The working environment of coal mine rotating machinery is harsh, and the actual collected rolling bearing vibration signals show the characteristics of obvious nonlinearity and non-stationarity. Therefore, it is difficult to extract bearing fault characteristics. The traditional rolling bearing fault identification method based on ‘manual characteristic extraction+pattern identification’ is influenced subjectively. In order to solve the above problems, a rolling bearing fault identification method based on multi hidden layers wavelet convolution extreme learning neural network (MHLWCELNN) is proposed. The method combines the advantages of 1D convolution neural network, auto-encoder, extreme learning machine and wavelet function. The local connection and weight sharing mechanism of 1D convolution neural network is used to reduce the parameters to be learned greatly. The auto-encoder makes the algorithm applicable to the unlabeled samples of bearing vibration signals. The extreme learning machine is applied to determine the output weight so as to avoid falling into local optimum and improve the training speed. The wavelet function is used as the activation function to improve the resolution of the bearing time and frequency domain signals, thus improving the fault identification rate. The experimental results show that compared with similar methods, MHLWCELNN has higher identification accuracy and smaller standard deviation, and can identify different fault types of rolling bearings more stably. The F1 value of MHLWCELNN is higher than that of similar methods, which verifies its effectiveness on unbalanced data sets. Gaussian wavelet has higher resolution in both time and frequency domains, and is suitable to act as an activation function. And it is more appropriate to set the proportion of training set as 80%.