Foreword:A novel method based on machine learning algorithms was proposed to realize the quality prediction and the process parameter optimization, in order to reuse the process information and the inspection information of machining parts effectively. A model-based definition (MBD) model which was integrated with process information and inspection information was treated as input. Process and inspection parameter extraction based on the MBD model was developed and the corresponding structured data set was established through the secondary development of three-dimensional modeling software. Several classifiers in machine learning were used to construct the quality prediction model based on process parameters and quality classification labels. Combining the information gain algorithm, after sorting all process parameters, the process parameter that had the greatest impact on quality was selected. Quality prediction and process parameter optimization tool set was developed to realize the optimization of the selected parameter by using the gradient boost decision tree algorithm. The validity and the reliability of the proposed method were verified by the milling experiment data provided by an aviation company. Results show that the proposed method can realize the quality prediction and process parameter optimization of machining parts effectively.
Model-based definition [1-3] (model-based definition, MBD) digital design and manufacturing technology has been adopted by many manufacturing companies. On the basis of MBD technology, digital thread (digital thread) and digital twin (digital twin, DT) ) has become the trend of the development of manufacturing informatization. As the key technology of cyber-physical system, digital twin [4-5] can digitally describe the physical world and effectively control the data of the whole life cycle of products. The production of digital twin Production and application is the further development of MBD technology, which provides new ideas for product development and production under the full three-dimensional development mode. In the “model-driven” product development link, the part process information model is constructed according to MBD technology, and the process is used.
The information model and process parameters can quickly build a process model to meet the requirements of the process planning process for information processing speed [6-8]; in the product inspection process, the process/process model is used to develop a three-dimensional digital inspection process system, and the inspection process model is constructed to solve traditional problems. Two-dimensional detection methods are slow, inefficient and low in accuracy [9-10]
The above models are all constructed using the data collected from the bottom layer. How to deeply utilize the process and inspection data carried by these models has gradually become the focus of current digital twin application research. In recent years
In recent years, some scholars have used the process information and inspection information collected in the manufacturing process to conduct data mining analysis, and used the analysis results to estimate the quality of parts and optimize the process parameters.
Tirkel[11] used machine learning algorithm to establish a prediction model of wafer processing cycle, which improved the efficiency of product manufacturing and supply. Zhu Xuechu et al. [12] studied industrial big data mining of wafer intelligent manufacturing system, and proposed a proposal based on the effective information obtained from mining. [13] designed a regression model based on massive wafer manufacturing data, and analyzed the relationship between 1 669 wafer processing parameters and order delivery indicators. Abajo et al. [14] proposed a tinplate quality diagnosis model based on process data by analyzing the quality diagnosis data of the new type of tinplate.
Based on the idea of digital twin and inspired by existing research, this study proposes a method for quality prediction and process parameter optimization of machined parts. The MBD model integrating process information and inspection information is the only.
One data source, using CATIA component application architecture (CAA) secondary development technology, to complete the automatic extraction of process information and inspection information and store it in the database; using logistic regression (LR), support vector machine (support vector machine) Machine, SVM), extreme gradient boosting (extreme gradient boosting, XGBoost) and other classifiers to build prediction models and complete data training to achieve part quality prediction based on process parameters;
The boost decision tree (GBDT) algorithm builds features to improve the accuracy of classification model prediction; based on the process parameters and quality classification labels of the parts, combined with the information gain algorithm, the process parameters that affect the quality of the parts are prioritized; finally, through the GBDT regression The training of the model optimizes the process parameters of the part.
Quality Prediction Method Of Machined Parts
According to the relevant theories of industrial engineering, the main factors affecting the quality of machined parts can be summarized into five aspects: processing methods, equipment, materials, environment and process parameters. Starting with the impact on the quality of the parts, the machine learning algorithm is used to build a part quality prediction model, and the quality of the parts is predicted by the process parameters selected in the machining process.
In actual industrial scenarios, there are various uncertainties and randomness in the influence of process parameter values on the quality of parts. Therefore, this study gives the structured data of general process information and inspection information.
Data format, process parameter prioritization interface, part quality model accuracy verification interface, and part quality prediction interface. All process parameters are prioritized based on the structured data format to assist process personnel in selecting effective process parameters; for the selected process Parameters, provide the accuracy verification interface of various parts quality prediction models, and display the fitting situation of the model to the process parameters; select the model with better fitting situation as the part quality prediction model to assist the technicians in predicting the part quality
Prioritization of process parameters affecting part quality
Given a training dataset and features, information entropy [15] represents the uncertainty in classifying the dataset; conditional entropy represents the uncertainty in classifying the dataset given the features. The information entropy and The difference of conditional entropy is the information gain, which represents the degree of uncertainty reduction when classifying the dataset because of the given features.
By adding the information gain of all process parameters of the part, the influence of different process parameters on the quality of the part can be judged, and the process parameter with the largest information gain has the greatest influence on the quality of the part.
By adding the information gain of all process parameters of the part, the influence of different process parameters on the quality of the part can be judged, and the process parameter with the largest information gain has the greatest influence on the quality of the part.
Part Quality Prediction Model Construction
The principle of the part quality prediction model is to use the processing parameters to realize the two-classification of the part quality, so it is a classification problem. Common classifications such as LR, SVM[16], XGBoost[17] and GBDT+LR[18-19] can be selected. Among them, LR, SVM and XGBoost are more versatile classifiers and can be used as basic classifiers.
GBDT belongs to the Boosting algorithm. The former tree structure can build features with discrimination for most samples, and the latter tree structure fits a small number of samples, so GBDT can be used to build features with discrimination. In special cases with few process parameters, In order to improve the accuracy of the part quality prediction model, effective features must be constructed to make up for the lack of human experience, and GBDT combined with LR can be used as the part quality prediction model in such cases.
Logistic regression model
The input of the part quality prediction model built using the logistic regression algorithm is the process parameters of the part, and the output is whether the quality of the part is qualified. For the sample with sample number i, the input of the model is the process parameters of the current numbered part, denoted as; logistic regression model The internal parameter of , that is, the weight of the process parameter, is recorded as. It can be seen that the closer the calculated value of , is to positive infinity,the closer the result is to 1, the part quality is classified as qualified; the closer the calculated value of is to negative infinity, the closer the classification result is to 0, that is, the part quality is classified as unqualified. For a given input instance, the sum of 2 Conditional probability, assigning instances to classes with higher probability.
GBDT model
When applying GBDT to fit the original data set and construct new features, it is necessary to splicing the new features and the original data set features, and then predict the quality of the part. Fit the part quality label. It means that the decision tree uses GBDT to fit the training data, and saves the trained classifier. Re-read all the data through the classifier and build new features. For the numbered samples, build features. Finally, set the The constructed new features and original features are spliced to form the final prediction feature, denoted as , and the process parameter weight of the logistic regression model is denoted as .
1.3.3 Ten-fold cross-validation In machine learning research, in order to verify the accuracy of the model, the original data is generally divided into a training set and a test set. With the effect of data testing algorithm, the data set is divided into 10 parts by randomly generating indexes, a total of 10 experiments are carried out, 1 part of the data is taken as test data each time, and the remaining 9 parts of data are used as training data for experiments.
The method is called 10-fold cross-validation
1.3.4 Part quality prediction process Figure 1 shows the construction process of the part quality prediction model, which is divided into four steps.
- Read the preprocessed data, and separate the process parameters and the part quality label into two variables;
- The user selects the model and generates the corresponding handle for model training;
- Ten-fold cross-validation training model, output the corresponding accuracy rate and save the training model;
- Load the pre-trained parts quality prediction model, read the prediction data, and output the model prediction results.
Process Parameter Optimization Method For Machined Parts
The principle of the process parameter optimization model is to use machine learning algorithms to combine process parameters and part quality labels to optimize a certain dimension of process parameters.
In the trial production stage, the information gain of all process parameters is calculated and sorted. When the remaining parameters remain unchanged, the process parameter optimization model is used to predict the one with the highest information gain (that is, the process parameter that has the greatest impact on the quality of the part) to achieve single-point optimization. .
Combined with the structured data format given in Section 1.1, the construction of the process parameter optimization model must select the process parameters to be optimized, and use the machine learning algorithm to build the regression model; then use the remaining process parameters and part quality labels to fit the process parameters to be optimized; Finally, a set of process parameters to be optimized is selected, and the part quality label is changed to the optimal design value, and the two are input into the model to obtain the optimization result.
Structured data format for process parameter optimization
The process parameter optimization model is aimed at a certain problem of a certain production line and has a certain pertinence. Therefore, the generalization ability of the model is not required, and the GBDT model can be directly selected for construction.
Build: For the sample number, take the dimension process parameters and the part quality label as new features, denoted as, input the GBDT model given in Section 1.3.2 for training, and save the trained model; take a certain piece of processing data, determine the to-be-optimized process parameters, and keep the rest of the process parameters unchanged, change the part quality label to the optimal value, rebuild the feature, record as, and input
into the trained model to get the optimized process parameters.
Instance Verification
In order to confirm the validity of the aforementioned parts quality prediction model and process parameter optimization model, the milling experimental data provided by an aviation company was used for verification. The experiment adopts the control variable method, and the experiments are carried out under different parameters of axial depth of cut and radial depth of cut. , the surface of aluminum alloy material is milled by means of variable speed, and the influence of three variables of machine tool spindle speed, axial depth of cut and radial depth of cut on the surface roughness of milling is studied by measuring and recording the surface roughness. The material is 7050-T7415 aviation aluminum alloy, which is processed by V1-2000-2T five-coordinate CNC gantry milling machine and 10 carbide end mill, and the surface roughness is measured by the TR200 measuring instrument of Beijing Jitai Company.
The process parameters of the same feature in the process information are added under the sub-nodes of the corresponding process steps according to the process steps. When adding inspection features, it will automatically generate whether the parts fit according to the theoretical value and tolerance value.
The nodes of the grid. The integration results of process information and inspection information in the MBD model are shown in Fig. 5.
The extraction process of process information and detection information is shown in Figure 6, which is divided into three steps.
- 1) Obtain the MBD model and traverse the process and detection process linked list;
- 2) Extract the cutting parameters and detection results of the specified features of the MBD model respectively;
- 3) The process information and detection information of the obtained MBD model are expressed in a structured manner to provide data basis for subsequent machine learning.
Quality prediction of machined parts
To validate the proposed method, a part quality prediction and process parameter optimization toolset user interface was developed using Python, as shown in Figure 7. All process parameters were analyzed using an information gain algorithm.
It can be seen from the results that the information gain of the spindle speed is the largest, that is, the impact of the spindle speed on the quality of the parts is the greatest. , SVM, XGBoost three classifiers, so as to realize the two-classification of the quality of the parts, and record the average accuracy of each classifier. Then, in order to further improve the accuracy of the model prediction, use GBDT to fit the original data and construct a new model. features, and spliced with the original process parameters and then input into the LR classifier. The final average accuracy of all models is shown in Table 4. From the data analysis in the table, it can be seen that the relatively simple.
- Process parameters affecting part quality
- Prioritize build parts
- Quality Prediction Model Gradient Descent
- optimization model
- Feedback to optimize process parameters
- MBD model
- Process information
- Detection information
- Process parameter extraction
- Detection information extraction
The prediction accuracy of a single LR classifier is low when the number of features is small, while the accuracy of the two complex classifiers SVM and XGBoost is greatly improved; the accuracy of GBDT combined with the LR classifier is 86.31%, which is higher than that of XGBoost. It improved by 3.43%, showing the best prediction accuracy. On the basis of the verified classifier accuracy, in order to further verify the reliability of the model, 5 pieces of raw data in the processed data were randomly selected and input into GBDT+ with the highest accuracy rate. The prediction is made in LR, and the results are shown in Table 5. From the data analysis in the table, it can be seen that the GBDT+LR model has a better fitting effect for the data that the rotational speed increases in turn when the radial depth of cut and the axial depth of cut are determined. , which verifies the reliability of the prediction model.
Process parameter optimization of machining parts
In order to verify the effectiveness of the process parameter optimization method, 450 pieces of data were randomly sampled from 500 pieces of experimental data.
Use the rotational speed as a label, use GBDT to fit it and save the fitted model; randomly sample 5 data from the remaining 50 data, change the surface roughness value to an ideal value before prediction, and output the optimized value. The spindle speed is shown in Figure 8. The optimized spindle speed is 14 750. By comparing the original machining data, it can be seen that the GBDT model has a better fitting effect on the process parameters and surface roughness data.
Conclusion
This study combines the “model-driven” product development model and the relevant theories of data mining to explore the application of machine learning algorithms to the quality prediction of machined parts and the optimization of process parameters.
The information gain algorithm is used to prioritize the process parameters, which provides an important basis for the optimization of the part machining process; the process parameters that have the greatest impact on the quality of the parts are screened out; the quantitative prediction of the parts based on the priority order of the process parameters realizes the uncertainty of the process parameters. Quality evaluation of quality; using measured processing data and information gain algorithm to optimize process parameters, effectively reducing its dependence on the experience of process personnel, and improving the intelligence level of process optimization work.
The example shows that the method proposed in this study and the corresponding tool set developed in this study can effectively and reliably realize the quality prediction and process parameter optimization of machined parts in practical applications, which is an important part of MBD.
The in-depth application and promotion of technology and digital twin ideas provide a solid technical foundation. The part quality prediction and process parameter optimization models constructed in this study are only suitable for solving the problem of single-point process parameters. The focus of the next research work will be to For the problem that the coupling of multiple parameters in the same process affects the quality of the parts, the process parameters of the previous process and the quality of the parts affect the problem of the next process, and other complex scenes in the actual processing process, carry out part quality prediction and process parameter optimization.