MES-based real-time prediction and dynamic optimization method for manufacturing process cycle

Abstract

This invention relates to the field of semiconductor manufacturing technology and discloses a method for real-time prediction and dynamic optimization of manufacturing process cycles based on MES (Manufacturing Execution System). The method includes the following steps: edge devices collect real-time status data and process progress data of production equipment; the edge devices compare and analyze the collected real-time data with locally stored historical data, and, in conjunction with process standard values, automatically adjust the execution order or parameters of subsequent processes if the deviation exceeds a preset threshold; after the adjustment is performed locally, the edge devices upload the decision results to the central system. By collecting and analyzing data in real-time from the edge devices, automatically adjusting the processes, and uploading the results to the central system, the system uses machine learning models to predict the cycle time and generate optimization strategies that are fed back to the edge devices. This achieves real-time prediction and dynamic optimization of the manufacturing process cycle, improving the low prediction accuracy problem caused by data acquisition delays in traditional centralized models.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, specifically to a method for real-time prediction and dynamic optimization of manufacturing process cycles based on MES. Background Technology

[0002] In semiconductor manufacturing, the collaborative production of multiple devices and processes places extremely high demands on the real-time performance and accuracy of data. Existing manufacturing processes typically connect equipment, sensors, and other components to the Manufacturing Execution System (MES) through the construction of an Industrial Internet platform, forming an industrial data acquisition and processing system to support intelligent scheduling and production optimization. However, real-time data acquisition in semiconductor manufacturing suffers from significant latency, especially in multi-device, multi-process scenarios. Lag or errors in data acquisition directly impact prediction accuracy, leading to the low prediction accuracy issues inherent in traditional centralized models due to data acquisition delays. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for real-time prediction and dynamic optimization of manufacturing process cycles based on MES, which improves the problem of low prediction accuracy caused by data acquisition delays in the traditional centralized mode.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for real-time prediction and dynamic optimization of manufacturing process cycles based on MES, comprising the following steps:

[0005] S1. Edge devices collect real-time status data and process progress data of production equipment;

[0006] S2. The edge device compares and analyzes the collected real-time data with the locally stored historical data, and combines the process standard values. If the deviation exceeds the preset threshold, it will automatically adjust the execution order or parameters of the subsequent processes.

[0007] S3. After the adjustment action is executed locally, the edge device uploads the decision results to the central system;

[0008] S4. The central system receives data uploaded by edge devices, processes the data using machine learning models, and predicts the manufacturing process cycle.

[0009] S5. Based on the prediction results, the central system generates optimization strategies and feeds them back to the edge devices. The edge devices then dynamically optimize the production process according to the optimization strategies.

[0010] By adopting the above technical solution, data is collected and compared in real time through edge devices. After the process is automatically adjusted, the data is uploaded to the central system. Machine learning models are used to predict the cycle and generate optimization strategies, which are then fed back to the edge devices. This enables real-time prediction and dynamic optimization of the manufacturing process cycle, improving the problem of low prediction accuracy caused by data collection delays in the traditional centralized mode.

[0011] Preferably, step S1 includes: the production equipment status data collected by the edge device, including equipment temperature and equipment rotation speed. and equipment pressure The process progress data represents the progress of the process, and the edge device preprocesses the collected data. The preprocessing operation includes removing invalid values ​​within the physical range and filling in missing values ​​to ensure data quality.

[0012] Preferably, step S2 includes: calculating the root mean square error between real-time data and historical data from the same period in the comparative analysis, using the formula: ,in The root mean square error, This is a real-time value. This is the historical average. The root mean square error (RMSE) formula is used to quantify the deviation between real-time data and historical data from the same period, based on the number of data samples involved in the calculation. Combined with the process standard value, when the RMSE exceeds the preset threshold, the execution order or parameters of subsequent processes are automatically adjusted.

[0013] Preferably, step S3 includes: the edge device compresses the decision result and uploads it to the central system. The compression algorithm adopts a content-based dynamic coding strategy, uses differential coding for high-frequency changing data, and uses lossless compression for low-frequency data.

[0014] Preferably, step S4 includes: the machine learning model is a combination of LSTM-autoencoder and Transformer; and the reconstruction error is calculated using the LSTM-autoencoder. ,in The raw data is input to the autoencoder. For the data reconstructed by the autoencoder, and The first and second data are the original data and the reconstructed data, respectively. The feature value is used to identify samples that exceed the quantile threshold of historical data as anomalies. After removing the anomaly data, it is then input into the Transformer model for manufacturing process cycle prediction.

[0015] Preferably, the LSTM layer of the LSTM-autoencoder contains neurons, and the encoder and decoder of the autoencoder adopt multi-dimensional mapping; the embedding layer of the Transformer model encodes timestamps, process types, and equipment identifiers, and the number of attention heads is multiple.

[0016] Preferably, the embedding layer of the Transformer model encodes timestamps, process types, and equipment IDs; the multi-head attention layer sets up attention heads; the feedforward neural network layer contains two fully connected layers; and it outputs the predicted remaining time of the manufacturing process cycle. The Huber loss function is employed. ,in , This represents the actual value of the manufacturing process cycle. The period value predicted by the model. This is the threshold parameter.

[0017] Preferably, step S5 includes: adjusting equipment parameters and changing process priority; if the manufacturing process cycle is predicted to be extended, then starting the relevant preheating process in advance or increasing the equipment operating power.

[0018] Preferably, after collecting data, the edge device first performs physical range verification, removes invalid values, and fills missing values ​​with the median; then it calculates the data fluctuation coefficient using a sliding window. ,in The standard deviation of the data sample. The mean of the data sample is used as the benchmark, and points exceeding 3 times the standard deviation are marked as suspicious. Finally, the correlation of the data is checked based on process knowledge, and records that violate prior knowledge are subject to manual review.

[0019] The MES-based real-time prediction and dynamic optimization system for manufacturing process cycles, used in the aforementioned MES-based real-time prediction and dynamic optimization method for manufacturing process cycles, includes the following modules:

[0020] The data acquisition module is used by edge devices to collect real-time status data and process progress data of production equipment;

[0021] The edge decision module is used by edge devices to compare and analyze real-time data collected with historical data stored locally. If the deviation exceeds a preset threshold, the execution order or parameters of subsequent processes will be automatically adjusted.

[0022] The data transmission module is used to upload the decision results of the edge device to the central system after the adjustment action is executed locally;

[0023] The predictive analytics module is used by the central system to receive data uploaded from edge devices, process the data using machine learning models, and predict the manufacturing process cycle.

[0024] The strategy optimization module is used to generate optimization strategies from the central system based on the prediction results and feed them back to the edge devices. The edge devices then dynamically optimize the production process according to the optimization strategies.

[0025] This invention provides a method for real-time prediction and dynamic optimization of manufacturing process cycles based on MES. It has the following beneficial effects:

[0026] 1. This invention collects and analyzes data in real time through edge devices, automatically adjusts the process, and uploads it to the central system. It uses machine learning models to predict the cycle and generate optimization strategies to feed back to the edge devices, thereby realizing real-time prediction and dynamic optimization of the manufacturing process cycle. This improves the problem of low prediction accuracy caused by data collection delays in the traditional centralized mode.

[0027] 2. This invention collects real-time status data and process progress data of production equipment through edge devices, thereby providing real-time data support for the analysis and optimization of the manufacturing process. This improves the problem of data acquisition delay in traditional technologies, which makes it impossible to grasp the production status in a timely manner due to the inability to obtain real-time data.

[0028] 3. This invention uses edge devices to compare and analyze real-time data collected with locally stored historical data. Combined with process standard values, if the deviation exceeds a preset threshold, the execution order or parameters of subsequent processes are automatically adjusted, thereby realizing local real-time decision-making, reducing dependence on the central system, and thus improving the problem of untimely decision-making caused by data transmission delays and heavy processing burden on the central system in traditional technologies.

[0029] 4. This invention receives data uploaded by edge devices through a central system, processes the data using a machine learning model, and predicts the manufacturing process cycle, thereby improving the accuracy of the manufacturing process cycle prediction. This addresses the problem in traditional technologies where prediction models are not intelligent enough and have insufficient ability to analyze historical and real-time data, resulting in low prediction accuracy. Attached Figure Description

[0030] Figure 1 This is a flowchart of the MES-based real-time prediction and dynamic optimization method for manufacturing process cycles proposed in this invention.

[0031] Figure 2 This is a system architecture diagram for real-time prediction and dynamic optimization of manufacturing process cycles based on MES proposed in this invention. Detailed Implementation

[0032] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see the appendix Figure 1 This invention provides a method for real-time prediction and dynamic optimization of manufacturing process cycles based on MES, including the following steps:

[0034] S1. Edge devices collect real-time status data and process progress data of production equipment;

[0035] S2. The edge device compares and analyzes the collected real-time data with the locally stored historical data, and combines the process standard values. If the deviation exceeds the preset threshold, it will automatically adjust the execution order or parameters of the subsequent processes.

[0036] S3. After the adjustment action is executed locally, the edge device uploads the decision results to the central system;

[0037] S4. The central system receives data uploaded by edge devices, processes the data using machine learning models, and predicts the manufacturing process cycle.

[0038] S5. Based on the prediction results, the central system generates optimization strategies and feeds them back to the edge devices. The edge devices then dynamically optimize the production process according to the optimization strategies.

[0039] Specifically, edge devices are used to collect real-time status data and process progress data of production equipment, and the collected data is preprocessed to reduce data acquisition delays and errors, providing a reliable data foundation for subsequent analysis and ensuring the real-time nature and accuracy of production status information.

[0040] Edge devices, based on the comparative analysis of real-time and historical data and combined with process standard values, automatically adjust the execution order or parameters of subsequent processes when the deviation exceeds a preset threshold. This achieves a local decision-making mode of "execute first, record later," avoiding reliance on the transmission delay of the central system, improving the response speed to time-sensitive processes, and optimizing the real-time scheduling efficiency of the production line.

[0041] After the edge device performs the adjustment locally, it uploads the decision results to the central system through a dynamic compression strategy, reducing the amount of data transmission, optimizing network bandwidth usage, and avoiding the central system processing bottleneck caused by the centralized transmission of large amounts of data, thereby improving the overall system responsiveness.

[0042] The central system uses a machine learning model combining LSTM-autoencoder and Transformer to process data uploaded from edge devices: it calculates reconstruction error through LSTM-autoencoder to identify and remove abnormal data, and then uses Transformer model combined with features such as timestamps and process types to predict the cycle. The prediction model is optimized by combining Huber loss function to achieve high-precision real-time prediction of manufacturing process cycle.

[0043] The central system generates optimization strategies based on the forecast results and feeds them back to the edge devices. The edge devices then dynamically optimize the production process accordingly, forming a closed-loop management of "data collection - local decision-making - central forecasting - strategy feedback". This continuously improves production efficiency and reduces the risk of production delays caused by equipment idleness or untimely material supply.

[0044] Edge devices perform three levels of data cleaning to ensure data quality; at the same time, through the collaborative architecture of edge computing and the central system, hardware investment costs are reduced, machine learning models are used to optimize production processes, reduce energy consumption and maintenance costs, and improve enterprise cost efficiency.

[0045] Step S1 includes: Step S1 includes: collecting production equipment status data, including equipment temperature, from the edge device. Equipment speed and equipment pressure The process progress data represents the progress of the process, and the edge device preprocesses the collected data. The preprocessing operation includes removing invalid values ​​within the physical range and filling in missing values ​​to ensure data quality.

[0046] Specifically, by collecting equipment operating parameters through edge devices, including equipment temperature, rotation speed, pressure, and process progress data such as process completion progress, it can achieve full-coverage monitoring of key production elements in semiconductor manufacturing, provide multi-dimensional data support for real-time assessment of production status, and ensure that the collected data is directly related to the needs of manufacturing process cycle prediction and dynamic optimization.

[0047] Edge devices preprocess the collected data by removing invalid values ​​within the physical range to prevent erroneous data caused by sensor failure or environmental interference from entering the subsequent processing flow; and by filling in missing values ​​to ensure the integrity of the data sequence, reducing analytical bias caused by missing data, and providing reliable basic data for subsequent comparative analysis and prediction models.

[0048] The preprocessing stage filters out invalid data and fills in missing values, reducing noise interference in the data. This makes the subsequent comparison and analysis of real-time data and historical data from edge devices and the processing of machine learning models in the central system more efficient, avoiding the waste of computing resources or model prediction deviations caused by raw data quality issues, and improving the overall data processing efficiency of the system.

[0049] Step S2 includes: calculating the root mean square error between real-time data and historical data from the same period in the comparative analysis, using the formula: ,in The root mean square error, This is a real-time value. This is the historical average. The root mean square error (RMSE) formula is used to quantify the deviation between real-time data and historical data from the same period, based on the number of data samples involved in the calculation. Combined with the process standard value, when the RMSE exceeds the preset threshold, the execution order or parameters of subsequent processes are automatically adjusted.

[0050] Specifically, by calculating the root mean square error between real-time data and historical data from the same period, the real-time status of the production process is quantitatively compared with the historical normal production status, accurately measuring the degree of deviation between the current production data and historical data, and providing objective numerical basis for judging whether the production process is abnormal.

[0051] Based on the calculation results of the root mean square error, edge devices can determine whether the deviation between real-time data and process standard values ​​exceeds a preset threshold. If the deviation exceeds the threshold, it indicates that there may be an anomaly in the production process or a step that needs optimization. Accordingly, the edge devices automatically adjust the execution sequence or parameters of subsequent processes, realizing intelligent dynamic scheduling based on data quantitative analysis. This avoids the subjectivity of human experience judgment and improves the scientific nature and accuracy of scheduling decisions.

[0052] The calculation of root mean square error incorporates historical data from the same period, reflecting the changing patterns of the production process over time. By continuously calculating this error, edge devices can track the fluctuation trends of production data in real time, identify gradual anomalies in the production process early, and ensure that dynamic scheduling decisions can respond promptly to subtle changes in the production process, maintaining the stability and consistency of the production flow.

[0053] Step S3 includes: the edge device compresses the decision results and uploads them to the central system. The compression algorithm adopts a content-based dynamic coding strategy, using differential coding for high-frequency changing data and lossless compression for low-frequency data.

[0054] Specifically, edge devices employ a content-based dynamic coding strategy to compress decision results. For high-frequency changing data, differential coding is used, transmitting only the difference between the data and the previous frame to reduce the amount of duplicate data transmission. For low-frequency data, a lossless compression algorithm is used to reduce the amount of data while ensuring data integrity, effectively improving network bandwidth utilization, avoiding network congestion caused by large amounts of data transmission, and ensuring the transmission efficiency of real-time data.

[0055] By compressing data, the amount of data uploaded to the central system is reduced, thereby lowering the central system's storage and computing resource consumption. This enables the central system to process core tasks more efficiently, avoids system response delays caused by excessive data volume, and enhances the central system's overall management capabilities of the production process.

[0056] The dynamic encoding strategy differentiates data based on its characteristics, ensuring the real-time transmission of critical data with high-frequency changes while ensuring the accuracy of low-frequency data through lossless compression. This enables the central system to acquire complete edge decision records in a timely manner, providing reliable data support for the overall optimization of the production process and achieving efficient collaboration between the edge and central systems.

[0057] Step S4 includes: using a machine learning model that combines an LSTM-autoencoder with a Transformer; and using the LSTM-autoencoder to calculate the reconstruction error. ,in The raw data is input to the autoencoder. For the data reconstructed by the autoencoder, and The first and second data are the original data and the reconstructed data, respectively. The feature value is used to identify samples that exceed the quantile threshold of historical data as anomalies. After removing the anomaly data, it is then input into the Transformer model for manufacturing process cycle prediction.

[0058] Specifically, an LSTM-autoencoder model is used to calculate the reconstruction error, and outlier samples are identified using historical data quantile thresholds. This model utilizes LSTM to capture the temporal dependencies of time-series data and combines this with the feature reconstruction capabilities of an autoencoder to effectively identify outlier data caused by sensor malfunctions, process fluctuations, and other factors, eliminating noise interference and providing high-quality data input for manufacturing process cycle prediction.

[0059] After removing outlier data, the data is input into the Transformer model for cycle prediction. The Transformer's multi-head attention mechanism is used to capture feature correlations in long-term time-series data. Combined with the encoding capabilities of the embedding layer for multi-dimensional features such as timestamps and process types, the predictive model's adaptability to complex production processes is improved. The introduction of the Huber loss function balances accurate fitting for small errors with robustness for large errors, making the prediction results closer to the actual production cycle.

[0060] The combination of LSTM autoencoder and Transformer forms a progressive processing flow of "anomaly detection - feature extraction - cycle prediction". The removal of anomaly data reduces the impact of noise on the prediction model, while the Transformer model's learning from historical and real-time data can feed back into the dynamic adjustment of the anomaly detection threshold, enabling the entire prediction system to continuously optimize as production data accumulates, thereby improving the long-term accuracy and reliability of manufacturing process cycle prediction.

[0061] The LSTM layer of the LSTM-autoencoder contains neurons, and the encoder and decoder of the autoencoder adopt multi-dimensional mapping; the embedding layer of the Transformer model encodes timestamps, process types, and equipment identifiers, and the number of attention heads is multiple.

[0062] Specifically, the LSTM layer of the LSTM-autoencoder uses neurons and a gating mechanism to capture the long-term and short-term temporal dependencies of production data, effectively extracting dynamic features from time-series data such as equipment status and process progress, providing more accurate temporal context information for feature reconstruction of the autoencoder, and improving the accuracy of abnormal data detection.

[0063] The autoencoder employs a multi-layered dimensionality mapping in its encoder and decoder. Through dimensionality reduction and reconstruction processes, it compresses high-dimensional production data into low-dimensional feature vectors, automatically learns key features in the data, and removes redundant information. This multi-layered mapping mechanism can uncover potential correlations between data points, making reconstruction error calculations more reflective of the inherent anomalies in the data and improving the robustness of anomaly detection.

[0064] The embedding layer of the Transformer model encodes timestamps, process types, and equipment identifiers, transforming the time, process, and equipment attributes of production data into a unified feature vector. Combined with the parallel computation of multiple attention heads, the model can simultaneously capture the correlation between features of different dimensions, improving the comprehensiveness and accuracy of manufacturing process cycle prediction.

[0065] The neuron setup of the LSTM layer, combined with the multi-dimensional mapping of the autoencoder, can adapt to the data complexity of multiple devices and processes in semiconductor manufacturing. The multi-head attention mechanism of the Transformer model supports the simultaneous processing of multi-source heterogeneous data, enabling the prediction model to better cope with complex scenarios such as variable coupling and operating condition switching in the production process, and enhancing the stability and reliability of cycle prediction.

[0066] The Transformer model's embedding layer encodes timestamps, process types, and equipment IDs; the multi-head attention layer sets up attention heads; and the feedforward neural network layer contains two fully connected layers, outputting the predicted remaining manufacturing cycle time. The Huber loss function is employed. ,in , This represents the actual value of the manufacturing process cycle. The period value predicted by the model. This is the threshold parameter.

[0067] Specifically, by encoding timestamps, process types, and equipment IDs through an embedding layer, the time attributes, process attributes, and equipment attributes in the production process are transformed into feature vectors of a unified dimension. This enables different types of production data to be correlated and analyzed in the same feature space, providing comprehensive feature inputs for manufacturing process cycle prediction.

[0068] The multi-head attention layer sets up multiple attention heads to calculate the correlation weights between different feature dimensions and time steps in parallel, enabling the model to simultaneously capture long-distance time dependencies and cross-correlation of multi-dimensional features in production data, thereby improving the prediction accuracy of complex production process cycles.

[0069] The two fully connected layers of the feedforward neural network perform nonlinear transformation on the features output by the attention mechanism to further extract deep feature representations. Combined with the Huber loss function, it not only ensures the prediction accuracy in small error scenarios, but also reduces the impact of outliers on the model, enabling the model to maintain stable prediction performance when facing sudden fluctuations or data noise in the production process.

[0070] Through this structure, the Transformer model can accurately predict the remaining cycle time of the manufacturing process based on historical production data and real-time data such as equipment status and process progress. The prediction results provide data support for the central system to generate optimization strategies, enabling the production process to be dynamically adjusted according to the prediction results, reducing production delays and improving overall production efficiency.

[0071] Step S5 includes: adjusting equipment parameters and changing process priorities; if the manufacturing process cycle is predicted to be extended, start the relevant preheating process in advance or increase the operating power of the equipment.

[0072] Specifically, by adjusting equipment parameters and changing process priorities, the production process can be optimized in real time based on manufacturing cycle forecasts. When the forecast cycle is extended, starting relevant preheating processes in advance or increasing equipment operating power can shorten the preparation time of key processes or speed up process execution, thereby offsetting the extended forecast cycle and ensuring that production progress meets planned requirements.

[0073] Dynamically adjusting equipment parameters and process priorities based on forecast results can optimize the allocation of production resources such as equipment and materials. For example, increasing the operating power of key equipment to speed up production, or prioritizing time-consuming processes to avoid bottlenecks, can make resource utilization more rational, reduce equipment idle time or material waiting time, and improve overall production efficiency.

[0074] To address the anticipated risk of extended production cycles, proactive measures such as initiating preheating processes can effectively mitigate the impact of unforeseen factors like equipment failures and process fluctuations on production cycles. This proactive optimization strategy reduces uncertainty in the production process, enhances the system's adaptability to abnormal situations, and maintains the stability of the production flow.

[0075] Step S5 translates the Transformer model's predictions into concrete production adjustments, forming a complete closed loop with local decision-making at edge devices and predictive analysis from the central system. Through this loop, the production system can continuously adjust its strategies based on real-time data, dynamically optimizing the manufacturing process cycle, thereby improving production efficiency and product quality while reducing production costs.

[0076] After collecting data, the edge device first performs physical range verification, removes invalid values, and fills in missing values ​​using the median; then it calculates the data fluctuation coefficient using a sliding window. ,in The standard deviation of the data sample. The mean of the data sample is used as the benchmark, and points exceeding 3 times the standard deviation are marked as suspicious. Finally, the correlation of the data is checked based on process knowledge, and records that violate prior knowledge are subject to manual review.

[0077] Specifically, after the edge device collects data, it first performs physical range verification to remove invalid values ​​that are outside the normal operating parameter range of the device, and fills missing values ​​with the median to prevent erroneous data caused by sensor failure or transmission abnormalities from entering the subsequent processing flow, thus ensuring the integrity and basic accuracy of the data and providing a reliable raw data foundation for subsequent analysis.

[0078] By calculating the data fluctuation coefficient using a sliding window, points exceeding three standard deviations are marked as suspicious data. This step effectively identifies sudden abnormal data fluctuations during production. Compared to simple threshold judgment, fluctuation coefficient analysis combined with data distribution characteristics is better suited to the dynamic changes in production data, improving the accuracy of abnormal data identification.

[0079] Based on process knowledge, data relevance is checked, and records that violate prior knowledge are subject to manual review. This step integrates the experience of process experts into the data processing workflow, compensating for the limitations of pure algorithm detection, ensuring that the data conforms to the process logic of semiconductor manufacturing, avoiding production decision errors caused by abnormal correlations in process parameters, and further improving the reliability and effectiveness of the data.

[0080] The three-tiered data cleaning mechanism filters data layer by layer based on three dimensions: data validity, volatility, and process rationality. This ensures the basic quality of the data while identifying potential process anomalies, forming a complete data governance chain. This system can reduce the interference of erroneous data on subsequent edge decisions and central forecasting, improving the stability and accuracy of the entire manufacturing process cycle prediction and dynamic optimization system.

[0081] Please see the appendix Figure 2 A real-time prediction and dynamic optimization system for manufacturing process cycles based on MES, used for real-time prediction and dynamic optimization methods for manufacturing process cycles based on MES, includes the following modules:

[0082] The data acquisition module is used by edge devices to collect real-time status data and process progress data of production equipment;

[0083] The edge decision module is used by edge devices to compare and analyze real-time data collected with historical data stored locally. If the deviation exceeds a preset threshold, the execution order or parameters of subsequent processes will be automatically adjusted.

[0084] The data transmission module is used to upload the decision results of the edge device to the central system after the adjustment action is executed locally;

[0085] The predictive analytics module is used by the central system to receive data uploaded from edge devices, process the data using machine learning models, and predict the manufacturing process cycle.

[0086] The strategy optimization module is used to generate optimization strategies from the central system based on the prediction results and feed them back to the edge devices. The edge devices then dynamically optimize the production process according to the optimization strategies.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for real-time prediction and dynamic optimization of manufacturing process cycles based on MES, characterized in that, Includes the following steps: S1. Edge devices collect real-time status data and process progress data of production equipment; S2. The edge device compares and analyzes the collected real-time data with the locally stored historical data, and combines the process standard values. If the deviation exceeds the preset threshold, it will automatically adjust the execution order or parameters of the subsequent processes. S3. After the adjustment action is executed locally, the edge device uploads the decision results to the central system; S4. The central system receives data uploaded by edge devices, processes the data using machine learning models, and predicts the manufacturing process cycle. S5. Based on the prediction results, the central system generates optimization strategies and feeds them back to the edge devices. The edge devices then dynamically optimize the production process according to the optimization strategies. Step S2 includes: calculating the root mean square error between real-time data and historical data during comparative analysis, using the formula: ,in The root mean square error, This is a real-time value. This is the historical average. The root mean square error formula is used to quantify the degree of deviation between real-time data and historical data of the same period, in order to determine the number of data samples involved in the calculation. Combined with the process standard value, when the root mean square error exceeds the preset threshold, the execution order or parameters of subsequent processes are automatically adjusted. Step S4 includes: using a machine learning model that combines an LSTM-autoencoder with a Transformer; and using the LSTM-autoencoder to calculate the reconstruction error. ,in The raw data is input to the autoencoder. For the data reconstructed by the autoencoder, and The first and second data are the original data and the reconstructed data, respectively. The feature value is dimensional. Samples that exceed the quantile threshold of historical data are judged as anomalies. The abnormal data is removed before being input into the Transformer model for manufacturing process cycle prediction. After collecting data, the edge device first performs physical range verification, removes invalid values, and fills in missing values ​​using the median; then it calculates the data fluctuation coefficient using a sliding window. ,in The standard deviation of the data sample. The mean of the data sample is used as the benchmark, and points exceeding 3 times the standard deviation are marked as suspicious. Finally, the correlation of the data is checked based on process knowledge, and records that violate prior knowledge are subject to manual review.

2. The method for real-time prediction and dynamic optimization of manufacturing process cycle based on MES according to claim 1, characterized in that, Step S1 includes: collecting production equipment status data, including equipment temperature, from the edge device. Equipment speed and equipment pressure The process progress data represents the progress of the process, and the edge device preprocesses the collected data. The preprocessing operation includes removing invalid values ​​within the physical range and filling in missing values ​​to ensure data quality.

3. The method for real-time prediction and dynamic optimization of manufacturing process cycle based on MES according to claim 1, characterized in that, The S3 step includes: the edge device compresses the decision results and uploads them to the central system. The compression algorithm adopts a content-based dynamic coding strategy, using differential coding for high-frequency changing data and lossless compression for low-frequency data.

4. The method for real-time prediction and dynamic optimization of manufacturing process cycle based on MES according to claim 1, characterized in that, The LSTM layer of the LSTM-autoencoder contains neurons, and the encoder and decoder of the autoencoder adopt multi-dimensional mapping; the embedding layer of the Transformer model encodes timestamps, process types, and equipment identifiers, and the number of attention heads is multiple.

5. The method for real-time prediction and dynamic optimization of manufacturing process cycle based on MES according to claim 4, characterized in that: The Transformer model's embedding layer encodes timestamps, process types, and device IDs; the multi-head attention layer sets up attention heads; and the feedforward neural network layer contains two fully connected layers, outputting the predicted remaining manufacturing cycle time. The Huber loss function is employed. ,in , This represents the actual value of the manufacturing process cycle. The period value predicted by the model. This is the threshold parameter.

6. The method for real-time prediction and dynamic optimization of manufacturing process cycle based on MES according to claim 1, characterized in that, The S5 step includes: adjusting equipment parameters and changing process priorities; if the manufacturing process cycle is predicted to be extended, the relevant preheating process is started in advance or the equipment operating power is increased.

7. A real-time prediction and dynamic optimization system for manufacturing process cycles based on MES, characterized in that, The method for real-time prediction and dynamic optimization of manufacturing process cycles based on MES as described in any one of claims 1-6 includes the following modules: The data acquisition module is used by edge devices to collect real-time status data and process progress data of production equipment; The edge decision module is used by edge devices to compare and analyze real-time data collected with historical data stored locally. If the deviation exceeds a preset threshold, the execution order or parameters of subsequent processes will be automatically adjusted. The data transmission module is used to upload the decision results of the edge device to the central system after the adjustment action is executed locally; The predictive analytics module is used by the central system to receive data uploaded from edge devices, process the data using machine learning models, and predict the manufacturing process cycle. The strategy optimization module is used to generate optimization strategies from the central system based on the prediction results and feed them back to the edge devices. The edge devices then dynamically optimize the production process according to the optimization strategies.

Images