矿井直流电阻率法具有施工效率高、成本低的特点，广泛应用于煤矿掘进工作面水害超前探测。随着应用的不断深入，对矿井直流电阻率法超前探测信号处理方法提出了更高的要求。基于有限元方法和数据重构所得响应信息库，采用Levenberg-Marquardt人工神经网络方法实现了矿井直流电阻率法超前预测。首先采用非结构网格剖分和BiCGSTAB迭代有限元法建立了直流电阻率法超前探测模型，基于EMD方法实现数值计算响应与实测响应匹配，并提出了基于异常体超前距离与其异常率的预测机制，得到3 000组23维重构响应；然后，采用L-M算法构建了人工神经网络模型；最后，基于训练好的网络对数值计算响应和实测数据进行预测分析。研究表明：针对数值模拟响应，基于L-M人工神经网络的矿井直流电阻率法超前预测方法可有效预测100 m范围内的水害异常，预测网络均方误差为0.002 47，测试样本显示最大距离误差小于0.6 m；针对实测数据发现，该神经网络在全区段预测准确率为67%，当预测异常位置在15~80 m时预测准确率超过85%，基于L-M算法的人工神经网络可应用于直流电阻率法超前预测。研究成果对煤矿井下掘进工作面直流电阻率法超前探测方法的完善与推广应用有重要意义。

The mine Direct Current (DC) resistivity method has the characteristics of high construction efficiency and low cost, which is widely used in the advanced detection of water hazards in coal mine heading face.With the continuous deepening of applications, higher requirements have been put forward for the advanced detection signal processing method of mine DC resistivity method. Based on the response information database obtained by finite element method and data reconstruction, the Levenberg-Marquardt (L-M) artificial neural network method is used to realize the advanced prediction of mine DC resistivity method. First, this study established an advanced prediction model of the DC resistivity method using unstructured mesh and iterative finite element method BiCGSTAB; matched the numerically simulated responses with the measured responses using the empirical mode decomposition (EMD) method; and proposed the prediction mechanism based on the advance detection distance and anomaly rate. As a result, 3 000 sets of 23-dimensional reconstructed responses were obtained. Subsequently, the artificial neural network model is constructed by L-M algorithm. Finally, using the trained artificial neural network, this study predicted the water hazards at the heading face based on the measured data and numerically simulated responses. The results of this study are as follows: (1) For the numerically simulated responses, the mine DC resistivity method based on the L-M artificial neural network can effectively detect the water hazard anomalies within the advance detection distance of 100 m. The mean square error in the detection was 0.002 47, and test samples exhibited that the maximum error of the advance detection distance was less than 0.6 m; (2) As shown by the measured data, the accuracy rate of the detection was 67% in the whole study and was higher than 85% for water hazard anomalies at the advance detection distance of 15-80 m. Therefore, the L-M artificial neural network can be applied to the advance prediction of water disasters based on the mine DC resistivity method. The research results are of great significance for the improvement and wide applications of the mine DC resistivity method-based advance detection of water hazards at the heading face of coal mines.