WO2021128577A1 - 一种基于sdae-dbn算法的零件表面粗糙度在线预测方法 - Google Patents

一种基于sdae-dbn算法的零件表面粗糙度在线预测方法 Download PDF

Info

Publication number
WO2021128577A1
WO2021128577A1 PCT/CN2020/077096 CN2020077096W WO2021128577A1 WO 2021128577 A1 WO2021128577 A1 WO 2021128577A1 CN 2020077096 W CN2020077096 W CN 2020077096W WO 2021128577 A1 WO2021128577 A1 WO 2021128577A1
Authority
WO
WIPO (PCT)
Prior art keywords
network
layer
surface roughness
data
signal
Prior art date
Application number
PCT/CN2020/077096
Other languages
English (en)
French (fr)
Inventor
刘阔
沈明瑞
秦波
黄任杰
牛蒙蒙
王永青
Original Assignee
大连理工大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 大连理工大学 filed Critical 大连理工大学
Priority to US15/734,940 priority Critical patent/US12026625B2/en
Publication of WO2021128577A1 publication Critical patent/WO2021128577A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/088Non-supervised learning, e.g. competitive learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/047Probabilistic or stochastic networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00Details relating to CAD techniques
    • G06F2111/10Numerical modelling
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/044Recurrent networks, e.g. Hopfield networks

Definitions

  • the invention belongs to the field of intelligent monitoring of mechanical processing processes, and specifically is an online prediction method for surface roughness of parts based on a stacked denoising autoencoder-deep belief network (Stacked Denoising Autoencoder-Deep Belief Network, SDAE-DBN).
  • SDAE-DBN Stacked Denoising Autoencoder-Deep Belief Network
  • Surface roughness is the main parameter to describe the surface microstructure of the part and measure the quality of the part. It not only affects the wear resistance, fatigue strength, corrosion resistance, sealing performance and stability of the part, but also affects the surface optical performance, electrical and thermal conductivity of the part. Performance and appearance also have an impact.
  • Traditional surface roughness measurement methods are mainly divided into contact measurement and non-contact measurement. Contact measurement because the probe tip is easy to wear and scratch the surface, which limits its application in high-precision testing. Non-contact measurement is more sensitive to dirt on the surface of the part. It needs to be cleaned before measurement, which reduces the efficiency of roughness measurement. . Therefore, how to accurately and efficiently realize the online measurement of the surface roughness of parts has become one of the key issues in the field of machining.
  • the research methods of surface roughness prediction mainly include: prediction models based on cutting theory and prediction models based on artificial intelligence.
  • the prediction method based on cutting theory has low roughness prediction accuracy due to many variable parameters.
  • the artificial intelligence-based prediction model is more and more used for surface quality prediction because of its good ability to approximate arbitrarily complex nonlinear systems.
  • the purpose of the present invention is to provide an online prediction method for surface roughness of parts based on SDAE-DBN, which solves the traditional methods of relying on manual experience to extract signal features and the need for a large amount of labeled data to train neural networks, and realizes the surface quality of parts. Online predictions.
  • the technical solution of the present invention is as follows: firstly, the three-way acceleration sensor is adsorbed to the rear bearing of the machine tool spindle through a magnetic seat, and a microphone is placed in front of the left front of the processed part to collect vibration and noise signals during the cutting process of the machine tool; Then, the polynomial least square method is used to eliminate the trend term of the dynamic signal, and the signal is smoothed by the five-point cubic smoothing method; secondly, the processing data is intercepted and normalized; then, a stacked denoising autoencoder is constructed , Use greedy layer-by-layer algorithm to train the network, and use the extracted features as the input of the deep belief network to train the network structure; finally, the real-time vibration and noise signals in the processing process are input into the deep network after data processing, and the network outputs the current The pros and cons of the surface roughness, so as to realize the real-time prediction of the surface roughness.
  • the first step is to collect vibration and noise signals during processing
  • the three-directional acceleration sensor is attached to the rear bearing of the machine tool spindle through a magnetic seat, and the microphone is placed on the left front of the workpiece, and the data acquisition software is used to collect the spindle vibration and cutting noise in real time during the machining process.
  • the second step is the preprocessing of the collected data
  • the polynomial least squares method is used to eliminate the trend term of the signal, and the processing signal collected for each processing is set as Select the M-order polynomial by formula (1) Fit the sampled signal.
  • the signal is smoothed by the five-point three-time smoothing method.
  • the calculation formula of the five-point triple smoothing method is:
  • the third step data interception and normalization
  • the dynamic signal of the cutting process is normalized according to formula (4), and normalized to the interval [0,1].
  • X′ is the normalized result of the collected data
  • X max and X min are the maximum and minimum values of the collected data during processing, respectively.
  • the normalized dynamic signal is divided into training set and test set according to the ratio of 10:3.
  • the fourth step is the construction and training of the stacked denoising autoencoder network
  • the denoising autoencoder learns abstract features from the original data with superimposed noise.
  • the learned features have better robustness and can avoid simply learning the same signal features.
  • the first coding layer of the stacked denoising self-encoder uses random mapping transformation q to "destroy" the normalized data X'into data X", and map it to the hidden layer according to formula (5).
  • W is the coding weight matrix of the first coding layer
  • b is the coding bias vector of the first coding layer
  • g is the activation function of the first coding layer
  • the first decoding layer of the stacked denoising autoencoder is reconstructed by mapping the implicit representation of the hidden layer data according to formula (6).
  • W′ is the bundling weight of the first decoding layer
  • b′ is the decoding bias vector of the first decoding layer
  • It is the decoding parameter of the first decoding layer.
  • the input data X′ is continuously reconstructed into X′′′, and the network weight and bias are constantly adjusted to achieve the goal of minimizing the network loss function.
  • denoising autoencoder two The working principle of denoising autoencoder two, denoising autoencoder three, and denoising autoencoder four is the same as that of denoising autoencoder one.
  • the fifth step the construction and training of the deep belief network
  • Deep Belief Network is a multi-hidden-layer generative structure graph model, which is composed of several layers of Restricted Boltzmann Machine (RBM) stacked.
  • RBM Restricted Boltzmann Machine
  • the restricted Boltzmann machine is an energy-based model, and its joint probability distribution is determined by formula (7).
  • f( ⁇ ) is the activation function of the network.
  • the deep belief network greedily trains each layer (from low to high) as an RBM by using the activation of the previous layer as input.
  • the specific training process is as follows: first fully train one RBM, secondly fix the weight and offset of the first RBM, use the state of its hidden neurons as the input vector of the second RBM; then fully train the second RBM After RBM, stack the second RBM on top of the first RBM, and repeat the above steps until the preset number of times is reached.
  • a Softmax classification layer is added to the top layer of the network to classify the surface roughness.
  • the probability value is calculated by formula (10).
  • the cost function of the Softmax regression model is:
  • 1 ⁇ is the indicator function, if the condition is true, it returns 1, otherwise it is 0.
  • the Softmax regression model is a supervised learning model that uses error back propagation algorithm to iteratively update the parameters to minimize the cost function and find the optimal parameters to fit the training set.
  • the model After training, use the test set to test the deep confidence network. When the test accuracy is higher than 90%, the model can be used to predict the surface roughness.
  • the sixth step real-time prediction of surface roughness during processing
  • the collected dynamic signals are preprocessed and input into the stacked denoising autoencoder network model to automatically extract signal features, and then the extracted features are used as the input of the deep confidence network to train the network model.
  • the output is the pros and cons of the surface quality. When the surface roughness is qualified, the output of the model is 0; when the surface roughness is unqualified, the output of the model is 1.
  • the stacked denoising autoencoder is used to extract the collected dynamic signal characteristics, which reduces the participation of manual and expert experience, avoids the interference caused by the introduction of human factors, and saves time and effort.
  • the deep belief network is used to predict the surface roughness of parts with high precision, which reduces the acquisition of tagged data, reduces the manpower, material and time costs for data acquisition, and avoids the traditional neural network from falling into local minima. Point the problem.
  • Figure 1 is the overall structure diagram of the online prediction method for surface roughness of parts based on SDAE-DBN.
  • Figure 2 shows the working principle diagram of the stacked denoising autoencoder.
  • Figure 3 is a diagram of the working principle of the deep confidence network.
  • Figure 4 shows the vibration and noise signals collected during the machining process.
  • Figure 5 is a graph of the prediction results of surface roughness.
  • the first step is to collect vibration and noise signals during processing
  • the three-directional acceleration sensor is adsorbed on the rear bearing of the machine tool spindle through a magnetic seat.
  • the sampling frequency of dynamic data is set to 1kHz, and vibration and noise signals are collected.
  • the second step is the preprocessing of the collected data
  • the third step data interception and normalization
  • the characteristics of the sudden change of the vibration signal of the cut-in and cut-out points during the processing are used to intercept the dynamic signals of the actual processing process, and the surface roughness of the processed surface is measured.
  • the measured surface roughness R a be divided into two cases qualified and unqualified.
  • the dynamic signal of the cutting process is normalized to the interval [0,1].
  • the normalized dynamic signal is divided into training set and test set according to the ratio of 10:3. Among them, the number of samples in the training set is 460, and the number of samples in the test set is 135.
  • the fourth step is the construction and training of the stacked denoising autoencoder network
  • Stack four denoising autoencoders to build a stacked denoising autoencoder network.
  • the number of neurons in the input layer and output layer of autoencoder 1 is 7500, and the number of neurons in the hidden layer is 3000.
  • the number of neurons in the input layer and output layer of autoencoder 2 is 3000, and the number of neurons in the hidden layer is 1000.
  • the number of neurons in the input and output layers of autoencoder three is 1000, the number of neurons in the hidden layer is 300, the number of neurons in the input and output layers of autoencoder four is 300, and the number of neurons in the hidden layer is 300.
  • the number is 100.
  • the fifth step the construction and training of the deep belief network
  • the training set features extracted by the denoising autoencoder are used as the input of the neural network to train the deep belief network model.
  • the number of hidden layers of the deep confidence network is set to 2
  • the number of nodes in the hidden layer are 60 and 20
  • the learning rate is set to 0.001
  • the number of iterations for pre-training is set to 1000
  • the number of iterations for fine-tuning is set to 1000.
  • Output the surface roughness of the processed part, and then use the test set to test the network.
  • the test accuracy is 91.1%.
  • the model can be used for online prediction of surface roughness.
  • the sixth step real-time prediction of surface roughness during processing
  • the vibration and noise signals collected during the actual processing are input into the stacked denoising autoencoder model to automatically extract features, and the extracted features are used as the input of the deep confidence network to determine the pros and cons of the surface roughness prediction.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Length Measuring Devices With Unspecified Measuring Means (AREA)

Abstract

一种基于SDAE-DBN算法的零件表面粗糙度在线预测方法,首先将三向加速度传感器通过磁座吸附在机床主轴后轴承处,在被加工零件左前方放置传声器,采集机床切削过程的振动和噪声信号;消除动态信号的趋势项,再对信号进行平滑处理;其次对加工过程的数据进行截取和归一化;接着构建堆叠去噪自编码器,采用贪婪逐层算法对网络进行训练,将提取的特征作为深度置信网络的输入训练网络结构;最后将加工过程中的实时振动和噪声信号经数据处理后输入到深度网络中,网络输出当前表面粗糙度的优劣情况,从而实现表面粗糙度的实时预测。本方法可减少人工和专家经验的参与,降低带标签数据的获取难度,并且能够提高表面粗糙度预测的准确性。

Description

一种基于SDAE-DBN算法的零件表面粗糙度在线预测方法 技术领域
本发明属于机械加工过程智能监测领域,具体为一种基于堆叠去噪自编码器-深度置信网络(Stacked Denoising Autoencoder-Deep Belief Network,SDAE-DBN)的零件表面粗糙度在线预测方法。
背景技术
表面粗糙度是描述零件表面微观形貌和衡量零件质量的主要参数,不仅影响零件的耐磨性、疲劳强度、抗腐蚀性、密封性和配合的稳定性,对零件的表面光学性能、导电导热性能和外观等也有影响。传统的表面粗糙度测量方法主要分为接触式测量和非接触式测量。接触式测量由于测尖易磨损且易划伤表面,限制了其在高精度检测中的应用,而非接触式测量对零件表面污物较敏感,测量前需清洗,降低了粗糙度测量的效率。因此,如何准确、高效地实现对零件表面粗糙的在线测量已经成为机械加工领域的关键问题之一。
目前,许多国内外学者对表面粗糙度预测技术进行了研究。关于表面粗糙度预测的研究方法主要包括:基于切削理论的预测模型和基于人工智能的预测模型。其中,基于切削理论的预测方法由于可变参数多导致粗糙度预测精度较低。而基于人工智能的预测模型由于具有良好的逼近任意复杂非线性系统的能力,被越来越多地用于表面质量预测。
在表面质量预测方面,国内学者已经申请了一些专利。在专利《一种硬质合金刀片化学机械抛光表面粗糙度的预测方法》(申请号:CN201710552078.8)中,采用基于高斯函数的异常检测算法对实验数据进行预处理,并建立了遗传算法优化的BP神经网络预测模型对表面粗糙度进行预测;在专利《一种机械加工表面粗糙度预测方法》(申请号:CN201810687856.9)中,通过Copula分布 估计算法优化网络权值及阈值,并基于BP算法进一步修正网络参数,从而实现表面粗糙度的预测;在专利《高速切削加工中工件表面粗糙度的预测方法》(申请号:CN201210426876.3)中,建立RBF神经网络模型,利用样本数据对模型进行训练来实现对表面粗糙度的预测;在专利《一种基于改进支持向量机算法的磨削表面粗糙度预测方法》(申请号:CN201910538299.9)中,利用交叉验证的思想将数据划分为训练集和测试集,并构建GOA-SVM预测模型,实现磨削表面粗糙度预测。
然而,上述专利采用的方法存在一些问题,如:(1)多依赖人工经验提取数据特征,使得数据处理过程繁琐化。(2)网络的训练需要大量带标签数据,限制了其在工业上的广泛使用。本专利将针对传统神经网络存在的问题,提出基于SDAE-DBN算法的表面粗糙度在线预测方法。
发明内容
本发明的目的为提供一种基于SDAE-DBN的零件表面粗糙度在线预测方法,解决传统方法存在的依赖人工经验提取信号特征和需要大量带标签数据训练神经网络的难题,实现了零件加工表面质量的在线预测。
为解决上述技术问题,本发明的技术方案为:首先,将三向加速度传感器通过磁座吸附在机床主轴后轴承处,在被加工零件左前方放置传声器,采集机床切削过程的振动和噪声信号;然后,采用多项式最小二乘法消除动态信号的趋势项,采用五点三次平滑法对信号进行平滑处理;其次,对加工过程的数据进行截取和归一化;接着,构建堆叠去噪自编码器,采用贪婪逐层算法对网络进行训练,将提取的特征作为深度置信网络的输入训练网络结构;最后,将加工过程中的实时振动和噪声信号经数据处理后输入到深度网络中,网络输出当前表面粗糙度的优劣情况,从而实现表面粗糙度的实时预测。
本发明的具体技术方案:
一种基于SDAE-DBN的零件表面粗糙度在线预测方法,具体步骤如下:
第一步,加工过程中的振动和噪声信号采集
将三向加速度传感器通过磁座吸附在机床主轴后轴承处,将传声器放置于被加工件的左前方,利用数据采集软件对加工过程中的主轴振动和切削噪声进行实时采集。
第二步,采集数据的预处理
对采集的振动和噪声数据进行消除趋势项以及平滑处理。
采用多项式最小二乘法消除信号的趋势项,设每次加工采集的加工信号为
Figure PCTCN2020077096-appb-000001
通过公式(1)选择M阶次多项式
Figure PCTCN2020077096-appb-000002
拟合采样信号。
Figure PCTCN2020077096-appb-000003
依据最小二乘法原理,选取合适的系数
Figure PCTCN2020077096-appb-000004
使
Figure PCTCN2020077096-appb-000005
Figure PCTCN2020077096-appb-000006
之间的误差平方和最小,即
Figure PCTCN2020077096-appb-000007
为满足e具有极小值,依次对系数
Figure PCTCN2020077096-appb-000008
求偏导数后其值为零,得到M+1个线性方程。解方程组,求出M+1个系数
Figure PCTCN2020077096-appb-000009
从而得到趋势项拟合曲线。
当M≥2时,趋势项为曲线趋势项,通常选取M=1~3对采样数据进行多项式趋势消除。
采用五点三次平滑法对信号进行平滑处理。五点三次平滑法的计算公式为:
Figure PCTCN2020077096-appb-000010
式中:j 1=3,4...,m-2,m为数据点数。
第三步,数据截取和归一化
利用加工过程中切入切出点振动信号突变的特点截取实际加工过程的动态信号,并对被加工表面的表面粗糙度进行测量,根据测量的表面粗糙度R a,将其划分为合格和不合格两种情况。
将切削过程的动态信号按照公式(4)进行归一化处理,归一化到[0,1]区间。
Figure PCTCN2020077096-appb-000011
其中,X′为采集数据归一化后的结果,X max和X min分别为加工过程中采集数据的最大值和最小值。
将归一化后的动态信号按照10:3的比例划分为训练集和测试集。
第四步,堆叠去噪自编码器网络的构建和训练
与自编码器相比,去噪自编码器从叠加噪声的原始数据中学习抽象特征,学习到的特征具有更好的鲁棒性,并且可以避免简单地学习相同的信号特征。
堆叠去噪自编码器的第一编码层通过随机的映射变换q,将归一化后的数据X′“破坏”为数据X″,并根据公式(5)将其映射到隐含层。
Figure PCTCN2020077096-appb-000012
其中,W为第一编码层的编码权值矩阵,b为第一编码层的编码偏置向量, g为第一编码层的激活函数,
Figure PCTCN2020077096-appb-000013
为第一编码层的编码参数。
堆叠去噪自编码器的第一解码层通过将隐含层数据的隐含表示根据公式(6)进行映射重构。
Figure PCTCN2020077096-appb-000014
其中,W′为第一解码层的捆绑权重,b′为第一解码层的解码偏置向量,
Figure PCTCN2020077096-appb-000015
为第一解码层的解码参数。
在去噪自编码器的训练过程中,持续将输入数据X′重构为X″′,并不断调整网络权重以及偏置,以达到最小化网络损失函数的目的。
去噪自编码器二、去噪自编码器三、去噪自编码器四的工作原理与去噪自编码器一的工作原理相同。
第五步,深度置信网络的构建和训练
将去噪自编码器提取的信号特征作为神经网络的输入对深度置信网络模型进行训练。深度置信网络是一种多隐层的生成性结构图模型,由若干层限制玻尔兹曼机(Restricted Boltzmann Machine,RBM)堆叠构成。
限制玻尔兹曼机是基于能量的模型,其联合概率分布由公式(7)确定。
Figure PCTCN2020077096-appb-000016
其中
Figure PCTCN2020077096-appb-000017
Figure PCTCN2020077096-appb-000018
是可见单元i 2和隐含单元j 2的二进制状态,θ 1={w,d,c}是模型的参数:
Figure PCTCN2020077096-appb-000019
是可见单元i 2和隐含单元j 2之间的权重,
Figure PCTCN2020077096-appb-000020
Figure PCTCN2020077096-appb-000021
是其偏置,V和H是可见单元和隐含单元个数。
由限制波尔兹曼机的结构性质可以得出,神经元的激活状态是条件独立的。当输入信号输入到可见层时,可见层将决定隐含层各神经元的状态。隐含层第j 2个神经元激活状态的概率通过式(8)计算:
Figure PCTCN2020077096-appb-000022
同理,可见层第i 2个神经元激活状态的概率通过式(9)计算:
Figure PCTCN2020077096-appb-000023
其中f(·)是网络的激活函数。
深度置信网络作为一种半监督深度学习算法,通过利用前一层的激活作为输入,贪婪地将每一层(从低到高)训练为RBM。具体训练过程如下所示:首先充分训练一个RBM,其次固定第一个RBM的权重和偏移量,使用其隐性神经元的状态,作为第二个RBM的输入向量;接着充分训练第二个RBM后,将第二个RBM堆叠在第一个RBM的上方,重复以上步骤直至达到预设的次数。多个RBM堆叠模型训练结束后,在网络上顶层增加一个Softmax分类层用来对表面粗糙度进行分类。
给定n 2个样本的k 2类训练数据
Figure PCTCN2020077096-appb-000024
其中样本集为
Figure PCTCN2020077096-appb-000025
标签集为
Figure PCTCN2020077096-appb-000026
Figure PCTCN2020077096-appb-000027
使用Softmax函数估算每一个类别的概率值。概率值通过公式(10)计算。
Figure PCTCN2020077096-appb-000028
其中j 3=1,2,...,k 2
Figure PCTCN2020077096-appb-000029
是Softmax回归模型的参数,
Figure PCTCN2020077096-appb-000030
这一项对概率分布进行归一化,使得所有概率之和为1。
Softmax回归模型的代价函数为:
Figure PCTCN2020077096-appb-000031
其中1{·}是指标函数,如果条件为真,则返回1,否则为0。Softmax回归模型是有监督学习模型,通过误差反向传播算法来迭代更新参数使得代价函数最小化,从而找到最优参数以适应训练集。
训练完毕后,采用测试集对深度置信网络进行测试,当测试准确率高于90%时,模型可用于表面粗糙度的预测。
第六步,加工过程表面粗糙度的实时预测
实际加工过程中,将采集的动态信号经预处理后输入到堆叠去噪自编码器网络模型中自动的提取信号特征,再将提取后的特征作为深度置信网络的输入对网络模型训练,模型的输出为表面质量的优劣情况。当表面粗糙度合格模型输出为0;当表面粗糙度不合格时,模型输出为1。
本发明的有益效果:
(1)为零件表面粗糙度预测提供了一种新方法,解决了依靠人工经验提取信号特征和需要大量带标签数据的难题。
(2)采用堆叠去噪自编码器提取采集到的动态信号特征,减少了人工和专家经验的参与,避免由于引入人为因素所造成的干扰,省时省力。
(3)采用深度置信网络对零件表面粗糙度进行高精度预测,减少了带标签数据的获取,降低了数据获取所消耗的人力、物力与时间成本,并且避免了传统神经网络容易陷入局部极小点的问题。
(4)提高了零件表面粗糙度预测的准确性。
附图说明
图1为基于SDAE-DBN的零件表面粗糙度在线预测方法的总体结构图。
图2为堆叠去噪自编码器工作原理图。
图3为深度置信网络工作原理图。
图4为加工过程采集的振动和噪声信号。
图5为表面粗糙度的预测结果图。
具体实施方式
为了使本发明的目的、技术方案和优点更加清晰明了,下面结合基于SDAE-DBN的零件表面粗糙度在线预测的具体实施方式并参照附图,对本发明作详细说明。本实施例以本发明的技术方案为前提给出了详细的实施方式和具体的操作过程,但本发明的保护范围不限于下述的实施例。
以一台三轴立式数控机床铣削加工为例,详细说明本发明的实施方式。
第一步,加工过程中的振动和噪声信号采集
将传声器通过支架固定在被加工件左前方,对准被加工表面。三向加速度传感器通过磁座吸附在机床主轴后轴承处。加工过程中,动态数据的采样频率设置为1kHz,采集振动与噪声信号。
第二步,采集数据的预处理
对采集的动态信号进行去趋势项和平滑处理。采用最小二乘法对采样数据进行多项式趋势消除,其中M=2。采用五点三次平滑法再对数据进行平滑处理。
第三步,数据截取和归一化
利用加工过程中切入切出点振动信号突变的特点截取实际加工过程的动态信号,并对被加工表面的表面粗糙度进行测量。根据测量的表面粗糙度R a,将其划分为合格和不合格两种情况。
将切削过程的动态信号进行归一化处理,归一化到[0,1]区间。将归一化后的动态信号按照10:3的比例划分为训练集和测试集。其中,训练集样本数为460,测试集样本数为135。
第四步,堆叠去噪自编码器网络的构建和训练
堆叠四个去噪自编码器构建堆叠去噪自编码器网络。自编码器一的输入层和输出层神经元个数均为7500,隐含层神经元个数为3000。自编码器二的输入层和输出层神经元个数为3000,隐含层神经元个数为1000。自编码器三的输入层和输出层神经元个数为1000,隐含层神经元个数为300,自编码器四的输入层和输出层神经元个数为300,隐含层神经元个数为100。利用数据对堆叠去噪自编码器进行训练,并输出通过自编码器提取的信号特征。
第五步,深度置信网络的构建和训练
将去噪自编码器提取的训练集特征作为神经网络的输入对深度置信网络模型进行训练。其中,深度置信网络的隐含层层数设为2,隐含层的节点数分别为60、20,学习率设置为0.001,预训练的迭代次数设置为1000,微调的迭代次数设置为1000,输出被加工零件的表面粗糙度情况,再用测试集对网络进行测试,测试准确度为91.1%,模型可以用于表面粗糙度的在线预测。
第六步,加工过程表面粗糙度的实时预测
将实际加工过程中采集的振动和噪声信号经过数据预处理后输入至堆叠去噪自编码器模型中自动提取特征,并将提取的特征作为深度置信网络的输入,对表面粗糙度的优劣进行预测。
应该说明的是,本发明的上述具体实施方式仅用于示例性阐述本发明的原理和流程,不构成对本发明的限制。因此,在不偏离本发明精神和范围的情况下所做的任何修改和等同替换,均应包含在本发明的保护范围内。

Claims (1)

  1. 一种基于SDAE-DBN算法的零件表面粗糙度在线预测方法,其特征在于:首先,将三向加速度传感器通过磁座吸附在机床主轴后轴承处,在被加工零件左前方放置传声器,采集机床切削过程的振动和噪声信号;然后,采用多项式最小二乘法消除动态信号的趋势项,采用五点三次平滑法对信号进行平滑处理;其次,对加工过程的数据进行截取和归一化;接着,构建堆叠去噪自编码器,采用贪婪逐层算法对网络进行训练,将提取的特征作为深度置信网络的输入训练网络结构;最后,将加工过程中的实时振动和噪声信号经数据处理后输入到深度网络中,网络输出当前表面粗糙度的优劣情况,从而实现表面粗糙度的实时预测;步骤如下:
    第一步,加工过程中的振动和噪声信号采集
    将三向加速度传感器通过磁座吸附在机床主轴后轴承处,将传声器放置于被加工件的左前方,对加工过程中的主轴振动和切削噪声进行实时采集;
    第二步,采集数据的预处理
    对采集的振动和噪声数据进行消除趋势项以及平滑处理;
    采用多项式最小二乘法消除信号的趋势项,设每次加工采集的加工信号为
    Figure PCTCN2020077096-appb-100001
    通过公式(1)选择M阶次多项式
    Figure PCTCN2020077096-appb-100002
    拟合采样信号;
    Figure PCTCN2020077096-appb-100003
    依据最小二乘法原理,选取合适的系数
    Figure PCTCN2020077096-appb-100004
    使
    Figure PCTCN2020077096-appb-100005
    Figure PCTCN2020077096-appb-100006
    之间的误差平方和最小,即
    Figure PCTCN2020077096-appb-100007
    为满足e具有极小值,依次对系数
    Figure PCTCN2020077096-appb-100008
    求偏导数后其值为零,得到M+1个线性方程;解方程组,求出M+1个系数
    Figure PCTCN2020077096-appb-100009
    从而得到趋势项拟合曲线;
    当M≥2时,趋势项为曲线趋势项,通常选取M=1~3对采样数据进行多项式趋势消除;
    采用五点三次平滑法对信号进行平滑处理;五点三次平滑法的计算公式为:
    Figure PCTCN2020077096-appb-100010
    式中:j 1=3,4...,m-2,m为数据点数;
    第三步,数据截取和归一化
    利用加工过程中切入切出点振动信号突变的特点截取实际加工过程的动态信号,并对被加工表面的表面粗糙度进行测量,根据测量的表面粗糙度R a,将其划分为合格和不合格两种情况;
    将切削过程的动态信号按照公式(4)进行归一化处理,归一化到[0,1]区间;
    Figure PCTCN2020077096-appb-100011
    其中,X′为采集数据归一化后的结果,X max和X min分别为加工过程中采集数据的最大值和最小值;
    将归一化后的动态信号按照10:3的比例划分为训练集和测试集;
    第四步,堆叠去噪自编码器网络的构建和训练
    堆叠去噪自编码器的第一编码层通过随机的映射变换q,将归一化后的数据X′“破坏”为数据X″,并根据公式(5)将其映射到隐含层;
    Figure PCTCN2020077096-appb-100012
    其中,W为第一编码层的编码权值矩阵,b为第一编码层的编码偏置向量,g为第一编码层的激活函数,
    Figure PCTCN2020077096-appb-100013
    为第一编码层的编码参数;
    堆叠去噪自编码器的第一解码层通过将隐含层数据的隐含表示根据公式(6)进行映射重构;
    Figure PCTCN2020077096-appb-100014
    其中,W′为第一解码层的捆绑权重,b′为第一解码层的解码偏置向量,
    Figure PCTCN2020077096-appb-100015
    为第一解码层的解码参数;
    在去噪自编码器的训练过程中,持续将输入数据X′重构为X″′,并不断调整网络权重以及偏置,以达到最小化网络损失函数的目的;
    去噪自编码器二、去噪自编码器三、去噪自编码器四的工作原理与去噪自编码器一的工作原理相同;
    第五步,深度置信网络的构建和训练
    将去噪自编码器提取的信号特征作为神经网络的输入对深度置信网络模型进行训练;深度置信网络是一种多隐层的生成性结构图模型,由若干层限制玻尔兹曼机堆叠构成;
    限制玻尔兹曼机是基于能量的模型,其联合概率分布由公式(7)确定;
    Figure PCTCN2020077096-appb-100016
    其中,
    Figure PCTCN2020077096-appb-100017
    Figure PCTCN2020077096-appb-100018
    是可见单元i 2和隐含单元j 2的二进制状态,θ 1={w,d,c}是模型的参数:
    Figure PCTCN2020077096-appb-100019
    是可见单元i 2和隐含单元j 2之间的权重,
    Figure PCTCN2020077096-appb-100020
    Figure PCTCN2020077096-appb-100021
    是其偏置,V和H是可见单元和隐含单元个数;
    由限制波尔兹曼机的结构性质得出,神经元的激活状态是条件独立的;当输入信号输入到可见层时,可见层将决定隐含层各神经元的状态;隐含层第j 2个神经元激活状态的概率通过式(8)计算:
    Figure PCTCN2020077096-appb-100022
    同理,可见层第i 2个神经元激活状态的概率通过式(9)计算:
    Figure PCTCN2020077096-appb-100023
    其中,f(·)是网络的激活函数;
    深度置信网络作为一种半监督深度学习算法,通过利用前一层的激活作为输入,贪婪地从低到高将每一层训练为RBM;具体训练过程如下所示:首先充分训练一个RBM,其次固定第一个RBM的权重和偏移量,使用其隐性神经元的状态,作为第二个RBM的输入向量;接着充分训练第二个RBM后,将第二个RBM堆叠在第一个RBM的上方,重复以上步骤直至达到预设的次数;多个RBM堆叠模型训练结束后,在网络上顶层增加一个Softmax分类层用来对表面粗糙度进行分类;
    给定n 2个样本的k 2类训练数据
    Figure PCTCN2020077096-appb-100024
    其中样本集为
    Figure PCTCN2020077096-appb-100025
    标签集为
    Figure PCTCN2020077096-appb-100026
    使用Softmax函数估算每一个类别的概率值;概率值通过公式(10)计算:
    Figure PCTCN2020077096-appb-100027
    其中j 3=1,2,...,k 2
    Figure PCTCN2020077096-appb-100028
    是Softmax回归模型的参数,
    Figure PCTCN2020077096-appb-100029
    这一项对概率分布进行归一化,使得所有概率之和为1;
    Softmax回归模型的代价函数为:
    Figure PCTCN2020077096-appb-100030
    其中1{·}是指标函数,如果条件为真,则返回1,否则为0;Softmax回归模型是有监督学习模型,通过误差反向传播算法来迭代更新参数使得代价函数最小化,从而找到最优参数以适应训练集;
    训练完毕后,采用测试集对深度置信网络进行测试,当测试准确率高于90%时,模型可用于表面粗糙度的预测;
    第六步,加工过程表面粗糙度的实时预测
    实际加工过程中,将采集的动态信号经预处理后输入到堆叠去噪自编码器网络模型中自动的提取信号特征,再将提取后的特征作为深度置信网络的输入对网络模型训练,模型的输出为表面质量的优劣情况;当表面粗糙度合格模型输出为0;当表面粗糙度不合格时,模型输出为1。
PCT/CN2020/077096 2019-12-24 2020-02-28 一种基于sdae-dbn算法的零件表面粗糙度在线预测方法 WO2021128577A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US15/734,940 US12026625B2 (en) 2019-12-24 2020-02-28 On-line prediction method of surface roughness of parts based on SDAE-DBN algorithm

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911346390.7A CN111177911B (zh) 2019-12-24 2019-12-24 一种基于sdae-dbn算法的零件表面粗糙度在线预测方法
CN201911346390.7 2019-12-24

Publications (1)

Publication Number Publication Date
WO2021128577A1 true WO2021128577A1 (zh) 2021-07-01

Family

ID=70655654

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/077096 WO2021128577A1 (zh) 2019-12-24 2020-02-28 一种基于sdae-dbn算法的零件表面粗糙度在线预测方法

Country Status (2)

Country Link
CN (1) CN111177911B (zh)
WO (1) WO2021128577A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114139579A (zh) * 2021-11-23 2022-03-04 吉林大学 一种基于方差分形的深度学习微地震事件检测方法
CN114239187A (zh) * 2021-12-30 2022-03-25 福州大学 一种基于改进bpnn-de算法的硅铝合金车削工艺参数优化方法
CN114297802A (zh) * 2021-12-31 2022-04-08 福州大学 一种薄板激光切割工艺参数的多目标优化方法
CN114697183A (zh) * 2022-03-31 2022-07-01 中国人民解放军国防科技大学 一种基于深度学习的信道同步方法
WO2023005700A1 (zh) * 2021-07-27 2023-02-02 天津大学 模型数据混合驱动的电网可靠性快速计算方法及装置
CN116740462A (zh) * 2023-07-05 2023-09-12 北京科技大学 一种基于特征增强的单分支粗糙度检测方法

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021174525A1 (zh) * 2020-03-06 2021-09-10 大连理工大学 一种基于多任务学习的零件表面粗糙度和刀具磨损预测方法
CN111595790A (zh) * 2020-05-30 2020-08-28 南京林业大学 基于高光谱图像的青梅糖酸度预测方法
CN111775051B (zh) * 2020-06-15 2021-08-10 德尔福柴油系统(烟台)有限公司 一种检测磨削产品是否合格的方法
CN114055249B (zh) * 2020-08-05 2022-11-11 北京福田康明斯发动机有限公司 镗孔表面粗糙度监测方法和系统以及存储介质
CN112069966B (zh) * 2020-08-31 2023-11-28 中国计量大学 基于改进型lstm网络的表面粗糙度预测方法
CN112907222B (zh) * 2021-03-25 2022-07-08 清华大学 一种多源异构电网运监业务数据融合方法
CN113095195B (zh) * 2021-04-03 2023-04-07 西北工业大学 一种基于表面形貌自特征的零件唯一识别方法
CN114357851A (zh) * 2021-10-23 2022-04-15 西北工业大学 一种基于dae-rnn的数控铣削表面粗糙度预测方法
CN114819311B (zh) * 2022-04-15 2024-06-21 武汉数字化设计与制造创新中心有限公司 一种数控加工表面粗糙度预测模型的构建方法
CN116578832B (zh) * 2023-07-13 2023-09-12 兰州理工大学 预测机械加工零件表面粗糙度的装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103761429A (zh) * 2014-01-10 2014-04-30 大连理工大学 铣削加工工件表面粗糙度的预测方法
CN108596158A (zh) * 2018-05-15 2018-09-28 同济大学 一种基于能耗的车削表面粗糙度预测方法
CN108920812A (zh) * 2018-06-28 2018-11-30 江苏大学 一种机械加工表面粗糙度预测方法
CN110059442A (zh) * 2019-05-07 2019-07-26 同济大学 一种基于零件表面粗糙度和功率信息的车削加工换刀方法

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108398268B (zh) * 2018-03-15 2020-06-09 哈尔滨工业大学 一种轴承性能退化评估方法
CN110561192A (zh) * 2019-09-11 2019-12-13 大连理工大学 一种基于堆叠自编码器的深孔镗刀状态监测方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103761429A (zh) * 2014-01-10 2014-04-30 大连理工大学 铣削加工工件表面粗糙度的预测方法
CN108596158A (zh) * 2018-05-15 2018-09-28 同济大学 一种基于能耗的车削表面粗糙度预测方法
CN108920812A (zh) * 2018-06-28 2018-11-30 江苏大学 一种机械加工表面粗糙度预测方法
CN110059442A (zh) * 2019-05-07 2019-07-26 同济大学 一种基于零件表面粗糙度和功率信息的车削加工换刀方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
YU, KUN: "Fault Diagnosis of Rolling Element Bearings Based on Deep Belief Network and Multiple Sensors Information Fusion", CHINA MASTER’S THESES FULL-TEXT DATABASE, 1 December 2016 (2016-12-01), pages 1 - 101, XP055824390, ISSN: 1674-0246 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023005700A1 (zh) * 2021-07-27 2023-02-02 天津大学 模型数据混合驱动的电网可靠性快速计算方法及装置
CN114139579A (zh) * 2021-11-23 2022-03-04 吉林大学 一种基于方差分形的深度学习微地震事件检测方法
CN114139579B (zh) * 2021-11-23 2024-04-30 吉林大学 一种基于方差分形的深度学习微地震事件检测方法
CN114239187A (zh) * 2021-12-30 2022-03-25 福州大学 一种基于改进bpnn-de算法的硅铝合金车削工艺参数优化方法
CN114297802A (zh) * 2021-12-31 2022-04-08 福州大学 一种薄板激光切割工艺参数的多目标优化方法
CN114697183A (zh) * 2022-03-31 2022-07-01 中国人民解放军国防科技大学 一种基于深度学习的信道同步方法
CN114697183B (zh) * 2022-03-31 2023-11-17 中国人民解放军国防科技大学 一种基于深度学习的信道同步方法
CN116740462A (zh) * 2023-07-05 2023-09-12 北京科技大学 一种基于特征增强的单分支粗糙度检测方法
CN116740462B (zh) * 2023-07-05 2024-03-15 北京科技大学 一种基于特征增强的单分支粗糙度检测方法

Also Published As

Publication number Publication date
CN111177911B (zh) 2022-09-20
CN111177911A (zh) 2020-05-19
US20210287098A1 (en) 2021-09-16

Similar Documents

Publication Publication Date Title
WO2021128577A1 (zh) 一种基于sdae-dbn算法的零件表面粗糙度在线预测方法
CN108803486B (zh) 基于并联深度学习网络的数控机床热误差预测与补偿方法
Wang et al. On line tool wear monitoring based on auto associative neural network
CN113051831B (zh) 机床热误差自学习预测模型建模方法及热误差控制方法
CN112435363B (zh) 一种刀具磨损状态实时监测方法
CN110561191B (zh) 基于pca与自编码器的数控机床刀具磨损数据处理方法
CN113239504A (zh) 一种基于优化神经网络的管道腐蚀缺陷预测方法
CN112069966B (zh) 基于改进型lstm网络的表面粗糙度预测方法
Yixu et al. An adaptive modeling method for a robot belt grinding process
Qin et al. Tool wear identification and prediction method based on stack sparse self-coding network
CN112926265A (zh) 基于遗传算法优化神经网络的大气多孔探针测量校准方法
CN112518425B (zh) 基于多源样本迁移强化学习的智能加工刀具磨损预测方法
CN111126255A (zh) 基于深度学习回归算法的数控机床刀具磨损值预测方法
Xu et al. Oversmoothing relief graph convolutional network-based fault diagnosis method with application to the rectifier of high-speed trains
CN114152442B (zh) 基于迁移卷积神经网络的滚动轴承跨工况故障检测方法
Ge et al. An improved PF remaining useful life prediction method based on quantum genetics and LSTM
CN112149896A (zh) 一种基于注意力机制的机械设备多工况故障预测方法
Yang et al. Crack detection in carbide anvil using acoustic signal and deep learning with particle swarm optimisation
CN115049627A (zh) 基于域自适应深度迁移网络的钢表面缺陷检测方法及系统
US12026625B2 (en) On-line prediction method of surface roughness of parts based on SDAE-DBN algorithm
You et al. Normalized variational auto-encoder with the adaptive activation function for tool setting in ultraprecision turning
CN114378653B (zh) 基于bp神经网络外圆磨削颤振在线识别和监测方法
CN109635358A (zh) 一种基于滑动窗口多尺度主元分析的机组故障检测方法
CN115901255A (zh) 基于参数优化vmd-svm的轴承故障诊断方法
Dong et al. Discrete Wavelet Transformation and Genetic Algorithm-Back Propagation Neural Network Applied in Monitoring Woodworking Tool Wear Conditions in the Milling Operation Spindle Power Signals.

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20906783

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20906783

Country of ref document: EP

Kind code of ref document: A1