The reflection spectrum of the near-infrared band was convenient to measure. It was not necessary to pretreat the sample, but also suitable for on-line analysis. In order to realize the identification of coal-rock in the automatic coal caving technology of fully mechanized caving mining based on the near-infrared spectroscopy, the fully mechanized caving work from a mine four kinds of typical massive coal-rock samples such as carbonaceous mudstone, sandy mudstone, sandstone and gas coal were collected, and the coal pile condition of the rear scraper conveyor was considered comprehensively. Near infrared diffuse reflectance spectra of four typical coal-rock with common detection distances (1.3, 1.4, 1.5 m) and detection angles (10, 20, 30, 40 and 90 degrees) were collected in the laboratory by the spectrometer. By analyzing the spectral characteristics of four typical coal-rock, it was found that the detection angle and distance have no significant influence on the spectral curve and the position of the absorption valley, but obviously affected the reflectivity of the spectral curve. Carbonaceous mudstone, sandy mudstone and sandstone all have obvious absorption valleys near the 1 400, 1 900 and 2 200 nm bands. In addition, sandstone and carbonaceous mudstones have double absorption valleys near the 2 200 nm band. The diffuse reflectance spectral curve of coal in the near-infrared region is generally horizontal, with no obvious absorption valley. At the detection distance of 1.3 m, the reflectance of the spectral curve increased with the increased of the detection angle; at the detection distance of 1.4 and 1.5 m, the reflectance of the spectral curve decreased with the increased of the detection angle. At the detection angles of 10°, 20° and 30°, the reflectivity of the spectral curve increased with the increased of the detection distance; at the detection angle of 40° and 90°, the reflectivity of the spectral curve increased with the detection distance. Three methods of first-order differential (FD), Savitzky-Golay convolution smoothing (SG convolution smoothing), and standard normal enthalpy switching (SNV) were used to enhanced spectral absorption characteristics and eliminated detection conditions for coal-rock diffuse reflection spectra. The effect of SG convolution smoothing on the premise of enhancing spectral absorption characteristics also effectively eliminated the influence of detection angle and height on the spectral curve. The qualitative analysis of coal-rock was carried out by using two models of cosine similarity and Pearson correlation coefficient. The results showed that the cosine similarity model based on S-G convolution smoothing was the best, and the correct classification rate was 100%. Obtaining the best pre-processing method and qualitative analysis model can provide reference for the rapid and qualitative identification of coal-rock by directly using the waveform of the reflection spectrum at different detection distances and detection angles.