Publicação
A comparison of machine learning approaches for predicting in-car display production quality
| Resumo: | In this paper, we explore eight Machine Learning (ML) approaches (binary and one-class) to predict the quality of in-car displays, measured using Black Uniformity (BU) tests. During production, the industrial manufacturer routinely executes intermediate assembly (screwing and gluing) and functional tests that can signal potential causes for abnormal display units. By using these intermediate tests as inputs, the ML model can be used to identify the unknown relationships between intermediate and BU tests, helping to detect failure causes. In particular, we compare two sets of input variables (A and B) with hundreds of intermediate quality measures related with assembly and functional tests. Using recently collected industrial data, regarding around 147 thousand in-car display records, we performed two evaluation procedures, using first a time ordered train-test split and then a more robust rolling windows. Overall, the best predictive results (92%) were obtained using the full set of inputs (B) and an Automated ML (AutoML) Stacked Ensemble (ASE). We further demonstrate the value of the selected ASE model, by selecting distinct decision threshold scenarios and by using a Sensitivity Analysis (SA) eXplainable Artificial Intelligence (XAI) method. |
|---|---|
| Autores principais: | Matos, Luís Miguel |
| Outros Autores: | Domingues, André; Moreira, Guilherme; Cortez, Paulo; Pilastri, André Luiz |
| Assunto: | Anomaly Detection One-class learning Automated Machine Learning Deep Learning Explainable artificial intelligence Supervised Learning |
| Ano: | 2021 |
| País: | Portugal |
| Tipo de documento: | comunicação em conferência |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | In this paper, we explore eight Machine Learning (ML) approaches (binary and one-class) to predict the quality of in-car displays, measured using Black Uniformity (BU) tests. During production, the industrial manufacturer routinely executes intermediate assembly (screwing and gluing) and functional tests that can signal potential causes for abnormal display units. By using these intermediate tests as inputs, the ML model can be used to identify the unknown relationships between intermediate and BU tests, helping to detect failure causes. In particular, we compare two sets of input variables (A and B) with hundreds of intermediate quality measures related with assembly and functional tests. Using recently collected industrial data, regarding around 147 thousand in-car display records, we performed two evaluation procedures, using first a time ordered train-test split and then a more robust rolling windows. Overall, the best predictive results (92%) were obtained using the full set of inputs (B) and an Automated ML (AutoML) Stacked Ensemble (ASE). We further demonstrate the value of the selected ASE model, by selecting distinct decision threshold scenarios and by using a Sensitivity Analysis (SA) eXplainable Artificial Intelligence (XAI) method. |
|---|