WO2020233207A1 - 一种基于半监督学习策略的高光谱数据分析方法 - Google Patents

一种基于半监督学习策略的高光谱数据分析方法 Download PDF

Info

Publication number
WO2020233207A1
WO2020233207A1 PCT/CN2020/079710 CN2020079710W WO2020233207A1 WO 2020233207 A1 WO2020233207 A1 WO 2020233207A1 CN 2020079710 W CN2020079710 W CN 2020079710W WO 2020233207 A1 WO2020233207 A1 WO 2020233207A1
Authority
WO
WIPO (PCT)
Prior art keywords
sample
loss function
layer
samples
network
Prior art date
Application number
PCT/CN2020/079710
Other languages
English (en)
French (fr)
Inventor
刘忆森
周松斌
李昌
韩威
黄可嘉
刘伟鑫
邱泽帆
Original Assignee
广东省智能制造研究所
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 广东省智能制造研究所 filed Critical 广东省智能制造研究所
Priority to US17/438,460 priority Critical patent/US20220129735A1/en
Publication of WO2020233207A1 publication Critical patent/WO2020233207A1/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
    • G01MEASURING; TESTING
    • G01NINVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/17Systems in which incident light is modified in accordance with the properties of the material investigated
    • G01N21/25Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G06N3/048Activation functions
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01NINVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/17Systems in which incident light is modified in accordance with the properties of the material investigated
    • G01N21/25Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
    • G01N21/31Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry
    • G01N21/35Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light
    • G01N21/359Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light using near infrared light

Definitions

  • the invention belongs to the technical field of hyperspectral nondestructive testing, and specifically relates to a hyperspectral data analysis method based on a semi-supervised learning strategy.
  • Hyperspectral-based quantitative analysis technology has a wide range of application scenarios, including food adulteration detection, fruit sugar content detection, microbial content detection, and organic matter content detection.
  • Commonly used hyperspectral quantitative analysis algorithms include partial least squares regression (PLSR), least squares support vector machine (LS-SVM), multiple linear regression (MLR) and other methods, but the model accuracy and model robustness of the quantitative analysis algorithm are still Need to be further improved.
  • PLSR partial least squares regression
  • LS-SVM least squares support vector machine
  • MLR multiple linear regression
  • the convolutional neural network has been proven to have very strong ability to analyze complex information, and the convolutional network has been well applied in the classification of remote sensing hyperspectral data, its application in the quantitative analysis of hyperspectral data still has Greater difficulty.
  • the main reason is that in practical applications, hyperspectral data samples, especially calibrated samples are difficult to obtain, and a small sample size brings a very large risk of overfitting.
  • the label value in the quantitative analysis is a continuous analog quantity, so the network structure design cannot be equal to the semi-supervised classification problem, and it is impossible to use the k+1 class or k classification to fight the network to achieve regression;
  • the classification problem Mainly use the generated data to help determine the boundary value of the classification, and in the regression problem, it is necessary to use the generated data to smooth the sample quantitative value distribution. Therefore, it is necessary to design a new network structure and loss function to realize the semi-supervised hyperspectral data quantitative analysis based on the generative countermeasure network, thereby improving the analysis accuracy.
  • the present invention proposes a hyperspectral data analysis method based on a semi-supervised learning strategy.
  • This method generates spectral samples by generating an adversarial network, and uses the generated samples to enhance the continuity of sample distribution and suppress over-fitting, thereby improving the accuracy of quantitative analysis of hyperspectral data.
  • the present invention solves the above-mentioned problems through the following technical means:
  • S3-1 Construct a generator network, which in turn consists of: fully connected layer-upsampling layer-convolutional layer-upsampling layer-convolutional layer-output layer, in which the number of nodes in the fully connected layer is 16*spectral band
  • the convolution layer is a one-dimensional convolution, the size of the convolution kernel is 1 ⁇ 5, the number of convolution kernels ranges from 16 to 128, the up-sampling layer is 2 times up-sampling, the number of nodes in the output layer and the number of spectral bands Same, except for the output layer, the nonlinear excitation function is ReLU, and the output layer nonlinear excitation is the sigmoid function;
  • S3-2 Construct a discriminator/regressor network, which in turn consists of: convolutional layer-pooling layer-convolutional layer-pooling layer-convolutional layer-pooling layer-output layer, of which the convolutional layer It is a one-dimensional convolution, the size of the convolution kernel is 1 ⁇ 5, the number of convolution kernels ranges from 16 to 128, the pooling layer is 1/2 down-sampling, and the output layer is two, one of which outputs the discriminator result , That is, the predicted value of the true and false of the spectral data, and the other output regressor result, that is, the predicted value of quantitative analysis. Except for the output layer, the nonlinear excitation function is leakyReLU, and the nonlinear excitation of the two output layers is the sigmoid function.
  • the loss function of the discriminator/regressor is the sum of the loss function of the labeled sample and the unlabeled sample
  • the unlabeled sample loss function is the sum of the loss function of the real unlabeled sample and the loss function of the generated sample:
  • the label sample loss function is the regression loss function, that is, the mean square error between the quantitative analysis predicted value of the label sample and the quantitative label value:
  • the unlabeled sample loss function is the cross entropy of the discriminator’s predicted value and the authenticity label:
  • the generated sample loss function is the cross entropy of the discriminator’s predicted value and the authenticity label:
  • the loss function of the generator is the sum of the loss function of the generated sample and the loss function of sample distribution matching:
  • the loss function of the generated sample (L' unlabel_fake ) is the cross-entropy of the predicted value of the generated sample through the discriminator and the authenticity label of the generator.
  • the authenticity label of the generated sample in the generator and the discriminator is opposite, so that the generator and the discriminator Formation of confrontation:
  • the sample distribution matching loss function (L distribution ) is the mean square error of the distribution of the unlabeled sample predicted by the regressor and the quantitative analysis value of the generated sample:
  • the sample distribution loss function exists to reduce the similarity between the existing sample and the quantitative analysis value distribution of the generated sample, so that the generated sample becomes a supplement to the existing unlabeled sample.
  • Adopt the gradient descent method to alternately train the discriminator/regressor and generator until the root mean square error of the quantitative analysis predicted value and label value of the training set samples converges to less than a certain threshold or the number of training steps is greater than a certain threshold;
  • the regressor in the trained generative confrontation network is used to obtain the quantitative analysis prediction value of the prediction set.
  • the present invention has the following advantages:
  • the present invention uses a generative adversarial network to generate samples, uses a sample distribution matching strategy to supplement the existing unlabeled sample set, and uses distributed matching generated samples to achieve a similar regularization effect to the regressor, thereby overcoming the quantitative analysis of deep learning
  • the prone to overfitting problem improves the accuracy of hyperspectral quantitative analysis.
  • FIG. 1 is a flowchart of the present invention.
  • Figure 2(a) is a scatter plot of the actual and predicted values of the active ingredients of tablets obtained by using partial least squares regression (PLSR).
  • PLSR partial least squares regression
  • Figure 2(b) is a scatter plot of the real value and predicted value of the effective ingredient of the tablet obtained by directly using the convolutional network (CNN) without using the generative confrontation network.
  • CNN convolutional network
  • Fig. 2(c) is a scatter plot of the real value and predicted value of the effective ingredient of the tablet obtained by the method of the present invention.
  • This example uses hyperspectroscopy to determine the active ingredients of acetaminophen tablets.
  • a hyperspectral data analysis method based on a semi-supervised learning strategy the specific steps are as follows:
  • S3-1 Construct a generator network, which in turn consists of: fully connected layer-upsampling layer-convolutional layer-upsampling layer-convolutional layer-output layer, in which the number of nodes in the fully connected layer is 64*45,
  • the convolution layer is a one-dimensional convolution, the size of the convolution kernel is 1 ⁇ 5, the number of convolution kernels is 64, the upsampling layer is 2 times upsampling, the number of nodes in the output layer is 180, and the non-linear excitation function except the output layer ReLU, the nonlinear excitation of the output layer is a sigmoid function;
  • S3-2 Construct a discriminator/regressor network, which in turn consists of: convolutional layer-pooling layer-convolutional layer-pooling layer-convolutional layer-pooling layer-output layer, of which the convolutional layer It is a one-dimensional convolution, the size of the convolution kernel is 1 ⁇ 5, the number of convolution kernels is 16, the pooling layer is 1/2 down-sampling, and the output layer is two, one of which outputs the result of the discriminator, that is, the true spectrum data Pseudo label, another output regressor result, that is, the effective component content value, except for the output layer, the nonlinear excitation function is leakyReLU, and the nonlinear excitation of the two output layers is the sigmoid function.
  • the loss function of the discriminator/regressor is the sum of the loss function of the labeled sample and the unlabeled sample
  • the unlabeled sample loss function is the sum of the loss function of the real unlabeled sample and the loss function of the generated sample:
  • the unlabeled sample loss function is the cross entropy of the discriminator’s predicted value and the authenticity label:
  • the generated sample loss function is the cross entropy of the discriminator’s predicted value and the authenticity label:
  • the loss function of the generator is the sum of the loss function of the generated sample and the loss function of sample distribution matching:
  • the loss function of the generated sample (L′ unlabel_fake ) is the cross entropy of the predicted value and the authenticity label of the sample generated by the generator through the discriminator.
  • the authenticity label of the generated sample of the discriminator and the generator is opposite to form a confrontation:
  • the sample distribution matching loss function (L distribution ) is the mean square error of the distribution of the unlabeled sample predicted by the regressor and the quantitative analysis value of the generated sample:
  • Quantitative analysis value distribution of unlabeled samples predicted by the regression Quantitatively analyze the value distribution of the generated sample predicted by the regressor.
  • the specific calculation method is: approximate the effective component value of the sample to a multinomial distribution of 100 items with a value of 0 to 1, and quantitatively analyze the value of the unlabeled sample of the current training batch predicted by the regressor and the generated sample with Approximate to the above 100 levels, and obtained through sample count statistics with
  • the gradient descent method is used to alternately train the discriminator/regressor and the generator.
  • the optimizer adopts the "Adam" optimizer, the learning rate is 0.0005, and the training step stops when the number of training steps reaches 1000;
  • the training set and the prediction set were randomly sampled 10 times for calculation, and partial least square regression and convolutional network were used as a comparison method for calculation.
  • the slice least squares regression method the number of principal components is determined by ten-fold cross-validation, and the parameters of the convolutional network are consistent with the regressor in the method of the present invention.
  • the root mean square prediction error RMSEC of the training set obtained by the partial least square method is 1.01 ⁇ 0.46%, and the root mean square prediction error RMSEP of the prediction set is 3.79 ⁇ 1.06%; the training set obtained by the convolutional network
  • the prediction error root mean square RMSEC is 1.45 ⁇ 0.71%, the prediction set root mean square error RMSEP is 5.84 ⁇ 1.77%;
  • the training set obtained by the method of the present invention has a prediction error root mean square RMSEC of 2.42 ⁇ 0.49%, and the prediction set The root mean square RMSEP of prediction error is 2.56 ⁇ 0.88%.
  • Figure 2 shows the scatter plot of the true value and the predicted value calculated by a random sampling.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Other Investigation Or Analysis Of Materials By Electrical Means (AREA)
  • Investigating Or Analysing Biological Materials (AREA)

Abstract

一种基于半监督学习策略的高光谱数据分析方法,其步骤如下:(1)获取高光谱样本数据;(2)构建样本训练集与预测集,其中无标签的预测集样本也用于半监督训练;(3)构建基于生成对抗网络的回归网络,包括一个生成样本的生成器网络,和一个同时具有判别样本真伪与输出定量分析值功能的判别器/回归器网络;(4)构建生成对抗网络的损失函数,包括判别器损失函数,回归器损失函数,及具备样本分布匹配功能的生成器损失函数;(5)训练生成对抗网络;(6)采用训练好的回归器进行定量分析值的预测。采用生成对抗网络来生成样本,并使用样本分布匹配策略来补充现有无标签样本集,提升了高光谱定量分析的精度。

Description

一种基于半监督学习策略的高光谱数据分析方法 技术领域:
本发明属于高光谱无损检测技术领域,具体涉及一种基于半监督学习策略的高光谱数据分析方法。
背景技术:
基于高光谱的定量分析技术应用场景广泛,包括食品掺假检测、水果糖度检测、微生物含量检测、有机质含量检测等。常用的高光谱定量分析算法包括偏最小二乘回归(PLSR)、最小二乘支持向量机(LS-SVM)、多元线性回归(MLR)等方法,但定量分析算法的模型精度与模型健壮性仍然有待于进一步提高。尤其是现实应用场景中,面临着有标签样本不均衡,或有标签样本少等情况,更增大了建模的难度。
当前,虽然卷积神经网络已被证明具备非常强的复杂信息解析能力,并且卷积网络在遥感高光谱数据分类应用中已获得了良好的应用,但其在高光谱数据定量分析中应用仍然具有较大的难度。最主要的原因是实际应用中,高光谱数据样本,尤其是已标定的样本难以获得,小样本量带来十分大的过拟合风险。
西安电子科技大学在发明专利文献“基于多类生成对抗网络的高光谱图像分类方法”(申请号:201810648520.4)及“协作生成对抗网络和空谱联合的高光谱图像分类方法”(申请号:201810977887.8)中,采用生成对抗网络(GAN)生成光谱曲线样本,从而缓解过拟合现象,提升了高光谱数据的分类精度。然而,以上文献方法皆是针对高光谱数据分类问题展开,而采用生成对抗网络进行高光谱数据的定量分析还是存在一定难度。首先,与分类问题的离散标签不同,定量分析中标签值为连续模拟量,因此网络结构设计不能等同于半监督分类问题,无法利用k+1类或k分类对抗网络实现回归;其次,分类问题中主要利 用生成数据来辅助确定分类的边界值,而在回归问题中,需要利用生成数据对样本定量值分布进行平滑。因此,需要设计全新的网络结构与损失函数来实现基于生成对抗网络的半监督高光谱数据定量分析,从而提升分析精度。
发明内容:
针对现有技术的上述不足,本发明提出了一种基于半监督学习策略的高光谱数据分析方法。该方法通过生成对抗网络生成光谱样本,利用生成样本来达到增强样本分布连续性、抑制过拟合的目的,从而提升高光谱数据定量分析的精度。
本发明通过以下技术手段解决上述问题:
一种基于半监督学习策略的高光谱数据分析方法,构建用于定量分析的生成对抗网络,其中生成器用于生成样本,判别器用于判别样本的真伪,同时输出定量分析结果;该方法包括如下步骤:
S1.获取有标签的高光谱样本数据及无标签的高光谱样本数据;
S2.构建样本训练集与预测集:
S2-1.将由标签样本作为训练集样本,无标签样本作为预测集样本,同时也用于半监督的训练;
S2-2.在每个训练集和预测集的高光谱样本数据块中随机取m次n个有效像素光谱曲线的均值,作为样本增强,得到的有标签和无标签平均光谱数据集记为D label和D unlabel
S3.构建基于生成对抗的回归网络:
S3-1.构建一个生成器网络,该网络依次由:全连接层—上采样层—卷积层—上采样层—卷积层—输出层构成,其中全连接层节点数为16*光谱波段数,卷积层为一维卷积,卷积核尺寸为1×5,卷积核个数取值范围为16~128,上采样层为2倍上采样,输出层节点数与光谱波段数相同,除输出层外非线性激励 函数为ReLU,输出层非线性激励为sigmoid函数;
S3-2.构建一个判别器/回归器网络,该网络依次由:卷积层—池化层—卷积层—池化层—卷积层—池化层—输出层构成,其中卷积层为一维卷积,卷积核尺寸为1×5,卷积核个数取值范围为16~128,池化层为1/2下采样,输出层为2个,其中一个输出判别器结果,即光谱数据真伪预测值,另一个输出回归器结果,即定量分析预测值,除输出层外非线性激励函数为leakyReLU,两个输出层非线性激励为sigmoid函数。
S4.构建生成对抗回归网络的损失函数:
S4-1.判别器/回归器的损失函数为有标签样本和无标签样本损失函数之和,其中无标签样本损失函数为真实无标签样本损失函数与生成样本损失函数之和:
L D=L supervised+L unsupervised
L unsupervised=L unlabel_real+L unlabel_fake
其中有标签样本损失函数为回归损失函数,即有标签样本的定量分析预测值与定量标签值的均方误差:
Figure PCTCN2020079710-appb-000001
无标签样本损失函数为判别器预测值与真伪标签的交叉熵:
Figure PCTCN2020079710-appb-000002
生成样本损失函数为判别器预测值与真伪标签的交叉熵:
Figure PCTCN2020079710-appb-000003
S4-2.生成器的损失函数为生成样本的损失函数与样本分布匹配损失函数之和:
L G=L′ unlabel_fake+L distribution
其中生成样本的损失函数(L′ unlabel_fake)为生成器生成样本经过判别器的预 测值与真伪标签的交叉熵,生成器与判别器中的生成样本真伪标签相反,从而使生成器与判别器形成对抗:
Figure PCTCN2020079710-appb-000004
样本分布匹配损失函数(L distribution)为回归器预测得到的无标签样本与生成样本定量分析值分布均方误差:
Figure PCTCN2020079710-appb-000005
其中
Figure PCTCN2020079710-appb-000006
为回归器预测得到的无标签样本定量分析值分布,
Figure PCTCN2020079710-appb-000007
为回归器预测得到的生成样本定量分析值分布。
Figure PCTCN2020079710-appb-000008
Figure PCTCN2020079710-appb-000009
的计算步骤为:
S4-2-1.由回归器预测当前训练批次的无标签样本与生成样本定量分析值,分别为
Figure PCTCN2020079710-appb-000010
Figure PCTCN2020079710-appb-000011
S4-2-2.将样本定量分析值的分布近似为k项的多项分布,将
Figure PCTCN2020079710-appb-000012
Figure PCTCN2020079710-appb-000013
近似到k个水平,然后通过计数得到概率分布
Figure PCTCN2020079710-appb-000014
Figure PCTCN2020079710-appb-000015
样本分布损失函数的存在是为了降低已有样本与生成样本定量分析值分布的相似性,从而使生成样本成为已有无标签样本的补充。
S5.训练生成对抗回归网络:
采用梯度下降方法,交替训练判别器/回归器与生成器,直到训练集样本的定量分析预测值与标签值的均方根误差收敛至小于一定阈值或训练步数大于一定阈值;
S6.采用训练好的生成对抗网络中的回归器得到预测集的定量分析预测值。
本发明对比现有技术具有以下优势:
本发明采用生成对抗网络来生成样本,并使用样本分布匹配策略来补充现有无标签样本集,利用分布匹配的生成样本来对回归器达到类似正则化效果,从而克服了深度学习在定量分析中容易产生的过拟合问题,提升了高光谱定量分 析的精度。
附图说明:
为了更清楚地说明本发明实施例中的技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本发明的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。
图1是本发明流程图。
图2(a)是采用偏最小二乘回归方法(PLSR)得到的药片有效成分真实值与预测值散点图。
图2(b)是不采用生成对抗网络,直接采用卷积网络(CNN)得到的药片有效成分真实值与预测值散点图。
图2(c)是采用本发明方法得到药片有效成分真实值与预测值散点图。
具体实施方式:
为使本发明的上述目的、特征和优点能够更加明显易懂,下面将结合附图和具体的实施例对本发明的技术方案进行详细说明。需要指出的是,所描述的实施例仅仅是本发明一部分实施例,而不是全部的实施例,基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本发明保护的范围。
本实施例为利用高光谱测定对乙酰氨基酚药片有效成分。如图1所示,一种基于半监督学习策略的高光谱数据分析方法,具体步骤如下:
S1.测定共437个对乙酰氨基酚药片的高光谱图像数据,其中有效成分(对乙酰氨基酚)的含量为0%~90%w/w,间隔为5%w/w,共19个水平,每个水平23个样本;高光谱波段为900-1700nm,其中前后100nm由于噪声高去除, 1000-1600nm的共180个波段用于数据分析。
S2.构建样本训练集与预测集:
S2-1.随机取10个样本作为训练集,即10个有标签样本,剩余427个样本作为预测集,用于评价模型精度,同时也作为无标签样本,用于半监督的训练;
S2-2.在每个训练集和预测集的高光谱样本数据块中随机取10次100个有效像素光谱曲线的均值,作为样本增强,得到的有标签和无标签平均光谱数据集记为D label和D unlabel
S3.构建基于生成对抗网络的回归网络:
S3-1.构建一个生成器网络,该网络依次由:全连接层—上采样层—卷积层—上采样层—卷积层—输出层构成,其中全连接层节点数为64*45,卷积层为一维卷积,卷积核尺寸为1×5,卷积核个数为64,上采样层为2倍上采样,输出层节点数为180,除输出层外非线性激励函数为ReLU,输出层非线性激励为sigmoid函数;
S3-2.构建一个判别器/回归器网络,该网络依次由:卷积层—池化层—卷积层—池化层—卷积层—池化层—输出层构成,其中卷积层为一维卷积,卷积核尺寸为1×5,卷积核个数为16,池化层为1/2下采样,输出层为2个,其中一个输出判别器结果,即光谱数据真伪标签,另一个输出回归器结果,即有效成分含量值,除输出层外非线性激励函数为leakyReLU,两个输出层非线性激励为sigmoid函数。
S4.构建生成对抗回归网络的损失函数:
S4-1.判别器/回归器的损失函数为有标签样本和无标签样本损失函数之和,其中无标签样本损失函数为真实无标签样本损失函数与生成样本损失函数之和:
L D=L supervised+L unsupervised
L unsupervised=L unlabel_real+L unlabel_fake
其中有标签样本损失函数为有标签样本的有效成分预测值与有效成分标签值的均方误差:
Figure PCTCN2020079710-appb-000016
无标签样本损失函数为判别器预测值与真伪标签的交叉熵:
Figure PCTCN2020079710-appb-000017
生成样本损失函数为判别器预测值与真伪标签的交叉熵:
Figure PCTCN2020079710-appb-000018
S4-2.生成器的损失函数为生成样本的损失函数与样本分布匹配损失函数之和:
L G=L′ unlabel_fake+L distribution
其中生成样本的损失函数(L′ unlabel_fake)为生成器生成样本经过判别器的预测值与真伪标签的交叉熵,其中判别器与生成器的生成样本真伪标签相反,从而形成对抗:
Figure PCTCN2020079710-appb-000019
样本分布匹配损失函数(L distribution)为回归器预测得到的无标签样本与生成样本定量分析值分布均方误差:
Figure PCTCN2020079710-appb-000020
其中
Figure PCTCN2020079710-appb-000021
为回归器预测得到的无标签样本定量分析值分布,
Figure PCTCN2020079710-appb-000022
为回归器预测得到的生成样本定量分析值分布。
Figure PCTCN2020079710-appb-000023
Figure PCTCN2020079710-appb-000024
具体计算方法为:将样本的有效成分值近似为取值为0~1的100项多项分布,将回归器预测得到的当前训练批次的无标签样本与生成样本定量分析值
Figure PCTCN2020079710-appb-000025
Figure PCTCN2020079710-appb-000026
近似到上述100个水平上,并通过样本计数统计得到
Figure PCTCN2020079710-appb-000027
Figure PCTCN2020079710-appb-000028
S5.训练生成对抗回归网络:
采用梯度下降方法,交替训练判别器/回归器与生成器,优化器采用“Adam”优化器,学习率为0.0005,训练步数达到1000时停止;
S6.采用训练好的生成对抗回归网络中的回归器得到预测集的药片有效成分预测值。
训练集与预测集随机抽样10次进行计算,偏最小二乘回归与卷积网络作为对比方法进行计算。其中片最小二乘回归方法中,主成分数由十折交叉验证决定,卷积网络的参数与本发明方法中的回归器保持一致。10次计算的结果为:偏最小二乘方法得到训练集的预测误差均方根RMSEC为1.01±0.46%,预测集的预测误差均方根RMSEP为3.79±1.06%;卷积网络得到的训练集的预测误差均方根RMSEC为1.45±0.71%,预测集的预测误差均方根RMSEP为5.84±1.77%;本发明方法得到的训练集的预测误差均方根RMSEC为2.42±0.49%,预测集的预测误差均方根RMSEP为2.56±0.88%。其中一次随机取样计算得到的真实值与预测值散点图见图2。
由计算结果可见,在小样本情况下,偏最小二乘回归与卷积网络均发生了较明显的过拟合现象,预测集精度远低于训练集精度。而本发明方法由于采用了生成对抗式网络结构,因此很好的缓解了过拟合问题,使预测集精度得到了显著提升。
以上所述实施例仅表达了本发明的几种实施方式,其描述较为具体和详细,但并不能因此而理解为对本发明专利范围的限制。应当指出的是,对于本领域的普通技术人员来说,在不脱离本发明构思的前提下,还可以做出若干变形和改进,这些都属于本发明的保护范围。因此,本发明专利的保护范围应以所附权利要求为准。

Claims (7)

  1. 一种基于半监督学习策略的高光谱数据分析方法,其特征在于,包括如下步骤:
    S1.获取有标签的高光谱样本数据及无标签的高光谱样本数据;
    S2.构建样本训练集与预测集:将有标签的高光谱样本数据集D label作为训练集,无标签的高光谱样本数据集D unlabel作为预测集;
    S3.构建基于生成对抗网络的回归网络:
    S3-1.构建一个由全连接层、卷积层、上采样层、输出层组成的生成器网络;
    S3-2.构建一个由卷积层、池化层、输出层组成的判别器/回归器网络,该网络具有判别数据真伪与定量分析值两个输出;
    S4.构建生成对抗回归网络的损失函数:
    S4-1.判别器/回归器的损失函数为有标签样本和无标签样本损失函数之和,其中无标签样本损失函数为真实无标签样本损失函数(L unlabel_real)与生成样本损失函数(L unlabel_fake)之和:
    L D=L supervised+L unsupervised
    L unsupervised=L unlabel_real+L unlabel_fake
    S4-2.生成器的损失函数为生成样本的损失函数(L′ unlabel_fake)与样本分布匹配损失函数(L distribution)之和:
    L G=L′ unlabel_fake+L distribution
    其中样本分布匹配损失函数,是为了使生成样本和真实样本在样本空间中互为补集,从而达到抑制过拟合效果;
    S5.训练基于生成对抗的回归网络:
    采用梯度下降方法,交替训练判别器/回归器与生成器,直到训练集样本的定量分析预测值与标签值的均方根误差收敛至小于一个阈值或训练步数大于一个阈值;
    S6.采用训练好的生成对抗网络中的回归器得到预测集的定量分析预测值。
  2. 根据权利要求1所述的基于半监督学习策略的高光谱数据分析方法,其特征在于,步骤S2中构建样本训练集与预测集步骤包括:
    S2-1.将由标签样本作为训练集样本,无标签样本作为预测集样本,同时也用于半监督的训练;
    S2-2.在每个训练集和预测集的高光谱样本数据块中随机取m次n个有效像素光谱曲线的均值,作为样本增强,得到的有标签和无标签平均光谱数据集记为D label和D unlabel
  3. 根据权利要求1所述的基于半监督学习策略的高光谱数据分析方法,其特征在于,步骤S3-1中所述生成器网络依次由:全连接层—上采样层—卷积层—上采样层—卷积层—输出层构成,其中全连接层节点数为16*光谱波段数,卷积层为一维卷积,卷积核尺寸为1×5,卷积核个数取值范围为16~128,上采样层为2倍上采样,输出层节点数与光谱波段数相同,除输出层外非线性激励函数为ReLU,输出层非线性激励为sigmoid函数。
  4. 根据权利要求1所述的基于半监督学习策略的高光谱数据分析方法,其特征在于,步骤S3-2中所述判别器/回归器网络依次由:卷积层—池化层—卷积层—池化层—卷积层—池化层—输出层构成,其 中卷积层为一维卷积,卷积核尺寸为1×5,卷积核个数取值范围为16~128,池化层为1/2下采样,输出层为2个,其中一个输出判别器结果,即光谱数据真伪预测值,另一个输出回归器结果,即定量分析预测值,除输出层外非线性激励函数为leakyReLU,两个输出层非线性激励为sigmoid函数。
  5. 根据权利要求1所述的基于半监督学习策略的高光谱数据分析方法,其特征在于,步骤S4-1中所述判别器/回归器的损失函数为有标签样本损失函数、无标签样本损失函数、和生成样本损失函数三个部分之和,其中有标签样本损失函数为有标签样本的定量分析预测值与定量标签值的均方误差:
    Figure PCTCN2020079710-appb-100001
    无标签样本损失函数为判别器预测值与真伪标签的交叉熵:
    Figure PCTCN2020079710-appb-100002
    生成样本损失函数为判别器预测值与真伪标签的交叉熵:
    Figure PCTCN2020079710-appb-100003
  6. 根据权利要求1所述的基于半监督学习策略的高光谱数据分析方法,其特征在于,步骤S4-2中所述生成器损失函数为生成样本的损失函数与样本分布匹配损失函数之和,其中生成样本的损失函数为生成器生成样本经过判别器的预测值与真伪标签的交叉熵;生成器与判别器中的生成样本真伪标签相反,从而使生成器与判别器形成对抗:
    Figure PCTCN2020079710-appb-100004
    样本分布匹配损失函数为:
    Figure PCTCN2020079710-appb-100005
    其中
    Figure PCTCN2020079710-appb-100006
    为回归器预测得到的当前训练批次的无标签样本定量分析值分布,
    Figure PCTCN2020079710-appb-100007
    为回归器预测得到的当前训练批次的生成样本定量分析值分布。
  7. 根据权利要求6所述的基于半监督学习策略的高光谱数据分析方法,其特征在于,无标签样本定量分析值分布与生成样本定量分析值分布计算步骤为:
    S4-2-1.由回归器预测当前训练批次的无标签样本与生成样本定量分析值,分别为
    Figure PCTCN2020079710-appb-100008
    Figure PCTCN2020079710-appb-100009
    S4-2-2.将样本定量分析值的分布近似为k项的多项分布,将
    Figure PCTCN2020079710-appb-100010
    Figure PCTCN2020079710-appb-100011
    近似到k个水平,然后通过计数得到概率分布
    Figure PCTCN2020079710-appb-100012
    Figure PCTCN2020079710-appb-100013
PCT/CN2020/079710 2019-05-20 2020-03-17 一种基于半监督学习策略的高光谱数据分析方法 WO2020233207A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/438,460 US20220129735A1 (en) 2019-05-20 2020-03-17 Semi-supervised Hyperspectral Data Quantitative Analysis Method Based on Generative Adversarial Network

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910420079.6A CN110320162B (zh) 2019-05-20 2019-05-20 一种基于生成对抗网络的半监督高光谱数据定量分析方法
CN201910420079.6 2019-05-20

Publications (1)

Publication Number Publication Date
WO2020233207A1 true WO2020233207A1 (zh) 2020-11-26

Family

ID=68113177

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/079710 WO2020233207A1 (zh) 2019-05-20 2020-03-17 一种基于半监督学习策略的高光谱数据分析方法

Country Status (3)

Country Link
US (1) US20220129735A1 (zh)
CN (1) CN110320162B (zh)
WO (1) WO2020233207A1 (zh)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112816499A (zh) * 2021-04-19 2021-05-18 征图新视(江苏)科技股份有限公司 一种高光谱结合深度学习工业检测系统
CN112949820A (zh) * 2021-01-27 2021-06-11 西安电子科技大学 基于生成对抗网络的认知抗干扰目标检测方法
CN112949422A (zh) * 2021-04-30 2021-06-11 西北工业大学 一种基于自监督光谱匹配框架的高光谱目标检测方法
CN113033595A (zh) * 2020-12-24 2021-06-25 重庆大学 基于生成对抗网络的多标签汽车造型生成方法
CN113298230A (zh) * 2021-05-14 2021-08-24 西安理工大学 一种基于生成对抗网络的不平衡数据集的预测方法
CN113762389A (zh) * 2021-09-08 2021-12-07 大连海事大学 一种基于动态诊断的高光谱影像小样本分类方法
CN114049567A (zh) * 2021-11-22 2022-02-15 齐鲁工业大学 自适应软标签生成方法以及在高光谱图像分类中的应用
CN114187477A (zh) * 2021-12-16 2022-03-15 西安交通大学 一种基于监督自对比学习的小样本高光谱图像分类方法
CN115015180A (zh) * 2022-07-20 2022-09-06 浙江理工大学 一种基于光谱反射率重建技术的水果糖度无损检测方法
CN115984635A (zh) * 2023-03-21 2023-04-18 自然资源部第一海洋研究所 多源遥感数据分类模型训练方法、分类方法和电子设备
CN116127345A (zh) * 2022-12-23 2023-05-16 北京科技大学 基于深度聚类生成对抗网络的转炉炼钢工艺模式设计方法

Families Citing this family (26)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110320162B (zh) * 2019-05-20 2021-04-23 广东省智能制造研究所 一种基于生成对抗网络的半监督高光谱数据定量分析方法
CN111046900B (zh) * 2019-10-25 2022-10-18 重庆邮电大学 基于局部流形正则化的半监督生成对抗网络图像分类方法
CN111127454A (zh) * 2019-12-27 2020-05-08 上海交通大学 基于深度学习生成工业缺陷样本的方法及系统
CN111794741B (zh) * 2020-08-11 2023-08-18 中国石油天然气集团有限公司 滑动定向钻井模拟机实现方法
CN111931062B (zh) * 2020-08-28 2023-11-24 腾讯科技(深圳)有限公司 一种信息推荐模型的训练方法和相关装置
CN112199637B (zh) * 2020-09-21 2024-04-12 浙江大学 基于回归注意力生成对抗网络数据增强的回归建模方法
CN112381756B (zh) * 2020-09-30 2022-12-16 广东省科学院智能制造研究所 一种基于块平滑神经网络的高光谱数据分析方法和系统
CN112232395B (zh) * 2020-10-08 2023-10-27 西北工业大学 一种基于联合训练生成对抗网络的半监督图像分类方法
CN112446429B (zh) * 2020-11-27 2022-06-21 广东电网有限责任公司肇庆供电局 一种基于cgan的巡检图像数据小样本扩充方法
CN112580547A (zh) * 2020-12-24 2021-03-30 武汉欧铭达科技有限公司 一种航空影像特征点自动识别方法和系统
CN113269228B (zh) * 2021-04-20 2022-06-10 重庆邮电大学 一种图网络分类模型的训练方法、装置、系统及电子设备
CN113657455B (zh) * 2021-07-23 2024-02-09 西北工业大学 一种基于三重网络与标注一致性正则化的半监督学习方法
CN113687610B (zh) * 2021-07-28 2023-03-24 国网江苏省电力有限公司南京供电分公司 一种gan-cnn电力监测系统终端信息防护法方法
CN113569951B (zh) * 2021-07-29 2023-11-07 山东科技大学 一种基于生成对抗网络的近红外定量分析模型构建方法
CN114002167A (zh) * 2021-11-02 2022-02-01 浙江大学 一种深度学习水果光谱分析模型更新方法
CN114220501A (zh) * 2021-11-24 2022-03-22 江苏大学 炒饭食味特性的快速定量评价方法
CN114511473B (zh) * 2022-04-19 2022-07-05 武汉大学 一种基于无监督自适应学习的高光谱遥感影像去噪方法
CN114826776B (zh) * 2022-06-06 2023-05-02 中国科学院高能物理研究所 一种用于加密恶意流量的弱监督检测方法及系统
CN114858782B (zh) * 2022-07-05 2022-09-27 中国民航大学 基于拉曼高光谱对抗判别模型的奶粉掺杂非定向检测方法
CN115314254B (zh) * 2022-07-07 2023-06-23 中国人民解放军战略支援部队信息工程大学 基于改进wgan-gp的半监督恶意流量检测方法
CN115310361B (zh) * 2022-08-16 2023-09-15 中国矿业大学 基于wgan-cnn煤矿井下粉尘浓度预测方法和系统
CN115829980B (zh) * 2022-12-13 2023-07-25 深圳核韬科技有限公司 一种眼底照片的图像识别方法、装置、设备及存储介质
CN116977667B (zh) * 2023-08-01 2024-01-26 中交第二公路勘察设计研究院有限公司 一种基于改进gain的隧道变形数据填补方法
CN117216724A (zh) * 2023-08-22 2023-12-12 北京石油化工学院 一种微生物培养过程中多种物质实时检测的方法
CN116821800B (zh) * 2023-08-31 2023-11-10 深圳市路桥建设集团有限公司 基于半监督生成对抗网络的结构状态分类方法及相关设备
CN117368121B (zh) * 2023-11-27 2024-05-07 安徽农业大学 一种茶叶含水率在线检测方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108764173A (zh) * 2018-05-31 2018-11-06 西安电子科技大学 基于多类生成对抗网络的高光谱图像分类方法
WO2018209894A1 (en) * 2017-05-19 2018-11-22 Huawei Technologies Co., Ltd. Semi-supervised regression with generative adversarial networks
CN109102029A (zh) * 2018-08-23 2018-12-28 重庆科技学院 信息最大化生成对抗网络模型合成人脸样本质量评估方法
CN109145992A (zh) * 2018-08-27 2019-01-04 西安电子科技大学 协作生成对抗网络和空谱联合的高光谱图像分类方法
CN109615632A (zh) * 2018-11-09 2019-04-12 广东技术师范学院 一种基于半监督条件生成式对抗网络的眼底图视盘和视杯分割方法
CN110320162A (zh) * 2019-05-20 2019-10-11 广东省智能制造研究所 一种基于生成对抗网络的半监督高光谱数据定量分析方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018209894A1 (en) * 2017-05-19 2018-11-22 Huawei Technologies Co., Ltd. Semi-supervised regression with generative adversarial networks
CN108764173A (zh) * 2018-05-31 2018-11-06 西安电子科技大学 基于多类生成对抗网络的高光谱图像分类方法
CN109102029A (zh) * 2018-08-23 2018-12-28 重庆科技学院 信息最大化生成对抗网络模型合成人脸样本质量评估方法
CN109145992A (zh) * 2018-08-27 2019-01-04 西安电子科技大学 协作生成对抗网络和空谱联合的高光谱图像分类方法
CN109615632A (zh) * 2018-11-09 2019-04-12 广东技术师范学院 一种基于半监督条件生成式对抗网络的眼底图视盘和视杯分割方法
CN110320162A (zh) * 2019-05-20 2019-10-11 广东省智能制造研究所 一种基于生成对抗网络的半监督高光谱数据定量分析方法

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113033595A (zh) * 2020-12-24 2021-06-25 重庆大学 基于生成对抗网络的多标签汽车造型生成方法
CN112949820A (zh) * 2021-01-27 2021-06-11 西安电子科技大学 基于生成对抗网络的认知抗干扰目标检测方法
CN112949820B (zh) * 2021-01-27 2024-02-02 西安电子科技大学 基于生成对抗网络的认知抗干扰目标检测方法
CN112816499A (zh) * 2021-04-19 2021-05-18 征图新视(江苏)科技股份有限公司 一种高光谱结合深度学习工业检测系统
CN112816499B (zh) * 2021-04-19 2021-06-29 征图新视(江苏)科技股份有限公司 一种高光谱结合深度学习工业检测系统
CN112949422B (zh) * 2021-04-30 2024-06-04 西北工业大学 一种基于自监督光谱匹配框架的高光谱目标检测方法
CN112949422A (zh) * 2021-04-30 2021-06-11 西北工业大学 一种基于自监督光谱匹配框架的高光谱目标检测方法
CN113298230A (zh) * 2021-05-14 2021-08-24 西安理工大学 一种基于生成对抗网络的不平衡数据集的预测方法
CN113298230B (zh) * 2021-05-14 2024-04-09 武汉嫦娥医学抗衰机器人股份有限公司 一种基于生成对抗网络的不平衡数据集的预测方法
CN113762389A (zh) * 2021-09-08 2021-12-07 大连海事大学 一种基于动态诊断的高光谱影像小样本分类方法
CN113762389B (zh) * 2021-09-08 2024-04-12 大连海事大学 一种基于动态诊断的高光谱影像小样本分类方法
CN114049567A (zh) * 2021-11-22 2022-02-15 齐鲁工业大学 自适应软标签生成方法以及在高光谱图像分类中的应用
CN114049567B (zh) * 2021-11-22 2024-02-23 齐鲁工业大学 自适应软标签生成方法以及在高光谱图像分类中的应用
CN114187477A (zh) * 2021-12-16 2022-03-15 西安交通大学 一种基于监督自对比学习的小样本高光谱图像分类方法
CN115015180B (zh) * 2022-07-20 2024-05-21 浙江理工大学 一种基于光谱反射率重建技术的水果糖度无损检测方法
CN115015180A (zh) * 2022-07-20 2022-09-06 浙江理工大学 一种基于光谱反射率重建技术的水果糖度无损检测方法
CN116127345B (zh) * 2022-12-23 2023-11-14 北京科技大学 基于深度聚类生成对抗网络的转炉炼钢工艺模式设计方法
CN116127345A (zh) * 2022-12-23 2023-05-16 北京科技大学 基于深度聚类生成对抗网络的转炉炼钢工艺模式设计方法
CN115984635A (zh) * 2023-03-21 2023-04-18 自然资源部第一海洋研究所 多源遥感数据分类模型训练方法、分类方法和电子设备

Also Published As

Publication number Publication date
CN110320162A (zh) 2019-10-11
CN110320162B (zh) 2021-04-23
US20220129735A1 (en) 2022-04-28

Similar Documents

Publication Publication Date Title
WO2020233207A1 (zh) 一种基于半监督学习策略的高光谱数据分析方法
Da Silva et al. Forecasting Brazilian and American COVID-19 cases based on artificial intelligence coupled with climatic exogenous variables
Thamrin et al. Predicting obesity in adults using machine learning techniques: an analysis of Indonesian basic health research 2018
Wang et al. OGRU: An optimized gated recurrent unit neural network
CN111460818B (zh) 一种基于增强胶囊网络的网页文本分类方法及存储介质
CN111340187B (zh) 基于对抗注意力机制的网络表征方法
CN109522953A (zh) 基于网络嵌入算法和cnn对图结构数据进行分类的方法
Licciardi et al. Spatiotemporal pattern recognition and nonlinear PCA for global horizontal irradiance forecasting
Johnpeter et al. Computer aided automated detection and classification of brain tumors using CANFIS classification method
Alakkari et al. Forecasting COVID-19 Infection Using Encoder-Decoder LSTM and Attention LSTM Algorithms.
Said et al. Skin cancer detection and classification based on deep learning
Dhiman et al. On wavelet transform based convolutional neural network and twin support vector regression for wind power ramp event prediction
Zhou et al. Research and improvement of encrypted traffic classification based on convolutional neural network
Sreedevi et al. Development of weighted ensemble transfer learning for tomato leaf disease classification solving low resolution problems
Zhu Convolutional neural networks based study and application for multicategory skin cancer detection
Adhinata et al. Prediction of Covid-19 daily case in indonesia using long short term memory method
CN115062686A (zh) 基于多角度特征的多元kpi时序异常检测方法和系统
Tian et al. Tropical cyclone maximum wind estimation from infrared satellite data with integrated convolutional neural networks
Liu Comparison of different Convolutional Neural Network models on Fruit 360 Dataset
Li et al. The research of disease spots extraction based on evolutionary algorithm
Ifeoluwapo et al. COVID-19 death risk assessment in Iran using Artificial Neural Network
Song et al. Modeling type 1 and type 2 diabetes mellitus incidence in youth: An application of Bayesian hierarchical regression for sparse small area data
Pal et al. A Deep learning based Safe Travel Advisory Tool for Covid-19
Hadi et al. Development of machine learning modelling and dengue risk mapping: a concept framework
Mohanty et al. An optimized KELM approach for the diagnosis of COVID-19 from 2D-SSA reconstructed CXR Images

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: 20809630

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: 20809630

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20809630

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 31/08/2022)

122 Ep: pct application non-entry in european phase

Ref document number: 20809630

Country of ref document: EP

Kind code of ref document: A1