Abstract:

Objective: Imported fixed-length cut-off saws are widely used in production lines of wood doors and windows. However, due to frequent adoption of closed proprietary control systems and communication protocols by these sawing machines, measurement data remains locked within the equipment. Furthermore, as the actual output sequence is random to ensure optimal wood utilization during production, it is difficult to accurately match the measurement data displayed by the equipment with the actual dimensions. This has resulted in a “data silo” problem that restricts wood door and window manufacturing enterprises from meeting automation upgrade demands, such as intelligent palletizing. To address this dilemma, this study aims to develop a low-cost, high-efficient mechanical online length measurement system for wooden door and window materials to support the intelligent transformation of production lines. Method: A lightweight retrofit solution based on existing equipment was proposed: a measurement shaft was added to the synchronous belt drive shaft of the push cart; displacement information was collected using high-precision rotary encoders, combined with photoelectric sensors to identify the start and end positions of wooden door and window materials, thus establishing a precise mapping relationship between encoder pulses and material length. Taking into account the systematic deviations introduced by factors such as signal delay, mechanical transmission, and elastic deformation of synchronous belts, the least squares linear regression (LSLR) and support vector machine regression (SVR) were employed to compensate for the measured values. Real-time data interaction between the upper computer and PLC was accomplished based on Python-SNAP7 communication protocol, enabling automatic data collection and feedback of error compensation results. To evaluate the error compensation performance of the two models, LSLR and SVR were employed to train on 200 measurement data samples. Polar coordinate error distribution plots and error data fitting curves were used to conduct visual analysis, and multiple statistical metrics including coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) were used to comprehensively evaluate compensation effects of the models. To ensure the reliability of model evaluation, industrial validation experiments were conducted on four typical dimensional specifications of door and window materials (500, 1 000, 1 500, and 2 000 mm), with 20 samples processed for each specification. A total of 80 independently collected data samples served as the model validation set. The mean value obtained from repeated manual tape measure readings served as the reference standard. This reference was compared with the measured values compensated by different models. Through normalization processing and observation of the trend in mean absolute error, the error compensation effects of different models were analyzed. Result: The mechanized online measurement system constructed in this study was able to achieve high-precision real-time detection of the length of wooden door and window materials across various specifications. Analysis results of 200 data sets demonstrated that the SVR model outperformed the LSLR model across all evaluation metrics: the coefficient of determination R² reached 0.999 8, the mean absolute error (MAE) was 0.412 7, and the root mean square error (RMSE) was 1.594 8, significantly superior to the LSLR model values of 0.989 5 (R²), 2.359 7 (MAE), and 1.047 4 (RMSE), respectively. Comparative experiments on samples of different-sized door and window materials revealed that the measurement system achieved the highest accuracy for 1 500 mm materials. After SVR compensation, data points nearly perfectly aligned with the ideal prediction line. For 500, 1 000, and 2 000 mm materials, the SVR-based system showed slight fluctuations but maintained strong robustness with few outliers. The mean absolute error after SVR model compensation remained below 0.5 mm across the entire measurement range, significantly outperforming both raw measurements and LSLR-compensated results. Conclusion: The proposed online measurement system provides an effective solution to bridge the data gap between fixed-length cross-cutting saws and subsequent stacking processes, and also lays a data foundation for subsequent positional palletizing and complete set matching of wooden door and window materials.

Key words: wooden door and window materials, length measurement, mechanical measurement, support vector regression, error compensation