최근 많은 연구자들이 머신러닝과 딥러닝이 적용된 다양한 분야들을 연구해 왔다. 저항점용접은 특히 자동차 생산에 널리 사용되는 접합 공정인데 다양한 공정변수들이 용접 품질에 직접적인 영향을 미친다. 공정변수 및 용접전류 제어시스템에서 용접강도를 보장하기위한 최적용접조건을 설정하는 과정은 인간의 전문지식에 의존하기 때문에 작업자의 실수에 취약하고 비효율적일 수 있다. 머신러닝과 딥러닝은 실시간 ...
최근 많은 연구자들이 머신러닝과 딥러닝이 적용된 다양한 분야들을 연구해 왔다. 저항점용접은 특히 자동차 생산에 널리 사용되는 접합 공정인데 다양한 공정변수들이 용접 품질에 직접적인 영향을 미친다. 공정변수 및 용접전류 제어시스템에서 용접강도를 보장하기위한 최적용접조건을 설정하는 과정은 인간의 전문지식에 의존하기 때문에 작업자의 실수에 취약하고 비효율적일 수 있다. 머신러닝과 딥러닝은 실시간 모니터링 신호를 이용한 비파괴검사기법을 개발하는데 유용하게 사용될 수 있다. 이미지 인식 기법 또한 용접 품질을 개선할 수 있는 유용한 도구이다. 본 논문은 TRIP 1180 고장력강판의 저항점용접에서 유지 시간에 따라 용접 품질이 달라지는 현상을 규명하는 것을 목적으로 하며, 프로파일 가압과 관련된 유지 시간의 민감도를 조사했다. 또한, 본 논문에서는 저항점용접공정 모니터링에 사용할 수 있는 머신러닝 학습 방법도 검토 하였다. 본 논문은 두 가지 주요 부분으로 구성되어 있다. 1. 프로파일 가압을 이용한 유지시간에 대한 용접강도의 민감도에 대한 연구 - 그리고 프로파일 가압을 하면서 상이한 유지 시간을 적용하여 용접 품질변화를 조사하였다. 2. ANN, 딥러닝을 사용하여 너깃 직경과 두께를 예측하는 데 동저항 신호를 사용하였다. 복잡한 다층겹판용접에서는 용접 품질을 향상시키기 위한 여러 방법 중 머신러닝분류(machine learning classification) 기법을 사용하여 용접품질을 판단하였다. 또한 용접공정변수의 변화를 이미지 인식기법을 통하여 정량적으로 산출하였다.
최근 많은 연구자들이 머신러닝과 딥러닝이 적용된 다양한 분야들을 연구해 왔다. 저항점용접은 특히 자동차 생산에 널리 사용되는 접합 공정인데 다양한 공정변수들이 용접 품질에 직접적인 영향을 미친다. 공정변수 및 용접전류 제어시스템에서 용접강도를 보장하기위한 최적용접조건을 설정하는 과정은 인간의 전문지식에 의존하기 때문에 작업자의 실수에 취약하고 비효율적일 수 있다. 머신러닝과 딥러닝은 실시간 모니터링 신호를 이용한 비파괴검사기법을 개발하는데 유용하게 사용될 수 있다. 이미지 인식 기법 또한 용접 품질을 개선할 수 있는 유용한 도구이다. 본 논문은 TRIP 1180 고장력강판의 저항점용접에서 유지 시간에 따라 용접 품질이 달라지는 현상을 규명하는 것을 목적으로 하며, 프로파일 가압과 관련된 유지 시간의 민감도를 조사했다. 또한, 본 논문에서는 저항점용접공정 모니터링에 사용할 수 있는 머신러닝 학습 방법도 검토 하였다. 본 논문은 두 가지 주요 부분으로 구성되어 있다. 1. 프로파일 가압을 이용한 유지시간에 대한 용접강도의 민감도에 대한 연구 - 그리고 프로파일 가압을 하면서 상이한 유지 시간을 적용하여 용접 품질변화를 조사하였다. 2. ANN, 딥러닝을 사용하여 너깃 직경과 두께를 예측하는 데 동저항 신호를 사용하였다. 복잡한 다층겹판용접에서는 용접 품질을 향상시키기 위한 여러 방법 중 머신러닝분류(machine learning classification) 기법을 사용하여 용접품질을 판단하였다. 또한 용접공정변수의 변화를 이미지 인식기법을 통하여 정량적으로 산출하였다.
In recent years, many researchers have been carried out different fields applied machine learning, deep learning. Resistance spot welding is a popular joining process, particularly in the automobile industry production. Weld quality level with a strength-based on process control, welding power sourc...
In recent years, many researchers have been carried out different fields applied machine learning, deep learning. Resistance spot welding is a popular joining process, particularly in the automobile industry production. Weld quality level with a strength-based on process control, welding power source and control system that various process factors directly impact weld quality. Many welding methods depend on human expertise to choose optimal settings, making them more susceptible to human mistakes and inefficiency. Deep learning, machine learning will be used to assess online monitoring signals and non-destructive testing approaches. Image recognition is a powerful tool for enhancing welding quality and reducing production and manufacturing time. This thesis aims to understand the relationship between various hold times and the weld quality of TRIP 1180 steel. Examine hold time sensitivity associated with profile force.Furthermore, the machine learning method being able to use in this thesis studied. The thesis consists of three main parts: 1. Study of relationship to hold-time sensitivity microstructure mechanical properties using profile force and increases different hold time control by profile force for enhancing weld quality. Monitoring signal of micro-spot-welding machine such as time, electrode movement that affects the quality assessment of micro, small scale spot welding. 2. Dynamic resistance signal used to predict the nugget diameter penetration by using ANN, deep learning. Complex stack-up material determines weldability of the sheet metal using the machine learning classification method above all methods for evaluating improving weld quality based on signal process, sheet metal. The investigated machine learning algorithms are decision tree random forest, artificial neural network. 3. To explore the development of a real-time angle measurement system for force compensation. DL CNN for identification of welding surface from force compensation with classification. In this study, hold time was determined final the weld structure increasingly becoming an essential factor by controlled force profile in weld nugget. Profile force involved as the variation of the force during the welding process. The level of the profile force influences the contact resistances that have strongly generated heat. The has been less previous evidence for hold time with profile force. The applied parameter welding of the current 7.5kA profile force was increased from 2.95kN to 4.4kNkgf, cycle squeeze time, 22 cycle welding time, and short hold time variations of 1, 2, 5, 10 cycles the long hold time of 30 cycles 60 cycles, respectively. The electrode used in this study were Cu-Cr-Zr dome radius type and diameter of electrode tip 6mm comparison between constant force and force profile respect different welding of holding time. The tensile shear strength test was accomplished in a universal testing machine. Coupon dimensions 150x50 mm single spot weld was made at the center of an overlapped area 50mm. All samples were performed at room temperature. According to AWS(ISO), standards peak load was extracted from the load-displacement curve that data recorded to evaluate the mechanical properties. The Vickers microhardness tester was being used to measure microhardness distribution under a force of 4.903 N. As shown, gradual hardness looked in the nugget zone CF10, CF90, PF10, PF90, the minimum value was 497 HV, and the maximum value was 522HV in strength. Comparison of the Tensile shear strength, Cross tension strength behavior of TRIP steel. It can be seen that TRIP steel has a good balance between strength and ductility at an appropriate hold time of 30 cycles. In the TRIP 1180 CTS test, a profile force of 30 cycles results in better energy absorption. Tensile Shear Strength: proportional to hold time. Cross-Tension Strength: inversely proportional to hold time. Profile force: no consistent and significant effect on strength for various hold times. CTS reduction seems to be due to the increase in hold time, increasing the cooling rate, resulting in a slight increase in maximum hardness. Due to the characteristics of resistance welding, both the nugget size and heat affected zone is composed of 100% martensite, regardless of the hold times, but even with the same 100% martensite, the hardness increases as the cooling rate increases. The second part of the dissertation focused on weld quality was investigated based on the non-destructive testing method as electrode movement and machine learning. SUS304 metal joining used with the same thickness layers. Three different states are observed of metal commonly used, and welding parameters in the same experiment condition and 2870 electrode movement samples were collected. The electrode movement value was implemented as the feature, and for this multi-class classification problem, an algorithm including a decision tree and multi-layer perceptron was used to classify the electrode movement using a dual accelerometer sensor. This method identifies the correct class; several possible weld quality outcomes from the electrode movement pattern are proposed in this paper. The electrode movement features are extracted from the voltage and current data. We compared the receiver operating characteristic curve metrics of the widely used feed-forward back-propagation ANN with the decision tree. The good experimental results were obtained in the decision tree algorithm as the classification. The highest classification detection accuracy rate achieves 93% decision tree accuracy of 83.23%. There are three different curves; class 2 obtained results better than other results MLP class 2 has a classification accuracy of 92.15%, 0.99 of AUC, 0.86, and 0.82, respectively. (Area under the ROC curve). Precision-Recall curve is class 2 indices close to the upper right corner that is best classification performances. The ANN model can predict diameter penetration good adjustment R 0.9987 of training and testing 0.92 significantly higher than two hidden layers ANN three layers model is a more reliable and accurate tool for predicting the overall outcomes. Predicted data very close to the actual data compared obtained from the deep learning model. Classification-based machine learning techniques were used to predict weld quality. Furthermore, they compared the classification accuracy of four models by ten material stack-up classes. There are four classification algorithms with two performances, Multi-layer Perceptron, Random Forest, KNN, and decision tree Precision-Recall, also known as ROC curve. Different classification algorithm's ROC and PR curves were evaluated. Accuracy is the performance measure that is being examined. Classification algorithms must be used to validate the efficacy of a feature selection algorithm-the better the results, the clearer the prediction. Following an examination of the prediction findings, it was evident that the Decision Tree approach succeeded for all this data. Group 2-1 five different sheet thickness is lower accuracy than other class. Three various dissimilar thickness performances achieved a good result. The third and last part of the dissertation is the study that recommended Convolutional Neural Network founded on multi-classification method welding surface apply VGG16, RESNET50, InceptionV3. A comparison of the various architecture of deep learning models selects the best model that guides the results found to improve performing the proper models using different deep learning optimization algorithms. InceptionV3 model provides the highest classification performance. This multi-classification network's application of non-destructive evaluation approaches has attracted intense interest in modern automotive industries.
In recent years, many researchers have been carried out different fields applied machine learning, deep learning. Resistance spot welding is a popular joining process, particularly in the automobile industry production. Weld quality level with a strength-based on process control, welding power source and control system that various process factors directly impact weld quality. Many welding methods depend on human expertise to choose optimal settings, making them more susceptible to human mistakes and inefficiency. Deep learning, machine learning will be used to assess online monitoring signals and non-destructive testing approaches. Image recognition is a powerful tool for enhancing welding quality and reducing production and manufacturing time. This thesis aims to understand the relationship between various hold times and the weld quality of TRIP 1180 steel. Examine hold time sensitivity associated with profile force.Furthermore, the machine learning method being able to use in this thesis studied. The thesis consists of three main parts: 1. Study of relationship to hold-time sensitivity microstructure mechanical properties using profile force and increases different hold time control by profile force for enhancing weld quality. Monitoring signal of micro-spot-welding machine such as time, electrode movement that affects the quality assessment of micro, small scale spot welding. 2. Dynamic resistance signal used to predict the nugget diameter penetration by using ANN, deep learning. Complex stack-up material determines weldability of the sheet metal using the machine learning classification method above all methods for evaluating improving weld quality based on signal process, sheet metal. The investigated machine learning algorithms are decision tree random forest, artificial neural network. 3. To explore the development of a real-time angle measurement system for force compensation. DL CNN for identification of welding surface from force compensation with classification. In this study, hold time was determined final the weld structure increasingly becoming an essential factor by controlled force profile in weld nugget. Profile force involved as the variation of the force during the welding process. The level of the profile force influences the contact resistances that have strongly generated heat. The has been less previous evidence for hold time with profile force. The applied parameter welding of the current 7.5kA profile force was increased from 2.95kN to 4.4kNkgf, cycle squeeze time, 22 cycle welding time, and short hold time variations of 1, 2, 5, 10 cycles the long hold time of 30 cycles 60 cycles, respectively. The electrode used in this study were Cu-Cr-Zr dome radius type and diameter of electrode tip 6mm comparison between constant force and force profile respect different welding of holding time. The tensile shear strength test was accomplished in a universal testing machine. Coupon dimensions 150x50 mm single spot weld was made at the center of an overlapped area 50mm. All samples were performed at room temperature. According to AWS(ISO), standards peak load was extracted from the load-displacement curve that data recorded to evaluate the mechanical properties. The Vickers microhardness tester was being used to measure microhardness distribution under a force of 4.903 N. As shown, gradual hardness looked in the nugget zone CF10, CF90, PF10, PF90, the minimum value was 497 HV, and the maximum value was 522HV in strength. Comparison of the Tensile shear strength, Cross tension strength behavior of TRIP steel. It can be seen that TRIP steel has a good balance between strength and ductility at an appropriate hold time of 30 cycles. In the TRIP 1180 CTS test, a profile force of 30 cycles results in better energy absorption. Tensile Shear Strength: proportional to hold time. Cross-Tension Strength: inversely proportional to hold time. Profile force: no consistent and significant effect on strength for various hold times. CTS reduction seems to be due to the increase in hold time, increasing the cooling rate, resulting in a slight increase in maximum hardness. Due to the characteristics of resistance welding, both the nugget size and heat affected zone is composed of 100% martensite, regardless of the hold times, but even with the same 100% martensite, the hardness increases as the cooling rate increases. The second part of the dissertation focused on weld quality was investigated based on the non-destructive testing method as electrode movement and machine learning. SUS304 metal joining used with the same thickness layers. Three different states are observed of metal commonly used, and welding parameters in the same experiment condition and 2870 electrode movement samples were collected. The electrode movement value was implemented as the feature, and for this multi-class classification problem, an algorithm including a decision tree and multi-layer perceptron was used to classify the electrode movement using a dual accelerometer sensor. This method identifies the correct class; several possible weld quality outcomes from the electrode movement pattern are proposed in this paper. The electrode movement features are extracted from the voltage and current data. We compared the receiver operating characteristic curve metrics of the widely used feed-forward back-propagation ANN with the decision tree. The good experimental results were obtained in the decision tree algorithm as the classification. The highest classification detection accuracy rate achieves 93% decision tree accuracy of 83.23%. There are three different curves; class 2 obtained results better than other results MLP class 2 has a classification accuracy of 92.15%, 0.99 of AUC, 0.86, and 0.82, respectively. (Area under the ROC curve). Precision-Recall curve is class 2 indices close to the upper right corner that is best classification performances. The ANN model can predict diameter penetration good adjustment R 0.9987 of training and testing 0.92 significantly higher than two hidden layers ANN three layers model is a more reliable and accurate tool for predicting the overall outcomes. Predicted data very close to the actual data compared obtained from the deep learning model. Classification-based machine learning techniques were used to predict weld quality. Furthermore, they compared the classification accuracy of four models by ten material stack-up classes. There are four classification algorithms with two performances, Multi-layer Perceptron, Random Forest, KNN, and decision tree Precision-Recall, also known as ROC curve. Different classification algorithm's ROC and PR curves were evaluated. Accuracy is the performance measure that is being examined. Classification algorithms must be used to validate the efficacy of a feature selection algorithm-the better the results, the clearer the prediction. Following an examination of the prediction findings, it was evident that the Decision Tree approach succeeded for all this data. Group 2-1 five different sheet thickness is lower accuracy than other class. Three various dissimilar thickness performances achieved a good result. The third and last part of the dissertation is the study that recommended Convolutional Neural Network founded on multi-classification method welding surface apply VGG16, RESNET50, InceptionV3. A comparison of the various architecture of deep learning models selects the best model that guides the results found to improve performing the proper models using different deep learning optimization algorithms. InceptionV3 model provides the highest classification performance. This multi-classification network's application of non-destructive evaluation approaches has attracted intense interest in modern automotive industries.
주제어
#Hold time Profile Force Machine Learning Classification Force Resistance Spot Welding
학위논문 정보
저자
라쉬 뭉흐치멕
학위수여기관
명지대학교 대학원
학위구분
국내박사
학과
기계공학과 Mechanical Engineering
지도교수
장희석
발행연도
2021
총페이지
xiii, 107p
키워드
Hold time Profile Force Machine Learning Classification Force Resistance Spot Welding
※ AI-Helper는 부적절한 답변을 할 수 있습니다.