Development of a quality prediction framework for industry 4.0 optoelectronics assembly lines

Abstract

As a result of the development of new technologies in the field of optoelectronics, a number of industries, including solar energy, LED, industrial lasers, and consumer electronics, are in the midst of a revolution. The production of optoelectronic goods must accommodate the development of new tools and processes, in addition to an increase in demand, complexity, and demand-side elements in respective order. Particularly in the laser sector, which makes use of high-end machinery and costly raw materials, flexibility and customization are of the utmost importance for the effectiveness of the production flow. In a manufacturing line like that, where the processes take place in micrometres and nanometres, there is a need to have better control over the whole production line in order to achieve both quality improvement and cost reduction. The motivation for this research was based on the gaps and challenges that were identified through the literature review that was conducted on the area of quality prediction on the field of optoelectronics . These challenges were used to form the research objectives that this study aims to achieve. Throughout the entirety of this investigation, the primary objective was to develop and deploy an accurate inline prediction framework for real-time prediction of the quality in complex, multi-step optoelectronic assembly processes that can generalise to similar products. To accomplish that, a use case involving a company that produces high-power multi-emitter laser modules was investigated. Due to the intricacy and number of steps involved in their assembly process, the defects that are inevitably produced might be challenging to locate and pinpoint. During the course of this research, a prediction framework was constructed with the intention of predicting the quality of the assembled module. This was accomplished by making predictions regarding the final output power of the assembled laser modules (expressed in W) at various points along the assembly process. First, to address the high-dimensionality of the dataset a hybrid feature selection approach (RReliefF-RFE) was developed to find the dataset's key information and lower its high dimensionality, which has a negative impact on the performance of the prediction models. This technique produces a subset of important variables on which three prediction algorithms were tested (XGBoost, Random Forest Regression (RFR) and Artificial Neural Network (ANN)) in order to select the most suitable candidate for quality prediction in the studied case. These three prediction methods were deemed suitable for the studied case based on their properties. ANN was chosen as the best candidate for the studied case as it performed well based on the evaluation metrics (RMSE and MAE) and also avoided overfitting. In the early stages of the prediction, the suggested framework was able to achieve a low error rate of prediction (5.2%), and as the assembly advanced to the later stages, this error rate continued to decline (to 3.5%). Finally, the proposed framework was tested on another product made by the same company that follows an assembly process that is similar to the studied case. The framework produced comparable errors to the originally studied case (4.8% in the early stages and 4% in the later stages), demonstrating the generalisation of the suggested framework in this type of assembly lines, prior to adjustments. By making early stage predictions of the quality throughout the assembly process, it became possible to monitor the entire assembly process, offer an accurate interpretation of it and the possible defects that it may yield, and, as a result, reduce the number of defects that are generated. The findings produced by this research are encouraging in terms of the development and implementation of a transferable quality prediction framework for the multi-step assembly processes of sophisticated laser modules.