Aiming at the fuzzy type characteristics between multi-frequency spectrums of vibration/acoustic signal and mill load parameters, the redundancy and complementarity in the multi-frequency spectrums , and the difficulty of using latent structural selected model in existing literatures to simulate the operation expert “listen” inference to identify the mill load parameters, a soft sensing model of mill load parameters based on adaptive extraction and selection of multi-scale vibration/acoustic spectrum characteristics is proposed. The method firstly acquires multi-scale spectrum, then extracts and selects the features of multi-scale vibration and acoustic spectrum adaptively using kernel partial least squares(KPLS) and mutual information(MI). Finally, the soft sensing model of mill load parameters is constructed by using online clustering, Madani fuzzy model, banch and bound(BB), adaptation weighted fusion(AWF) and modeling of selective sets. The experiments are carried out in the experimental wet mill. The experimental results show that the proposed model can simulate the expert's reasoning mechanism and has better modeling accuracy.