In trace gas detection based on laser absorption spectroscopy, the overlap of different gas absorption spectra affects the extraction of the characteristics of the absorption spectrum, which consequently introduce the error of the deduced concentration. In this paper, a trace gas detector by the combination of a multipass cell and direct absorption spectroscopy is present. BP neural network model and PLS model are utilized, respectively, to restrict the spectral overlap. In order to simplify the training progress, simulated spectral models have been used as training set. The laser frequency are calibrated with the help of the transmission peaks of an F-P cavity and then introduced into the simulation model. As a result, the accuracy of the simulation has been improved. Then, the measured data is acted as test set. The linearity of the system's response to the concentration is greater than 0.99, and the relative error is less than 0.21%. Finally, the influence of etalon noise to the two algorithm has been analyzed and the result shows the concentration error with PLS model is less than 4.4×10-7, which is more than five times that by BP neural network.