Abstract: In order to realize the identification of rock hardness when the roadheader cuts rock wall under different working conditions, a rock hardness identification method based on the feature preference and the STOA-random forest(RF) is proposed. A test bench is built for the rock hardness identification when a roadheader cuts under multiple working conditions, obtaining some motor current and torque signals from cutting different hardness rocks under different working conditions. The method first decomposes the experimentally obtained current and torque signals by using the complete ensemble empirical modal decomposition of adaptive noise(CEEMDAN) to obtain the eigenmodal component(IMF) and calculates the sample entropy(SE) of the IMF,and the IMF with the highest SE is subjected to the quadratic decomposition of the variational modal decomposition(VMD). After calculating the fuzzy entropy of the secondary decomposited IMF,the current and torque signals are reconstructed, and the time-frequency features of the reconstructed signals are calculated, and the time-frequency-entropy features of the current and torque signal samples are composed with the fuzzy entropy. In order to avoid too many features affecting the recognition model, the method of Relief-F combined with Pearson correlation coefficient is proposed to select the features for dimensionality reduction. Finally, the maximum number of features and the number of decision tree in RF are optimized by the STOA method with the highest recognition accuracy as the fitness function, and the establishment of different rock hardness recognition models under different truncation conditions is completed. The results show that(1) the CEEMDAN decomposition and then the VMD decomposition of current and torque signals can reduce the randomness and volatility of original signals, and the accuracy of rock hardness identification can be improved by 15.2% and 23.9%,respectively, compared with the primary decomposition of CEEMDAN and VMD.(2) The time domain features of current signals account for the greatest weight of rock hardness identification accuracy under different cutting conditions.(3) The proposed Relief-F combined with Pearson correlation coefficient feature selection method has obviously clustered the current and torque signal features of four hardness rocks cut under three working conditions.(4) The STOA has optimized the selection of RF key coefficients, and the number of iterations of the algorithm is less, compared with the traditional particle swarm optimization algorithm, the recognition accuracy of different hardness rocks cut under different working conditions is improved by 7.2%.