Rolling Bearing Fault Diagnosis Method Based on Parallel QPSO-BPNN Under Spark-GPU Platform

Facing the massive rolling bearing vibration data, how to improve training efficiency, diagnosis and accuracy of fault model is a challenge. Considering that Spark-GPU platform provides powerful distributed parallel computing capabilities back propagation neural network (BPNN) optimized by quantum particle swarm optimization (QPSO) algorithm has characteristics low computational complexity high accuracy, method based on QPSO-BPNN under proposed. First, parallelization realized, which can efficiency in big data environment. Second, order convergence speed model, parameter update strategy suitable for designed. At each iteration during training, local parameters worker node are collected master node, global updated according weights synchronized node. Third, combination multiple models ensemble learning The weighted voting adopted combine output results different obtain best result sample, certain extent. Experimental show proposed quickly perform large-scale reaches 98.73%.