基于改进SOM神经网络的数控机床振动故障自动检测系统设计

摘要:

数控机床是制造业中的关键设备,其稳定运行对于提高生产效率至关重要。由于数控机床信号在传输线上传播时,会产生线路的阻抗不匹配现象,导致信号的衰减问题,降低信号强度,难以准确检测出振动信号。故设计一种基于改进SOM神经网络的数控机床振动故障自动检测系统。采用RS6103型号振动传感器实时采集数控机床的振动信号,改进振动信号调理电路,将传感器输出的衰减电流信号转换为稳定的电压信号,去除噪声干扰,提高振动信号的准确性,并将EP4CE10型号FPGA芯片作为核心控制器,完成硬件部分的设计。软件部分采用免疫遗传算法对SOM神经网络进行改进,优化连接权重寻优过程,提高神经网络的训练速度和振动故障检测精度。从采集的振动信号中提取特征向量,包括均值、峰峰值、峰值因子等,将提取的特征向量输入改进后的SOM神经网络中,通过竞争学习原理实现振动故障检测。实验结果显示:设计系统应用后振动信号调理结果信噪比更高,整体质量更好,提取的振动信号偏度特征与实际特征趋于一致,振动故障检测效果较好。

Abstract:

Numerical control machine tools are key equipment in the manufacturing industry, and their stable operation is crucial for improving production efficiency. Due to the impedance mismatch of the transmission line caused by the propagation of signals from CNC machine tools, signal attenuation problems occur, reducing signal strength and making it difficult to accurately detect vibration signals. Therefore, a CNC machine tool vibration fault automatic detection system based on improved SOM neural network is designed. The RS6103 vibration sensor is used to collect real-time vibration signals from CNC machine tools. The vibration signal conditioning circuit is improved to convert the attenuated current signal output by the sensor into a stable voltage signal, remove noise interference, and improve the accuracy of the vibration signal. The EP4CE10 FPGA chip is used as the core controller to complete the hardware design. The software part adopts immune genetic algorithm to improve the SOM neural network, optimize the process of finding connection weights, and improve the training speed and vibration fault detection accuracy of the neural network. Extract feature vectors from collected vibration signals, including mean, peak to peak value, peak factor, etc., and input the extracted feature vectors into an improved SOM neural network to achieve vibration fault detection through the principle of competitive learning. The experimental results show that after the application of the designed system, the signal-to-noise ratio of the vibration signal conditioning results is higher, the overall quality is better, the extracted vibration signal skewness features tend to be consistent with the actual features, and the vibration fault detection effect is better.