Abstract: Aiming at the problem of the difficulty of feature extraction of the vibration signal of the cut-off arm of the roadheader, it is proposed to decompose the vibration signal of the cut-off arm of the roadheader based on the wavelet packet singular values, and take the information gain of the singular values of each order as the evaluation index, and select the Relief filtered feature selection method for feature selection. The singular values of the time-frequency matrix of the input vibration signal are used as the input of the PNN neural network coal rock identification, in order to solve the problem of the sensitivity of the selection of the initial prototype vector of the neuron center of the hidden layer of the PNN network, the LVQ algorithm is proposed, and in order to overcome the clustering defects of the LVQ algorithm, the selection of the initial prototype vector of the LVQ is optimized with the use of the LOF algorithm, and the analysis of the experiment based on the parameters of the neural network structure is obtained as follows feature vector dimension (number of neuron nodes in the input layer) and optimal smoothing factor. The experimental results show that LOF-LVQ-PNN has high recognition accuracy when the feature vector dimension (number of neuron nodes in the input layer) is 10 and the smoothing factor σ is 0.046, and its recognition rate is 93.6% for 400 test samples and 97.8% for 400 training samples. Therefore, LOF-LVQ-PNN has higher recognition accuracy in rock hardness recognition of road header cut-off.