Explainable Product Quality Assessment in a Medical Device Assembly Pilot Line

New technologies and data analysis tools such as deep learning models can be beneficial for product quality assessment purposes. However, these black box models can be challenging due to uncertainty and lack of explainability in sensitive pharmaceutical processes. Therefore, different interpretable algorithms have been proposed to overcome the challenges in complex machine learning models. This paper presents an explainable deep-leaning-based fault detection method for quality assessment in an industrial medical device assembly line. This methodology consists of a multi-layer perceptron model that classifies the samples. Then a layer-wise relevance propagation algorithm seeks to explain the logic behind the prediction. Finally, the heatmap pertaining to relevance propagation visualizes the main contributors to the output prediction. Due to the small industrial dataset, a public dataset associated with a robot-driven screwdriving process assists in evaluating the current methodology. The final results show that the classifier can diagnose different fault classes, and the LRP algorithm can highlight the essential input features and visualize the decision-making process. Furthermore, the LRP algorithm can be beneficial for diagnosing unknown abnormal samples due to the different distribution of contributing features in the heatmap figure. Moreover, a more reliable dimension reduction method can be applied by employing the LRP algorithm and selecting corresponding input data points with higher relevance.

Automated summary

This paper presents an explainable deep-learning-based fault detection method for quality assessment in an industrial medical device assembly line. The method includes a multi-layer perceptron model and a layer-wise relevance propagation algorithm, which can diagnose different fault classes and visualize the decision-making process.