[Objective] Laser-induced breakdown spectroscopy(LIBS) holds significant potential for application in the field of coal property analysis

due to its advantages of eliminating complex sample pretreatment

enabling multi-parameter synchronous detection

and offering rapid analysis. However

discrepancies in spectral responses exist among different instruments. These discrepancies cause severe accuracy degradation when a quantitative model trained on spectra acquired by a master instrument is applied to slave instruments. Therefore

this study constructed cross-instrument LIBS quantitative analysis models of coal property by integrating TrAdaBoost transfer learning with various machine-learning algorithms. [Methods] Two LIBS-based coal analyzers were designated as the master and slave instruments respectively

and LIBS spectra were collected from different numbers of coal samples on both devices. Random forest(RF)

support-vector regression(SVR)

and their TrAdaBoost-enhanced counterparts(TrA-RF and TrA-SVR) were employed to build quantitative analysis models. Model performance was evaluated by predicting the coal properties of unknown coal samples on the slave instrument. [Results] The results indicated that both TrA-RF and TrA-SVR models significantly outperformed their non-transfer counterparts. The Tr A-RF model achieved the highest accuracy for calorific value

ash content

and carbon content. Compared with RF model

the mean absolute errors decreased from 1.390 MJ/kg

4.774 %

and 3.826 % to 0.654 MJ/kg

2.338%

and 1.927%

respectively. TrA-SVR model yielded the highest accuracy for volatile matter prediction. Compared with the SVR model

the mean absolute error decreased from 2.722%(SVR) to 2.524%. [Conclusion] These findings demonstrate that coupling transfer learning with an appropriate base learner markedly enhances the adaptability of LIBS-based coal property models across different instruments.