JP-7857447-B2 - Method for controlling an elasticity imparting process, method for training a control network, and related apparatus.

Inventors

• ポン ワン
• シャンタオ ポン
• イーポー チウ
• チュンリャン チン
• チョンリャン ウー
• ターコー リー
• フォン シュイ

Assignees

• チョーチアン ヘンイー ペトロケミカル カンパニー，リミテッド
• ハイニン ヘンイー ニュー マテリアルズ カンパニー リミテッド

Dates

Publication Date

20260512

Application Date

20250122

Priority Date

20240125

Claims (16)

1. A method for controlling the process of imparting elasticity, The process involves sequentially collecting predetermined control parameters at multiple points in time during the elasticity imparting process flow to obtain a first control parameter sequence, The first control parameter sequence is input into a parameter prediction model to predict the parameters at multiple future time points, Selecting the target parameter for the target time from the prediction parameters of the aforementioned multiple time points, Based on control parameters at multiple time points before and after the aforementioned target time point, a second control parameter sequence including the aforementioned target parameter is constructed. The process involves processing the second control parameter sequence based on the indicator prediction model and obtaining the indicator prediction result. If the indicator prediction result does not meet the desired value, the indicator prediction result is input into the data generation model to obtain the desired control parameter for the target time point , The aforementioned indicator prediction results include at least one of the grade, pass rate, full winding rate, and dyeing uniformity of the wound yarn package , and are a method for controlling the elasticity imparting process.
2. Processing the second control parameter sequence based on the aforementioned indicator prediction model and obtaining the indicator prediction result is: Based on the control parameters at each point in time of the second control parameter sequence, a corresponding elasticity-granting process architecture chart is generated, and multiple elasticity-granting process architecture charts are obtained. The method according to claim 1, further comprising inputting the plurality of elasticity imparting process architecture charts into the indicator prediction model to obtain the indicator prediction results.
3. Based on the control parameters at each point in time of the second control parameter sequence, a corresponding elasticity-granting process architecture chart is generated, and multiple elasticity-granting process architecture charts are obtained. This includes performing the following operations on each control parameter at each point in the second control parameter sequence to generate a corresponding elasticity-granting process architecture chart: The aforementioned operation is, The process involves performing a normalization operation on each of the multiple sub-parameters in the control parameter, mapping the sub-parameters to a predetermined value interval, and obtaining the normalized values of the sub-parameters. This includes scaling the processing path in the POY yarn elasticity imparting process proportionally to the initialization image, and obtaining an elasticity imparting process architecture chart corresponding to the said time point. The method according to claim 2, wherein the values of pixel points other than the processing path in the elasticity imparting process architecture chart are set to default values, the points on the processing path corresponding to the subparameters in the control parameters are set to the normalized value of the subparameters, and the points on the processing path other than the subparameters in the control parameters are set to target values different from the default values.
4. Inputting the aforementioned multiple elasticity-granting process architecture charts into the indicator prediction model to obtain the indicator prediction results is, Based on a short-term feature extraction module constructed by the attention mechanism in the indicator prediction model, features are extracted within a first feature extraction range in the plurality of elasticity-granting process architecture charts to obtain short-term features, wherein the first feature extraction range includes elasticity-granting process architecture charts at n time points centered on the elasticity-granting process architecture chart at the target time point, where n is a positive integer, and the short-term feature extraction module uses the elasticity-granting process architecture chart at the target time point as a query feature, and the elasticity-granting process architecture charts at the n time points excluding the target time point as key features and value features, to obtain the short-term features, and Based on a long-term feature extraction module constructed by the attention mechanism in the indicator prediction model, features are extracted within a second feature extraction range in the plurality of elasticity-granting process architecture charts to obtain long-term features, wherein the second feature extraction range includes m elasticity-granting process architecture charts at various time points centered on the elasticity-granting process architecture chart at the target time point, where m is a positive integer greater than n, and the long-term feature extraction module divides the m elasticity-granting process architecture charts at various time points using a sliding window mechanism, determines pooling information within each sliding window, constructs key features and value features necessary for the attention mechanism based on the pooling information of the plurality of sliding windows, and obtains the long-term features based on the key features, value features, and query features. Using a fusion module, a fusion operation is performed on the long-term features and the short-term features to obtain the fused features. The process involves processing the fused features based on the prediction module in the indicator prediction model to obtain the indicator prediction result, The method according to claim 2, including the method described in claim 2.
5. The aforementioned indicator prediction model further includes a supplementary feature extraction module, The above method further, The following operations are performed based on the aforementioned supplementary feature extraction module to obtain supplementary features, The aforementioned operation is, The data is analyzed for the second control parameter sequence, and the cumulative difference of the parameters at each time point in the second control parameter sequence relative to the first time point in the second control parameter sequence is obtained to obtain a cumulative difference sequence. Based on the cumulative difference sequence, the time length relative to the first time point at which the cumulative difference is greater than a predetermined threshold is determined, and a time length sequence is obtained. Features are extracted from the second control parameter sequence, the cumulative difference sequence, and the time length sequence, and the supplementary features are obtained. Includes, Performing a fusion operation on the long-term features and the short-term features using the fusion module to obtain fused features is: The method according to claim 4, comprising performing a fusion operation on the long-term features, the short-term features and the supplementary features using a fusion module to obtain the fused features.
6. A method for training a control network for an elasticity-granting process, The control network includes a parameter prediction subnetwork, an indicator prediction subnetwork, and a data generation subnetwork. The aforementioned method, Obtaining a first training sample containing predetermined control parameters at multiple time points in the elasticity imparting process, The first training sample is input to the parameter prediction subnetwork to obtain the parameter prediction value for at least one future time point, The process involves processing the parameter prediction values based on the aforementioned indicator prediction subnetwork to obtain the predicted indicator, The training loss is determined based on the difference between the predicted indicator and the actual indicator, and the difference between the predicted parameter value and the actual parameter value. Adjusting the parameter prediction subnetwork and the indicator prediction subnetwork based on the training loss, If the training convergence condition is met, the parameter prediction model corresponding to the parameter prediction subnetwork and the index prediction model corresponding to the index prediction subnetwork are obtained. Based on the parameter prediction model and the indicator prediction model, the data generation subnetwork is trained to obtain a data generation model. Includes, The data generation model is for generating desired control parameters for the elasticity imparting machine based on the index prediction results output from the index prediction model. A method for training a control network for an elasticity imparting process , wherein the indicator prediction results include at least one of the grade, pass rate, full winding rate, and dye uniformity of the wound yarn package .
7. The aforementioned data generation subnetwork includes a training queue generator and a training queue discriminator, To obtain a data generation model by training the data generation subnetwork based on the parameter prediction model and the indicator prediction model, This includes performing multiple training sessions on the aforementioned data generation subnetwork. Each of the aforementioned training sessions will involve performing the following operations: The aforementioned operation is, The true control parameter is input to the training-waiting discriminator, and a first discrimination result for the true control parameter is obtained by the training-waiting discriminator. The training wait generator is used to generate false control parameters, The false control parameters are input to the training-waiting classifier to obtain a second classification result. Adjusting the model parameters of the training-waiting classifier based on the first and second classification results, The model parameters of the aforementioned training wait discriminator are fixed, and the aforementioned training wait generator is trained. The method according to claim 6, including the method described in claim 6 .
8. A control device for the elasticity imparting process, A collection unit for sequentially collecting predetermined control parameters at multiple time points in the elasticity imparting process flow to obtain a first control parameter sequence, A parameter prediction unit for inputting the first control parameter sequence into a parameter prediction model to predict predictable parameters at multiple future time points, A selection unit for selecting a target parameter for a target time from the multiple prediction parameters at the aforementioned time points, A construction unit for constructing a second control parameter sequence including the target parameter based on control parameters at multiple time points before and after the target time point, An indicator prediction unit for processing the second control parameter sequence based on an indicator prediction model and obtaining indicator prediction results, If the indicator prediction result does not meet the desired value, the data generation unit inputs the indicator prediction result into a data generation model to obtain the desired control parameter for the target time point , The indicator prediction results include at least one of the grade, pass rate, full winding rate, and dyeing uniformity of the wound yarn package , and are control devices for an elasticity imparting process.
9. The aforementioned indicator prediction unit is A generation subunit for generating a corresponding elasticity-granting process architecture chart based on the control parameters at each point in time of the second control parameter sequence, and for obtaining multiple elasticity-granting process architecture charts, A prediction subunit for inputting the multiple elasticity-granting process architecture charts into the indicator prediction model and obtaining the indicator prediction results, The apparatus according to claim 8, including the apparatus described in claim 8 .
10. The generation subunit is used to generate the corresponding elasticity-granting process architecture chart by performing the following operations on the control parameters at each point in the second control parameter sequence: The aforementioned operation is, The process involves performing a normalization operation on each of the multiple sub-parameters in the control parameter, mapping the sub-parameters to a predetermined value interval, and obtaining the normalized values of the sub-parameters. This includes scaling the processing path in the POY yarn elasticity imparting process proportionally to the initialization image to obtain an elasticity imparting process architecture chart corresponding to the said time point, The apparatus according to claim 9, wherein the values of pixel points other than the processing path in the elasticity imparting process architecture chart are set to default values, the points on the processing path corresponding to the subparameters in the control parameters are set to the normalized value of the subparameters, and the points on the processing path other than the subparameters in the control parameters are set to target values different from the default values.
11. The aforementioned prediction subunit is used to perform the following operations: The aforementioned operation is, Based on a short-term feature extraction module constructed by the attention mechanism in the indicator prediction model, features are extracted within a first feature extraction range in the plurality of elasticity-granting process architecture charts to obtain short-term features, wherein the first feature extraction range includes elasticity-granting process architecture charts at n time points centered on the elasticity-granting process architecture chart at the target time point, where n is a positive integer, and the short-term feature extraction module uses the elasticity-granting process architecture chart at the target time point as a query feature, and the elasticity-granting process architecture charts at the n time points excluding the target time point as key features and value features, to obtain the short-term features, and Based on a long-term feature extraction module constructed by the attention mechanism in the indicator prediction model, features are extracted within a second feature extraction range in the plurality of elasticity-granting process architecture charts to obtain long-term features, wherein the second feature extraction range includes m elasticity-granting process architecture charts at various time points centered on the elasticity-granting process architecture chart at the target time point, where m is a positive integer greater than n, and the long-term feature extraction module divides the m elasticity-granting process architecture charts at various time points using a sliding window mechanism, determines pooling information within each sliding window, constructs key features and value features necessary for the attention mechanism based on the pooling information of the plurality of sliding windows, and obtains the long-term features based on the key features, value features, and query features. Using a fusion module, a fusion operation is performed on the long-term features and the short-term features to obtain the fused features. The apparatus according to claim 9 , further comprising processing the fused features based on a prediction module in an indicator prediction model to obtain the indicator prediction result.
12. The aforementioned indicator prediction model further includes a supplementary feature extraction module, The aforementioned prediction subunit is further used to perform the following processing: The aforementioned process is, Based on the aforementioned supplementary feature extraction module, the following operations are performed to obtain supplementary features: This includes performing a fusion operation on the long-term features, short-term features, and supplementary features using a fusion module to obtain the fused features, The aforementioned operation is, The data is analyzed for the second control parameter sequence, and the cumulative difference of the parameters at each time point in the second control parameter sequence relative to the first time point in the second control parameter sequence is obtained to obtain a cumulative difference sequence. Based on the cumulative difference sequence, the time length relative to the first time point at which the cumulative difference is greater than a predetermined threshold is determined, and a time length sequence is obtained. The apparatus according to claim 11 , further comprising extracting features from the second control parameter sequence, the cumulative difference sequence, and the time length sequence, and obtaining the supplementary features.
13. A training device for a control network of an elasticity imparting process, The control network includes a parameter prediction subnetwork, an indicator prediction subnetwork, and a data generation subnetwork. The training device is An acquisition unit for acquiring a first training sample including predetermined control parameters at multiple time points in the elasticity imparting process, An input unit for inputting the first training sample into the parameter prediction subnetwork to obtain a parameter prediction value for at least one future time point, A processing unit for processing the parameter prediction values based on the indicator prediction subnetwork and obtaining the predicted indicator, A training loss unit for determining the training loss based on the difference between the predicted indicator and the actual indicator, and the difference between the predicted parameter value and the actual parameter value, An adjustment unit for adjusting the parameter prediction subnetwork and the indicator prediction subnetwork based on the training loss, If the training convergence condition is met, a first generation unit for obtaining a parameter prediction model corresponding to the parameter prediction subnetwork and an index prediction model corresponding to the index prediction subnetwork, The system includes a second generation unit for training the data generation subnetwork based on the parameter prediction model and the index prediction model to obtain a data generation model, The data generation model generates desired control parameters for the elasticity imparting machine based on the index prediction results output from the index prediction model . The aforementioned indicator prediction results include at least one of the grade, pass rate, full winding rate, and dyeing uniformity of the wound yarn package, and are used in training the control network of the elasticity imparting process.
14. The aforementioned data generation subnetwork includes a training queue generator and a training queue discriminator, The second generation unit is used to perform multiple training cycles on the data generation subnetwork. Each of the aforementioned training sessions will involve performing the following operations: The aforementioned operation is, The true control parameter is input to the training-waiting discriminator, and a first discrimination result for the true control parameter is obtained by the training-waiting discriminator. The training wait generator is used to generate false control parameters, The false control parameters are input to the training-waiting classifier to obtain a second classification result. Adjusting the model parameters of the training-waiting classifier based on the first and second classification results, The apparatus according to claim 13 , further comprising fixing the model parameters of the training waiting discriminator and training the training waiting generator.
15. It is an electronic device, At least one processor, Includes memory that is communicably connected to at least one processor, An electronic device having memory that stores instructions executable by at least one processor, the instructions being executed by at least one processor to cause at least one processor to perform the method according to any one of claims 1 to 5.
16. A non-temporary, computer-readable storage medium storing computer commands, wherein the computer commands are used to cause a computer to perform the method described in any one of claims 1 to 5.

Description

This disclosure relates to the technical field of data processing, and more particularly to a method for controlling an elasticity-granting process, a method for training a control network, and related apparatus. In the elasticity-granting (false twist) process, for the convenience of subsequent transportation and management, POY (Pre-Oriented Yarn), abbreviated as POY yarn, is processed into DTY (Draw Textured Yarn). The DTY yarn is then wound onto paper spools to form a DTY yarn package. The elasticity imparting process involves many controlled components, and the parameter settings and control of these components significantly impact the final quality of the wound yarn package. This disclosure provides a method for controlling an elasticity-granting process, a method for training a control network, and related apparatus for improving the elasticity-granting process flow. According to one aspect of this disclosure, a method for controlling an elasticity imparting process is provided, and the control method is: The process involves sequentially collecting predetermined control parameters at multiple points in time during the elasticity imparting process flow to obtain a first control parameter sequence, The first control parameter sequence is input into a parameter prediction model to predict the parameters at multiple future time points, Selecting target parameters for a target time period from prediction parameters at multiple time periods, Based on control parameters at multiple time points before and after the target time point, a second control parameter sequence including the target parameter is constructed, The process involves processing a second control parameter sequence based on the indicator prediction model to obtain the indicator prediction result, and If the indicator prediction results do not meet the desired values, this includes inputting the indicator prediction results into a data generation model to obtain the desired control parameters for the target time point. According to one aspect of this disclosure, a method for training a control network of an elasticity-granting process is provided, wherein the control network includes a parameter prediction subnetwork, an indicator prediction subnetwork, and a data generation subnetwork, and the method is: Obtaining a first training sample containing predetermined control parameters at multiple time points in the elasticity imparting process, The first training sample is input into the parameter prediction subnetwork to obtain the parameter prediction values for at least one future time point, The process involves processing parameter prediction values based on the indicator prediction subnetwork to obtain the predicted indicator, and The training loss is determined based on the difference between the predicted indicator and the actual indicator, and the difference between the predicted parameter value and the actual parameter value. Adjusting the parameter prediction subnetwork and the indicator prediction subnetwork based on the training loss, If the training convergence condition is met, the parameter prediction model corresponding to the parameter prediction subnetwork and the index prediction model corresponding to the index prediction subnetwork are obtained. This includes training a data generation subnetwork based on a parameter prediction model and an indicator prediction model to obtain a data generation model, The data generation model is designed to generate desired control parameters for the elasticity enhancer based on the indicator prediction results output from the indicator prediction model. According to another aspect of this disclosure, a control device for an elasticity imparting process is provided, the control device is, A collection unit for sequentially collecting predetermined control parameters at multiple time points in the elasticity imparting process to obtain a first control parameter sequence, A parameter prediction unit inputs a first control parameter sequence into a parameter prediction model to predict predict parameters at multiple future time points, A selection unit for selecting target parameters for a target time point from prediction parameters at multiple time points, A construction unit for constructing a second control parameter sequence including the target parameter based on control parameters at multiple points in time before and after the target point in time, An indicator prediction unit for processing a second control parameter sequence based on an indicator prediction model and obtaining indicator prediction results, The system includes a data generation unit that, if the indicator prediction results do not meet the desired values, inputs the indicator prediction results into a data generation model to obtain the desired control parameters for the target time point. According to one aspect of this disclosure, a training device for a control network of an elasticity-granting process is provide