Penetration detection is an important link to realize the on-line control of high power laser welding quality. However, because the penetration region has the characteristics of low radiation and mesoscopic scale, which is generated from the bottom of the laser keyhole, its signal is completely masked by the ejected materials from the keyhole and the surrounding interference signals. As the penetration state is difficult to be directly obtained, the conventional measurements mainly use indirect measurement. In this paper, a coaxial synergistic extraction method of laser welding penetration characteristic signal is proposed by combining technologies of spectral perspective, infrared microscopic imaging, photoelectric sensing and spatial location extraction. It uses the fluorescence source as the direct detection signal, which is excited by high power laser in the inner wall of the keyhole, and by using the spectral characteristics of different luminescent materials in the infrared spectrum, separates and suppresses the strong interference signals such as plasma, metal vapor flame, particle cluster and so on over the keyhole. So as to enhance the infrared fluorescence signal in the keyhole effectively and realize the effect of spectral fluoroscopy. At the same time, by using the self-developed microscopic optical system with long focus in coaxial laser welding, which is according to the principle of infrared microscopic imaging, the real images of the stimulated radiation behavior which inside the keyhole are extracted. On this basis, the correlation between penetration state of high power laser welding and internal image features of keyhole is studied. The phenomenon of low radiation value and its existing position directly related to penetration state are found. After that, by means of visual assisted positioning adjustment and penetration characteristic position test experiments, that can improve the positioning accuracy successively, until the sensor photoelectric sensor chip is positioned to the penetration characteristic region in the fluorescence radiation real image with high precision. Therefore, by layer-by-layer optical separation method of spectral perspective, microscopic imaging and mesoscopic location signal extraction, accurate extraction and maximization of keyhole penetration data can be realized. The experiments results show that the coaxial synergistic extraction method of penetration characteristics of high power solid laser welding, which is based on composite applications of multiple spectral and optical processing methods, is effective and can be used as a new method for on-line detecting penetration of high power laser welding.