| dc.description.abstract | As traditional manufacturing industries face the challenges of product diversification and rapidly changing production demands, the need for flexibility in production lines is increasingly growing. Although fully automated equipment can enhance efficiency, its high costs make the return on investment unsatisfactory in low-volume production scenarios. Additionally, with rising labor costs, highly adaptable human labor is becoming insufficient to meet the demands of production lines. Consequently, companies are turning to deploy adaptable robotic arms capable of performing complex assembly tasks. However, frequent production line changes require the robotic arms to be recalibrated each time they are repositioned, a process that is both time-consuming and tedious. Using a teaching pendant for repositioning is also a time-intensive process, and the more calibration points required, the longer the time needed. Moreover, this process requires experienced and skilled personnel to ensure calibration accuracy; otherwise, product quality may be compromised.
This study proposes a novel method that combines visible light positioning with industrial robotic arms. By utilizing the characteristics of visible light and the Received Signal Strength (RSS) fingerprinting technique, along with neural network models, this method aims to achieve the repositioning of robotic arms. The method is based on a Multiple Input Multiple Output (MIMO) system and obtains precise location information through the RSS fingerprinting technique. In the fingerprinting process, appropriate signal processing and data processing methods are applied to prepare the data needed for model training. The processed data is then input into various models, including Neural Networks, Kolmogorov-Arnold Networks, and Random Forest Regressor, for training and evaluation, analyzing their characteristics and performance under different conditions to improve positioning accuracy and stability. The research results show that the system achieves an average accuracy error of less than 5 mm.
During the data processing, filters are used to preemptively eliminate ambient light and other noise, and the received signal strength is analyzed using the Signal-to-Noise Ratio (SNR) for feature selection, minimizing SNR errors in the data features and improving the model′s fit and predictive capabilities. Through the SNR feature selection method proposed in this study, overall accuracy was improved by 55.65%. The integration of these techniques aims to reduce the time and manpower required for robotic arm calibration, thereby enhancing the flexibility and efficiency of production lines. | en_US |