Electromagnetic radiation (EMR), as an effective monitoring technology, has been applied to monitoring and early warning of coal-rock dynamic disasters such as rockburst, coal and gas outburst and so on. However, due to the complexity of the EMR signal generation mechanism, it is susceptible to interference from the underground environment and thus affects the accuracy of disaster risk monitoring and warning. Accurate identification of effective EMR signals induced by coal-rock rupture is the key to the application and promotion of this technology. For this purpose, EMR monitoring experiments under uniaxial compression conditions of coal-rock were carried out. The time-domain, frequency-domain and fractal characteristic variability of effective and interfering signals of EMR were analyzed. On this basis, machine learning methods such as linear discriminant method, support vector machine and integrated learning method were utilized to establish the intelligent identification models of effective and interfering signals of EMR, respectively. Moreover, the recognition accuracy of different models was compared and analyzed. The results show that the fractal box dimension, average frequency, count and peak frequency characteristics are more obvious to distinguish between effective and interfering signals of EMR, and the identification accuracy of single characteristic is above 70%. Both the signal characteristic set and the machine learning method have an impact on the recognition accuracy of both effective and interfering signals. Comparative analysis yields the highest recognition accuracy of the integrated learning method based on all feature sets, with an average recognition accuracy of 95.4% for the two types of signals. It can meet the needs of electromagnetic radiation monitoring and early warning applications.