Abstract: In order to accurately assess the psychological load level of personnel in the process of coal mine hoist operation and effectively reduce coal mine safety accidents triggered by excessive psychological load of personnel, firstly, by simulating coal mine hoist operation and combining with the N-back auditory subtask, three difficult task scenarios of low, medium and high difficulty were designed, and physiological signals (ECG, EDA and RESP) were collected from 18 experimental personnel, and at the same time, the NASA task coad index (NASA-TLX) subjective scores and the number of sub-task errors, and the normality test and hypothesis test were used to obtain the significance indicators. Secondly, based on the significance indexes, three machine learning methods, namely, support vector machine, random forest and K nearest neighbor, were used to construct a mental load assessment model for operators. Finally, the classification accuracy of different physiological signals and combinations under the three machine learning algorithms is compared. The results showed that: with the increase of sub-task difficulty, the subjective scores increased significantly, and the number of sub-task errors increased significantly; there were significant differences in the mean heart rate value (meanHR), the standard deviation of the heart interval (SDNN) and the ratio of the NN50 divided by the total number of NN intervals (PNN50) among the ECG indexes, and skin conductivity (SC) among the dermatoelectric indexes, and the respiratory indexes did not have a significant differences. The classification and recognition effects showed that: the accuracy of the constructed mental load assessment models differed in different physiological modalities; the accuracy of the classification model based on bimodal physiological signals was generally higher than that of the unimodal physiological signal classification model; and the random forest classification model in the bimodal modality (ECG + RESP) had the highest accuracy, with a recognition accuracy rate of 86.5%.