Calibration and application of a large-scale LIBS project based on transfer learning in the online quantitative analysis of coal

Abstract

In this paper, four identical online LIBS coal quality analysis systems were developed, and these systems were installed on four production lines numbered 1, 2, 3, and 4, respectively, to perform real-time detection and statistical analysis of the moisture (Mad), ash content (Ad), sulfur (St), and calorific value (Q) in raw coal blocks. To calibrate the four devices efficiently, we propose the transfer learning method and part of the parameters locked in the PLS model. The proposed method is time-saving and does not require extensive sample preparation, making LIBS measurement suitable for real-time monitoring. Compared to the traditional laboratory, the result of online measurement based on the LIBS system improves the sampling frequency by about 3600 times, the measurement accuracy does not decrease, and the mean absolute error (MAE) of moisture, ash content, sulfur and calorific value are found to be less than 0.55 wt%, 1.50 wt%, 0.10 wt% and 1.0 MJ kg⁻¹ respectively. The implementation of these systems in actual production lines has provided a robust framework for continuous monitoring and quality control, leading to significant operational improvements and cost savings. This method has greatly enhanced the efficiency of LIBS device calibration, laying the foundation for the large-scale application of LIBS online systems and opening up new application scenarios for LIBS.