A high-end equipment product manufacturing cycle prediction method and system based on machine learning

By using machine learning algorithms and transfer learning of Bayesian ridge regression models, a method for predicting the manufacturing cycle of high-end equipment products is constructed, which solves the accuracy problem of traditional prediction methods and achieves accurate manufacturing cycle prediction and optimized decision support.

CN122616786BActive Publication Date: 2026-11-10QINGDAO HUISHITONG INFORMATION TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610689443.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-05-19
Publication Date
2026-11-10
Estimated Expiration
2046-05-19

AI Technical Summary

Technical Problem

Traditional forecasting of manufacturing cycles for high-end equipment products relies on human experience, resulting in significant discrepancies between the forecasted and actual production cycles. This lack of accuracy makes it difficult to provide effective assistance for rational production scheduling and optimization decisions.

Method used

By employing machine learning algorithms and deeply mining historical process data, a manufacturing cycle prediction method based on self-learning capability is constructed. Using the transfer learning algorithm of Bayesian ridge regression model, combined with factors such as product characteristics, production line capacity, and production factors, process-level and whole vehicle manufacturing cycle prediction models are constructed.

Benefits of technology

It enables accurate prediction of the manufacturing cycle of high-end equipment products, conforms to the actual production cycle, provides support for reasonable product scheduling and optimization decisions, and improves the accuracy and adaptability of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122616786B_ABST
    Figure CN122616786B_ABST
Patent Text Reader

Abstract

The application discloses a high-end equipment product manufacturing cycle prediction method and system based on machine learning, and the system comprises: a historical project data table used for recording product historical production data of each historical project; a historical project matching module, a process path screening module, a historical data set construction module, a key process identification module, a process time determination module and a whole vehicle manufacturing cycle prediction model construction module. The migration learning algorithm capability of the Bayesian ridge regression model is introduced, and the problems that manufacturing enterprises face in the multi-working-condition environment, such as data heterogeneity and loss and great difference between process routes, are innovatively solved, the product manufacturing cycle prediction result is more accurate, and the result is closely consistent with the actual production cycle situation, so that the application has the innovation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of manufacturing cycle prediction technology, and specifically to a method and system for predicting the manufacturing cycle of high-end equipment products based on machine learning. Background Technology

[0002] The production of high-end equipment is characterized by complex processes, long production cycles, and diverse operating conditions. Consequently, the manufacturing cycle is significantly affected by various uncertainties. In the face of increasingly fierce competition, high-quality, on-time delivery has become a crucial issue for enterprises' refined production management. Traditionally, the manufacturing cycle of high-end equipment relies on extensive forecasting based on human experience. The factors considered in this forecasting process are highly correlated with individual experience and familiarity with the production site, resulting in significant discrepancies between the predicted and actual production cycles. Therefore, accurately predicting the product manufacturing cycle, and thus providing support for rational production scheduling and optimized decision-making, is of great significance. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for predicting the manufacturing cycle of high-end equipment products based on machine learning. By deeply mining massive amounts of data such as historical process data, production cycles, and production resources, and introducing machine learning algorithms, a manufacturing cycle prediction method and system based on self-learning capabilities are constructed. This allows for accurate production cycle prediction, providing auxiliary support for rational product scheduling and optimization decisions.

[0004] The technical solution adopted in this invention is as follows:

[0005] This invention provides a method for predicting the manufacturing cycle of high-end equipment products based on machine learning, comprising the following steps:

[0006] Step S1: Obtain the historical project data table, which is used to record the historical production data of each historical project.

[0007] Step S2: Given information on new projects for which product manufacturing cycle prediction is required, at least one historical project that is most similar is matched among the historical projects based on relevant matching rules.

[0008] Step S3: For each matched historical project, based on the historical project data table, count its process paths and filter out the Top N process paths with the most process types.

[0009] Step S4: For each selected process path, based on the start and end times in the historical work reporting data, the matching process time is obtained. Each process path and the matching process time form a structured historical dataset that can be used for modeling and analysis.

[0010] Step S5: Based on the historical dataset and the historical project data table, determine the difference in actual process time for each product under the same process path, identify the process path with a large difference as a key process, and identify the process path with a small difference as a normal process.

[0011] Step S6: For each identified key process, construct a corresponding process-level regression prediction model, and train the model using a subset of historical data corresponding to the key process to obtain a trained process-level regression prediction model; use the trained process-level regression prediction model to perform regression prediction on the process time of the key process to obtain the predicted process time; fuse the matched process time and the predicted process time of the key process to obtain the updated process time value of the key process; for each ordinary process, directly use the matched process time as its process time.

[0012] Step S7: Construct a vehicle manufacturing cycle prediction model; take the updated process time values ​​of each key process, the process time of each ordinary process, and the historical dataset used for modeling and analysis as inputs to train the vehicle manufacturing cycle prediction model to obtain the trained vehicle manufacturing cycle prediction model.

[0013] The trained vehicle manufacturing cycle prediction model is used to predict the vehicle manufacturing cycle for the new input project.

[0014] Furthermore, the historical project data table records the historical production data of each historical project, including the start and end times of each process, and has undergone data cleaning processing. The specific data cleaning method is as follows:

[0015] If the start or end time of the process is null, it will be rejected directly.

[0016] If the process ends earlier than it begins, it is rejected outright.

[0017] Furthermore, step S2 includes:

[0018] Step S21, the new project information includes product platform, product material and product model;

[0019] Step S22: Prioritize product platform as the first matching item, and product material and product model as the second priority matching items. Use a gradually relaxed matching criteria approach to match similar historical items. Specifically:

[0020] In each historical project, check if there is a historical project that simultaneously meets the requirements of product platform, product material, and product model; if so, obtain the matching historical project.

[0021] If not, relax the criteria and match whether there are historical projects that simultaneously meet the requirements of product platform and product model; if so, obtain the matching historical projects.

[0022] If not, relax the criteria and match whether there are historical projects that simultaneously meet the requirements of product platform and product material consistency; if there are, obtain the matching historical projects.

[0023] If not, the conditions will be relaxed to include historical projects that are consistent only with the product platform.

[0024] Furthermore, in step S4, the method for determining the matching process time for each process path is as follows:

[0025] Based on historical work reporting data, obtain the most recent historical production record of the product using the process path. Based on the start and end times of the process path recorded in the product's historical production record, obtain the matching process time. If the recorded start or end time of the process path is missing, or the matching process time is abnormal, the corresponding standard working hours are used as a substitute.

[0026] Furthermore, step S6 includes:

[0027] Step S61: Determine the fusion feature vector of the key process; the fusion feature vector includes the project period of the key process, the historical average construction period of the key process, the process category of the key process, and the interruption time of the key process.

[0028] Step S62: Analyze the historical dataset to obtain the fusion feature dataset corresponding to the fusion feature vector of the key process; use the fusion feature dataset to train the process-level regression prediction model to obtain the trained process-level regression prediction model.

[0029] Step S63, assuming a total of The key processes will then lead to the corresponding construction. Each process-level regression prediction model; for each process-level regression prediction model , Determine its mean square error in the fused feature dataset. The regression prediction model for each process stage is determined using the following formula. weight :

[0030] ;

[0031] in: and , which are the maximum and minimum mean square errors in the regression prediction models for each process level, respectively;

[0032] Step S64, the calculation model for the process time update value of the key process is as follows:

[0033] ;

[0034] in: Key processes The updated value of the process time; To match the process time; This represents the predicted process time.

[0035] Furthermore, the trained process-level regression prediction model is used to perform regression prediction on the process time of key processes, obtaining the predicted process time, including:

[0036] Step S61, focusing on key processes The corresponding historical data subset is used as the source domain data; the relevant dataset of new projects that require product manufacturing cycle prediction is used as the target domain data.

[0037] Step S62: Encode and standardize the source domain data and target domain data respectively to obtain source domain samples and target domain samples;

[0038] Step S63: Train the Bayesian ridge regression model using source domain samples, and extract the source domain model parameters, including the source domain model regression coefficients. Source domain model intercept ;

[0039] A Bayesian ridge regression model is trained using samples from the target domain, and the target domain model parameters, including the target domain model regression coefficients, are extracted. Intercept of target domain model ;

[0040] Step S64: The source domain model parameters and the target domain model parameters are fused using a linear weighting method to construct a transfer learning model. ;in, ; ;in, These are the regression coefficients for the transfer learning model; The intercept of the transfer learning model;

[0041] Step S65: Input the target domain data into the transfer learning model to obtain the key processes. Predicted process time .

[0042] Furthermore, the inputs to the vehicle manufacturing cycle prediction model specifically include the following dimensions: project period; busy / slack period of the project; process complexity of the project, i.e., the number of key processes and the number of ordinary processes; historical average lead time for each process; updated process time values ​​for key processes; and process time for ordinary processes.

[0043] This invention also provides a machine learning-based system for predicting the manufacturing cycle of high-end equipment products, comprising:

[0044] Historical project data table, used to record the historical production data of each historical project;

[0045] The historical project matching module is used to match at least one most similar historical project among various historical projects based on relevant matching rules, given information on new projects that require product manufacturing cycle prediction.

[0046] The process path filtering module is used to, for each matched historical project, count its process paths based on the historical project data table, and filter out the Top N process paths with the most types of processes.

[0047] The historical dataset construction module is used to obtain the matching process time for each selected process path based on the start and end times in the historical work reporting data. Each process path and the matching process time form a structured historical dataset that can be used for modeling and analysis.

[0048] The key process identification module is used to determine the difference in actual process time for each product under the same process path based on the historical dataset and the historical project data table, and to identify process paths with large differences as key processes and process paths with small differences as ordinary processes.

[0049] The process time determination module is used to construct a corresponding process-level regression prediction model for each identified key process, and train the model using a subset of historical data corresponding to the key process to obtain a trained process-level regression prediction model; using the trained process-level regression prediction model, the process time of the key process is regressed and predicted to obtain the predicted process time; the matched process time and the predicted process time of the key process are fused to obtain the updated process time value of the key process; for each ordinary process, the matched process time is directly used as its process time.

[0050] The whole vehicle manufacturing cycle prediction model construction module is used to construct a whole vehicle manufacturing cycle prediction model; the process time update values ​​of each key process, the process time of each ordinary process, and the historical dataset used for modeling analysis are used as inputs to train the whole vehicle manufacturing cycle prediction model to obtain a trained whole vehicle manufacturing cycle prediction model; the trained whole vehicle manufacturing cycle prediction model is used to predict the whole vehicle manufacturing cycle of the new project currently input.

[0051] The invention provides a method and system for predicting the manufacturing cycle of high-end equipment products based on machine learning, which has the following advantages:

[0052] 1. This invention introduces the transfer learning algorithm capability of the Bayesian Ridge Regression model, which innovatively solves the problems faced by manufacturing enterprises in multi-working-condition environments, such as data heterogeneity and missing data, and large differences in process routes. The product manufacturing cycle prediction results are more accurate and closely match the actual production cycle, which is a pioneering innovation.

[0053] 2. In the process of designing the manufacturing cycle prediction algorithm, this invention innovatively introduces factors such as product characteristics (product model, material), production line capacity, and production elements (equipment, personnel, materials, processes, etc.), while taking into account the differences in the state of relevant factors under different working conditions. This method innovatively introduces the high-end discrete equipment manufacturing process and has strong novelty.

[0054] 3. The manufacturing cycle prediction algorithm fits the relevant factors affecting the manufacturing cycle under various historical working conditions, innovatively characterizes the historical manufacturing cycle, and forms a prediction of the future manufacturing cycle through reasoning ability. Its prediction results are more in line with reality. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 A flowchart of the machine learning-based high-end equipment product manufacturing cycle prediction method provided by the present invention;

[0057] Figure 2 A schematic diagram of the transfer learning manufacturing cycle prediction framework based on Bayesian ridge regression model provided by this invention;

[0058] Figure 3 The single-condition manufacturing cycle model based on the Bayesian ridge regression model provided by this invention is directly used for the new sample data of condition 2, as shown in the result figure.

[0059] Figure 4 The result diagram of the transfer learning model based on the Bayesian ridge regression model provided by this invention. Detailed Implementation

[0060] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0061] This invention is based on the complex manufacturing process, long production cycle, and numerous interfering factors affecting the manufacturing cycle of discrete high-end equipment products. It utilizes machine learning algorithms and incorporates factors such as product characteristics (product model, material), production line capacity, and production elements (equipment, personnel, materials, processes, etc.) to innovatively construct a product manufacturing cycle prediction method. At the same time, it integrates the impact analysis of various factors on the manufacturing cycle under different working conditions, and finally forms a manufacturing cycle prediction method for high-end discrete products, which has significant innovation and originality.

[0062] This invention mainly addresses the following key issues:

[0063] 1. This invention introduces the transfer learning algorithm capability of the Bayesian Ridge Regression model, which innovatively solves the problems faced by manufacturing enterprises in multi-working-condition environments, such as data heterogeneity and missing data, and large differences in process routes. The product manufacturing cycle prediction results are more accurate and closely match the actual production cycle, which is a pioneering innovation.

[0064] 2. In the process of designing the manufacturing cycle prediction algorithm, this invention innovatively introduces factors such as product characteristics (product model, material), production line capacity, and production elements (equipment, personnel, materials, processes, etc.), while taking into account the differences in the state of relevant factors under different working conditions. This method innovatively introduces the high-end discrete equipment manufacturing process and has strong novelty.

[0065] 3. The manufacturing cycle prediction algorithm fits the relevant factors affecting the manufacturing cycle under various historical working conditions, innovatively characterizes the historical manufacturing cycle, and forms a prediction of the future manufacturing cycle through reasoning ability. Its prediction results are more in line with reality.

[0066] Specifically, the goal of the manufacturing cycle prediction model is to collect and analyze historical production data of completed products through statistical analysis, machine learning, and other methods. By cleaning data and identifying patterns, it aims to build manufacturing cycle prediction models for different product platform projects, thereby solving the problem of insufficient process data support for medium- and long-term production planning caused by insufficient process planning cycles.

[0067] The challenges in constructing this model include the following aspects:

[0068] (1) Accuracy of historical data: Since the historical production data of completed products obtained from the MES system contains a lot of noise data and the data fluctuates greatly (such as being affected by the accuracy of the record, the level of production busyness, the skill level and status of workers, etc.), it is necessary to evaluate the accuracy of the data through statistical analysis. (2) Generalization of the prediction model: Due to the continuous innovation of the production system and production process, the historical completion data cannot fully predict the future completion time. It is necessary to pay attention to the generalization ability and adaptive update ability of the model in the model construction. (3) Correspondence between the overall manufacturing cycle and the process completion time: The main purpose of this model is to predict the overall manufacturing cycle. Based on the analysis of the time of each underlying process, the overall manufacturing cycle needs to be divided into a certain granularity to give prediction results at different granularities (overall manufacturing cycle of a vehicle, production time of process path).

[0069] This invention provides a method for predicting the manufacturing cycle of high-end equipment products based on machine learning. (See attached document.) Figure 1 The steps are as follows:

[0070] Step S1: Obtain the historical project data table, which is used to record the historical production data of each historical project.

[0071] The historical project data table records the historical production data of each historical project, including the start and end times of each process. The data has been cleaned, specifically as follows: if there are null values ​​for the start or end time of a process, they are directly removed; if the end time of a process is earlier than the start time, it is directly removed.

[0072] This step specifically involves data acquisition and verification. To ensure the accuracy and effectiveness of data analysis and subsequent modeling, each data table undergoes a thorough, step-by-step review. This process involves a comprehensive assessment of data availability for each process across all product platforms, ensuring data integrity and usability. By meticulously verifying the timeframe of the data, and ensuring sufficient data volume to support subsequent statistical analysis and model building, this step ensures the data is available for all necessary purposes.

[0073] The database table names used in the manufacturing cycle prediction model are shown in Table 1:

[0074] Table 1: Database Table Names

[0075] Data Name Data Description aps_capacity Basic Data - Production Capacity and Production Line Configuration aps_capacity_relation Production Line-Workstation Relationship Table aps_month_plan_show_prefix Monthly Progression Plan Prefix Table aps_month_plan_show_suffix Monthly Progression Plan Suffix Table aps_project Project Information aps_sub_capacity Small production line information aps_tc_routing TC process path data aps_work_report APS work reporting information sys_dict dictionary table sys_dict_item dictionary entries

[0076] During the review of the aforementioned data tables, visual charts were used to intuitively display the statistical results and identify anomalies, missing data, or inconsistencies. For example, bar charts clearly show the data distribution across different product platforms or processes, allowing for assessment of data uniformity and identifying instances where data is insufficient for certain platforms or processes, thus failing to meet the needs of model training.

[0077] If key data required for the model is found to be missing during the inspection process, it is necessary to promptly explore the reasons for the missing data, assess the possibility of supplementing the data, and formulate corresponding solutions, such as model adjustment, adding constraints, constructing features, and re-collecting data.

[0078] Data cleaning:

[0079] Machine learning algorithms rely on data to make predictions. Data quality directly impacts model performance and predictive ability. High-quality data reduces noise interference during model training and improves the model's generalization ability. Furthermore, in-depth analysis of raw data helps understand its distribution, correlations, and underlying patterns, providing strong support for feature engineering and model selection. Therefore, preliminary exploration of the raw data and assessment of its quality are necessary. Potential data quality issues include missing values, outliers, and incorrect data formats.

[0080] 1. Filtering for missing work reporting times

[0081] Data with null values ​​for process start and end times are filtered out and discarded.

[0082] 2. Filtering based on the reporting time of process abnormalities

[0083] Abnormal data where the process end time is earlier than the process start time is directly rejected.

[0084] Step S2: Given information on new projects for which product manufacturing cycle prediction is required, at least one historical project that is most similar is matched among the historical projects based on relevant matching rules.

[0085] Step S21, the new project information includes product platform, product material and product model;

[0086] Step S22: Prioritize product platform as the first matching item, and product material and product model as the second priority matching items. Use a gradually relaxed matching criteria approach to match similar historical items. Specifically:

[0087] In each historical project, check if there is a historical project that simultaneously meets the requirements of product platform, product material, and product model; if so, obtain the matching historical project.

[0088] If not, relax the criteria and match whether there are historical projects that simultaneously meet the requirements of product platform and product model; if so, obtain the matching historical projects.

[0089] If not, relax the criteria and match whether there are historical projects that simultaneously meet the requirements of product platform and product material consistency; if there are, obtain the matching historical projects.

[0090] If not, the conditions will be relaxed to include historical projects that are consistent only with the product platform.

[0091] Specifically, in the process of manufacturing cycle prediction and time analysis, for new project information (including product platform, material, and model), a matching algorithm (such as matching based on product type, project type, etc.) is used to find the most similar projects in historical data. This algorithm prioritizes precise matching based on product platform, material, and model. If no results are found, the conditions are gradually relaxed to match only some key fields, and finally, a set of historical project numbers (wbs) most relevant to the new project is selected from the historical data.

[0092] The specific process is as follows:

[0093] Given new project information (product platform, material, model), this task will perform the following functions:

[0094] Search for similar historical items in the aps_tc_routing database based on relevant matching rules.

[0095] The matching rules are as follows:

[0096] (1) First, find the historical project number (wbs) that is completely consistent with the new project in terms of "product platform", "material" and "model".

[0097] (2) If no exact match is found, the criteria are relaxed to match only “Product Platform” and “Model”.

[0098] (3) If no exact match is found, the criteria are relaxed to match only “Product Platform” and “Material”.

[0099] (4) If it is still not found, then only match “product platform”.

[0100] (5) Finally, a set of matching historical item numbers (wbs) is obtained.

[0101] Step S3: For each matched historical project, based on the historical project data table, count its process paths and filter out the Top N process paths with the most process types.

[0102] For example, for the matched historical project number (wbs), count the process paths of all vehicles under each project, and select the Top N process paths (currently 2) with the most process types in each project.

[0103] Step S4: For each selected process path, based on the start and end times in the historical work reporting data, the matching process time is obtained. Each process path and the matching process time form a structured historical dataset that can be used for modeling and analysis.

[0104] The matching process time for each process path is determined as follows: Based on historical work reporting data, obtain the most recent historical production record of the product using the process path. Based on the start and end times of the process path recorded in the product's historical production record, obtain the matching process time. If the recorded start or end time of the process path is missing, or the matching process time is abnormal, the corresponding standard working time is used as a substitute.

[0105] As a specific method, based on the matched process path, the start and end times of the closest process path (based on the start time) to the current time are matched in the work report table (aps_work_report), and the actual process time is calculated based on the start and end times. If the start and end times are missing or the calculated actual process time is abnormal, the time matched for that process is replaced by the standard working hours in aps_tc_routing. This step provides detailed and reliable data support for model training and subsequent analysis.

[0106] Specifically, the start and end times of each process for each vehicle need to be matched in the work report (aps_work_report), and the actual process time is calculated accordingly. For cases where start or end times are missing, or where the actual process time is abnormal, the standard working hours (vgw02) from the aps_tc_routing table are used as a substitute to ensure data integrity and reasonableness. Through this process, a structured historical dataset suitable for modeling and analysis is ultimately formed, laying a solid data foundation for the development and optimization of manufacturing cycle prediction models.

[0107] Step S5: Based on the historical dataset and the historical project data table, determine the difference in actual process time for each product under the same process path, identify the process path with a large difference as a key process, and identify the process path with a small difference as a normal process.

[0108] In the process of building predictive models, selecting key processes is crucial. Furthermore, preliminary research indicates an uneven distribution of product platform categories, necessitating the use of data-driven techniques to identify key processes for each product platform.

[0109] Based on the modeling dataset constructed in step S4, each vehicle is grouped according to the process path, the variance of the process time under the same process path is calculated, and the top 50 processes with the largest variance are selected as key processes.

[0110] Step S6: For each identified key process, construct a corresponding process-level regression prediction model, and train the model using a subset of historical data corresponding to the key process to obtain a trained process-level regression prediction model; use the trained process-level regression prediction model to perform regression prediction on the process time of the key process to obtain the predicted process time; fuse the matched process time and the predicted process time of the key process to obtain the updated process time value of the key process; for each ordinary process, directly use the matched process time as its process time.

[0111] Step S61: Determine the fusion feature vector of the key process; the fusion feature vector includes the project period of the key process, the historical average construction period of the key process, the process category of the key process, and the interruption time of the key process.

[0112] Step S62: Analyze the historical dataset to obtain the fusion feature dataset corresponding to the fusion feature vector of the key process; use the fusion feature dataset to train the process-level regression prediction model to obtain the trained process-level regression prediction model.

[0113] Step S63, assuming a total of The key processes will then lead to the corresponding construction. Each process-level regression prediction model; for each process-level regression prediction model , Determine its mean square error in the fused feature dataset. The regression prediction model for each process stage is determined using the following formula. weight :

[0114] ;

[0115] in: and , which are the maximum and minimum mean square errors in the regression prediction models for each process level, respectively;

[0116] Step S64, the calculation model for the process time update value of the key process is as follows:

[0117] ;

[0118] in: Key processes The updated value of the process time; To match the process time; This represents the predicted process time.

[0119] Using the trained process-level regression prediction model, the process time of key processes is predicted through regression analysis. (See reference...) Figure 2 ,include:

[0120] Step S61, focusing on key processes The corresponding historical data subset is used as the source domain data; the relevant dataset of new projects that require product manufacturing cycle prediction is used as the target domain data.

[0121] Step S62: Encode and standardize the source domain data and target domain data respectively to obtain source domain samples and target domain samples;

[0122] Step S63: Train the Bayesian ridge regression model using source domain samples, and extract the source domain model parameters, including the source domain model regression coefficients. Source domain model intercept ;

[0123] A Bayesian ridge regression model is trained using samples from the target domain, and the target domain model parameters, including the target domain model regression coefficients, are extracted. Intercept of target domain model ;

[0124] Step S64: The source domain model parameters and the target domain model parameters are fused using a linear weighting method to construct a transfer learning model. ;in, ; ;in, These are the regression coefficients for the transfer learning model; The intercept of the transfer learning model;

[0125] Step S65: Input the target domain data into the transfer learning model to obtain the key processes. Predicted process time .

[0126] One specific way to implement this step is as follows:

[0127] For each key process, a linear regression model was constructed that incorporates features such as "project period" and "historical average duration." The features constructed by the model are shown in Table 2.

[0128] Table 2: Meaning of Features

[0129] Feature Name type Definition method Project period Categorical variables Based on the data from this project, the earliest start time distribution of the process for each vehicle was manually divided. Historical average construction period Continuous variables Search across all projects. For each process step of each vehicle, take the times of the three most recent historical processes for the same model and material, naming them his_mes_1, his_mes_2, and his_mes_3. Then, group and sum the times for each vehicle (pspid) to obtain his_mes_1_sum, his_mes_2_sum, and his_mes_3_sum for the three most recent historical manufacturing times for each vehicle. Finally, calculate his_mes_1_sum, his_mes_2_sum, and his_mes_3_sum as the historical average duration (his_mes_avg). Note that if there are no recent historical processes, the value is 0. locomotive and carriage Categorical variables Based on the vehicle numbering rules, identify the vehicle type (cab or carriage) corresponding to each sample. Determine the type based on the number after the last "." in the pspid: if it's 01, it's a cab; otherwise, it's a carriage. Interruption time Continuous variables Based on the blueprint information provided earlier, calculate the "interruption time" caused by lunch break and nighttime rest during the work period. Lunch break: 12:00-13:00, fixed at 1 hour. Nighttime rest: 16:40-8:00 the next day, totaling 15 hours and 20 minutes.

[0130] The constructed features are used as input to train the process-level regression prediction model. After training all process-level regression prediction models, the originally matched process times are updated. The specific update steps are as follows:

[0131] First, for key processes with large differences in process time identified in the original matching, the minimum and maximum values ​​of the MSE of all process-level regression prediction models are calculated. A normalization function is constructed using the Min-Max approach, and its weights are... The calculation formula is as follows:

[0132]

[0133] For a key process, the update method is as follows:

[0134]

[0135] in: Key processes The updated value of the process time; To match the process time; This represents the predicted process time.

[0136] For ordinary processes with small differences in the originally matched process time, the predicted value is directly adopted from the matched process time.

[0137] Summarizing the above update steps, the overall process time prediction formula can be abstracted as follows:

[0138]

[0139] In this invention, the transfer learning-based manufacturing cycle prediction framework based on the Bayesian ridge regression model is as follows: Figure 2 As shown, this method transfers posterior knowledge from the source domain data to the target domain to improve prediction performance in the target domain. The method constructs a transfer learning model by linearly weighting the posterior means of the source and target domains and evaluates it on the target domain test set. Finally, the model is used to predict new samples in the target domain.

[0140] The specific steps are as follows:

[0141] Data preprocessing: First, missing value analysis was performed on the raw data. The results showed that there were no missing values ​​in the raw data. One-hot encoding was performed on the categorical variable `order_type`. To ensure consistency of units, the remaining numerical variables were standardized. All data from the Case 1 samples were used for model training. Data from the Case 2 samples were divided using the hold-out method, with 60% used as the training set and 40% as the test set.

[0142] Source domain and target domain model training:

[0143] Train a Bayesian ridge regression model on the source domain data to obtain the parameters of the source domain model: regression coefficients source_coef ( ) and intercept source_intercept ( ).

[0144] Train a Bayesian ridge regression model on the target domain data to obtain the parameters of the target domain model: regression coefficients target_coef ( ) and intercept target_intercept ( ).

[0145] Parameter fusion: Parameters from the source and target domains are fused using a linear weighting method to construct a transfer learning model. Transfer weights are then used. Weight the parameters of the source and target domains: ; Migration weight in this embodiment Set it to 0.1.

[0146] Transfer evaluation: Predictions are made using the fused parameters, and model performance is evaluated on the target domain test set. Evaluation metrics for the model are calculated, and a scatter plot comparing predicted and actual values ​​is plotted.

[0147] The predictive performance of transfer learning under new working conditions:

[0148] Table 3 shows the comparison of the prediction performance of transfer learning under the new working conditions. Figure 3 , Figure 4 .

[0149] Table 3 Comparison of the prediction effects of transfer learning under new working conditions

[0150] RMSE <![CDATA[R 2 ]]> The single-condition manufacturing cycle model based on the Bayesian ridge regression model was directly applied to the new sample data of condition 2. 10323.8 0.6281 Transfer learning model based on Bayesian ridge regression model 7047.8 0.8267

[0151] It can be seen that the RMSE and R-value of the single-condition model based on Bayesian ridge regression (i.e., trained only with data from condition 1 and directly transferred to the new samples from condition 2) are significantly lower. 2 The values ​​are 10323.8 and 0.6281 respectively, indicating that the model has a large error on the new data of working condition 2, and its fitting ability is average. Meanwhile, from... Figure 3It can be seen that although the overall trend is consistent with the true value, the point distribution is relatively scattered, and some samples deviate from the fitted line, indicating that the model has limited fitting ability on working condition 2.

[0152] The transfer learning model based on the Bayesian ridge regression model showed reduced error and improved fitting ability on the new data from Working Condition 2. Figure 4 It can be seen that its sample size is compared to Figure 3 The concentration of data near the diagonal indicates that the model's predictions for the new samples in Working Condition 2 are more accurate, with better fitting and significantly reduced error. This demonstrates that the transfer learning model effectively utilizes information from both the source domain (Working Condition 1) and the target domain (Working Condition 2), improving prediction accuracy and generalization ability on the new samples in Working Condition 2, outperforming the single-working-condition model that relies on direct transfer.

[0153] Step S7: Construct a vehicle manufacturing cycle prediction model; take the updated process time values ​​of each key process, the process time of each ordinary process, and the historical dataset used for modeling and analysis as inputs to train the vehicle manufacturing cycle prediction model to obtain the trained vehicle manufacturing cycle prediction model.

[0154] The trained vehicle manufacturing cycle prediction model is used to predict the vehicle manufacturing cycle for the new input project.

[0155] The inputs to the vehicle manufacturing cycle prediction model specifically include the following dimensions: project period; busy / slack period of the project; process complexity of the project, i.e., the number of key processes and the number of ordinary processes; historical average lead time for each process; updated process time values ​​for key processes; and process time for ordinary processes.

[0156] Specifically, based on the updated timeline and modeling dataset for key processes, the corrected data is integrated into the vehicle manufacturing cycle prediction model. The constructed features are shown in Table 4.

[0157] Table 4: Meaning of Features

[0158] Feature Name type Definition method Busyness Categorical variables Based on data from all projects, if the number of vehicles produced in a certain month of a certain year is greater than 80, then that month is considered "busy"; otherwise, it is considered "slack". Project period Categorical variables Based on the data from this project, the earliest start time distribution of the process for each vehicle was manually divided. process complexity Continuous variables The number of processes involved in this vehicle Historical average construction period Continuous variables Search across all projects. For each process step of each vehicle, take the times of the three most recent historical processes for the same model and material, naming them his_mes_1, his_mes_2, and his_mes_3. Then, group and sum the times for each vehicle (pspid) to obtain his_mes_1_sum, his_mes_2_sum, and his_mes_3_sum for the three most recent historical manufacturing times for each vehicle. Finally, calculate his_mes_1_sum, his_mes_2_sum, and his_mes_3_sum as the historical average duration (his_mes_avg). Note that if there are no recent historical processes, the value is 0.

[0159] The constructed features are used as input to train the vehicle manufacturing cycle prediction model, ultimately providing the vehicle manufacturing cycle. Simultaneously, by combining the model's baseline value and mean squared error (MSE), a reasonable fluctuation range at a 95% confidence level is determined, using an interval calculation method of mean ± 2 standard deviations. .

[0160] This invention also provides a machine learning-based system for predicting the manufacturing cycle of high-end equipment products, comprising:

[0161] Historical project data table, used to record the historical production data of each historical project;

[0162] The historical project matching module is used to match at least one most similar historical project among various historical projects based on relevant matching rules, given information on new projects that require product manufacturing cycle prediction.

[0163] The process path filtering module is used to, for each matched historical project, count its process paths based on the historical project data table, and filter out the Top N process paths with the most types of processes.

[0164] The historical dataset construction module is used to obtain the matching process time for each selected process path based on the start and end times in the historical work reporting data. Each process path and the matching process time form a structured historical dataset that can be used for modeling and analysis.

[0165] The key process identification module is used to determine the difference in actual process time for each product under the same process path based on the historical dataset and the historical project data table, and to identify process paths with large differences as key processes and process paths with small differences as ordinary processes.

[0166] The process time determination module is used to construct a corresponding process-level regression prediction model for each identified key process, and train the model using a subset of historical data corresponding to the key process to obtain a trained process-level regression prediction model; using the trained process-level regression prediction model, the process time of the key process is regressed and predicted to obtain the predicted process time; the matched process time and the predicted process time of the key process are fused to obtain the updated process time value of the key process; for each ordinary process, the matched process time is directly used as its process time.

[0167] The whole vehicle manufacturing cycle prediction model construction module is used to construct a whole vehicle manufacturing cycle prediction model; the process time update values ​​of each key process, the process time of each ordinary process, and the historical dataset used for modeling analysis are used as inputs to train the whole vehicle manufacturing cycle prediction model to obtain a trained whole vehicle manufacturing cycle prediction model; the trained whole vehicle manufacturing cycle prediction model is used to predict the whole vehicle manufacturing cycle of the new project currently input.

[0168] 1. This invention focuses on protecting the prediction model algorithm and implementation ideas for the high-end equipment manufacturing cycle. In terms of model algorithm construction, it fully considers the related factors affecting the high-end equipment manufacturing cycle, and the related factors are scalable and can be added or reduced according to the actual situation.

[0169] 2. The algorithm model constructed in this invention fits historical manufacturing cycles, shutdowns caused by external factors in key production links, and unexplained shutdowns into the model algorithm, making the model algorithm highly generalizable and versatile, and can be transferred and replicated to the prediction of manufacturing cycles in other related industries or products.

[0170] 3. This invention focuses on protecting the process path selection method, the manufacturing cycle of each production stage, and the prediction algorithm for the product manufacturing cycle.

[0171] 4. This invention is an overall implementation idea for manufacturing cycle prediction, which is indivisible. The overall implementation steps of manufacturing cycle prediction are also an important protection point of this invention.

[0172] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the manufacturing cycle of high-end equipment products based on machine learning, characterized in that, Includes the following steps: Step S1: Obtain the historical project data table, which is used to record the historical production data of each historical project. Step S2: Given information on new projects for which product manufacturing cycle prediction is required, at least one historical project that is most similar is matched among the historical projects based on relevant matching rules. Step S3: For each matched historical project, based on the historical project data table, count its process paths and filter out the top N process paths with the most process types. Step S4: For each selected process path, based on the start and end times in the historical work reporting data, the matching process time is obtained. Each process path and the matching process time form a structured historical dataset that can be used for modeling and analysis. Step S5: Based on the historical dataset and the historical project data table, determine the difference in actual process time for each product under the same process path, identify the process path with a large difference as a key process, and identify the process path with a small difference as a normal process. Step S6: For each key process identified, construct a corresponding process-level regression prediction model, and use a subset of historical data corresponding to the key process to train the model, thereby obtaining the trained process-level regression prediction model. The trained process-level regression prediction model is used to perform regression prediction on the process time of key processes, and the predicted process time is obtained. By combining the matched process time and the predicted process time of the key process, the updated process time value of the key process is obtained; for each ordinary process, the matched process time is directly used as its process time. Step S7: Construct a vehicle manufacturing cycle prediction model; take the updated process time values ​​of each key process, the process time of each ordinary process, and the historical dataset used for modeling and analysis as inputs to train the vehicle manufacturing cycle prediction model to obtain the trained vehicle manufacturing cycle prediction model. The trained vehicle manufacturing cycle prediction model is used to predict the vehicle manufacturing cycle for the new input project. in: The trained process-level regression prediction model is used to predict the process time of key processes, resulting in the predicted process time, including: Key processes The corresponding historical data subset is used as the source domain data; the relevant dataset of new projects that require product manufacturing cycle prediction is used as the target domain data. The source domain data and target domain data are encoded and standardized respectively to obtain source domain samples and target domain samples; The Bayesian ridge regression model is trained using source domain samples, and the source domain model parameters, including the source domain model regression coefficients, are extracted. Source domain model intercept ; A Bayesian ridge regression model is trained using samples from the target domain, and the target domain model parameters, including the target domain model regression coefficients, are extracted. Intercept of target domain model ; A transfer learning model is constructed by fusing source domain model parameters and target domain model parameters using a linear weighting method. ;in, ; ;in, These are the regression coefficients for the transfer learning model; The intercept of the transfer learning model; For migration weights; Inputting the target domain data into the transfer learning model yields key processes. Predicted process time .

2. The method for predicting the manufacturing cycle of high-end equipment products based on machine learning according to claim 1, characterized in that, The historical project data table records the historical production data of each historical project, including the start and end times of each process, and has undergone data cleaning processing. The specific data cleaning method is as follows: If the start or end time of the process is null, it will be rejected directly. If the process ends earlier than it begins, it is rejected outright.

3. The method for predicting the manufacturing cycle of high-end equipment products based on machine learning according to claim 1, characterized in that, Step S2 includes: Step S21, the new project information includes product platform, product material and product model; Step S22: Prioritize product platform as the first matching item, and product material and product model as the second priority matching items. Use a gradually relaxed matching criteria approach to match similar historical items. Specifically: In each historical project, check if there is a historical project that simultaneously meets the requirements of product platform, product material, and product model; if so, obtain the matching historical project. If not, relax the criteria and match whether there are historical projects that simultaneously meet the requirements of product platform and product model; if so, obtain the matching historical projects. If not, relax the criteria and match whether there are historical projects that simultaneously meet the requirements of product platform and product material consistency; if there are, obtain the matching historical projects. If not, the conditions will be relaxed to include historical projects that are consistent only with the product platform.

4. The method for predicting the manufacturing cycle of high-end equipment products based on machine learning according to claim 1, characterized in that, In step S4, the method for determining the matching process time for each process path is as follows: Based on historical work reporting data, obtain the most recent historical production record of the product using the process path. Based on the start and end times of the process path recorded in the product's historical production record, obtain the matching process time. If the recorded start or end time of the process path is missing, or the matching process time is abnormal, the corresponding standard working hours are used as a substitute.

5. The method for predicting the manufacturing cycle of high-end equipment products based on machine learning according to claim 1, characterized in that, Step S6 includes: Step S61: Determine the fusion feature vector of the key process; the fusion feature vector includes the project period of the key process, the historical average construction period of the key process, the process category of the key process, and the interruption time of the key process. Step S62: Analyze the historical dataset to obtain the fusion feature dataset corresponding to the fusion feature vector of the key process; use the fusion feature dataset to train the process-level regression prediction model to obtain the trained process-level regression prediction model. Step S63, assuming a total of The key processes will then lead to the corresponding construction. Each process-level regression prediction model; for each process-level regression prediction model , Determine its mean square error in the fused feature dataset. The regression prediction model for each process stage is determined using the following formula. weight : ; in: and , which are the maximum and minimum mean square errors in the regression prediction models for each process level, respectively; Step S64, the calculation model for the process time update value of the key process is as follows: ; in: Key processes The updated value of the process time; To match the process time; This represents the predicted process time.

6. The method for predicting the manufacturing cycle of high-end equipment products based on machine learning according to claim 1, characterized in that, The inputs to the vehicle manufacturing cycle prediction model specifically include the following dimensions: project period; busy / slack period of the project; process complexity of the project, i.e., the number of key processes and the number of ordinary processes included; historical average lead time for each process; updated process time values ​​for key processes. Process time for conventional processes.

7. A machine learning-based high-end equipment product manufacturing cycle prediction system that implements the machine learning-based high-end equipment product manufacturing cycle prediction method according to any one of claims 1 to 6, characterized in that, include: Historical project data table, used to record the historical production data of each historical project; The historical project matching module is used to match at least one most similar historical project among various historical projects based on relevant matching rules, given information on new projects that require product manufacturing cycle prediction. The process path filtering module is used to, for each matched historical project, count its process paths based on the historical project data table, and filter out the top N process paths with the most types of processes. The historical dataset construction module is used to obtain the matching process time for each selected process path based on the start and end times in the historical work reporting data. Each process path and the matching process time form a structured historical dataset that can be used for modeling and analysis. The key process identification module is used to determine the difference in actual process time for each product under the same process path based on the historical dataset and the historical project data table, and to identify process paths with large differences as key processes and process paths with small differences as ordinary processes. The process time determination module is used to construct a corresponding process-level regression prediction model for each identified key process, and to train the model using a subset of historical data corresponding to the key process to obtain the trained process-level regression prediction model. The trained process-level regression prediction model is used to perform regression prediction on the process time of key processes, and the predicted process time is obtained. By combining the matched process time and the predicted process time of the key process, the updated process time value of the key process is obtained; for each ordinary process, the matched process time is directly used as its process time. The whole vehicle manufacturing cycle prediction model construction module is used to construct a whole vehicle manufacturing cycle prediction model; the process time update values ​​of each key process, the process time of each ordinary process, and the historical dataset used for modeling analysis are used as inputs to train the whole vehicle manufacturing cycle prediction model to obtain a trained whole vehicle manufacturing cycle prediction model; the trained whole vehicle manufacturing cycle prediction model is used to predict the whole vehicle manufacturing cycle of the new project currently input.

Citation Information

Patent Citations

  • Integrated circuit process parameter optimization method and system based on machine learning

    CN119067028A

  • Industrial user short-term load dynamic prediction method based on historical load data

    CN121414179A