WO2021138982A1 - 一种基于elm-sdae算法的刀具状态监测方法 - Google Patents

一种基于elm-sdae算法的刀具状态监测方法 Download PDF

Info

Publication number
WO2021138982A1
WO2021138982A1 PCT/CN2020/077092 CN2020077092W WO2021138982A1 WO 2021138982 A1 WO2021138982 A1 WO 2021138982A1 CN 2020077092 W CN2020077092 W CN 2020077092W WO 2021138982 A1 WO2021138982 A1 WO 2021138982A1
Authority
WO
WIPO (PCT)
Prior art keywords
network
tool
elm
data
vibration
Prior art date
Application number
PCT/CN2020/077092
Other languages
English (en)
French (fr)
Inventor
刘阔
沈明瑞
秦波
厉大维
黄任杰
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
Publication of WO2021138982A1 publication Critical patent/WO2021138982A1/zh

Links

Images

Classifications

    • BPERFORMING OPERATIONS; TRANSPORTING
    • B23MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
    • B23QDETAILS, COMPONENTS, OR ACCESSORIES FOR MACHINE TOOLS, e.g. ARRANGEMENTS FOR COPYING OR CONTROLLING; MACHINE TOOLS IN GENERAL CHARACTERISED BY THE CONSTRUCTION OF PARTICULAR DETAILS OR COMPONENTS; COMBINATIONS OR ASSOCIATIONS OF METAL-WORKING MACHINES, NOT DIRECTED TO A PARTICULAR RESULT
    • B23Q17/00Arrangements for observing, indicating or measuring on machine tools
    • B23Q17/09Arrangements for observing, indicating or measuring on machine tools for indicating or measuring cutting pressure or for determining cutting-tool condition, e.g. cutting ability, load on tool
    • B23Q17/0952Arrangements for observing, indicating or measuring on machine tools for indicating or measuring cutting pressure or for determining cutting-tool condition, e.g. cutting ability, load on tool during machining
    • B23Q17/0971Arrangements for observing, indicating or measuring on machine tools for indicating or measuring cutting pressure or for determining cutting-tool condition, e.g. cutting ability, load on tool during machining by measuring mechanical vibrations of parts of the machine
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B23MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
    • B23QDETAILS, COMPONENTS, OR ACCESSORIES FOR MACHINE TOOLS, e.g. ARRANGEMENTS FOR COPYING OR CONTROLLING; MACHINE TOOLS IN GENERAL CHARACTERISED BY THE CONSTRUCTION OF PARTICULAR DETAILS OR COMPONENTS; COMBINATIONS OR ASSOCIATIONS OF METAL-WORKING MACHINES, NOT DIRECTED TO A PARTICULAR RESULT
    • B23Q17/00Arrangements for observing, indicating or measuring on machine tools
    • B23Q17/09Arrangements for observing, indicating or measuring on machine tools for indicating or measuring cutting pressure or for determining cutting-tool condition, e.g. cutting ability, load on tool
    • B23Q17/0952Arrangements for observing, indicating or measuring on machine tools for indicating or measuring cutting pressure or for determining cutting-tool condition, e.g. cutting ability, load on tool during machining
    • B23Q17/0957Detection of tool breakage

Definitions

  • the present invention belongs to the technical field of tool condition monitoring, and specifically is an Extreme Learning Machine-Stacked Denoising Autoencoders (Extreme Learning-Stacked Denoising Autoencoders,
  • the purpose of the present invention is to provide a tool condition monitoring method based on the ELM-SDAE algorithm, which solves the problem that the existing monitoring method relies on manual experience to extract signal features, the training process takes a long time and is easy to fall into the local optimum, and realizes the real-time tool status monitor.
  • the technical scheme of the present invention is as follows: firstly, the vibration signal in the processing process is collected by a three-way acceleration sensor; then, the missing value is interpolated by the linear interpolation method, and simple random under-sampling and white addition are used.
  • the method of noise is to balance the data and superimpose the vibration signal to obtain the effective value of the main vibration direction; then, construct the ELM-SDAE network, and use the greedy layer-by-layer method to train the network model; finally, the actual processing process
  • the real-time vibration is input into the ELM-SDAE network after data preprocessing, and the network outputs the current status of the tool, so as to realize the real-time monitoring of the tool status.
  • the first step is to collect vibration information during processing
  • a three-way acceleration sensor is used to collect the vibration signal during the processing of the CNC machine tool, and the vibration signal corresponding to the tool to be monitored is intercepted from the three-way vibration signal, and divided into K segments of vibration signals corresponding to the processing of the K workpieces .
  • the second step is the preprocessing of the collected data
  • the vibration signal is data cleaned.
  • the linear interpolation method is used to interpolate the missing values.
  • the third step the construction and training of the ELM-SDAE network
  • the number of nodes in the input layer and output layer of the first ELM-DAE network are both The number of hidden layer nodes is First, set the vibration signal V to 0 or other values according to a certain ratio, so as to obtain the "destroyed" vibration signal And replace the expected output of the network with the vibration signal V; then, in order to enhance the generalization performance of the autoencoder network, orthogonalized hidden layer parameters are randomly generated.
  • Equation (2) The relationship between the input and output of the denoising autoencoder network is represented by equation (2).
  • Is the output matrix of the hidden layer Is the output weight vector of the hidden layer node
  • W 1 is the input weight vector of the hidden layer node in the feature extraction stage
  • b 1 is the bias vector of the node in the feature extraction stage
  • g( ⁇ ) is the activation function of the network.
  • C 1 is the regularization coefficient of the network in the feature extraction stage, which affects the generalization performance of the network in the feature extraction stage.
  • connection weight ⁇ 1 between the hidden layer and the output layer can be calculated according to formula (4):
  • the output of the kth network that is, the features extracted from the vibration signal, can be calculated by equation (5):
  • the activation function adopts a linear function, otherwise, a nonlinear activation function is adopted.
  • the output weight of the hidden layer can be calculated by formula (7).
  • W 2 is the input weight vector of the hidden layer in the classification stage
  • b 2 is the bias vector of the node in the classification stage
  • C 2 is the regularization coefficient of the classification stage, which affects the classification performance of the network in the classification stage.
  • the fourth step real-time monitoring of tool status
  • the real-time data is input into the trained ELM-SDAE network model after data preprocessing, and the model outputs the real-time state of the tool.
  • the model output is 0; when the tool status is broken, the model output is 1; when the tool status is blunt, the model output is 2.
  • the present invention has the advantages of:
  • the extreme learning machine algorithm can avoid the situation that the training process may fall into the local optimum, the learning rate is slow, etc., which can lead to model training failure and insufficient generalization ability.
  • the ELM-SDAE algorithm abandons the traditional iterative error adjustment strategy, changes it to randomly set the hidden layer weights and offsets, and solves the output layer weight matrix, the calculation time has been greatly improved reduce.
  • Figure 1 is a flowchart of tool condition monitoring based on ELM_SDAE algorithm.
  • Figure 2 is a schematic diagram of the sensor layout of a deep hole boring machine.
  • Figure 3 (a) is the Y-direction vibration time-domain diagram of the three-way acceleration sensor in the broken state; (b) is the Y-direction vibration time-domain diagram of the three-way acceleration sensor in the blunt state; (c) is the Y-direction vibration of the normal state three-way acceleration sensor Vibration time domain diagram.
  • Figure 4 (a) is the synthesized vibration data before degradation; (b) is the synthesized vibration data after degradation.
  • FIG. 5 is a schematic diagram of the ELM-SDAE network.
  • Figure 6 shows the prediction results of the network.
  • the first step the vibration signal acquisition during the deep hole boring process
  • the #1 three-way acceleration sensor 6 and #2 three-way acceleration sensor 7 are adsorbed on the side of the bearing bush of the tool holder holder of the deep hole boring machine through the magnetic suction seat.
  • the sampling frequency of the data acquisition software is set to 1000 Hz, and the vibration during processing is collected. According to the change of the tool state during the machining process, the vibration signal is divided into 6469 segments of vibration signals, including 6360 segments of normal status data, 87 segments of broken tool status data, and 22 segments of blunt status data.
  • the second step is the preprocessing of the collected data
  • the vibration signals V x , V y , and V z in the three directions collected by the two three-directional acceleration sensors are vector superimposed to obtain the effective value of the main vibration direction.
  • the third step the construction and training of the ELM-SDAE network
  • the ELM-SDAE network consisting of four denoising autoencoders based on the ELM algorithm and one ELM classifier.
  • the degradation rate of the four denoising autoencoders is 0.1
  • the regularization coefficient C 1 is 1000
  • the regularization coefficient C 2 is 1e7
  • the activation function is the hyperbolic tangent function
  • the first denoising autoencoder input layer and
  • the number of neurons in the output layer is 2000
  • the number of neurons in the hidden layer is 1200.
  • the number of neurons in the input layer and output layer of the second denoising autoencoder is 1200
  • the number of neurons in the hidden layer is 400.
  • the number of neurons in the input layer and output layer of the third denoising autoencoder is 400, and the number of neurons in the hidden layer is 150.
  • the number of neurons in the input layer and output layer of the fourth denoising autoencoder is 150, and the number of neurons in the hidden layer is 50.
  • the data is divided into training set and test set in a 4:1 ratio, where the number of samples in the training set is 452 and the number of samples in the test set is 123.
  • 5960 unused normal state data are used to train the autoencoder
  • the training set data is input into the trained stacked denoising autoencoder to obtain deep features, and the network classifier is trained.
  • the three states of the tool are output. Then use the test set data to test the network, the test accuracy rate is 93.5%, and the model can be used to monitor the tool status.
  • the fourth step real-time monitoring of deep hole boring tool status
  • the real-time vibration collected in the actual machining process is preprocessed and input into the ELM-SDAE network to monitor the tool status during the boring process.

Landscapes

  • Engineering & Computer Science (AREA)
  • Mechanical Engineering (AREA)
  • Automatic Control Of Machine Tools (AREA)
  • Numerical Control (AREA)

Abstract

一种基于ELM-SDAE算法的刀具状态监测方法,先通过三向加速度传感器采集加工过程中的振动信号;再采用线性插值的方法对缺失值进行插补,采用朴素随机欠采样和添加白噪声的方法,对数据进行平衡处理,将振动信号进行矢量叠加获取主振动方向的有效值;接着构建ELM-SDAE网络,采用贪婪逐层方法对网络模型进行训练;最后将实际加工过程中的实时振动经数据预处理后输入ELM-SDAE网络中,网络输出刀具的当前状态,实现刀具状态的实时监测。该方法减少人工和专家经验的参与,避免训练过程可能陷入局部最优、学习率慢等能导致模型训练失败和泛化能力不足的情况,大幅度缩减网络的训练时间。

Description

一种基于ELM-SDAE算法的刀具状态监测方法 技术领域
本发明属于刀具状态监测技术领域,具体为一种基于极限学习机的堆叠去噪自编码器(Extreme Learning Machine-Stacked Denoising Autoencoders,
ELM-SDAE)算法的刀具状态监测方法。
背景技术
在机械加工领域,刀具状态直接影响被加工零件的加工精度和表面质量。在常规切削加工中,一般由有经验的加工人员根据加工噪声、切屑颜色、切削振动以及刀具的切削时间来估计磨损情况。但由于工人经验的局限性,无法精确判断刀具是否磨损,可能发生磨损刀具仍在继续加工的情况,降低了加工质量;也可能发生尚未磨损的刀具被提前换掉的情况,造成浪费,提高了生产成本。因此,如何准确、高效地实现对刀具状态的在线监测已经成为机械加工领域的关键问题之一。
目前,国内外学者在刀具状态监测方面进行了一定的研究。在专利“Method and apparatus for efficient use of CNC machine shaping tool including cessation of use no later than the onset of tool deterioration by monitoring audible sound during shaping”(WO/2017/192821)中,通过将采集的声信号进行降噪处理后,进行特征提取和数据平衡,并将主成分向量与标记的状态数据进行比较,从而判断当前刀具所处状态。在专利“基于小波包分析与RBF神经网络的刀具磨损在线监测方法”(CN108356606A)中,采集数控机床铣削加工时的切削力信号并测量加工后刀具的后刀面磨损量,通过小波包分析和滤波处理获取刀具磨损特征值后,利用RBF神经网络实现刀具后刀面磨损量的预测。
综上所述,目前在刀具状态监测方面的研究仍然很少且存在一些问题,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)网络的训练过程耗时长且可能出现陷入局部最优、学习率慢等导致模型训练失败和泛化能力不足的情况。本发明针对传统神经网络监测刀具状态存在的问题,提出一种基于ELM-SDAE算法的刀具状态监测方法。
发明内容
本发明的目的为提供一种基于ELM-SDAE算法的刀具状态监测方法,解决现有监测方法依赖人工经验提取信号特征、训练过程耗时长且容易陷入局部最优的难题,实现了刀具状态的实时监测。
为解决上述技术问题,本发明的技术方案为:首先,通过三向加速度传感器采集加工过程中的振动信号;然后,采用线性插值的方法对缺失值进行插补,采用朴素随机欠采样和添加白噪声的方法,对数据进行平衡处理,将振动信号进行矢量叠加获取主振动方向的有效值;接着,构建ELM-SDAE网络,采用贪婪逐层方法对网络模型进行训练;最后,将实际加工过程中的实时振动经数据预处理后输入ELM-SDAE网络中,网络输出刀具的当前状态,从而实现刀具状态的实时监测。
一种基于ELM-SDAE算法的刀具状态监测方法,具体步骤如下:
第一步,加工过程中的振动信息采集
采用三向加速度传感器采集数控机床加工过程中的振动信号,从该三向振动信号中截取与待监测刀具对应的振动信号,并将其分成与K个加工工件的加工过程对应的K段振动信号。
第二步,采集数据的预处理
首先,考虑到信号采集过程中出现的数据异常情况,对振动信号进行数据清洗。采用线性插值方法对缺失值进行插补计算。
假设已知两点坐标(x 0,y 0)和(x 1,y 1),区间[x 0,x 1]内某一位置的纵坐标y通过式(1)线性插值后计算得到。
Figure PCTCN2020077092-appb-000001
其次,针对实际加工过程中,刀具多处于正常状态的情况,采用朴素随机欠采样降低正常状态刀具数据量,采用添加高斯白噪声的方式对磨钝或断刀状态数据进行数据增强。然后,对采集的加工过程振动信号进行规整处理,使各段振动信号长度一致。最后,将三个方向的振动信号V x、V y、V z进行矢量叠加,获取主振动方向的有效值V。
第三步,ELM-SDAE网络的构建和训练
在特征提取阶段,假设样本个数为N,第一个ELM-DAE网络的输入层和输出层节点数均为
Figure PCTCN2020077092-appb-000002
隐含层节点数为
Figure PCTCN2020077092-appb-000003
首先,将振动信号V按照一定比例置为0或其他值,从而获得“破坏”后的振动信号
Figure PCTCN2020077092-appb-000004
并将网络的期望输出替换成振动信号V;然后,为增强自编码器网络的泛化性能,随机生成正交化的隐含层参数。去噪自编码器网络输入与输出的关系由式(2)表示。
Figure PCTCN2020077092-appb-000005
其中,
Figure PCTCN2020077092-appb-000006
为隐含层的输出矩阵,
Figure PCTCN2020077092-appb-000007
为隐含层节点的输出权值向量,W 1为特征提取阶段隐含层节点的输入权值向量,b 1为特征提取阶段节点的偏置向量,g(·)为网络的激活函数。
为了提高网络的泛化能力,在损失函数中添加正则项,如式(3)所示。
Figure PCTCN2020077092-appb-000008
其中,C 1为特征提取阶段网络的正则化系数,影响特征提取阶段网络的泛化性能。
隐含层和输出层之间的连接权重β 1可根据式(4)进行计算:
Figure PCTCN2020077092-appb-000009
对于由k个ELM-DAE网络堆叠而成的神经网络,第k个网络的输出即从振动信号中提取的特征可以通过式(5)进行计算:
Figure PCTCN2020077092-appb-000010
其中,当k为1时,
Figure PCTCN2020077092-appb-000011
在堆叠后的ELM-SAE网络中,如果第k层神经元的个数与第k-1层神经元个数相同,则激活函数采用线性函数,否则采用非线性激活函数。
在分类阶段,假设训练集样本个数为M,ELM网络的输入层神经元节点数为L 2,输出层神经元节点数即刀具状态类别为L 3,隐含层节点数为N 2,网络的期望输出为
Figure PCTCN2020077092-appb-000012
随机生成正交化隐含层参数后,隐含层的输出权值可由公式(7)计算。
Figure PCTCN2020077092-appb-000013
Figure PCTCN2020077092-appb-000014
其中,W 2为分类阶段隐含层的输入权值向量,b 2为分类阶段节点的偏置向量,C 2为分类阶段正则化系数,影响分类阶段网络的分类性能。
第四步,刀具状态的实时监测
在实际加工过程中,将实时数据经数据预处理后输入至训练后的ELM-SDAE网络模型中,模型输出刀具的实时状态。当刀具状态正常时,模型输出为0;当刀具状态为断刀时,模型输出为1;当刀具状态为磨钝时,模型输出为2。
本发明的有益效果:通过该方法,实现了刀具状态的实时监测,避免了特 征工程专家的参与,缩短了网络的训练时间,降低了对机床操作工人经验的依赖,有利于提高加工效率,降低生产成本。
本发明与现有技术相比,其优点在于:
(1)与反馈神经网络相比,极限学习机算法可以避免训练过程可能陷入局部最优、学习率慢等能够导致模型训练失败和泛化能力不足的情况。
(2)与传统堆叠去噪自编码器相比,ELM-SDAE算法摒弃了传统的迭代误差调整策略,改成随机设置隐层权重与偏置,并求解输出层权重矩阵,计算时间得到了大幅缩减。
(3)通过基于极限学习机的堆叠去噪自编码器网络自动提取信号的深层次特征,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。
附图说明
图1为基于ELM_SDAE算法的刀具状态监测流程图。
图2为深孔镗床传感器布置示意图。
图3(a)为断刀状态三向加速度传感器Y向振动时域图;(b)为磨钝状态三向加速度传感器Y向振动时域图;(c)为正常状态三向加速度传感器Y向振动时域图。
图4(a)为退化前的合成振动数据;(b)为退化后的合成振动数据。
图5为ELM-SDAE网络原理图。
图6为网络的预测结果图。
图中:1工件保持架;2工件;3电机;4机床齿轮箱;5床身;6-1#三向加速度传感器;7-2#三向加速度传感器;8刀杆;9刀杆保持架。
具体实施方式
为了使本发明的技术方案和有益效果更加清晰明了,下面结合深孔镗刀状态监测的具体实施方式并参照附图,对本发明作详细说明。本实施例是以本发 明的技术方案为前提进行的,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。
以一台卧式深孔镗床加工深孔为例,详细说明本发明的实施方式。
第一步,深孔镗削过程中的振动信号采集
通过磁性吸座将#1三向加速度传感器6和#2三向加速度传感器7吸附在深孔镗床刀杆保持架轴瓦侧面。数据采集软件的采样频率设置为1000Hz,采集加工过程中的振动。根据加工过程中刀具状态的变化将振动信号划分成6469段振动信号,其中正常状态数据6360段,断刀状态数据87段,磨钝状态数据22段。
第二步,采集数据的预处理
首先,利用线性插值方法对原始数据进行缺失值插补,其次,对正常状态的刀具数据进行朴素随机欠采样,获得400段振动信号数据,对磨钝状态数据进行加噪,添加三种不同强度的高斯白噪声,从而获得共88段振动信号数据,然后,对采集的加工过程振动信号进行规整处理,使各段振动信号长度为1000采样点。最后,两个三向加速度传感器采集的三个方向的振动信号V x、V y、V z进行矢量叠加,获取主振动方向的有效值。
第三步,ELM-SDAE网络的构建和训练
堆叠由四个基于ELM算法的去噪自编码器和一个ELM分类器组成的ELM-SDAE网络。其中,四个去噪自编码器的退化率为0.1,正则化系数C 1为1000,正则化系数C 2为1e7,激活函数为双曲正切函数,第一个去噪自编码器输入层和输出层神经元个数均为2000,隐层神经元个数为1200。第二个去噪自编码器输入层和输出层神经元个数均为1200,隐层神经元个数为400。第三个去噪自编码器输入层和输出层神经元个数均为400,隐层神经元个数为150。第四个去噪自编码器输入层和输出层神经元个数均为150,隐层神经元个数为50。 将数据按照4:1的比例划分训练集和测试集,其中训练集样本数为452,测试集样本数为123。基于贪婪学习的方法,使用未被利用的5960个正常状态数据对自编码器进行训练,将训练集数据输入训练后的堆叠去噪自编码器获得深层次特征,并对网络分类器进行训练,输出刀具的三种状态。再用测试集数据对网络进行测试,测试准确率达93.5%,模型可以用于刀具状态的监测。
第四步,深孔镗刀状态的实时监测
将实际加工过程采集的实时振动经数据预处理后输入至ELM-SDAE网络中,对镗削加工过程中的刀具状态进行监测。
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ELM-SDAE算法的刀具状态监测方法,其特征在于:首先,通过三向加速度传感器采集加工过程中的振动信号;然后,采用线性插值的方法对缺失值进行插补,采用朴素随机欠采样和添加白噪声的方法,对数据进行平衡处理,将振动信号进行矢量叠加获取主振动方向的有效值;接着,构建ELM-SDAE网络,采用贪婪逐层方法对网络模型进行训练;最后,将实际加工过程中的实时振动经数据预处理后输入ELM-SDAE网络中,网络输出刀具的当前状态,从而实现刀具状态的实时监测;具体步骤如下:
    第一步,加工过程中的振动信息采集
    采用三向加速度传感器采集数控机床加工过程中的振动信号,从该三向振动信号中截取与待监测刀具对应的振动信号,并将其分成与K个加工工件的加工过程对应的K段振动信号;
    第二步,采集数据的预处理
    首先,考虑到信号采集过程中出现的数据异常情况,对振动信号进行数据清洗;采用线性插值方法对缺失值进行插补计算;
    假设已知两点坐标(x 0,y 0)和(x 1,y 1),区间[x 0,x 1]内某一位置的纵坐标y通过式(1)线性插值后计算得到:
    Figure PCTCN2020077092-appb-100001
    其次,针对实际加工过程中,刀具多处于正常状态的情况,采用朴素随机欠采样降低正常状态刀具数据量,采用添加高斯白噪声的方式对磨钝或断刀状态数据进行数据增强;然后,对采集的加工过程振动信号进行规整处理,使各段振动信号长度一致;最后,将三个方向的振动信号V x、V y、V z进行矢量叠加,获取主振动方向的有效值V;
    第三步,ELM-SDAE网络的构建和训练
    在特征提取阶段,假设样本个数为N,第一个ELM-DAE网络的输入层和输出层节点数均为
    Figure PCTCN2020077092-appb-100002
    隐含层节点数为
    Figure PCTCN2020077092-appb-100003
    首先,将振动信号V按照比例置为0或其他值,从而获得“破坏”后的振动信号
    Figure PCTCN2020077092-appb-100004
    并将网络的期望输出替换成振动信号V;然后,为增强自编码器网络的泛化性能,随机生成正交化的隐含层参数;去噪自编码器网络输入与输出的关系由式(2)表示:
    Figure PCTCN2020077092-appb-100005
    其中,
    Figure PCTCN2020077092-appb-100006
    为隐含层的输出矩阵,
    Figure PCTCN2020077092-appb-100007
    为隐含层节点的输出权值向量,W 1为特征提取阶段隐含层节点的输入权值向量,b 1为特征提取阶段节点的偏置向量,g(·)为网络的激活函数;
    为了提高网络的泛化能力,在损失函数中添加正则项,如式(3)所示:
    Figure PCTCN2020077092-appb-100008
    其中,C 1为特征提取阶段网络的正则化系数,影响特征提取阶段网络的泛化性能;
    隐含层和输出层之间的连接权重β 1根据式(4)进行计算:
    Figure PCTCN2020077092-appb-100009
    对于由k个ELM-DAE网络堆叠而成的神经网络,第k个网络的输出即从振动信号中提取的特征通过式(5)进行计算:
    Figure PCTCN2020077092-appb-100010
    其中,当k为1时,
    Figure PCTCN2020077092-appb-100011
    在堆叠后的ELM-SAE网络中,如果第k层神经元的个数与第k-1层神经元个数相同,则激活函数采用线性函数,否则采用非线性激活函数;
    在分类阶段,假设训练集样本个数为M,ELM网络的输入层神经元节点数 为L 2,输出层神经元节点数即刀具状态类别为L 3,隐含层节点数为N 2,网络的期望输出为
    Figure PCTCN2020077092-appb-100012
    随机生成正交化隐含层参数后,隐含层的输出权值由公式(7)计算:
    Figure PCTCN2020077092-appb-100013
    Figure PCTCN2020077092-appb-100014
    其中,W 2为分类阶段隐含层的输入权值向量,b 2为分类阶段节点的偏置向量,C 2为分类阶段正则化系数,影响分类阶段网络的分类性能;
    第四步,刀具状态的实时监测
    在实际加工过程中,将实时数据经数据预处理后输入至训练后的ELM-SDAE网络模型中,模型输出刀具的实时状态;当刀具状态正常时,模型输出为0;当刀具状态为断刀时,模型输出为1;当刀具状态为磨钝时,模型输出为2。
PCT/CN2020/077092 2020-01-10 2020-02-28 一种基于elm-sdae算法的刀具状态监测方法 WO2021138982A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010027781.9 2020-01-10
CN202010027781.9A CN111136509B (zh) 2020-01-10 2020-01-10 一种基于elm-sdae算法的刀具状态监测方法

Publications (1)

Publication Number Publication Date
WO2021138982A1 true WO2021138982A1 (zh) 2021-07-15

Family

ID=70524444

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/077092 WO2021138982A1 (zh) 2020-01-10 2020-02-28 一种基于elm-sdae算法的刀具状态监测方法

Country Status (2)

Country Link
CN (1) CN111136509B (zh)
WO (1) WO2021138982A1 (zh)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111694320B (zh) * 2020-06-10 2021-04-20 大连理工大学 一种虑及颤振抑制的机床自适应控制方法
CN113628403A (zh) * 2020-07-28 2021-11-09 威海北洋光电信息技术股份公司 基于多核支持向量机的光纤振动传感周界安防入侵行为识别算法
CN112734000A (zh) * 2020-11-11 2021-04-30 江西理工大学 一种入侵检测方法、系统、设备及可读存储介质
CN113103068B (zh) * 2021-04-19 2022-03-29 大连理工大学 一种基于深度迁移学习的刀具状态监测方法
CN113458873B (zh) * 2021-07-01 2022-03-11 太原科技大学 一种刀具磨损量和剩余寿命预测的方法
CN113352149B (zh) * 2021-08-11 2021-11-02 西安博澜数字科技有限公司 一种刀具工况参数监控方法和装置
CN116551467B (zh) * 2023-07-05 2023-09-01 南京维拓科技股份有限公司 基于物联网监控数据的刀具磨损状态识别方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170320182A1 (en) * 2016-05-06 2017-11-09 Massachusetts Institute Of Technology Method and Apparatus for Efficient Use of CNC Machine Shaping Tool Including Cessation of Use No Later than the Onset of Tool Deterioration By Monitoring Audible Sound During Shaping
CN110153802A (zh) * 2019-07-04 2019-08-23 西南交通大学 一种基于卷积神经网络和长短时记忆神经网络联合模型的刀具磨损状态辨识方法
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
CN110576336A (zh) * 2019-09-11 2019-12-17 大连理工大学 基于ssae-lstm模型的深孔加工刀具磨损量监测方法
KR20190142909A (ko) * 2018-06-19 2019-12-30 덕흥 주식회사 스크류 가공장치 및 스크류 가공장치의 절삭 툴 모니터링 방법

Family Cites Families (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20000065123A (ko) * 1996-04-30 2000-11-06 안자키 사토루 공작기계의공구이상검출방법및장치
JP4321581B2 (ja) * 2006-11-30 2009-08-26 パナソニック電工株式会社 工作機械総合監視装置
CN101847270B (zh) * 2010-05-14 2013-07-03 东南大学 基于虚拟样机的四轴重型货车悬架系统控制模拟方法
CN104786101A (zh) * 2015-04-29 2015-07-22 常州信息职业技术学院 立铣切削颤振的监测方法
CN106002490B (zh) * 2016-05-12 2017-12-26 西北工业大学 基于刀位轨迹和冗余剔除的铣削工件粗糙度监测方法
CN107016241A (zh) * 2017-04-05 2017-08-04 重庆交通大学 基于加噪样本增强深度自编码学习网络的旋转机械寿命阶段识别方法
CN108154187B (zh) * 2018-01-04 2022-03-01 湘潭大学 一种基于振动信号的深孔啄钻加工质量检测方法
US10921777B2 (en) * 2018-02-15 2021-02-16 Online Development, Inc. Automated machine analysis
CN109141884A (zh) * 2018-08-17 2019-01-04 江南大学 基于eemd-ar模型和dbn的轴承故障诊断方法
CN109571141A (zh) * 2018-11-01 2019-04-05 北京理工大学 一种基于机器学习的刀具磨损状态监测方法
CN109858509A (zh) * 2018-11-05 2019-06-07 杭州电子科技大学 基于多层随机神经网络单分类器异常检测方法
CN110006650A (zh) * 2019-03-18 2019-07-12 华中科技大学 一种基于栈式剪枝稀疏去噪自动编码器的故障诊断方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170320182A1 (en) * 2016-05-06 2017-11-09 Massachusetts Institute Of Technology Method and Apparatus for Efficient Use of CNC Machine Shaping Tool Including Cessation of Use No Later than the Onset of Tool Deterioration By Monitoring Audible Sound During Shaping
KR20190142909A (ko) * 2018-06-19 2019-12-30 덕흥 주식회사 스크류 가공장치 및 스크류 가공장치의 절삭 툴 모니터링 방법
CN110153802A (zh) * 2019-07-04 2019-08-23 西南交通大学 一种基于卷积神经网络和长短时记忆神经网络联合模型的刀具磨损状态辨识方法
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
CN110576336A (zh) * 2019-09-11 2019-12-17 大连理工大学 基于ssae-lstm模型的深孔加工刀具磨损量监测方法

Also Published As

Publication number Publication date
CN111136509B (zh) 2021-06-08
CN111136509A (zh) 2020-05-12

Similar Documents

Publication Publication Date Title
WO2021138982A1 (zh) 一种基于elm-sdae算法的刀具状态监测方法
CN108356606B (zh) 基于小波包分析与rbf神经网络的刀具磨损在线监测方法
CN110355608B (zh) 基于自注意力机制与深度学习的刀具磨损量预测方法
US9508042B2 (en) Method for predicting machining quality of machine tool
CN111716150B (zh) 一种刀具状态智能监测的进化学习方法
Marani et al. Prediction of cutting tool wear during a turning process using artificial intelligence techniques
Li et al. Hybrid learning for tool wear monitoring
CN113927371A (zh) 一种基于多传感器特征融合的刀具磨损预测方法
Wang et al. Chatter detection based on probability distribution of wavelet modulus maxima
CN102929210A (zh) 基于特征的数控加工过程控制和优化系统及方法
WO2024000720A1 (zh) 基于切削力成分解耦的变工况刀具磨损监测方法及系统
CN106112697A (zh) 一种基于3σ准则的铣削颤振自动报警阈值设定方法
Liu et al. A hybrid health condition monitoring method in milling operations
CN110561192A (zh) 一种基于堆叠自编码器的深孔镗刀状态监测方法
Jemielniak Contemporary challenges in tool condition monitoring
CN111331429A (zh) 基于小波包能量分析的刀具磨损状态监测方法及装置
WO2020143203A1 (zh) 一种深孔镗削加工颤振的在线监测与抑制方法
CN112247673B (zh) 一种基于遗传bp神经网络的木工刀具磨损状态诊断方法
CN104669114B (zh) 磨床颤振监测系统及其磨削故障信号的提取和诊断方法
Tonshoff et al. Application of fast Haar transform and concurrent learning to tool-breakage detection in milling
CN106363463A (zh) 基于占能比的铣削加工颤振在线监测方法
CN114378653B (zh) 基于bp神经网络外圆磨削颤振在线识别和监测方法
CN106881630B (zh) 基于自适应滤波与ar模型的高速铣削颤振在线识别方法
Zhou et al. Feature extraction and analysis method of spindle vibration signal for tool wear monitoring
Gui et al. Unsupervised detection of tool breakage: A novel approach based on time and sensor domain data analysis

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: 20911401

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: 20911401

Country of ref document: EP

Kind code of ref document: A1