基于表征学习的滚珠丝杠副系统状态监测与性能评估技术研究

发布时间：2018-01-01 17:02

本文关键词：基于表征学习的滚珠丝杠副系统状态监测与性能评估技术研究　出处：《西南交通大学》2016年博士论文　论文类型：学位论文

【摘要】：滚珠丝杠副系统是一类将旋转运动转换为直线运动的机械传动装置,广泛应用于精密机械定位与测量系统。其在运行过程中不可避免的会出现磨损,结构松动等现象。而滚珠丝杠副系统发生故障或者出现性能退化会严重影响机械设备的精度与运行安全。因此研究滚珠丝杠副系统的状态监测与性能评估对于提高制造业水平,减少企业经济损失非常重要。滚珠丝杠副系统的振动信号包含了丰富的设备状态与性能信息。但是,采集的振动信号中往往包含有大量的噪声成分,这对有效信息提取造成了一定困扰。同时,如何有效提取与表达信号中的信息成分也是信号处理领域中的研究热点与难点。论文结合表征学习在滚珠丝杠副系统状态监测与性能评估方面展开了深入的研究,具体内容如下：(1)以字典学习和深度学习为代表系统研究了表征学习的基本算法与结构。揭示了表征学习的内涵属性是不引入先验知识的情况下从原始信号中学习得到其本质特征表示。同时,通过图像信号处理说明了表征学习可以学习得到信号的基本组成成分,并可广泛应用于信号去噪与分类。(2)研究了基于字典学习和稀疏编码的机械振动信号去噪技术。针对滚珠丝杠副支撑轴承振动信号的特点,构建了由固定字典函数和轴承振动信号构成的复合异构训练样本集合。探索了振动信号在学习字典域内稀释表示进行去噪的方法。提出利用在线字典学习来增强算法计算实时性。并通过仿真信号和实验数据验证了算法的可行性和有效性。(3)设计了滚珠丝杠副系统丝杠变速工况下的故障定位算法。通过多项式调频小波方法,计算出丝杠旋转瞬时频率。提出了基于幅值阀值和导数阀值的时域信号异常检测方法定位丝杠故障时间点。基于丝杠瞬时旋转频率和故障时间点计算获得丝杠在不同工况下的故障位置点。通过实验平台研究了两种不同工况下的丝杠多故障点定位。实验结果表明,提出的算法可有效判断丝杠故障并计算出准确位置,定位误差小于两个丝杠导程。(4)研制了滚珠丝杠副系统综合加速性能退化实验台。根据丝杠寿命公式,分析了影响丝杠运行寿命的因素,设计了滚珠丝杠副系统丝杠性能退化实验方案。通过改变运行速度,负载大小模拟滚珠丝杠副系统的不同运动工况,并采集了不同工况下的振动、扭矩等传感器信号,为滚珠丝杠副系统的准确性能评估提供了数据保障。(5)提出了基于深度神经网络的滚珠丝杠副系统性能评估方法。对实验采集的滚珠丝杠副系统振动信号进行时域、频域和时频域特征提取,设计了频域谱峭度相关系数这一新的频域特征值,可分辨能力强。深度神经网络学习提取深层次的特征参数并降低特征维数,提高可识别性。引入去噪深度学习算法,将输入向量中某些位置的值用随机噪声代替,增强了模型的鲁棒性。(6)基于以上理论和实验,进行了滚珠丝杠副系统状态监测与性能评估系统的实用化研究。基于Socket原理设计了数据采集端与上位机客户端的数据通信协议,实现了系统的远程化操作。研究了基于图形处理单元的算法并行化加速方案,缩短了模型计算时间。基于WxPython开发了图像化显示界面,便于可视化操作与显示。本文在滚珠丝杠副系统性能退化实验,数据表征学习,状态监测与性能评估等方面进行了深入研究。探索了基于表征学习的振动信号处理新方法,提出了一种振动信号自适应表达新思路。对滚珠丝杠副系统的状态监测与性能评估进行了实用化研究,对系统的工业化推广起到了积极作用。
[Abstract]:The ball screw system is a kind of rotary motion into linear motion of the mechanical transmission device, is widely used in precision mechanical positioning and measuring system. It is inevitable in the process of operation will be worn, loose structure and so on. And the ball screw system failure or performance degradation will seriously affect the precision and safety operation mechanical equipment. So the research of ball screw system for condition monitoring and performance evaluation to improve the manufacturing level, reduce the economic loss is very important. The vibration signal of the ball screw system includes the equipment status and performance information rich. However, vibration signals often contain a lot of noise, which caused some problems for effective information extraction. At the same time, research how to extract information in the composition and expression of signal is in the field of signal processing With difficulty. Combined with the characterization of learning conducted in-depth research on the ball screw system for condition monitoring and performance evaluation, the specific contents are as follows: (1) the dictionary learning and deep learning of the basic algorithm and structure characterization of learning as the representative system. To reveal the connotation of the attribute Xi is not to introduce prior knowledge in the case of the essential characteristics of the original signal from the learned representation. At the same time, the image signal processing shows the characterization of learning can get basic signal components of learning, and can be widely used in the signal denoising and classification. (2) studied the mechanical vibration signal denoising technology based on dictionary learning and sparse encoding. According to the characteristics of the ball screw bearing vibration signal, to construct a complex heterogeneous training sample consists of fixed dictionary function and the bearing vibration signal collection of vibration signal in the study are discussed. The dictionary learning domain representation method for denoising dilution. Put forward to enhance the learning algorithm in real-time using the online dictionary. And the feasibility and effectiveness of the algorithm is verified by simulation signals and experimental data. (3) the fault location algorithm for variable speed screw ball screw system design. Through polynomial frequency modulation wavelet method. Calculate the screw rotation of instantaneous frequency is proposed. The anomaly detection method of positioning screw failure time point time domain signal amplitude threshold and threshold based on derivative calculation of the fault location point screw under different working conditions of the screw instantaneous rotation frequency and fault point. Based on the experimental platform is studied under the two different conditions of multi fault locating screw. The experimental results show that the proposed algorithm can effectively determine the screw fault and calculate the accurate position, the positioning error is less than two screw (4) is developed. The ball screw system accelerated performance degradation experiment. According to the screw life formula, analyzed the influence factors of service life of screw, the design of ball screw screw system performance degradation experiments. By changing the running speed, load different motion conditions of size simulation of ball screw system, and vibration under different conditions of the acquisition, torque the sensor signal, provides data protection for accurate performance evaluation of ball screw system. (5) performance evaluation method of ball screw system based on neural network is proposed. Time domain vibration signal of ball screw system experiment data extraction, frequency domain and time-frequency domain features, the design of frequency domain spectral kurtosis and correlation coefficient a new spectral feature, distinguishing ability. The depth of the neural network feature extraction parameters of deep and reduce the feature dimension and improve recognition The introduction of deep learning. Denoising algorithm, the input vector values in certain positions with random noise instead, enhances the robustness of the model. (6) based on the above theory and experiment, the practical research of ball screw system for condition monitoring and performance evaluation system. Based on the designing principle of Socket data communication protocol data collection terminal and PC client, realize the remote operating system. Research on parallel acceleration scheme algorithm based on graphics processing unit, shorten the calculation time. WxPython model is developed based on the image display interface, easy operation and visual display. The degradation experiment in the performance of ball screw system, data representation learning. In-depth research on condition monitoring and performance evaluation and so on. To explore a new method for characterizing vibration signal processing based on learning, the adaptive expression of a vibration signal A practical study on the state monitoring and performance evaluation of the ball screw system has been carried out, which has played an active role in the industrialization of the system.

【学位授予单位】：西南交通大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TH132;TH17

【参考文献】