WO2021128576A1 - 基于生成式对抗网络的刀具状态监测数据集增强方法 - Google Patents

基于生成式对抗网络的刀具状态监测数据集增强方法 Download PDF

Info

Publication number
WO2021128576A1
WO2021128576A1 PCT/CN2020/077095 CN2020077095W WO2021128576A1 WO 2021128576 A1 WO2021128576 A1 WO 2021128576A1 CN 2020077095 W CN2020077095 W CN 2020077095W WO 2021128576 A1 WO2021128576 A1 WO 2021128576A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
generated
discriminator
generator
tool
Prior art date
Application number
PCT/CN2020/077095
Other languages
English (en)
French (fr)
Inventor
王永青
牛蒙蒙
刘阔
秦波
沈明瑞
厉大维
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 US16/970,301 priority Critical patent/US20210197335A1/en
Publication of WO2021128576A1 publication Critical patent/WO2021128576A1/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
    • B23Q17/0957Detection of tool breakage
    • 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
    • B23Q17/0976Detection or control of chatter

Definitions

  • the invention belongs to the field of machining state monitoring, and specifically is a tool state monitoring data set enhancement method based on a generative confrontation network.
  • Tool wear is a common problem in metal cutting.
  • the processing of the material makes the cutting edge of the tool dull, increases the friction between the tool and the workpiece, and also increases the power consumption. If the tool wear status cannot be judged in time, it will affect the processing quality and processing efficiency.
  • GANs Generative Adversarial Networks
  • the present invention provides a tool condition monitoring data set enhancement method based on a generative confrontation network.
  • the generator and discriminator in the generative confrontation network are both multi-layer perceptron structures, and the method of confrontation training is adopted between the two to complete the process of establishing the generative confrontation network model.
  • the technical scheme of the present invention a tool condition monitoring data set enhancement method based on a generative confrontation network.
  • a sensor acquisition system is used to obtain the vibration signal and noise signal during the cutting process of the tool;
  • the noise data of the prior distribution will be obeyed
  • Input to the generator to generate data, and input the generated data and the collected real sample data to the discriminator for identification.
  • the generator and the discriminator conduct confrontation training until the training is completed; then, use the trained generator to generate Sample data, and determine whether the distribution of the generated sample data and the real tool status sample data is similar;
  • the accuracy of the deep learning network model to predict the tool status is used to check the usability of the generated data; the specific steps are as follows:
  • the first step is to collect the vibration and sound signals during the cutting process of the tool
  • the second step is to establish a generative confrontation network model and conduct confrontation training
  • the generative confrontation network framework used in this method is composed of a generator and a discriminator; both the generator and the discriminator have a multi-layer perceptron structure, where the generator is responsible for generating fake data with the same dimensions as the real data, and the discriminator is responsible for distinguishing Real data and generated data; during the adversarial training process, the generator tries to fool the discriminator with the generated fake data to make it true, and the discriminator improves its ability to distinguish between the generated data and the real data, both The game finally reaches the Nash equilibrium state, that is, the sample data generated by the generator is indistinguishable from the real sample data, and the discriminator cannot distinguish the generated sample data from the real sample data;
  • the number of tool state samples collected by this method is l, and the dimension of the vibration signal is 6000, which is set as among them
  • the dimension of the noise data set is 1000, set among them
  • Data set of tool status among them u 7000; the tool state data set of the input discriminator
  • the maximum-minimum method is used for normalization processing, so that the input data is converted into a number between [0,1], and after the sample data is generated, the denormalization processing is performed.
  • the normalization function used in the form is as in formula (1) As shown, the form of the denormalization function is shown in formula (2):
  • tool (i) is the original data of the tool state
  • tool (i)' is the data after normalization, Is the smallest number in the data sequence, Is the largest number in the sequence;
  • f is the activation function
  • the activation function of the hidden layer adopts the ReLU function, and the function form is shown in formula (4):
  • the activation function of the output layer adopts the Sigmoid function, and the function form is shown in formula (5):
  • the output of the discriminator is a two-class situation, the last layer uses the Sigmoid function, and the output probability value is shown in equation (6):
  • P data (x) is the tool state data set
  • the data distribution of P z (z) is a priori noise distribution
  • D(x) means that x comes from The probability of
  • D(G(z)) represents the probability that G(z) comes from the generated data, where G(z) is the data sample generated by the generator from the noise data that obeys the prior distribution
  • Means x comes from Expectations of data distribution, Indicates the expectation that z comes from the noise distribution
  • the goal of the discriminator is to maximize the error function to distinguish between real data and generated data, and the generator is to minimize the error function to generate data samples that are closer to the real sample data distribution;
  • the Adam optimization algorithm is used to update the parameters
  • the training steps of the generative confrontation network are as follows:
  • the generator generates p fake tool status data samples from random noise
  • Steps (1) to (3) are a training period. After completing a period, the training process starts again from (1); after repeating multiple cycles of training the discriminator and generator, save the generator network parameters;
  • the third step is to compare the similarity between the generated data and the real data
  • the fourth step is to verify the availability of the generated sample data
  • the original unbalanced data set and the enhanced data set are used to train the deep learning network model, test the prediction accuracy of the two, and verify the usability of the generated data; there is no intersection between the training set and the test set, and the test set is composed of real data.
  • the present invention has the following beneficial effects:
  • the generative confrontation network model adopted in the present invention can learn the distribution of data, generate sample data with the same distribution as the original data, and can effectively enhance the training data set.
  • the present invention uses the enhanced data set to train the deep network model, which can effectively improve the accuracy of tool state monitoring.
  • Figure 1 is a flowchart of a tool condition monitoring data set enhancement method based on a generative confrontation network.
  • Figure 2 is a schematic diagram of the sensor installation position.
  • Figure 3 is a structural diagram of the generative confrontation network used in the present invention.
  • Figure 4 (a) is a time-domain diagram, (b) is a spectrum diagram.
  • Figure 5 (a) is the training process of the deep learning network, (b) is the prediction result of the deep learning network.
  • Two three-directional acceleration sensors are attached to the two cage bearings of the deep-hole boring bar through the magnetic base, and the sound sensor is placed at one end of the inner hole of the workpiece to collect the tool bar vibration and cutting noise during the machining process.
  • the sensor installation position is shown in Figure 2.
  • the three types of sample data collected are shown in Table 1. Each sample contains 7000 data points (6000 data points for vibration signals and 1000 data points for noise signals):
  • the sample data in the blunt state in Table 1 is significantly less than the sample data in the normal state and the broken knife state, so we generate the sample data in the blunt state.
  • both the generator and the discriminator adopt a three-layer fully connected neural network model, and the number of neurons in the hidden layer of the generator and the discriminator is set to 125, and the input of the generator The number of neurons is 100.
  • the network structure is shown in Figure 3.
  • the learning rate is set to 0.001, the batch size is 12, the number of iterations is set to 100, and the input noise distribution obeys a uniform distribution in the interval [-1,1].
  • the ratio of the real sample data and the generated sample data in the blunt state is 1:3.
  • the deep learning network adopts a deep belief network model, and the parameters are set as follows: the learning rate is 0.001; the number of iterations of the unsupervised training process is 100, and the number of iterations of the fine-tuning process is 200.
  • the hidden layer has three layers, and the number of neurons in each layer is 100, 60, and 30, respectively. Since the momentum gradient descent method is better than the gradient descent method, we use the momentum gradient descent method to optimize the parameters, and the momentum term is 0.9.
  • the sample data is shown in Table 2. The original unbalanced data set and the enhanced data set were divided into training set and test set according to the ratio of 4:1. Use the training set to train the network and test it on the test set.

Landscapes

  • Engineering & Computer Science (AREA)
  • Mechanical Engineering (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)

Abstract

本发明提供了一种基于生成式对抗网络的刀具状态监测数据集增强方法,属于机械加工状态监测技术领域。首先,采用传感器采集系统获取刀具切削过程中的振动和声音信号;其次,将服从先验分布的噪声数据输入到生成器生成数据,并将生成数据和采集的真实样本数据输入到鉴别器进行鉴别,生成器和鉴别器两者之间进行对抗训练,直到训练完成;然后,利用训练好的生成器生成样本数据,并判断生成的样本数据和真实的刀具状态样本数据的分布是否相似;最后,结合深度学习网络模型预测刀具状态的准确性检验生成数据的可用性。该方法的最大优点能够增强刀具状态数据集,提高深度学习网络模型预测刀具状态的准确性。

Description

基于生成式对抗网络的刀具状态监测数据集增强方法 技术领域
本发明属于机械加工状态监测领域,具体为一种基于生成式对抗网络的刀具状态监测数据集增强方法。
背景技术
刀具磨损是金属切削加工中存在的一种常见问题。对材料的加工使刀具刃口钝化,增加了刀具与工件之间的摩擦,也增加了功率的消耗。若不能及时判断刀具磨损状态,会影响加工质量和加工效率。
得益于深度学习技术的发展,采用深度学习网络对刀具状态进行间接监测成为一种非常有效的方法。然而,这些方法都是以加工过程大数据为基础。在大多数机械加工过程中,刀具通常在正常状态下工作,可以收集到的异常状态下的数据很少,容易出现数据集不平衡的问题。异常状态样本数据缺乏以及数据不平衡的问题严重影响深度学习网络的预测精度。传统扩充样本数据集的方式是过采样,但过采样只是重复利用仅有的少量样本信息,不能自动学习样本的数据分布特性。因此,如何获取异常状态的样本数据成为亟待解决的问题。
生成对抗网络(Generative Adversarial Networks,GANs)作为2014年提出的无监督学习模型,在增强数据集、加工状态监测领域有着广阔的应用前景。它可以通过对少量的样本分布进行学习,从而产生大量的样本数据。这种特点非常适合解决加工状态监测中缺乏平衡样本数据集的问题。
发明内容
针对刀具状态监测数据集不平衡导致深度学习网络预测准确性难以提高的问题,本发明提供一种基于生成式对抗网络的刀具状态监测数据集增强方法。生成式对抗网络中生成器和鉴别器均为多层感知器结构,两者之间采用对抗训 练的方式,完成生成式对抗网络模型的建立过程。利用训练好的生成器生成样本数据,并结合深度学习网络预测模型验证生成样本数据的可用性。
本发明的技术方案:一种基于生成式对抗网络的刀具状态监测数据集增强方法,首先,采用传感器采集系统获取刀具切削过程中的振动信号和噪声信号;其次,将服从先验分布的噪声数据输入到生成器生成数据,并将生成数据和采集的真实样本数据输入到鉴别器进行鉴别,生成器和鉴别器两者之间进行对抗训练,直到训练完成;然后,利用训练好的生成器生成样本数据,并判断生成的样本数据和真实的刀具状态样本数据的分布是否相似;最后,结合深度学习网络模型预测刀具状态的准确性检验生成数据的可用性;具体步骤如下:
第一步,采集刀具切削过程中的振动和声音信号
将两个加速度传感器分别安装在主轴的鼻端和主轴前轴承处,分别采集刀具加工过程中的振动信号;将声音传感器安装在工作台上,采集加工过程中的切削噪声信号;
第二步,建立生成式对抗网络模型并进行对抗训练
本方法采用的生成式对抗网络框架由一个生成器和一个鉴别器构成;生成器和鉴别器均为多层感知器结构,其中生成器负责生成和真实数据维度相同的伪数据,鉴别器负责区分真实数据和生成数据;在对抗训练过程中,生成器试图用生成的伪数据去愚弄鉴别器,使其鉴别为真,而鉴别器通过提高自己的鉴别能力分辨生成数据和真实数据,两者进行博弈,最终达到纳什平衡状态,即生成器生成的样本数据与真实的样本数据无差别,鉴别器也无法区分生成的样本数据和真实的样本数据;
本方法采集的刀具状态样本个数为l,振动信号的维度为6000,设为
Figure PCTCN2020077095-appb-000001
其中
Figure PCTCN2020077095-appb-000002
噪声数据集的维度为1000,设为
Figure PCTCN2020077095-appb-000003
其中
Figure PCTCN2020077095-appb-000004
刀具状态的数据集
Figure PCTCN2020077095-appb-000005
其中
Figure PCTCN2020077095-appb-000006
u=7000;对输入鉴别器的刀具状态数据集
Figure PCTCN2020077095-appb-000007
采用最大最小法进行归一化处理,使输入数据转化为[0,1]之间的数,并在生成样本数据之后进行反归一化处理,采用的归一化函数形式如式(1)所示,反归一化函数形式如式(2)所示:
Figure PCTCN2020077095-appb-000008
Figure PCTCN2020077095-appb-000009
式中,tool (i)是刀具状态的原始数据,tool (i)'是归一化之后的数据,
Figure PCTCN2020077095-appb-000010
为数据序列中的最小数,
Figure PCTCN2020077095-appb-000011
为序列中的最大数;
生成器和鉴别器均采用三层全连接神经网络,输入数据集为归一化之后的数据集
Figure PCTCN2020077095-appb-000012
输入层到隐含层以及隐含层到输出层的映射公式如式(3)所示:
h i=f θ(w*tool (i)'+b)       (3)
式中,f为激活函数,θ={w,b}是网络的参数矩阵,其中w是输入层、隐含层和输出层神经元之间的连接权值,b是隐含层和输出层神经元的阈值;
隐含层的激活函数采用ReLU函数,函数形式如式(4)所示:
Figure PCTCN2020077095-appb-000013
输出层的激活函数采用Sigmoid函数,函数形式如式(5)所示:
Figure PCTCN2020077095-appb-000014
鉴别器的输出是二分类情况,最后一层采用Sigmoid函数,输出的概率值如式(6)所示:
Figure PCTCN2020077095-appb-000015
本方法设置的目标函数如式(7)所示:
Figure PCTCN2020077095-appb-000016
鉴别器的目标函数和最优解如式(8)和(9)所示:
Figure PCTCN2020077095-appb-000017
Figure PCTCN2020077095-appb-000018
生成器的目标函数如式(10)所示:
Figure PCTCN2020077095-appb-000019
式中,P data(x)是刀具状态数据集
Figure PCTCN2020077095-appb-000020
的数据分布,P z(z)是一个先验噪声分布;D(x)表示x来自
Figure PCTCN2020077095-appb-000021
的概率;D(G(z))表示G(z)来自生成数据的概率,其中G(z)是生成器由服从先验分布的噪声数据生成的数据样本;
Figure PCTCN2020077095-appb-000022
表示x来自
Figure PCTCN2020077095-appb-000023
的数据分布的期望,
Figure PCTCN2020077095-appb-000024
表示z来自噪声分布的期望;鉴别器的目标是最大化误差函数,以区分真实数据和生成数据,生成器则是最小化误差函数,生成和真实的样本数据分布更接近的数据样本;
基于目标函数,采用亚当优化算法更新参数;
生成式对抗网络的训练步骤如下:
(1)生成器从随机噪声产生p个假的刀具状态数据样本
Figure PCTCN2020077095-appb-000025
(2)将生成的样本数据
Figure PCTCN2020077095-appb-000026
标签为0和原始的样本数据
Figure PCTCN2020077095-appb-000027
标签为1混合在一起输入到鉴别器中;基于损失函数,固定生成器的参数不变,只更新鉴别器的参数,并且训练鉴别器以提高鉴别器对真假样本的分辨能力;
(3)训练鉴别器之后,将生成样本
Figure PCTCN2020077095-appb-000028
的标签设为1;基于损失函数,误差反向传播,在这个阶段,鉴别器的参数被冻结,不能被更新,只能更新生成器中的参数,并且训练发生器以产生更真实的数据样本;
(4)步骤(1)~(3)为一个训练时期,完成一个时期之后训练过程再次从(1)开始;重复多个周期训练鉴别器和生成器之后,保存生成器网络参数;
第三步,对比生成数据和真实数据的相似性
利用训练好的生成器生成样本数据,将生成的刀具状态样本数据
Figure PCTCN2020077095-appb-000029
和真实的刀具状态样本数据
Figure PCTCN2020077095-appb-000030
的时频图进行对比分析,判断生成的样本数据和真实的样本数据的分布是否相同;如果相同,则将生成的样本数据进行反归一化,
Figure PCTCN2020077095-appb-000031
为反归一化之后的生成的刀具状态样本数据,并将
Figure PCTCN2020077095-appb-000032
添加到原始的不平衡数据集
Figure PCTCN2020077095-appb-000033
中,增强的数据集为
Figure PCTCN2020077095-appb-000034
如果不相同,则返回到生成式对抗网络中继续进行对抗训练,直到生成的样本数据和真实的样本数据分布相同为止;
第四步,验证生成样本数据的可用性
采用原始的不平衡数据集和增强的数据集分别训练深度学习网络模型,测试两者的预测精度,验证生成数据的可用性;训练集和测试集无任何交集,并且测试集由真实数据组成。
与现有技术相比,本发明的有益效果在于:
1.本发明采用的生成式对抗网络模型能够学习数据的分布,生成与原始数据分布相同的样本数据,可有效增强训练数据集。
2.本发明利用增强数据集训练深度网络模型,可有效提高刀具状态监测的准确性。
附图说明
图1为基于生成式对抗网络的刀具状态监测数据集增强方法的流程图。
图2为传感器安装位置示意图。
图3为本发明所采用的生成式对抗网络的结构图。
图4(a)为时域图,(b)为频谱图。
图5(a)为深度学习网络的训练过程,(b)为深度学习网络的预测结果。
图中:1工件保持架;2工件;3机床齿轮箱;4传声器;5床身;6 1#三向加速度传感器;7刀杆;8 2#三向加速度传感器;9刀杆保持架。
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,结合附图1以国产某型号深孔镗床镗削加工为例,详细说明本发明的实施方式。
将两个三向加速度传感器通过磁座吸附粘贴在深孔镗杆的两个保持架轴承上,声音传感器放置于工件内孔的一端,对加工过程中的刀杆振动以及切削噪声进行采集。传感器安装位置见附图2所示。采集的三类样本数据如表1所示,每个样本包含7000个数据点(振动信号的数据点为6000个,噪声信号的数据点为1000个):
表1样本数量
Figure PCTCN2020077095-appb-000035
表1中磨钝状态的样本数据明显少于正常状态和断刀状态的样本数据,因此我们对磨钝状态的样本数据进行生成。
本发明采用的生成式对抗网络模型中,生成器和鉴别器都采用三层全连接神经网络模型,其中生成器和鉴别器的隐含层的神经元个数设置为125个,生成器的输入神经元个数为100个。网络结构见附图3所示。学习率设为0.001,批量大小为12个,迭代次数设置为100次,输入的噪声分布服从区间为[-1,1] 的均匀分布。磨钝状态真实样本数据和生成样本数据的比例为1:3。
利用训练好的生成器生成样本数据,并用MATLAB做出真实样本数据和生成样本数据的时频图,见附图4(a)和(b)所示。从时域图和频谱图可以看出,真实的样本数据和生成的样本数据分布相似度较高。
深度学习网络采用深度置信网络模型,参数设置如下:学习速率为0.001;无监督训练过程的迭代次数为100,微调过程的迭代次数为200。隐含层为三层,每一层的神经元的个数分别为100、60、30。由于动量梯度下降法优于梯度下降法,因此我们采用动量梯度下降法来优化参数,动量项为0.9。样本数据如表2所示。将原始的不平衡数据集和增强的数据集分别按照4:1的比例分成训练集和测试集。利用训练集对网络进行训练,并在测试集上测试。
由结果可知,不平衡数据集测试的准确率为97.1%,误差率为2.9%;增强数据集测试的准确率为99.2%,误差率为0.8%。两者比较可知,深度学习网络模型预测的准确性提高了2.9%,而误差率下降了三倍以上。由此验证了生成的样本数据的可用性。增强数据集在深度学习网络上的训练过程和训练结果见附图5(a)和(b)所示。
表2样本数量
Figure PCTCN2020077095-appb-000036

Claims (1)

  1. 一种基于生成式对抗网络的刀具状态监测数据集增强方法,首先,采用传感器采集系统获取刀具切削过程中的振动信号和噪声信号;其次,将服从先验分布的噪声数据输入到生成器生成数据,并将生成数据和采集的真实样本数据输入到鉴别器进行鉴别,生成器和鉴别器两者之间进行对抗训练,直到训练完成;然后,利用训练好的生成器生成样本数据,并判断生成的样本数据和真实的刀具状态样本数据的分布是否相似;最后,结合深度学习网络模型预测刀具状态的准确性检验生成数据的可用性;其特征在于,步骤如下:
    第一步,采集刀具切削过程中的振动和声音信号
    将两个加速度传感器分别安装在主轴的鼻端和主轴前轴承处,分别采集刀具加工过程中的振动信号;将声音传感器安装在工作台上,采集加工过程中的切削噪声信号;
    第二步,建立生成式对抗网络模型并进行对抗训练
    本方法采用的生成式对抗网络框架由一个生成器和一个鉴别器构成;生成器和鉴别器均为多层感知器结构,其中生成器负责生成和真实数据维度相同的伪数据,鉴别器负责区分真实数据和生成数据;在对抗训练过程中,生成器试图用生成的伪数据去愚弄鉴别器,使其鉴别为真,而鉴别器通过提高自己的鉴别能力分辨生成数据和真实数据,两者进行博弈,最终达到纳什平衡状态,即生成器生成的样本数据与真实的样本数据无差别,鉴别器也无法区分生成的样本数据和真实的样本数据;
    本方法采集的刀具状态样本个数为l,振动信号的维度为6000,设为
    Figure PCTCN2020077095-appb-100001
    其中
    Figure PCTCN2020077095-appb-100002
    噪声数据集的维度为1000,设为
    Figure PCTCN2020077095-appb-100003
    其中
    Figure PCTCN2020077095-appb-100004
    刀具状态的数据集
    Figure PCTCN2020077095-appb-100005
    其中
    Figure PCTCN2020077095-appb-100006
    对输入鉴别器的刀具状态数据集
    Figure PCTCN2020077095-appb-100007
    采用最大最小法进行归一化处理,使输 入数据转化为[0,1]之间的数,并在生成样本数据之后进行反归一化处理,采用的归一化函数形式如式(1)所示,反归一化函数形式如式(2)所示:
    Figure PCTCN2020077095-appb-100008
    Figure PCTCN2020077095-appb-100009
    式中,tool (i)是刀具状态的原始数据,tool (i)'是归一化之后的数据,
    Figure PCTCN2020077095-appb-100010
    为数据序列中的最小数,
    Figure PCTCN2020077095-appb-100011
    为序列中的最大数;
    生成器和鉴别器均采用三层全连接神经网络,输入数据集为归一化之后的数据集
    Figure PCTCN2020077095-appb-100012
    输入层到隐含层以及隐含层到输出层的映射公式如式(3)所示:
    h i=f θ(w*tool (i)'+b)    (3)
    式中,f为激活函数,θ={w,b}是网络的参数矩阵,其中w是输入层、隐含层和输出层神经元之间的连接权值,b是隐含层和输出层神经元的阈值;
    隐含层的激活函数采用ReLU函数,函数形式如式(4)所示:
    Figure PCTCN2020077095-appb-100013
    输出层的激活函数采用Sigmoid函数,函数形式如式(5)所示:
    Figure PCTCN2020077095-appb-100014
    鉴别器的输出是二分类情况,最后一层采用Sigmoid函数,输出的概率值如式(6)所示:
    Figure PCTCN2020077095-appb-100015
    本方法设置的目标函数如式(7)所示:
    Figure PCTCN2020077095-appb-100016
    鉴别器的目标函数和最优解如式(8)和(9)所示:
    Figure PCTCN2020077095-appb-100017
    Figure PCTCN2020077095-appb-100018
    生成器的目标函数如式(10)所示:
    Figure PCTCN2020077095-appb-100019
    式中,P data(x)是刀具状态数据集
    Figure PCTCN2020077095-appb-100020
    的数据分布,P z(z)是一个先验噪声分布;D(x)表示x来自
    Figure PCTCN2020077095-appb-100021
    的概率;D(G(z))表示G(z)来自生成数据的概率,其中G(z)是生成器由服从先验分布的噪声数据生成的数据样本;
    Figure PCTCN2020077095-appb-100022
    表示x来自
    Figure PCTCN2020077095-appb-100023
    的数据分布的期望,
    Figure PCTCN2020077095-appb-100024
    表示x来自噪声分布的期望;鉴别器的目标是最大化误差函数,以区分真实数据和生成数据,生成器则是最小化误差函数,生成和真实的样本数据分布更接近的数据样本;
    基于目标函数,采用亚当优化算法更新参数;
    生成式对抗网络的训练步骤如下:
    (1)生成器从随机噪声产生p个假的刀具状态数据样本
    Figure PCTCN2020077095-appb-100025
    (2)将生成的样本数据
    Figure PCTCN2020077095-appb-100026
    标签为0和原始的样本数据
    Figure PCTCN2020077095-appb-100027
    标签为1混合在一起输入到鉴别器中;基于损失函数,固定生成器的参数不变,只更新鉴别器的参数,并且训练鉴别器以提高鉴别器对真假样本的分辨能力;
    (3)训练鉴别器之后,将生成样本
    Figure PCTCN2020077095-appb-100028
    的标签设为1;基于损失函数,误差反向传播,在这个阶段,鉴别器的参数被冻结,不能被更新,只能更新生成器中的参数,并且训练发生器以产生更真实的数据样本;
    (4)步骤(1)~(3)为一个训练时期,完成一个时期之后训练过程再次从(1)开始;重复多个周期训练鉴别器和生成器之后,保存生成器网络参数;
    第三步,对比生成数据和真实数据的相似性
    利用训练好的生成器生成样本数据,将生成的刀具状态样本数据
    Figure PCTCN2020077095-appb-100029
    和真实的刀具状态样本数据
    Figure PCTCN2020077095-appb-100030
    的时频图进行对比分析,判断生成的样本数据和真实的样本数据的分布是否相同;如果相同,则将生成的样本数据进行反归一化,
    Figure PCTCN2020077095-appb-100031
    为反归一化之后的生成的刀具状态样本数据,并将
    Figure PCTCN2020077095-appb-100032
    添加到原始的不平衡数据集
    Figure PCTCN2020077095-appb-100033
    中,增强的数据集为
    Figure PCTCN2020077095-appb-100034
    如果不相同,则返回到生成式对抗网络中继续进行对抗训练,直到生成的样本数据和真实的样本数据分布相同为止;
    第四步,验证生成样本数据的可用性
    采用原始的不平衡数据集和增强的数据集分别训练深度学习网络模型,测试两者的预测精度,验证生成数据的可用性;训练集和测试集无任何交集,并且测试集由真实数据组成。
PCT/CN2020/077095 2019-12-26 2020-02-28 基于生成式对抗网络的刀具状态监测数据集增强方法 WO2021128576A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/970,301 US20210197335A1 (en) 2019-12-26 2020-02-28 Data Augmentation Method Based On Generative Adversarial Networks In Tool Condition Monitoring

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911361333.6 2019-12-26
CN201911361333.6A CN111037365B (zh) 2019-12-26 2019-12-26 基于生成式对抗网络的刀具状态监测数据集增强方法

Publications (1)

Publication Number Publication Date
WO2021128576A1 true WO2021128576A1 (zh) 2021-07-01

Family

ID=70239929

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/077095 WO2021128576A1 (zh) 2019-12-26 2020-02-28 基于生成式对抗网络的刀具状态监测数据集增强方法

Country Status (2)

Country Link
CN (1) CN111037365B (zh)
WO (1) WO2021128576A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114378653A (zh) * 2022-01-27 2022-04-22 上海机床厂有限公司 基于bp神经网络外圆磨削颤振在线识别和监测方法

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111639461B (zh) * 2020-05-26 2022-10-18 浙江大学 针对工业不均衡数据的刀具磨损状态检测方法
CN113970170A (zh) * 2020-07-24 2022-01-25 中国移动通信集团浙江有限公司 中央空调系统能耗预测方法、装置及计算设备
CN112091727A (zh) * 2020-08-12 2020-12-18 上海交通大学 一种基于虚拟样本生成的刀具破损识别方法、装置和终端
CN111993158B (zh) * 2020-08-28 2022-08-23 深圳市玄羽科技有限公司 刀具检测模型的生成方法、检测方法、系统、设备及介质
CN112348253B (zh) * 2020-11-06 2023-04-07 中国石油大学(华东) 一种基于全连接神经网络的固态发酵效用预测方法
CN112508239A (zh) * 2020-11-22 2021-03-16 国网河南省电力公司电力科学研究院 一种基于vae-cgan的储能出力预测方法
CN112328588B (zh) * 2020-11-27 2022-07-15 哈尔滨工程大学 一种工业故障诊断非平衡时序数据扩充方法
CN112668227B (zh) * 2020-12-31 2024-02-02 华中科技大学 一种薄壁件让刀变形误差预测模型建立方法及其应用
CN113369994A (zh) * 2021-06-30 2021-09-10 温州大学 一种高速铣削过程刀具状态监测方法
CN113537031B (zh) * 2021-07-12 2023-04-07 电子科技大学 基于多鉴别器条件生成对抗网络的雷达图像目标识别方法
CN114510618B (zh) * 2021-12-31 2022-12-20 安徽郎溪南方水泥有限公司 基于智慧矿山的处理方法及装置
CN114346761B (zh) * 2022-01-06 2023-04-28 中国科学技术大学 一种基于改进条件生成对抗网络的刀具磨损状况检测方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104741638A (zh) * 2015-04-20 2015-07-01 江苏师范大学 一种车削刀具磨损状态监测系统
CN108107838A (zh) * 2017-12-27 2018-06-01 山东大学 一种基于云知识库和机器学习的数控装备刀具磨损检测方法
CN109158953A (zh) * 2018-09-04 2019-01-08 温州大学激光与光电智能制造研究院 一种刀具磨损状态在线监测方法及系统
CN109977464A (zh) * 2019-02-18 2019-07-05 江苏科技大学 一种基于bp神经网络的活塞切削加工变形量的预测方法

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE3902840A1 (de) * 1989-01-31 1990-08-23 Fraunhofer Ges Forschung Verfahren und vorrichtung zur ermittlung der oberflaechenstruktur
JPH1128647A (ja) * 1997-07-07 1999-02-02 Toshiba Mach Co Ltd 工具交換時期判定方法および装置
CN108763857A (zh) * 2018-05-29 2018-11-06 浙江工业大学 一种基于相似度生成对抗网络的过程软测量建模方法
CN108961217B (zh) * 2018-06-08 2022-09-16 南京大学 一种基于正例训练的表面缺陷检测方法
CN109165664B (zh) * 2018-07-04 2020-09-22 华南理工大学 一种基于生成对抗网络的属性缺失数据集补全与预测方法
CN109508717A (zh) * 2018-10-09 2019-03-22 苏州科达科技股份有限公司 一种车牌识别方法、识别装置、识别设备及可读存储介质
CN109543742A (zh) * 2018-11-16 2019-03-29 聚时科技(上海)有限公司 一种基于GAN和Self-Attention的图像局部信息迁移方法
CN109993229A (zh) * 2019-04-02 2019-07-09 广东石油化工学院 一种严重不平衡数据分类方法
CN110188824B (zh) * 2019-05-31 2021-05-14 重庆大学 一种小样本植物病害识别方法及系统
US10426442B1 (en) * 2019-06-14 2019-10-01 Cycle Clarity, LLC Adaptive image processing in assisted reproductive imaging modalities

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104741638A (zh) * 2015-04-20 2015-07-01 江苏师范大学 一种车削刀具磨损状态监测系统
CN108107838A (zh) * 2017-12-27 2018-06-01 山东大学 一种基于云知识库和机器学习的数控装备刀具磨损检测方法
CN109158953A (zh) * 2018-09-04 2019-01-08 温州大学激光与光电智能制造研究院 一种刀具磨损状态在线监测方法及系统
CN109977464A (zh) * 2019-02-18 2019-07-05 江苏科技大学 一种基于bp神经网络的活塞切削加工变形量的预测方法

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114378653A (zh) * 2022-01-27 2022-04-22 上海机床厂有限公司 基于bp神经网络外圆磨削颤振在线识别和监测方法

Also Published As

Publication number Publication date
CN111037365B (zh) 2021-08-20
CN111037365A (zh) 2020-04-21

Similar Documents

Publication Publication Date Title
WO2021128576A1 (zh) 基于生成式对抗网络的刀具状态监测数据集增强方法
US20210197335A1 (en) Data Augmentation Method Based On Generative Adversarial Networks In Tool Condition Monitoring
US11761930B2 (en) Prediction method of part surface roughness and tool wear based on multi-task learning
Qin et al. The optimized deep belief networks with improved logistic sigmoid units and their application in fault diagnosis for planetary gearboxes of wind turbines
CN107316046B (zh) 一种基于增量补偿动态自适应增强的故障诊断方法
CN111366123B (zh) 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
CN110070060B (zh) 一种轴承设备的故障诊断方法
CN112069310B (zh) 基于主动学习策略的文本分类方法及系统
CN112529341B (zh) 一种基于朴素贝叶斯算法的钻井漏失机率预测方法
CN111562108A (zh) 一种基于cnn和fcmc的滚动轴承智能故障诊断方法
CN108256482A (zh) 一种基于卷积神经网络进行分布学习的人脸年龄估计方法
CN108958217A (zh) 一种基于深度学习的can总线报文异常检测方法
CN102201237B (zh) 基于模糊支持向量机的可靠性检测的情感说话人识别方法
CN114358123B (zh) 一种基于深度对抗迁移网络的广义开放集故障诊断方法
CN113927371A (zh) 一种基于多传感器特征融合的刀具磨损预测方法
CN110232415B (zh) 一种基于生物信息特征的列车转向架故障识别方法
CN103077720A (zh) 一种说话人识别方法及系统
Li et al. Tool breakage detection using deep learning
CN109594967A (zh) 一种基于录井大数据的卡钻事故检测预警方法
CN109214444B (zh) 基于孪生神经网络和gmm的游戏防沉迷判定系统及方法
CN110659682A (zh) 一种基于MCWD-KSMOTE-AdaBoost-DenseNet算法的数据分类方法
CN113537609A (zh) 一种基于文本智能挖掘的政策热点预测方法
JP5453107B2 (ja) 音声セグメンテーションの方法および装置
CN106251861A (zh) 一种基于场景建模的公共场所异常声音检测方法
CN116561692A (zh) 一种动态更新的实时量测数据检测方法

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: 20907264

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: 20907264

Country of ref document: EP

Kind code of ref document: A1