameters of the encoder and decoder, as well as the standardized processing parameters (mean and standard deviation) used in data preprocessing.

[0095] (5.2) Preprocess the production parameters collected online in real time: The standardization calculation formula standardizes the operating parameters of each process, fills the parameters of different processes to the specified time length, and splices the processed operating parameters into complete time series data.

[0096] (5.3) The data to be processed is input into the Seq2seq model for prediction, the predicted value is obtained, and then the predicted value is restored to the dimension of the original measured value.

[0097] Through the above steps, real-time virtual measurement of products on the production line can be achieved, and the measurement values ​​of each point can be predicted in time, providing a basis for quality control of the production process. The predicted value can be displayed and monitored through the production management system. When the predicted value exceeds the preset threshold, the production management system can promptly alarm and remind relevant personnel to handle it.

[0098] The following describes the detailed steps of the embodiment of the present invention by taking the product measurement of a liquid crystal panel as an example:

[0099] S10: Obtaining the operating parameters of the production machines and the measurement values ​​of the measurement machines in the LCD panel production chain.

[0100] During the production process of liquid crystal panel products, a measuring machine needs to be set up behind the production machine to inspect the processing conditions of the previous process to ensure product quality. The embodiment of the present invention needs to measure certain key characteristic values ​​of multiple measuring points of the product to ensure product quality.

[0101] The operating parameters of the liquid crystal panel described in the embodiment of the present invention include process parameters and processing parameters. The process parameters include but are not limited to the pressure applied when pressing the optical film, panel cutting parameters, baking time, baking temperature, ambient temperature, developer dosage, etc.; the processing parameters include but are not limited to the parameters of the processing tools, such as the rotation speed of the etching tool, the speed of the cutting knife, etc. The key characteristic values ​​of a measurement point include but are not limited to film thickness, brightness, and optical constants. The data collected in the embodiment of the present invention includes 239 liquid crystal panels, each of which undergoes 9 processes, 27 process parameters for each process, and a total of 900 to 1000 time steps. Each product on the measuring machine has measurement values ​​at 28 measurement points. Some examples of operating data of the production machine are shown in Table 1, and examples of measurement data of the product are shown in Table 2.

[0102]

[0103] Table 1 is an example of some operating data of a product on a production machine

[0104]

[0105] Table 2 is an example of measurement data of multiple measurement points

[0106] S20: Standardize, fill in, concatenate and combine the operating parameters of the production machine and the measurement values ​​of the measurement machine to obtain a sample set.

[0107] After collecting the operating parameters of the product's production machine, the parameters need to be processed in a series of steps to form a complete sample set. The specific process is as follows:

[0108] S201: Standardize the operation parameters of each process using a standardized calculation formula.

[0109] S202: Since the processing time of each product in each process is different and the number of operating parameters selected in each process is different, the time series needs to be filled to the same length and the same dimension; the specific process is as follows:

[0110] ① Calculate the longest time series length among all products and the maximum number of parameters .

[0111] ②For the length less than the time series length A certain number of 0s are filled at the end of the sequence. Other methods include padding and truncation. Truncation may lose some information, so we just fill it here.

[0112] ③For data dimensions less than the maximum number of parameters A sequence filled with 0s to the maximum number of parameters .

[0113] S203: Construct sample data. Organize the processed data in the following format: The dimension of input data X is ,in, is the batch size, is the length of the time series after alignment, is the number of process parameters, and the dimension of the output data Y is , It is the number of measuring points of the measuring machine.

[0114] Through the above steps, the original data set is converted into a sample set in a standard format that can be used for model training.

[0115] S30: Construct a Seq2seq model, wherein the Seq2seq model includes an encoder and a decoder.

[0116] The encoder involved in the embodiment of the present invention is used to extract the operating parameter representation, and the decoder is used to learn the correlation relationship between the points. Specifically, the encoder and the decoder both adopt a bidirectional LSTM structure, which can extract deeper feature representation.

[0117] S40: Perform model training on the Seq2seq model based on the sample set to obtain a trained Seq2seq model.

[0118] Due to the limited amount of data in the sample set, in order to fully demonstrate the effectiveness of the model, the 5-fold cross-validation method is used for model training and verification. The specific steps are as follows:

[0119] ① Sample set division: The entire sample set is randomly divided into 5 subsets of equal size. In each round of cross-validation, 80% of the data from 4 subsets is used as the training set (153 samples), 20% of the data is used as the validation set (38 samples), and the remaining 1 subset is used as the test set (48 samples); this can make full use of limited data while ensuring the reliability of model evaluation.

[0120] ② Model training: For each fold of model training process, the model is trained on the training set, and the model performance is evaluated on the validation set after each model training process. An early stopping strategy is adopted. When the loss corresponding to the validation set has not improved for 10 consecutive training processes, the training is stopped and the model parameters with the best performance on the validation set are recorded.

[0121] ③Model evaluation: For each fold of the test process, use the saved optimal model parameters to make predictions on the validation set, calculate the evaluation indicators on the test set, and record the evaluation results of each test.

[0122] S50: Real-time virtual measurement based on the trained Seq2seq model.

[0123] Use the trained model to make predictions on the validation set to obtain virtual measurement values. In addition, in order to verify the effectiveness of the method involved in the embodiment of the present invention, it is compared with the traditional LSTM model and the seq2seq model without autoregression:

[0124] (1) LSTM model: The structure of the model is as follows Figure 3 The encoder structure on the left.

[0125] (2) Seq2seq model: This model has the same structure as the model involved in the embodiment of the present invention, but does not adopt an autoregressive approach.

[0126] The three models are evaluated using the root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) evaluation indicators. The results are shown in Table 3:

[0127]

[0128] Table 3 shows the performance comparison of different models (average value of 5-fold cross validation)

[0129] As can be seen from the table, the autoregression-based seq2seq model proposed in the embodiment of the present invention is superior to the traditional LSTM model and the ordinary seq2seq model in various evaluation indicators. Specifically, in terms of the RMSE indicator, the autoregression-based seq2seq model is 7.36 lower than the LSTM model and 3.488 lower than the ordinary seq2seq model; in terms of the MAE indicator, the autoregression-based seq2seq model is 5.866 lower than the LSTM model and 2.606 lower than the ordinary seq2seq model, and the predicted MAPE is also reduced from 1.76% to 1.56%. This shows that the seq2seq structure used in the embodiment of the present invention combines the correlation relationship of the points, which can improve the accuracy of virtual measurement, and the use of autoregression can further improve the accuracy of virtual measurement.

[0130] The prediction results of the autoregressive seq2seq model proposed in the embodiment of the present invention are obtained by randomly selecting 4 products from the validation set. Figure 4 , Figure 4 This is a comparison chart of the prediction results of 4 products, in which the horizontal axis represents the point number and the vertical axis represents the measurement value. It can be seen from the figure that the model proposed in the embodiment of the present invention can better fit the measurement value of each measurement point, and the change trends of multiple measurement points are basically the same, indicating that the model proposed in the embodiment of the present invention can better capture the correlation between different measurement points.

[0131] In summary, the present invention extracts the time dependence of production parameters through the encoder, and uses the decoder to capture the correlation between multiple point measurement values, thereby improving the prediction accuracy of multi-point measurement, and can achieve end-to-end virtual prediction. It has high prediction accuracy and real-time performance, which can effectively help factories perform real-time full inspections and improve the yield rate of final products.

[0132] In addition, in one embodiment, based on the same inventive concept as the above embodiment, the embodiment of the present invention provides a multi-point virtual measurement system, the system corresponds one-to-one with the method of embodiment 1, and the system includes:

[0133] A data acquisition unit, the data acquisition unit is used to acquire the operating parameters of the production machine and the measurement values ​​of the measurement machine;

[0134] A data processing unit, the data processing unit is used to perform standardization processing, data filling, data splicing and data combination on the operating parameters of the production machine and the measurement values ​​of the measurement machine to obtain a sample set;

[0135] A model building unit, wherein the model building unit is used to build a Seq2seq model, wherein the Seq2seq model includes an encoder and a decoder;

[0136] A model training unit, wherein the model training unit performs model training on the Seq2seq model based on the sample set to obtain a trained Seq2seq model;

[0137] A virtual measurement unit realizes real-time multi-point virtual measurement based on a trained Seq2seq model.

[0138] It should be noted that each unit in the multi-point virtual measurement system in this embodiment corresponds one-to-one to each step in the multi-point virtual measurement method in the aforementioned embodiment. Therefore, the specific implementation method and technical effects achieved in this embodiment can refer to the implementation method of the aforementioned multi-point virtual measurement method, which will not be repeated here.

[0139] In addition, in one embodiment, the present application also provides a computer device, which includes a processor, a memory, and a computer program stored in the memory, and the computer program implements the method in the aforementioned embodiment when executed by the processor.

[0140] In addition, in one embodiment, the present application further provides a computer storage medium, on which a computer program is stored, and the computer program implements the method in the aforementioned embodiment when executed by a processor.

[0141] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or various devices including one or any combination of the above memories. The computer may be various computing devices including intelligent terminals and servers.

[0142] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0143] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file storing other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0144] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0145] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0146] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0147] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a multimedia terminal device (which can be a mobile phone, a computer, a television receiver, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0148] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A multi-point virtual measurement method, characterized in that: The method The following processes are included: Obtain the operating parameters of the production machine and the measurement values ​​of the measuring machine; Standardize, fill, splice and combine the operating parameters of the production machine and the measurement values ​​of the measuring machine to obtain a sample set; Constructing a Seq2seq model, wherein the Seq2seq model includes an encoder and a decoder; The Seq2seq model is trained based on the sample set to obtain a trained Seq2seq model; Realize real-time multi-point virtual measurement based on the trained Seq2seq model; The structural principle of the Seq2seq model is as follows: The operating parameters of the production machine are embedded in words through a one-dimensional convolutional neural network, and combined with a process encoding about time steps; The encoder is used to represent and extract the operating parameters of the word embedding and process encoding to obtain the hidden layer state, memory cell state and feature output of each time step, and the feature vector output of all time steps is performed through the linear layer; The decoder uses the hidden layer state and memory cell state of the encoder at the last time step as the initial state, and gradually predicts the measurement value of each point of the measuring machine through autoregression, and outputs the predicted value through the linear layer; The feature vector output of the linear layer of the encoder and the predicted value output of the linear layer of the decoder are added through linear mapping by means of residual connection to obtain the final predicted value.

2. A multi-point virtual measurement method according to claim 1, characterized in that: The encoder and decoder of the Seq2seq model are bidirectional LSTMs respectively.

3. The multi-point virtual measurement method according to claim 1, characterized in that: The process of standardizing, filling, splicing and combining the operating parameters of the production machine and the measurement values ​​of the measuring machine is as follows: Standardize the operating parameters of production machines according to the process; Perform data splicing on the operating parameters of different processes, and fill the spliced ​​operating parameters with fixed time length and dimension to obtain the final operating parameters of the production machine; The operating parameters of the final production machine and the measured values ​​of the measuring points of the measuring machine are combined to obtain training samples, and a sample set is constructed based on a number of training samples.

4. A multi-point virtual measurement method according to claim 3, characterized in that: The calculation formula for standardizing the operating parameters of the production machine according to the process is as follows: Among them, X ik represents the operating parameters of the production machine corresponding to the k-th process of the i-th product, represents the mean value of the operating parameters of the production machine corresponding to the kth process, Represents the standard deviation of the operating parameters of the production machine corresponding to the kth process.

5. A multi-point virtual measurement method according to claim 4, characterized in that: The calculation formula of the decoder to gradually predict the measurement value of each point of the measuring machine by autoregression is as follows: Where, j (j = 1, 2, ..., M) represents the point number of the measuring machine, h j and c j denote the hidden layer state and memory cell state of the decoder at the jth time step, respectively. j represents the feature output of the decoder at the jth time step, LSTM represents the decoder, and Linear represents the linear layer.

6. A multi-point virtual measurement method according to claim 1, characterized in that: The process of model training for the Seq2seq model based on the sample set is as follows: The Seq2seq model is trained based on the sample set, and the optimizer is used to optimize the model during the model training stage; The Seq2seq model is selected and its performance is tested based on the sample set, and the root mean square error is used as the loss function to evaluate the model performance during the model training phase.

7. A multi-point virtual measurement method according to claim 6, characterized in that: The calculation formula for evaluating the performance of the model using the root mean square error as the loss function is as follows: Where L represents the loss function value, n represents the number of products, M represents the number of points of the measuring machine, and y ij represents the true value of the measurement value of the jth measuring machine of the i-th product, It represents the predicted value of the measurement value of the jth measuring machine for the i-th product.

8. A multi-point virtual measurement system, characterized in that: The system comprises: A data acquisition unit, the data acquisition unit is used to acquire the operating parameters of the production machine and the measurement values ​​of the measurement machine; A data processing unit, the data processing unit is used to perform standardization processing, data filling, data splicing and data combination on the operating parameters of the production machine and the measurement values ​​of the measurement machine to obtain a sample set; A model building unit, wherein the model building unit is used to build a Seq2seq model, wherein the Seq2seq model includes an encoder and a decoder; A model training unit, wherein the model training unit performs model training on the Seq2seq model based on the sample set to obtain a trained Seq2seq model; A virtual measurement unit, wherein the virtual measurement unit realizes real-time multi-point virtual measurement based on a trained Seq2seq model; The structural principle of the Seq2seq model is as follows: The operating parameters of the production machine are embedded in words through a one-dimensional convolutional neural network, and combined with a process encoding about time steps; The encoder is used to represent and extract the operating parameters of the word embedding and process encoding to obtain the hidden layer state, memory cell state and feature output of each time step, and the feature vector output of all time steps is performed through the linear layer; The decoder uses the hidden layer state and memory cell state of the encoder at the last time step as the initial state, and gradually predicts the measurement value of each point of the measuring machine through autoregression, and outputs the predicted value through the linear layer; The feature vector output of the linear layer of the encoder and the predicted value output of the linear layer of the decoder are added through linear mapping by means of residual connection to obtain the final predicted value.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the multi-point virtual measurement method described in any one of claims 1-7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the multi-point virtual measurement method described in any one of claims 1 to 7 is implemented.

Images