Abstract: When the long-term real-time monitoring of blasting vibration is carried out, the monitoring sensor is usually installed with fixed type. At this time, the main direction of the sensor is often inconsistent with the direction of the explosion source, which leads to the difference between the monitored vibration speed and the real vibration speed. In order to evaluate the influence of the main direction of the sensor on the blasting vibration velocity monitoring, the vibration velocity monitoring field experiment is carried out relying on the underground blasting of a metal mine. It is found that when multiple sensors placed at the same measurement point direct at different directions, the relative error of the peak vibration velocity could reach 78.68% at most, and the error decreases with the decrease of the detonation center distance and the increase of the maximum charge of a single stage. Based on this, this paper proposes a correction method of blasting vibration velocity monitoring data based on BP neural network, and constructs a data set containing 90 groups of field monitoring data. The BP neural network model is trained with the burst distance, the maximum loading of a single stage, the direction deviation angle and the peak vibration velocity as the input parameters and the real vibration velocity as the output parameters. The optimal configuration of Levenberg-Marquardt training algorithm, single hidden layer and 15 hidden layer neurons is determined by orthogonal experiments. Based on the established BP neural network prediction model, the error correction experiment of the fixed installation monitoring results of blasting vibration monitoring equipment is carried out. It is found that the difference between the corrected monitoring value and the actual value is within 1 cm/s, and the relative errors are less than 15%. The results show that the modified model meets the needs of the mine fixed blasting vibration monitoring, and can improve the accuracy and reliability of the data.