采用LWD-QPSO-SOMBP神经网络的拖拉机柴油机故障诊断

针对目前拖拉机柴油机故障诊断中单BP（Back Propagation）神经网络模型的局限性，该研究提出一种LWD-QPSO-SOMBP（Linear Weight Decrease-Quantum Particle Swarm Optimization-Self Organizing Maps Back Propagation）神经网络的拖拉机柴油机故障诊断模型。首先，将SOM（Self Organizing Maps）神经网络和BP神经网络结合，重置网络结构并利用LWD-QPSO（Linear Weight Decrease-Quantum Particle Swarm Optimization）算法对网络的权值和阈值进行优化；然后，分析拖拉机柴油机的故障机理，确定反映故障发生的数据信号；最后，确定LWD-QPSO-SOMBP神经网络模型的结构参数，基于CAN（Controller Area Network）总线技术采集潍柴WP6型拖拉机柴油机传感器信号数据对LWD-QPSO-SOMBP神经网络的性能进行测试，并将测试结果与BP神经网络、SOMBP（Self Organizing Maps Back Propagation）神经网络、PSO-SOMBP（Particle Swarm Optimization-Self Organizing Maps Back Propagation）神经网络、LWD-PSO-SOMBP（Linear Weight Decrease-Particle Swarm Optimization-Self Organizing Maps Back Propagation）神经网络及改进量子粒子群（Improved Quantum Particle Swarm Optimization，IQPSO）算法优化后的SOMBP神经网络的测试结果进行对比。试验结果表明， LWD-QPSO-SOMBP神经网络输出总误差为0.111 8、平均相对误差为0.005 8、均方误差为0.000 3，相比于其他5种神经网络均为最低。LWD-QPSO-SOMBP神经网络充分发挥并有效综合了SOM神经网络在数据预处理及PSO算法在优化BP神经网络初始权值阈值方面的优势，实现了拖拉机柴油机的高精度故障诊断。LWD-QPSO-SOMBP神经网络由于使用SOM神经网络结构对输入数据进行预处理，网络收敛速度大幅度提升，相比单BP神经网络，迭代次数由2 431次降为63次，下降了97.40%；同时采取LWD-QPSO算法对BP神经网络的初始权值阈值进行优化，降低了传统PSO算法的粒子适应度，进一步提高了网络的收敛精度和收敛速度，相比传统PSO算法，粒子适应度从0.15降为0.11，下降了26.67%，网络训练误差由0.004降为0.000 6，下降了85.00%；LWD-QPSO-SOMBP神经网络的故障诊断准确率大幅度提升，相比于单BP神经网络，输出总准确率由85.00%上升至99.44%。研究结果可为高精度拖拉机柴油机故障诊断提供参考。

Fault diagnosis of tractor diesel engine based on LWD-QPSO-SOMBP neural network

A diesel engine, the power source of a tractor, most directly determines the performance and safety of the tractor. Many efforts have been made on the faults of diesel engines in the agricultural field, due mainly to the complexity of the mechanism, diversity of faults, and the concurrency of multiple faults. Furthermore, fault diagnosis of diesel engines is developing towards artificial intelligence in recent years. Among them, back propagation (BP) neural network with excellent non-linear mapping has widely been used in fault diagnosis of tractor diesel engines. However, BP neural network tends to fall into local the minimum and slow convergence in engineering practical application. In this study, a modified fault diagnosis model was proposed for the tractor diesel engine using Linear Weight Decrease-Quantum Particle Swarm Optimization-Self Organizing Maps Back Propagation (LWD-QPSO-SOMBP) neural network. Firstly, A Self Organizing Maps (SOM) neural network was used to process the input data of the BP neural network. A composite network model was proposed to combine the SOM and BP neural network, in order to alleviate the training pressure of the BP neural network. Secondly, the network structure was modified to optimize the initial network weights, where the Linear Weight Decrease-Quantum Particle Swarm Optimization (LWD-QPSO) was proposed for the network weights and thresholds. Thirdly, the failure mechanism of the tractor diesel engine was analyzed to determine 8 kinds of data signals for the failure. Finally, the structure parameters were determined for the LWD-QPSO-SOMBP neural network model. A fault diagnosis test was then carried out using Controller Area Network (CAN) bus technology. The CAN bus was used to collect and analyze the sensor signal data of the Weichai WP6 tractor diesel engine, thereby evaluating the performance of the LWD-QPSO-SOMBP neural network. A comparison was also made on several neural networks to verify the accuracy of fault diagnosis and performance of LWD-QPSO-SOMBP neural network, including BP, Self Organizing Maps Back Propagation (SOMBP), Particle Swarm Optimization-Self Organizing Maps Back Propagation (PSO-SOMBP), and Linear Weight Decrease-Particle Swarm Optimization-Self Organizing Maps Back Propagation (LWD-PSO-SOMBP), and SOMBP neural network optimized by Improved Quantum Particle Swarm Optimization (IQPSO). The test results show that the LWD-QPSO-SOMBP neural network effectively integrated the SOM neural network in data preprocessing and the PSO in optimizing the initial weight threshold of the BP neural network, compared with the rest. As such, a high-precision fault diagnosis of tractor diesel engines was thus achieved. The LWD-QPSO-SOMBP neural network greatly improved the convergence rate of the framework using the SOM neural network to pre-process the network input data. The iteration times were reduced 97.40% from 2431 to 63, compared with single BP neural network. At the same time, LWD-QPSO was adopted in the LWD-QPSO-SOMBP neural network to optimize the initial weight threshold of the network. It was found that the particle fitness of traditional PSO was reduced greatly further to improve the convergence accuracy and speed of the network. The PSO particle fitness decreased by 26.67% from 0.15 to 0.11, while, the convergence error of the network decreased by 85.00% from 0.004 to 0.0006, compared with the traditional. The diagnostic accuracy of the LWD-QPSO-SOMBP neural network model was greatly improved, while, the training accuracy increased from 85.00% to 99.44%, compared with the single BP network. Consequently, the LWD-QPSO-SOMBP neural network model presented an excellent diagnostic performance. This finding can provide a sound reference for high-precision intelligent fault diagnosis of tractor diesel engines.