A machine learning-based production line flaw risk prediction method and system

Abstract

This invention provides a machine learning-based method and system for predicting production line defect risks. The method includes collecting historical operating data and product quality inspection data from each workstation on the production line, preprocessing this data to construct a feature dataset; building and training a fusion machine learning model, and deploying it to the edge terminal of the production line; real-time acquisition of workstation operating data as input to the model, outputting defect risk levels, synchronously locating associated workstations and parameters, and generating and pushing early warning information. The system includes modules for historical data acquisition, data preprocessing, and model building and training, which operate collaboratively through a control system. This invention achieves dynamic prediction of production line defect risks and accurate location of associated workstations, providing data support for production line process optimization.

Description

Technical Field

[0001] This invention relates to the field of product processing defect risk prediction technology, specifically a method and system for predicting production line defect risks based on machine learning. Background Technology

[0002] In the operation of industrial production lines, the detection and traceability of product defects are crucial for ensuring product quality. Traditional production line defect management often adopts a post-production inspection model, identifying defective products through manual sampling or end-of-line inspection equipment. This method cannot provide early warning of defect risks during production. Once abnormal process parameters occur, a large number of defective products can be generated, significantly increasing the company's production costs. At the same time, manually determining the workstations and parameters associated with defects is limited by the experience level of personnel, making it difficult to accurately pinpoint the core factors causing defects. This results in a lack of targeted process adjustments and fails to reduce the defect rate at its root.

[0003] With the gradual application of machine learning technology in the industrial field, some production lines have begun to introduce simple models to assist in defect analysis. However, these applications have significant limitations. Existing models are mostly trained on data from single product categories, lacking adaptability to multi-product mixed-line production scenarios. When the production line switches product categories, the model's prediction accuracy drops significantly. Furthermore, the data acquisition process in existing systems lacks a precise correlation between workstation operation data and product quality data. The data preprocessing workflow is simplistic, with invalid and noisy data directly input into the model, leading to significant deviations in prediction results. Simultaneously, most models are deployed on cloud servers, lacking real-time performance and failing to meet the dynamic adjustment needs of the production line. Overall, the application effect has not met expectations. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a machine learning-based method and system for predicting production line defects, enabling dynamic prediction of production line defects and precise location of associated workstations.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A machine learning-based method for predicting production line defect risk includes the following steps:

[0007] S1. Collect historical operating data and corresponding product quality inspection data from each workstation on the production line;

[0008] S2. Preprocess the collected historical data to complete data cleaning and feature integration;

[0009] S3. Build a machine learning model to establish the correlation between historical operating data and product defect results;

[0010] S4. Divide the dataset to train and validate the model, and determine the optimal structure and parameters of the model;

[0011] S5. Deploy the trained model to the production line monitoring system, collect the operating data of each workstation in real time and input it into the model;

[0012] S6. The model outputs the product defect risk prediction results and simultaneously locates the workstations associated with the defects and their corresponding operating parameters.

[0013] S7. Generate defect risk warning information and complete the single defect risk prediction and workstation parameter analysis process.

[0014] Preferably, in step S1, the equipment operating parameters, material input parameters, and environmental parameters of each workstation are collected, and the defect type data, defect location data, and defect judgment result data of the product are collected simultaneously. Data association identifiers are established according to production batches to ensure that the workstation data corresponds one-to-one with the product quality data.

[0015] Preferably, in step S2, missing and outlier values ​​in the data are removed, key data gaps are filled by interpolation, parameters of different dimensions are standardized, data features that are potentially related to product defects are extracted, and a feature dataset is constructed.

[0016] Preferably, in step S3, a fusion model is constructed by selecting a decision tree algorithm, a random forest algorithm, and a convolutional neural network algorithm. The feature dataset is used as the input layer, and the product defect results are used as the output layer. The hierarchical structure and number of neurons of the model are set, and a nonlinear correlation between the workstation parameters and the defect results is established.

[0017] Preferably, in step S4, the feature dataset is divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used to drive iterative learning of the model. The model's hierarchical structure and parameters are adjusted through the validation set. The prediction accuracy of the model is verified using the test set. The optimal model whose prediction accuracy meets the preset requirements is retained.

[0018] Preferably, in step S5, the optimal model is packaged and deployed to the edge computing terminal of the production line. The dynamic operation data of each workstation is collected in real time through sensors, and the data is transmitted synchronously according to a preset collection frequency. The dynamic operation data is converted into an input format that the model can recognize and then input into the model.

[0019] Preferably, in step S6, the model outputs the product defect risk level through real-time running data inference, locates the workstations highly correlated with the defects through feature reverse tracing, extracts the operating parameters of the workstations that exceed the normal fluctuation range, and forms a corresponding list of defect-related workstations and parameters.

[0020] A machine learning-based production line defect risk prediction system includes a historical data acquisition module, a data preprocessing module, a model building and training module, a model deployment module, a real-time data acquisition module, a risk prediction and analysis module, an early warning information generation module, and a control system module. The historical data acquisition module collects historical operating data and product quality inspection data from each workstation. The data preprocessing module performs data cleaning, imputation, and feature integration. The model building and training module constructs a fusion model and completes training and validation. The model deployment module encapsulates and deploys the optimal model to an edge computing terminal. The real-time data acquisition module collects dynamic operating data from each workstation. The risk prediction and analysis module outputs the defect risk level and identifies associated workstations and parameters. The early warning information generation module generates risk warning information. The control system module is electrically connected to each module and controls the modules to collaboratively complete the prediction and analysis process.

[0021] Preferably, the historical data acquisition module includes a workstation data acquisition unit, a quality data acquisition unit, and a data association unit; the workstation data acquisition unit is used to collect data related to equipment, materials, and environment at each workstation; the quality data acquisition unit is used to collect data related to product defect types, locations, and judgment results; the data association unit is used to add a unified identifier to the two types of data according to production batches; the early warning information generation module includes an early warning rule configuration unit and an early warning information push unit; the early warning rule configuration unit is used to set early warning trigger conditions corresponding to different risk levels; the early warning information push unit is used to push early warning information to the terminal devices of the corresponding workstations.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] The historical data acquisition module sets up a data association unit, which can add a unified identifier to the workstation operation data and quality inspection data according to the production batch or product category, ensuring that the two types of data correspond one-to-one, providing accurate data association for model training, and avoiding model prediction deviations caused by data misalignment.

[0024] The data preprocessing module performs cleaning, imputation, standardization, and feature extraction operations, which can effectively remove invalid data such as missing values ​​and outliers, and integrate them into a high-quality feature dataset, improving the efficiency and accuracy of model training and laying a data foundation for subsequent risk prediction.

[0025] The model building and training module selects multiple algorithms to build a fusion model, which can give full play to the advantages of different algorithms and establish a non-linear correlation between workstation parameters and product defects. Compared with a single algorithm model, it has higher prediction accuracy and generalization ability.

[0026] The model deployment module deploys the optimal model to the edge computing terminal, enabling localized calculation of production data, significantly improving the real-time performance of data processing and risk prediction, and meeting the needs of dynamic monitoring of the production line.

[0027] The risk prediction and analysis module uses feature-based reverse tracing to locate associated workstations and parameters, which can accurately identify core workstations and abnormal parameters that cause defect risks, providing a clear direction for production line process adjustments and reducing the defect rate at its source.

[0028] The early warning information generation module sets up early warning rule configuration and information push unit, which can set early warning trigger conditions according to the actual needs of the production line and accurately push early warning information to the corresponding workstation terminal to realize timely intervention of defect risks.

[0029] The system's modules work together through the control system to automate the entire process of defect risk prediction, reduce manual intervention, improve the efficiency of production line quality management, and adapt to the diverse needs of single-category and multi-category mixed production lines. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention;

[0031] Figure 2 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0032] The invention will now be further described with reference to the accompanying drawings.

[0033] like Figure 1 As shown, a machine learning-based method for predicting production line defect risk includes the following steps:

[0034] S1. Collect historical operating data and corresponding product quality inspection data from each workstation on the production line:

[0035] Collect equipment operating parameters, material input parameters, and environmental parameters from each workstation. Simultaneously collect product defect type data, defect location data, and defect judgment result data. Establish data association identifiers according to production batches to ensure that workstation data corresponds one-to-one with product quality data.

[0036] S2. Preprocess the collected historical data to complete data cleaning and feature integration:

[0037] Remove missing and outlier values ​​from the data, fill in key data gaps using interpolation, standardize parameters of different dimensions, extract data features that are potentially related to product defects, and construct a feature dataset.

[0038] S3. Build a machine learning model to establish the correlation between historical operating data and product defect results:

[0039] A fusion model was constructed by selecting decision tree algorithm, random forest algorithm, and convolutional neural network algorithm. The feature dataset was used as the input layer and the product defect results were used as the output layer. The hierarchical structure and number of neurons of the model were set, and a nonlinear correlation between the workstation parameters and the defect results was established.

[0040] S4. Divide the dataset to train and validate the model, and determine the optimal structure and parameters of the model:

[0041] The feature dataset is divided into training set, validation set and test set according to a preset ratio. The training set is used to drive the model to learn iteratively. The model's hierarchical structure and parameters are adjusted through the validation set. The prediction accuracy of the model is verified through the test set. The optimal model with prediction accuracy that meets the preset requirements is retained.

[0042] S5. Deploy the trained model to the production line monitoring system, collect real-time operating data from each workstation, and input it into the model:

[0043] The optimal model is packaged and deployed to the edge computing terminal of the production line. The dynamic operation data of each workstation is collected in real time through sensors. The data is transmitted synchronously according to the preset collection frequency. The dynamic operation data is converted into an input format that the model can recognize and then input into the model.

[0044] S6. The model outputs product defect risk prediction results and simultaneously locates the workstations associated with the defects and their corresponding operating parameters:

[0045] The model outputs the product defect risk level through real-time running data inference, and locates the workstations that are highly correlated with the defects by reverse tracing the features, extracts the operating parameters of the workstations that exceed the normal fluctuation range, and forms a corresponding list of defect-related workstations and parameters.

[0046] S7. Generate defect risk warning information and complete the single defect risk prediction and workstation parameter analysis process.

[0047] like Figure 2As shown, a machine learning-based production line defect risk prediction system includes a historical data acquisition module, a data preprocessing module, a model building and training module, a model deployment module, a real-time data acquisition module, a risk prediction and analysis module, an early warning information generation module, and a control system module. The historical data acquisition module collects historical operating data and product quality inspection data from each workstation. The data preprocessing module performs data cleaning, imputation, and feature integration. The model building and training module constructs a fusion model and completes training and verification. The model deployment module encapsulates and deploys the optimal model to an edge computing terminal. The real-time data acquisition module collects dynamic operating data from each workstation. The risk prediction and analysis module outputs the defect risk level and identifies associated workstations and parameters. The early warning information generation module generates risk early warning information. The control system module is electrically connected to each module and controls the modules to collaboratively complete the prediction and analysis process.

[0048] Furthermore, the historical data acquisition module includes a workstation data acquisition unit, a quality data acquisition unit, and a data association unit. The workstation data acquisition unit is used to collect data related to equipment, materials, and the environment at each workstation. The quality data acquisition unit is used to collect data related to product defect types, locations, and judgment results. The data association unit is used to add a unified identifier to the two types of data according to production batches. The early warning information generation module includes an early warning rule configuration unit and an early warning information push unit. The early warning rule configuration unit is used to set early warning trigger conditions corresponding to different risk levels. The early warning information push unit is used to push early warning information to the terminal devices of the corresponding workstations.

[0049] Example 1

[0050] This embodiment is applied to defect risk prediction and workstation parameter analysis for a single product category production line. The specific process is as follows:

[0051] The historical data acquisition module collects equipment operating parameters, material input parameters, and environmental parameters for each workstation on the production line, and simultaneously collects corresponding product defect type data, defect location data, and defect judgment result data. The data association unit adds a unified identifier to the two types of data according to the production batch, ensuring a one-to-one correspondence between workstation data and product quality data.

[0052] The data preprocessing module processes the collected historical data, removes missing and outlier values, fills in key data gaps using interpolation, standardizes parameters of different dimensions, extracts data features that are potentially related to product defects, and constructs a feature dataset.

[0053] The model building and training module selects decision tree, random forest, and convolutional neural network algorithms to construct a fusion model. The feature dataset is divided into training, validation, and test sets according to a preset ratio. The training set drives iterative model learning, the validation set adjusts the model's hierarchical structure and parameters, and the test set verifies the model's prediction accuracy. The optimal model with prediction accuracy meeting the preset requirements is retained.

[0054] The model deployment module encapsulates and deploys the optimal model to the edge computing terminal of the production line. The real-time data acquisition module is activated, using sensors to collect dynamic operating data from each workstation in real time. The data is transmitted synchronously at a preset acquisition frequency, and then converted into an input format recognizable by the model before being input into the model.

[0055] The risk prediction and analysis module outputs the product defect risk level. It then uses feature-based reverse tracing to locate workstations highly correlated with the defect, extracting operating parameters from those workstations that exceed normal fluctuation ranges, and creating a list of defect-related workstations and their corresponding parameters. The early warning information generation module generates early warning information based on preset early warning rules and pushes the information to the terminal devices of the corresponding workstations via a push unit, completing a single defect risk prediction and workstation parameter analysis process.

[0056] Example 2

[0057] This embodiment is applied to defect risk prediction and workstation parameter analysis in multi-product mixed production lines. The specific process is as follows:

[0058] The historical data acquisition module collects equipment operating parameters, material input parameters, and environmental parameters for different product categories at each workstation of the mixed-line production line. Simultaneously, it collects defect type data, defect location data, and defect judgment result data for each product category. The data association unit adds dual identifiers to the data according to product category and production batch, establishing a correspondence between multi-category workstation data and product quality data.

[0059] The data preprocessing module classifies and processes the collected historical data of multiple categories, divides the data into subsets according to product categories, removes missing and outlier values, fills in key data gaps, completes parameter standardization for each subset, extracts defect-related data features specific to each product category, and constructs a multi-category feature dataset.

[0060] The model building and training module constructs a fusion model for each product category based on its technological characteristics. The feature datasets for each category are divided into training, validation, and test sets for model training and validation. After training, a multi-category model library is established, and model calling rules are set to ensure the system can automatically match the corresponding model based on the product category being produced in real time.

[0061] The model deployment module encapsulates and deploys a multi-category model library to the edge computing terminal of the production line. The real-time data acquisition module is activated to collect dynamic operating data from each workstation and information on the product categories currently being produced. Data is transmitted at a preset frequency and converted into a model-recognizable format. The system automatically calls the corresponding model based on the product category information.

[0062] The risk prediction and analysis module outputs the defect risk level of the current product category, identifies workstations and parameters highly correlated with defects in this category, and generates a category-specific defect association list. The early warning information generation module generates differentiated early warning information based on the early warning rules for this category and pushes it to the corresponding workstation terminal devices. The system continuously monitors the production process; when production categories change, the system automatically switches models and repeats the above prediction and analysis process, achieving dynamic prediction of defect risks in multi-category mixed-line production.

Claims

1. A machine learning-based production line defect risk prediction method, characterized by, Includes the following steps: S1. Collect historical operating data of each workstation on the production line and corresponding product quality inspection data; S2. Preprocess the collected historical data to complete data cleaning and feature integration; S3. Build a machine learning model to establish the correlation between historical operating data and product defect results; S4. Divide the dataset to train and validate the model, and determine the optimal structure and parameters of the model; S5. Deploy the trained model to the production line monitoring system to collect real-time operating data from each workstation and input it into the model; S6. The model outputs the product defect risk prediction results and simultaneously locates the workstations associated with the defects and their corresponding operating parameters. S7. Generate defect risk warning information and complete the single defect risk prediction and workstation parameter analysis process.

2. The machine learning based production line defect risk prediction method of claim 1, wherein, In step S1, the equipment operating parameters, material input parameters, and environmental parameters of each workstation are collected. Simultaneously, the defect type data, defect location data, and defect judgment result data of the product are collected. Data association identifiers are established according to production batches to ensure that the workstation data corresponds one-to-one with the product quality data.

3. The machine learning based production line defect risk prediction method of claim 1, wherein, In step S2, missing and outlier values ​​in the data are removed, key data gaps are filled by interpolation, parameters of different dimensions are standardized, data features that are potentially related to product defects are extracted, and a feature dataset is constructed.

4. The machine learning based production line defect risk prediction method of claim 1, wherein, In step S3, a fusion model is constructed by selecting decision tree algorithm, random forest algorithm, and convolutional neural network algorithm. The feature dataset is used as the input layer, and the product defect results are used as the output layer. The hierarchical structure and number of neurons of the model are set, and a nonlinear correlation between the workstation parameters and the defect results is established.

5. The machine learning based production line defect risk prediction method of claim 1, wherein, In step S4, the feature dataset is divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used to drive iterative learning of the model. The model's hierarchical structure and parameters are adjusted through the validation set. The prediction accuracy of the model is verified using the test set. The optimal model whose prediction accuracy meets the preset requirements is retained.

6. The machine learning based production line defect risk prediction method of claim 1, wherein, In step S5, the optimal model is packaged and deployed to the edge computing terminal of the production line. The dynamic operation data of each workstation is collected in real time through sensors, and the data is transmitted synchronously according to the preset collection frequency. The dynamic operation data is converted into an input format that the model can recognize and then input into the model.

7. The machine learning based production line defect risk prediction method of claim 1, wherein, In step S6, the model outputs the product defect risk level through real-time running data inference, locates the workstations highly correlated with the defects through feature reverse tracing, extracts the operating parameters of the workstations that exceed the normal fluctuation range, and forms a corresponding list of defect-related workstations and parameters.

8. A machine learning based production line flaw risk prediction system for performing the method of any one of claims 1 to 7, characterized by, It includes a historical data acquisition module, a data preprocessing module, a model building and training module, a model deployment module, a real-time data acquisition module, a risk prediction and analysis module, an early warning information generation module, and a control system module; the historical data acquisition module is used to collect historical operating data and product quality inspection data from each workstation. The data preprocessing module is used to perform data cleaning, imputation, and feature integration; The model building and training module is used to build the fusion model and complete training and validation. The model deployment module is used to encapsulate and deploy the optimal model to the edge computing terminal; the real-time data acquisition module is used to collect dynamic operating data of the workstation; the risk prediction and analysis module is used to output the defect risk level and locate the associated workstation and parameters; the early warning information generation module is used to generate risk early warning information; the control system module is electrically connected to each module and is used to control each module to work together to complete the prediction and analysis process.

9. The machine learning based production line defect risk prediction system of claim 8, wherein, The historical data acquisition module includes a workstation data acquisition unit, a quality data acquisition unit, and a data association unit; the workstation data acquisition unit is used to collect relevant data on equipment, materials, and environment at each workstation. The quality data acquisition unit is used to collect data related to product defect types, locations, and judgment results; the data association unit is used to add unified identifiers to the two types of data according to production batches; the early warning information generation module includes an early warning rule configuration unit and an early warning information push unit. The early warning rule configuration unit is used to set the early warning trigger conditions corresponding to different risk levels; The early warning information push unit is used to push early warning information to the terminal equipment of the corresponding workstation.

Images